Representing Irish Mutations in Distributed Morphology and Optimality Theory

Jack Pruett

This volume is the first handbook devoted entirely to the multitude of frameworks adopted in the field of morphology, including Minimalism, Optimality Theory, Network Morphology, Cognitive Grammar, and Canonical Typology. Following an introduction from the editors, the first part of the volume offers critical discussions of the main theoretical issues within morphology, both in word formation and in inflection, as well as providing a short history of morphological theory. In the core part of the handbook, part II, each theory is introduced by an expert in the field, who guides the reader through its principles and technicalities, its advantages and disadvantages, and its points of agreement and disagreement with alternative theories. Chapters in part III explore the bigger picture, connecting morphological theory to other subdisciplines of linguistics, such as diachronic change, language acquisition, psycholinguistics, and sign language theory. The handbook is intended as a guide for morphologists from all theoretical backgrounds who want to learn more about frameworks other than their own, as well as for linguists in related subfields looking for theoretical connections with the field of morphology.

In this paper I argue that initial consonant mutation in the Celtic languages does not involve synchronically derived phonological alternation, nor is it the product of full lexical listing of alternant wordforms. Rather, Celtic initial mutation involves associations of consonants represented in the lexicon which relate a specific initial consonant of a radical form to its associated mutation reflexes. Together with subcategorisation, which ensures that the correct mutation reflex of a wordform appears in the correct environment, the appropriate initial consonant is selected from an association of consonants extracted by the speaker from the recurring patterns in the language. It is these consonant associations, not full alternant wordforms, that are listed in the lexicon. The advantages to this approach are threefold: broad patterns are identified which full lexical listing obscures, the participation of neologisms in mutation is accounted for straightforwardly, and some of the motivation for diachronic reanalysis within the mutation system is made more apparent.

Standard Irish is the outcome of language planning and as such it significantly diverges from the three main spoken dialects of the language (or traditional Gaeltacht varieties) that provided the basis for its creation. It is also expected to differ, in its codified form, from the way it is actually employed within the small, usually urban communities of bilinguals who employ standard Irish and not some form of Gaeltacht Irish as a second language. The reason why such difference is expected is that the language planners codified as part of the standard many complex structures that had already been abandoned in the spoken dialects, basing their reconstruction on historicity rather than actual usage (especially as the actual usages were far from uniform). In this article, which presents part of the work involved in my currently ongoing research, some such complexities are presented which pertain to grammatical gender.

CLS 58 2022 Proceedings of the Fifty-eighth Annual Meeting of the Chicago Linguistic Society Edited by: Lucas Fagen Sam Gray Quain Stephanie Reyes Irene Tang CLS 58 Proceedings of the fifty-eighth annual meeting of the Chicago Linguistic Society First Edition; Published 2023 ISSN 0577-7240 ISBN 978-0-914203-88-9 Cover design by Emre Hakgüder For information on other publications by the Chicago Linguistic Society, contact: T HE C HICAGO L INGUISTIC S OCIETY 1115 E. 58th St. Rosenwald Hall, Room 229 Chicago, IL 60637 Tel: (773) 702-8522 Email: clsbooks@gmail.com http://chicagolinguisticsociety.org/ ©2023 by T HE C HICAGO L INGUISTIC S OCIETY. All rights reserved. No part of this book may be reproduced in any form, by print, photoprint, microfilm, or by any other means, without written permission from the publisher. Acknowledgments The Chicago Linguistic Society would sincerely like to thank the following sponsors for their generous financial support of this year’s meeting: • The Center for Leadership and Involvement • The Franke Institute for the Humanities • The University of Chicago Graduate Council The 58th annual meeting of the Chicago Linguistic Society was held April 2224, 2022 on the University of Chicago campus. As such, it was the first in-person CLS meeting following the COVID-19 pandemic. We’d like to thank everyone who took part in this event as we navigated the return to in-person functionality. We are especially grateful to our keynote speakers for presenting their research and to all of the session presenters for sharing their work with us. We appreciate the considerable effort involved in preparing for, traveling to, and presenting at our conference, not to mention the demanding work that their research entails. We are also very grateful to all who submitted an abstract for consideration but did not have an opportunity to present. With a three-day schedule and around 250 submissions, it was necessary to leave out from the program many outstanding pieces of research. If the Chicago Linguistic Society conference maintains high standards, it is due in large measure to the notable authors who submit their work for consideration. Special thanks must also go to the faculty and students in the Department of Linguistics who offered their valuable advice and support from the very first to the very last stages of conference preparation. In addition, we want to express our deep gratitude to all of our anonymous abstract reviewers — external and internal — who provided helpful feedback to submitting authors. Thanks, too, to all the student volunteers who helped us with the conference logistics. Finally, we would like to thank all conference attendees for joining us here at the 58th annual meeting of the Chicago Linguistic Society, continuing the tradition of excellence. CLS 58 Lucas Fagen Sam Gray Quain Stephanie Reyes Irene Tang Table of Contents A SIER A LCÁZAR Modal particles and force 1 A DINA C AMELIA B LEOTU , D EBORAH F OUCAULT AND T OM ROEPER 13 Long short leaves are ‘leaves that are short that are long’: The Recursive Set-Subset Principle orders both adjectives and relative clauses A DINA C AMELIA B LEOTU AND RODICA I VAN 29 SE figure reflexives alternate with Voice in Romanian: Insights from a semi-artificial denominal paradigm P ENELOPE DANIEL 45 A new analysis of interpretive properties of the verb and subject in Spanish DOM K ATHRYN DAVIDSON Semiotic distinctions in compositional semantics 59 M ATTEO F IORINI A pragmatic characterization of po in Camuno 79 B OER F U UR underlearning of Mandarin Chinese tone 3 sandhi words 89 L INGHUI E VA G AN Question answer pairs in Hong Kong Sign Language (HKSL) 103 E MILY G ASSER 119 VRK mutation in West Papua: Phonological variation across time and space J ULIE G ONCHAROV AND G URMEET K AUR Polarity, homogeneity, and authority 139 A NNA G RABOVAC Maximization of the concord domain 153 J OHANNES H EIN , C ORY B ILL , I MKE D RIEMEL , AURORE G ONZALEZ , 167 I VONA I LI Ć AND PALOMA J ERETI Č Negative concord in the acquisition of English and German: Some results from a corpus study S HIORI I KAWA , A KITAKA YAMADA AND YOICHI M IYAMOTO 183 Japanese clausal argument ellipsis and embedded clause periphery ROBIN J ENKINS Covert raising and finite ECM in Japanese, Turkish and Uyghur 199 JAMES P. K IRBY 211 Comparative tonal text-setting in Mandarin and Cantonese popular song K AZUYA K UDO AND KOJI S HIMAMURA 231 On the adjectivalizer -si in the reduplicated and deverbal adjectives in Japanese E LANGO K UMARAN Murrinhpatha number mismatch as partial Agree 243 S I K AI L EE On syntactic tenselessness in Singlish: Evidence from eventivity 255 S OO -H WAN L EE Linearly adjacent allomorphs and syntactic copies 267 M INGYA L IU AND S TEPHANIE ROTTER 277 Modal concord in American and British English: A register-based experimental study A IDAN M ALANOSKI Genre formation as enregisterment 289 A NDREW M C I NNERNEY 301 Two strategies for PP-fronting from certain island environments in English J OHN M IDDLETON Predicate landing sites in verb-initial languages 315 YOSHO M IYATA 329 Argument for verb-stranding ellipsis by blocking verb-raising in Japanese A NDA N EAGU 345 Optionality and variation in Camuno interrogatives: a syntactic analysis JACK P RUETT 355 Representing Irish mutations in Distributed Morphology and Optimality Theory JACOB R EED Text-setting tradeoffs in English-language popular incorporating musical constraints into OT analysis 373 music: C RISTINA RUIZ A LONSO 385 On the article el heading finite clauses in Spanish: a semantic proposal and its contribution to the clause S TARR S ANDOVAL , DANIEL G REESON AND M ARCIN M ORZYCKI Instrument terms, bare singulars, and event kinds 403 YOSUKE S ATO 415 PF-LF domain mismatches under ellipsis and the non-simultaneous transfer hypothesis S HU - HAO S HIH 427 Segmental and suprasegmental cues in competition: A nonce word experiment N IANPO S U AND H ELENA A PARICIO NPI licensing and intrusion effects in Japanese 443 M AHO TAKAHASHI AND G RANT G OODALL 455 Island sensitivity with relativization in Japanese: The case of double relatives T OMOYA TANABE AND RYOICHIRO KOBAYASHI Remarks on verb echo answers and head movement in Japanese 467 FANGHUA Z HENG 479 A featural analysis of Mandarin classifiers and plural morpheme men and its implications on the semi-lexicality of Mandarin Chinese Z IHUI Z HU AND A N N GUYEN 501 The interaction between structure, discourse, and prosody in whquestions in English L AL Z IMMAN 517 What is a gendered voice? Intersectional perspectives on ethics, empiricism, and epistemologies in the sociophonetics of gender Proceedings of CLS 58 (2022), 1-13 © Chicago Linguistic Society 2023. All rights reserved. 1 Modal particles and force1 Asier Alcázar University of Missouri-Columbia 1 Introduction German modal particles (MPs) have a dual function as markers of clause type and modifiers of its illocutionary force (Coniglio & Zegrean 2012: 8, and references therein). MPs relate to the middle field of German syntax and each particle contributes complex pragmatic meaning (Cardinaletti 2001). For interrogatives, MP meaning can be approximated or paraphrased in certain interpretations2 (schon, 1 cf. Bayer & Obenauer 2008: 5, ex. 27; [den]n, 2 cf. Bayer 2014: 15, ex. 32b). (1) Wo wird Klaus schon sein where will Klaus Prt be ‘Where will Klaus be? (There is no doubt that he is in…)’ (2) Homna-n däi aa a Haus? have -N they also a house ‘…Do they also have a house? (I am wondering)’ Facets of the meaning of these interrogatives is recognizable. In (1), the answer is known and is thus interpreted as rhetorical, while in (2) the speaker wonders.3 But MP meaning is diverse and elusive. Beyond German, MPs are found in languages with or without a middle field, such as Dutch, mainland Scandinavian languages, Italian and Romanian (Abraham 1991, Cardinaletti 2011, Coniglio & Zegrean 2012). Epistemic and evaluative modality are common (Munaro & Poletto 2003). MPs introduce several puzzles. This paper will focus on two of them. First, MPs call for representation of force in syntax (Rizzi 1997, Cinque 1999, pace Portner & Zanuttini 2003, Portner 2004) because their distribution is clause-sensitive and they are illocutionary modifiers (1-2). Second, MPs antagonize our understanding of grammaticalization as generalization of meaning. Although MP meaning bleaches relative to their lexical sources, as expected, this is nonetheless “coupled with a [paradoxical] rise in pragmatic complexity” (Abraham 1991: 371).4 In this paper, I propose to view these two properties as related. I assume that force is represented in syntax as a verbal projection (with Speas & Tenny 2003, Alcázar & Saltarelli 2014). I propose that illocutionary modification is a syntactic 1 My thanks to the anonymous reviewers for their comments and suggestions, and to the organizers for kindly allowing hybrid sessions during the transition into a new normal. 2 On special meanings in questions and special marking, see Obenauer (2004), Alcázar (2017). 3 Other facets are not. For instance, Bayer (2014) explains denn denotes the speaker is concerned. 4 As, for example, the nuanced pragmatic meaning in interrogatives (1-2) and other clause types. 2 ASIER ALCÁZAR operation, namely adverb incorporation (see Rivero 1992). Note that, in German & Italian, the lexical source of MPs is primarily adverbial (Cardinaletti 2011). This paper presents data from a Spanish adverb (3: acaso ‘maybe’) that functions like an MP. Acaso can work as a clause-sensitive illocutionary modifier (4: rhetorical Q marker, see Escandell Vidall 1999: 3971, and Sánchez López 1999: 2608). (3) varios Estados -como Canadá, Brasil, Polonia, acaso Italia junto a España- librarán la batalla diplomática ‘several nations—such as Canada, Brazil, Poland, maybe Italy with Spain—will fight it out diplomatically’5 (4) ¿Se justifica acaso tales ausencias si se observa quienes en cambio, de su mismo ámbito, han sido distinguidos? Obviamente, no. ‘Can you justify [x]6 such absences if you consider those who, by contrast, in their same field, have been recognized? Obviously, no.’7 Corpus data reveals acaso (4) shares some of the characteristic properties of MPs. These include a relation to finiteness, clause-typing and being a root phenomenon. I propose that adverb incorporation changes the intension of force due to gaining additional features from the adverb, which in turn restricts the pragmatic contexts to those only compatible with said features. For example, acaso can restrict the interpretation of questions to contexts where the speaker’s epistemic stance is high (4) or low certainty (wondering Q, 8-9). By contrast, neutral Qs would be incompatible with Force when combined with these epistemic features. This analysis is inspired in part by Moyna’s (2011) account of Spanish ADV+V compounds (e.g., malvender [badly.sell] ‘sell off cheap’). ADV+V compounds present limitations in pragmatic interpretation that will be illustrated later. Yet the same adverb in a V ADV sequence (vender mal ‘sell off cheap’) does not show these restrictions. I argue this contrast generally parallels that between unmarked interrogatives, which are unrestricted in interpretation (e.g., information-seeking, rhetorical, wondering…) vs. MP-marked ones, which observe limitations in their pragmatic interpretation (1-2, 4; and 8-9 below). This paper presents a corpus study of acaso sampled from the journalistic prose section of CREA, a publicly available corpus. It reviews the functions attributed to acaso in the grammatical tradition. Then it introduces the corpus and annotation. Acaso is shown to behave as an MP in certain contexts. Adverb incorporation is briefly discussed, along with ADV+V compounds. The analysis relates the rise in pragmatic complexity to illocutionary modification via adverb incorporation. 5 Ex. 287. El País, Feb. 20, 1980, “La vuelta de EEUU a la OIT,...” [The return of USA to ILO…]. In corpus exs. glosses will be spared due to limitations in space. 6 [x] indicates the approximate position of acaso. Acaso’s position in questions is flexible. 7 Ex. 557. El Cultural, Nov. 21, 2003: Premios ¿Nacionales? [National? Awards]. The writer denounces Plácido Domingo and Alfredo Kraus had received no national awards yet. Modal particles and force 3 2 Acaso in the descriptive tradition Acaso is used as a stand-alone noun, within prepositional phrases, as well as an epistemic adverb (Bello 1988: 329, Alarcos Llorach 2000: 162, Quilis, Hernández, de la Concha 1972: 122, 132-3, Hernández Alonso 1986: 208, Kovacci 1999: 705, 755-6). As a noun, acaso means ‘chance, random, unforeseen event’. It can present within the PP por un acaso ‘by chance’. Although both may be obsolete relative to the type of text studied. More frequently, acaso is found in the PP por si acaso ‘in case’, in the adverbial si acaso ‘at most’, and the epistemic adverb acaso ‘maybe’. The original status of acaso as a noun is debatable, but several sources point to caso ‘case’, through various mechanisms. In its dictionary, the Royal Spanish Academy assumes the etymology for acaso is caso (DRAE 2022, acaso), but it does not explain why. Montolío (1999: 3677) proposes that acaso relates to caso by analogy with final clauses: te lo digo para el caso de que no lo sepas ‘I tell you should you not know’ [lit. “for the case that you don’t know”]. Alarcos Llorach (2000: 133) lists acaso as an adverbial (‘locución adverbial’), formed by the preposition ai ‘to’ and noun caso, like the prepositions enfrente ‘in front’ [in+front] or encima ‘on’ [in+(hill)top]. Finally, Alcázar (2017) views acaso as a phonological erosion of an earlier, clausal form of ‘in case’: por si hiciera/viniera al caso [lit. “for if (it) made/came to.the case”] ‘in the case that it were fitting/necessary’. ‘A-’ would be the preposition (‘to’) selected by either verb. What lexical form(s) of acaso lead to MP uses is a non-trivial question. Sources could be multiple. Por un acaso ‘by chance’ is a good literal fit for one of acaso’s meanings as a Q marker, similar to the by any chance question in English, where a comparable adverbial marks or identifies the question.8 A possible development may have occurred where acaso extended from by any chance Qs to other types of questions where the speaker has low certainty, such as wondering Qs (8-9). At the same time, because epistemic modality figures prominently in interrogative uses, the epistemic adverb ‘maybe’ seems a fitting candidate. The epistemic adverb can express low certainty. On the other hand, the use of acaso to mark high certainty (4) does not appear to have a direct possible link. This is unfortunate as high certainty (rhetorical Qs, biased Qs) is by far the most frequent use (>80%). While a direct or unique lexical source may not be established for MP uses, both candidates present similarities with the source of German & Italian MPs. In Spanish too an adverb acaso ‘maybe’ and an adverbial por un acaso ‘by any chance’ are the likely sources of illocutionary modifiers in interrogatives. 3 Corpus and annotation The corpus sample was extracted from CREA9, an online corpus published by the Royal Spanish Academy. The sample consists of 1262 sentences representing journalistic prose (cf. periódicos ‘newspapers’). Examples date from 1976 to 2004. 8 The by any chance question type has special properties in English (see Sadock 1971). CREA stands for Corpus de Referencia del Español Actual ‘Modern Spanish Reference Corpus’. It can be accessed in the following link: https://corpus.rae.es/creanet.html. 9 4 ASIER ALCÁZAR The sample is restricted to documents published in Spain. The sample encompasses all available examples within the search parameters (journalistic prose, Spain). I developed software in Python 3.9 to support this project. Software includes an annotation tool, scripts to automatically download and preprocess corpus data, capture metadata (year, author, topic…), and various other tools to postprocess examples, correlate variables, obtain frequency counts, etc. Due to limitations in space, annotation categories and their values will not be listed exhaustively. Some include grammatical/lexical function, sentence type (declarative, exclamative…), clause type (main, coordinate, subordinate…), sentence polarity, polarity of the presumed answer, scope, speaker orientation and verb morphology, to include finiteness and mood. 4 Lexical vs. grammatical uses in corpus data The corpus sample is a combination of lexical and grammatical uses reported in the descriptive tradition, but grammatical uses also include roles that may be described as an MP use. The boundaries between lexical and grammatical uses can be murky, subject to dialectal variation and interpretation. Within lexical uses, we find an instance of acaso as ‘chance’, 112 cases of acaso in the PP por si acaso ‘in case’ and 46 examples of si acaso ‘at most’. While the latter is considered lexical (see Montolío above, or Diccionario Panhispánico de Dudas 2005), it shares characteristics with acaso in questions, in that it points to probability extremes based on the speaker’s epistemic stance. Informally stated, si acaso ‘at most, at best’ points to the maximum level at which a proposition will be true, if at all. This level is presumed to be low, at best. (5) No vamos a solucionar nada, si acaso empeorar las cosas. ‘We are not going to solve anything, make things worse at best.’10 Si acaso is speaker oriented, it evaluates relative to the beliefs of the speaker of the utterance (also epistemic adv.: 3, interrogative uses: 4, 8-9). It could be an MP.11 Within grammatical uses, we find the epistemic adv. acaso ‘maybe’ in declaratives (n = 702). In Spain, acaso can have a neutral to low probability reading, like quizá(s) or tal vez ‘maybe’. This feature helps distinguish the epistemic adv. use of acaso from other uses that indicate a high or low certainty. By contrast, in some dialects of Basque Country Spanish, acaso as an epistemic adv. has shifted towards a lower probability. This lower probability is reminiscent of si acaso ‘at most, at best’ (5), as well as low certainty Qs (8-9). In conditionals (n = 56), acaso must immediately follow si ‘if’. This is unusual because, as an epistemic adv., acaso enjoys a flexible syntactic position. This use 10 J. L. Alonso de Santos. 1981. La estanquera de Vallecas. Madrid: Antonio Machado. Quoted in Diccionario Panhispánico de Dudas (2005, acaso). 11 MPs may appear in more than one type of clause and share pragmatic features (see §5). Modal particles and force 5 is homophonous with si acaso ‘at most’. 12 If the epistemic value of acaso is low, it may be better paraphrased as ‘at all’ than ‘maybe’ (6). (6) “Respecto a "para qué sirve el teatro" si acaso sirve para algo, recuerdo un día en que le hice la misma pregunta a usted y me respondió: "Si yo lo supiera, no haría teatro".” ‘Regarding “what is theater good for” if [x] it is good for anything at all, I remember a day when I asked you the same question and you answered: “If I knew it, I wouldn’t do theater.” ’13 Conditional si acaso licenses a form of ellipsis (7) that is implicitly affirmative ‘if at all’. In these cases, ‘maybe’ does not seem to be a good translation either. (7) Lo celebrará, si acaso, con Brad Pitt. ‘He will celebrate it, if [he wins]/if at all, with Brad Pitt.’14 I chose to separate conditionals from the epistemic adv. due to a different syntactic distribution, elliptical properties, and meaning. That said, in Basque Country Spanish, both the epistemic adv. and conditional uses may point to low or null probabilities. For this reason, it is harder to distinguish them in these dialects. In questions (n = 347), the corpus sample presents a series of contexts that can be classified as high or low certainty. As noted, the corpus sample is exhaustive in that it contains all instances of acaso for journalistic prose, for documents published in Spain, from 1976 to 2004. It is thus significant that acaso, in its role as a question marker, is not utilized to mark other types of questions. Of the absent types, it is especially noteworthy that acaso is not able to mark plain information-seeking questions, that is, questions where the epistemic state of the speaker is not explicitly disclosed. Having said that, at least one Q subtype marked by acaso could be construed as neutral (9: wondering-ignorance Qs, see below). Within the first set, we find rhetorical and biased Qs. Rhetorical Qs are by far the most frequent (n = 239). A lower percentage of questions associate with contexts of low certainty (<20%). Within these, the most frequent forms are wondering questions. These may be subdivided into two groups relative to the certainty the speaker may show in context. In some wondering Qs, the speaker weakly favors some options over others (vs. a biased Q, where preference is strong). This subtype may be regarded a wonderingspeculative Q. For example, in (8), Andy Warhol wonders in his diary why Lou Reed did not invite him to his wedding. 12 Si acaso ‘at most’ may arise as intra-clausal grammaticalization: the antecedent of a conditional clause is reanalyzed as a non-clausal adverbial within a declarative (see ex. 7 in main text). 13 Ex. 638. ABC Cultural, Nov. 22, 1996. ¿Modificar, qué? [Modify, what?]. 14 Ex. 122. El País/El País de las Tentaciones, May 30, 2003: El cuerpo de ‘Indi’ [‘Indi’s body’]. The ex. refers to World Stunt Awards, which recognize stuntmen. 6 ASIER ALCÁZAR (8) "¿Acaso temía -escribió- que me presentara con demasiados amigos o que comiera o bebiera demasiado?" ‘[x] Did he fear –[Warhol] wrote- that I would show up with too many friends or that I ate or drunk too much?’15 In other instances, the speaker may absolutely ignore the answer to the question. In this respect, this type could be alternatively analyzed as a neutral Q, since the speaker has no epistemic preference. In (9) the main character in the police drama Voces ‘Voices’ wonders about a whodunit. Acaso marks the first and last Q.16 (9) ¿se trata acaso del amigo con aspecto de macarra?, ¿de la hermana con problemas psiquiátricos?, ¿del compañero de ésta, alcohólico y violento?, ¿del padrastro arquitecto y escultor?, ¿o acaso?... ‘is it [x] the friend that looks like a thug? the sister with psychiatric problems? The sister’s significant other, an alcoholic and violent? The stepfather who is an architect and sculptor? Or [x] …?’17 There are three arguments in favor of viewing wondering-ignorance Qs as a special type. First, acaso provides a wondering quality that a neutral Q would not impose. Second, the use of acaso underscores speaker uncertainty about what the answer might be. Third, in the larger context of the corpus, acaso Qs associate with high or low speaker certainty over 95% of the time.18 If wondering-ignorance Qs were neutral Qs, they would constitute a minority of cases. An explanation would be required for acaso to mark high certainty (>80%) and low certainty. In sum, acaso presents a combination of lexical and grammatical uses in the corpus. In interrogatives, acaso identifies the epistemic stance of the speaker as high or low certainty. These values associate with rhetorical Qs and wondering Qs, among other less frequent types, and thus acaso could be seen as an illocutionary modifier. In declaratives–si acaso ‘at most’, conditional si acaso–share some properties with acaso in interrogatives in that they are speaker-oriented and can point to a low or potentially null probability, relative to the dialect. The next section explores further arguments for an analysis of acaso as an MP noting characteristics in questions similar to German & Italian MPs. 5 Arguments for acaso to feature an MP role The introduction mentioned that MPs have a dual role as clause-type markers and illocutionary modifiers. These roles are more easily assessed against established 15 Ex. 1003. La Vanguardia, Dec. 3, 1994: Potencialmente peligroso [Potentially dangerous]. Though attested, acaso rarely marks constituent Qs (n = 3). It is therefore not surprising that the constituent Qs in between are not marked by acaso. 17 Ex. 34. La Vanguardia, Jun. 6, 1995: [no title information], M. Ángeles Cabre. 18 The remaining set comprises doubtful uses that could be explained by various factors, such as acaso being not a question marker but the epistemic adv. instead. 16 Modal particles and force 7 grammatical categories or labels, such as “rhetorical question”. This section will present three arguments in favor of viewing the role of acaso in interrogatives as an MP. These arguments concern clause-typing, being a root CP phenomenon, and a strong relation to finiteness. This is done by comparison to the epistemic adv. ‘maybe’ in declaratives, though certain uses in declaratives require further study. First, German & Italian MPs can be limited to specific clauses or just one type. For example, in Bavarian German -n (phonologically eroded denn) is an MP in polar Qs (2), which can also exceptionally be found in certain embedded contexts (see Bayer 2014). The Italian MP pur(e), for example, can be found in imperatives as well as declaratives (see Coniglio & Zagrean 2012). For additional examples see Munaro & Poletto (2003), Cardinaletti (2011). Corpus data shows acaso is a clause type marker as well. Acaso is restricted to polar Qs. There is a single instance of a pronominal Q, but it is a likely an editorial mistake (346 vs. 1). Within polar Qs, acaso marks the singleton type. By singleton type I mean a question that does not present a disjunction (vs. the alternative Q: do you want coffee or tea?) or partial Q (e.g., a constituent Q). Alternative and constituent polar Qs are attested in the sample, with 6 and 3 cases respectively. But, by comparison, the question is of a singleton type over 95% of cases. Second, in German & Italian, MPs are considered a root phenomenon, as in root interrogatives or imperatives (Abraham 1991). Although there are examples of MPs in subordinate clauses, their syntactic distribution may still be limited by root properties (central vs. peripheral adverbial clauses, Haegeman 2002; see Coniglio & Zagrean 2012 for discussion). In corpus data, there is a stark difference between acaso in declaratives vs. interrogatives with respect to root contexts. In declaratives, there is a small preference for root contexts (n = 466) over subordinate contexts (n = 366). While in interrogatives there are only three cases of subordinate contexts (one of them is a quoted question). In interrogatives acaso clearly marks root Qs. Third, a relationship between force and finiteness is proposed for German & Italian MPs (Coniglio & Zegrean 2012, Bayer & Obenauer 2012, Bayer 2014). In corpus data, this alignment is visible if we compare declaratives and interrogatives again. In declaratives, finite and non-finite/verbless structures are about even (51.5% vs. 48.5%, respectively). In interrogatives, by contrast, the percentage of finite verbs is very high: 97.4%. In sum, acaso in interrogatives shares three additional characteristic properties of MPs in German & Italian beyond its role as an illocutionary modifier in Qs. These characteristics can find an explanation if acaso in interrogatives is an MP. 6 Adverb incorporation Adverbs and adverbials are sources for MPs in German, Italian & Spanish. This begs the question of why said sources would be so successful. There is a syntactic operation known as adverb incorporation that may help explain the prevalence of adverbs and adverbials. Furthermore, the existence of this operation may provide a syntactic account for illocutionary modification on the added assumption that Force 8 ASIER ALCÁZAR is a verbal projection. This section introduces adverb incorporation. The following one presents a pragmatic characteristic of adverb incorporation in Spanish. This second characteristic may offer clues regarding the ‘pragmatic complexity’ of MPs that contradicts our understanding of grammaticalization as bleaching. Adverb incorporation (Rivero 1992, Alexiadou 1997) is a process analogous to noun incorporation and preposition incorporation (Baker 1988), where an argumental head within VP merges with V. For example, in (11) the noun seuan ‘man’ incorporates into V in Southern Tiwa. (10) is the same example, without incorporation, and it has the same meaning as (11). Similarly, in (13) the adverb anápoda ‘upside down’ incorporates into V in Modern Greek (4-5 cf. Rivero 1992: 292, ex. 2; 289, ex. 1). Ex. (12) is the same example without incorporation, but it has the same meaning as (13). (10) Seuan-ide timu-ban Man SUF 1sS/AO see-PAST ‘I saw the/a man.’ (11) tiseuan-mu-ban 1sS/AO man-see-PAST ‘I saw the/a man.’ (12) I María tha to girísi the Mary will it turn ‘Mary will turn it upside down’ (13) I María tha to anapodo-girísi the Mary will it turn upside+down ‘Mary will turn it upside down’ anápoda upside+down Relative to the language studied, adverb incorporation can be limited to adverbs that are subcategorized by V and merge within VP (Modern Greek, Nahuatl; e.g. manner, directional and aktionsart adverbs, Rivero 1992). Later studies showed adverb incorporation can be seemingly more productive, inclusive of higher-level projections, such as aspect and epistemic modality (Eskimo, Classical Tagalog, Alexiadou 1997, Ch. 5 and references therein). At the time of writing, I am not aware of cases of incorporation into projections of the left periphery, such as complementizers, for example. I would argue that illocutionary modification is a sign that adverbs may incorporate that high in the structure by considering the pragmatic parallels between MPs and a certain set of adverbs in Spanish, which are discussed next. 7 Pragmatic differences in adverb incorporation in Spanish Val Álvaro (1999: 4824), and specially Moyna’s (2011) monograph, a diachronic corpus study of Spanish compounding, restrict ADV+V compounds to manner Modal particles and force 9 adverbs—Moyna additionally notes a few examples of menos ‘less, under-‘, an adverb of quantity (p. 100). Manner adverbs include mal ‘poorly, wrongly, badly, mal-, mis-’ and bien ‘well’. The most productive is mal. English features mal in French borrowings, such as malfunction or maladapt (mis- has a similar meaning: misappropriate). Earlier exs. with bien could appear undipthongized, as in bendecir [well.say] ‘bless’ (Moyna, p. 100). From the standpoint of the adverb incorporation cases Rivero/Alexiadou discussed, it is significant that the Spanish compounds belong to the manner class. Manner adverbs are subcategorized by V and occur within VP. Relative to their adverbial class, then, manner adverbs can constitute a canonical case of adverb incorporation. Moyna makes the observation that all of Vendler’s aspectual classes are represented in ADV+V compounds (p. 100), and that there are no changes to predicate-argument structure, as would be expected, with a few exceptions. Thus, while the adverbial class is restricted to manner adverbs, the predicates that combine with them have no restrictions. Moyna notes that “The presence of [Adv + V] compounds across the Romance family is evidence of the pattern's early origin and inherited status [from Latin cf. Brunet 2005].” (2011: 103). My claims on incorporation and its meaning are limited to Spanish for this paper but they could extend to other Romance languages (see Moyna: 103-4, references therein on Adv + V in Romance & Latin). Val Álvaro and Moyna see ADV+V morphologically, as compounds. Moyna specifically views ADV+V as a special type with syntactic properties (they can create negative-intensive pairs, as in syntax: negation-emphatic do). Moyna’s argumentation for ADV+V to share properties with syntax is generally compatible with the idea that they stem from a syntactic operation: adverb incorporation. I will now briefly review two of Moyna’s arguments. First, in other types of compounds “the head can be on either side of the non-head” (p. 102). For example, you could find N+ADJ compounds (hierbabuena [herb.good] ‘mint’) along with ADJ+N compounds (buenaventura [good.luck] ‘good luck’). Nonetheless, a distinguishing property of ADV+V compounds is that they lack V+ADV counterparts. A second argument is that ADV+V compounds continue to arise past the historical point where manner adverbs could not be positioned preverbally. Only V ADV is available, but it does not yield comparable V+ADV compounds. This fact alone strongly suggests that the origin of ADV+V compounds is syntactic. Incorporation of the adverb into V would be independent of the relative position of manner adverbs over the years. Spanish ADV+V compounds resemble those studied by Rivero/Alexiadou in that the meaning of the incorporated and unincorporated form is the same (10=11: Southern Tiwa, 12=13: Modern Greek). While these also holds true of the Spanish data, Moyna observes two differences between ADV+V vs. V ADV. The first concerns the interpretation of these forms relative to context. “the compound always has a more restricted meaning than the phrase. If we compare vender mal and malvender, for example, the former has a much wider 10 ASIER ALCÁZAR range of possible meanings than the latter, including the possibility of selling an item at the wrong time, using the wrong procedure, illegally, and so on. […] By contrast, malvender ‘badly-sell’ is restricted to the meaning of selling for too little in return.” (Moyna 2011: 102) Moyna proposes that a solution to this difference in semantic narrowing can be found if the adverb modifies the intension of the verb in the compound form. In the V ADV phrasal sequence, this restriction would not be present. The second is that ADV+V can create pairs of negative-intensive compounds that do not arise in V ADV. For example, malherir [badly.injure] ‘to injure seriously’ conveys intensity, but its V ADV counterpart herir mal ‘to injure wrongly’ does not (see pp. 101-2). If negation and intensity are verbal properties, adverb incorporation would allow these to arise in ADV+V. 8 Analysis I propose that adverb incorporation is a plausible syntactic explanation for the unusual pragmatic interpretation of sentences marked with MPs. At the same time, adverb incorporation may provide an independent rationale for the source of MPs to be adverbs and adverbials disproportionately. Adverb incorporation can also relate illocutionary modification to the paradoxical rise in ‘pragmatic complexity’. The analysis rests on the following assumptions. First, Force is a syntactic projection (Rizzi 1997, Cinque 1999, Coniglio & Zegrean 2012; pace Portner 2004, Portner & Zanuttini 2003), positioned either in the left periphery, or high in the inflectional domain, with one or multiple positions, relative to its implementation. Second, Force is a verbal projection (Speas & Tenny 2003, Alcázar & Saltarelli 2014), where speaker and addressee variables are hosted. Third, there is an independently necessary syntactic operation known as adverb incorporation [into V] (Rivero 1992, Alexiadou 1997). While this operation is not attested in the left periphery in the literature examined, illocutionary modification of the sentence by adverbs and adverbials (MPs) may be seen as evidence of it. The second assumption views ForceP as a vP[VP] shell structure, which would contain a VP layer. The reason for adverbs and adverbials to be MPs in such high numbers could be explained by their ability to modify Force adverbially. On the other hand, illocutionary modification would follow from the feature specification of Force being enriched with the features of the adverb or adverbial. In the case of acaso, as an MP it contributes epistemic features where the speaker attributes either a high or low probability for a proposition to be true. This prevents normal informationseeking questions from being compatible with acaso. Finally, the gains in ‘pragmatic complexity’ that run counter to grammaticalization as semantic bleaching can also find a rationale under adverb incorporation. The apparent gains in ‘pragmatic complexity’ are the direct result of feature enrichment of Force with the features of the adverb or adverbial. This theory would predict that the pragmatic meaning of MPs would be diverse and difficult to predict, since it would be a corollary of the incorporation of the adverb or adverbial into Force. Modal particles and force 11 This theory is compatible with observations like that of Bayer (2014) on Bavarian -n (denn). -n has become a pronominal Q marker, while it remains an MP in polar Qs. If the MP incorporates into Force, it may eventually reanalyze as it. 9 Conclusion This paper has presented data from a corpus study of acaso. It has argued that acaso presents similarities with German & Italian modal particles. In interrogatives, acaso is an illocutionary modifier and clause type marker, it is a root phenomenon, and it has gained a relation to finiteness. This paper has proposed that two puzzling properties of MPs may be related, namely illocutionary modification and MP’s ‘pragmatic complexity’ (Abraham 1991). The proposed relation is articulated via a syntactic operation: adverb incorporation (Rivero 1992). Adverb incorporation is a process in part analogous to noun incorporation and preposition incorporation (Baker 1988). In Spanish, ADV+V compounds have special structural properties that make them a different type of compound (Moyna 2011). I have proposed to view these compounds as cases of adverb incorporation. Said compounds in Spanish have pragmatic restrictions in the interpretation of the adverb that phrasal sequences of V ADV do not observe, in spite of sharing the same meaning. This paper has proposed that, if Force is viewed as a verbal projection (Speas & Tenny 2003), adverb incorporation can change its intension. This would be responsible at once for both the ‘pragmatic complexity’ gained, and the illocutionary modification function of MPs, both resulting from the feature contribution of the adverb to Force. References Abraham, W. 1991. The grammaticization of the German modal particles. In Approaches to Grammaticalization, Volume II, ed. by E. C. Traugott & B. Heine, 331–380. Amsterdam: Benjamins. Alarcos Llorach, E. 1999. Gramática de la Lengua Española. Madrid: Real Academia Española. Alcázar, A. 2017. A syntactic analysis of rhetorical questions. In Proceedings of the 34th West Coast Conference on Formal Linguistics, ed. by A. Kaplan, A. S. Kaplan, M. K. McCarvel, & E. J. Rubin, 32–41. Somerville, MA: Cascadilla Proceedings Project. Alcázar, A., & M. Saltarelli. 2014. The Syntax of Imperatives. Cambridge: Cambridge University Press. Alexiadou, A. 1997. Adverb Placement: A Case Study in Antisymmetric Syntax. Amsterdam: John Benjamins. Baker, M. C. 1988. Incorporation: A theory of grammatical function changing. Chicago: University of Chicago Press. Bayer, J. 2014. From modal particle to interrogative marker: A study of German denn. In Functional Heads. Papers Presented to Guglielmo Cinque on the Occasion of his 60th Birthday, ed. by L. Brugè, A. Cardinaletti, G. Giusti, N. Munaro, & C. Poletto. New York: Oxford University Press. Bayer, J., & H.-G. Obenauer. 2011. Discourse particles, clause structure, and question types. The Linguistic Review 28.449–491. Bello, A. 1988. Gramática de la Lengua Castellana Destinada al Uso de los Americanos. Madrid: Editorial Arco. Bosque, I., & V. Demonte. 1999. Gramática Descriptiva de la Lengua Española, Vols. 1-3. Real Academia Española. 12 ASIER ALCÁZAR Brunet, C. 2005. Bene facere: Une ou deux lexies? In La composition et la préverbation en latin, ed. by C. Moussy, 197–207. Paris: Presses de l'Université Paris-Sorbonne. Cardinaletti, A. 2011. German and Italian modal particles and clause structure. The Linguistic Review 28.493–531. Cinque, G. 1999. Adverbs and Functional Heads. New York: Oxford University Press. Coniglio, M., & I. Zegrean. 2012. Splitting up Force: Evidence from Discourse Particles. In Main Clause Phenomena: New Horizons, ed. by L. Aelbrecht, L. Haegeman, & R. Nye, 229-255. John Benjamins, Amsterdam. Escandell Vidal, M. V. 1999. Los Enunciados Interrogativos: Aspectos Semánticos y Pragmáticos. In Bosque & Demonte, Vol. 3, 3929-3992. Haegeman, L. 2002. Anchoring to speaker, adverbial clauses and the structure of CP. In Georgetown University Working Papers in Theoretical Linguistics 2, 117-180. Hernández Alonso, C. 1986. Gramática Funcional del Español (2ª ed.). Gredos: Madrid. Kovacci Conicet, O. 1999. El Adverbio. In Bosque & Demonte, Vol. 1, 705–86. Montolío Durán, E. 1999. Las Construcciones Condicionales. In Bosque & Demonte, Vol. 3, 3643– 3738. Moyna, M. I. 2011. Compound Words in Spanish: Theory and History. Amsterdam: John Benjamins. Munaro, N. & C. Poletto. 2003. Ways of clausal typing. Rivista di grammatica generativa 27.87– 105. Obenauer, H.-G. 2004. Nonstandard wh-questions and alternative checkers in Pagotto. In The syntax and semantics of the left periphery, ed. by H. Lohnstein and S. Trissler, 343–384. Berlin: Mouton de Gruyter. Portner, P. 2004. The semantics of imperatives within a theory of clause types. In Proceedings of Semantics and Linguistic Theory 14, ed. by K. Watanabe & R.B. Young, 235–52. Ithaca, NY: CLC Publications. Quilis, A., C. Hernández, & V. G. de la Concha. 1972. Lengua Española: Iniciación Universitaria. Valladolid: Industrial Litográfica. Rivero, M. L. 1992. Adverb Incorporation and the Syntax of Adverbs in Modern Greek. Linguistics and Philosophy 15(3).289–331. Rizzi, L. 1997. The Fine Structure of the Left Periphery. In Elements of Grammar, ed. by L. Haegeman, 289–330. Dordrecht: Kluwer. Real Academia Española. 2005. Diccionario Panhispánico de Dudas. Madrid: Real Academia Española, Asociación de Academias de la Lengua Española y Santillana Ediciones Generales. Real Academia Española. Diccionario de la lengua española, 23.ª ed., [versión 23.6 en línea]. <https://dle.rae.es> [December 2022]. Sánchez López, C. 1999. La Negación. In Bosque & Demonte, Vol. 2, 2561-2634. Sadock, J. 1971. Queclaratives. In Papes from the Seventh Regional Meeting of the Chicago Linguistic Society, 223–232. Chicago: CLS. Speas, M. & C. Tenny. 2003. Configurational properties of point of view roles. In Asymmetry in Grammar, ed. by A. DiSciullo, 311–43. Amsterdam: John Benjamins. Val Álvaro, J. F. 1999. La composición. In Bosque & Demonte, Vol. 3, 4757–4842 Zanuttini, R. & P. Portner. 2003. Exclamative Clauses: At the syntax-semantics interface. Language 79(1).39–81. CREA corpus Banco de datos (CREA). Corpus de referencia del español actual. Madrid: Real Academia Española. https://corpus.rae.es/creanet.html Proceedings of CLS 58 (2022), 13-28 © Chicago Linguistic Society 2023. All rights reserved. 13 Long short leaves are ‘leaves that are short that are long’ The Recursive Set-Subset Principle Orders Both Adjectives and Relative Clauses Adina Camelia Bleotu1, Deborah Foucault2 and Tom Roeper2 University of Bucharest1 and University of Massachusetts Amherst2 1 Introduction Our study investigates how adjectival and relative clause modifiers referring to sets and subsets are ordered. We argue for the existence of a Recursive Set-Subset Ordering Principle (RSSO), which directly derives from Merge, and which applies to adjectives (Adjs), relative clauses (RCs), other modifiers, or any combination thereof. (1) Recursive Set-Subset Ordering Principle (RSSO): Set modifiers are merged earlier to the noun than Subset modifiers. The RSSO automatically provides a compositional set-subset interpretation for the hierarchical recursive syntax. We investigate this hypothesis experimentally by testing native American English speakers’ understanding of: (2) a. adjectives: [longsubset [shortset leaves]] b. RCs: [leaves [that are shortset ][that are long]subset] c. adjective and RC: [[shortset leaves] that are longsubset ] Importantly, the RSSO is a hierarchical constraint, not a constraint on linear order. It can account for the ordering of pre- and postnominal recursive modifiers within the same language (English), since, given the prenominal attachment of adjectives in English, their linear ordering will be [Adjsubset [Adjset Noun]], and, given the postnominal attachment of RCs, their linear ordering will be [[Noun RCset] RCsubset]. When a noun combines with an adjective and a RC (as in short leaves that are long), two orders should in principle be possible: [[Adjsubset [Noun RCset]] or [[Adjset Noun] RCsubset]. The syntax of RCs is complex and varies across languages (see de Vries 2002, Cinque 2020, for instance). In many cases, it relates to the determiner phrase. However, in cases where both an adjective and an RC are present, we need to test experimentally whether the RSSO applies to the adjective 14 BLEOTU, FOUCAULT AND ROEPER or the RC. If the noun is open to direct adjective modification, but the RC must attach to a NP, then the ordering should be [[Adjset Noun] RCsubset], associating the adjective with a set interpretation and the RC with a subset interpretation (see 3): (3) NP 2 NP RCsubset 2 that are long APset N short leaves If both the adjective and the RC can directly attach to a noun, then both (3) and (4) should be possible structures associated with short leaves that are long: (4) NP 2 APsubset NP short 2 N RCset leaves that are long Interestingly, the RSSO accounts for the ordering of modifiers not just intralinguistically but also cross-linguistically. In Romance languages, for instance, where both Adjs and RCs are postnominal, the linear ordering will be [Noun Modifierset ] Modifiersubset] for all modifiers. 2 Background 2.1 Cognitive Adjective Ordering Restrictions Research on the ordering of adjectives has focused on cognitive Adjective Ordering Restrictions (AORs), which order adjectives according to the (kind of) property they express. Two examples of cognitive AORs are given below in (5a, b): an example of a fine-grained ordering, where several properties are considered (5a), and an example of a more general ordering, where the driving property of adjective orders is considered to be subjectivity (5b), with more subjective adjectives being placed further away from the noun than less subjective ones. (5) a. Quality> Size> Shape>Color>Provenance (Sproat & Shih 1991, Dixon 1982, Scott 2002) b. Subjective>Non-Subjective (Scontras et al. 2017) AORs have been claimed to be (i) syntactic (Cinque 1995, 2010, 2015), (ii) semantic (Sproat & Shih 1991, Dixon 1982, Scott 2002), or (iii) cognitive, guided by subjectivity (Scontras et al. 2017). Interestingly, cognitive orderings may have Long short leaves 15 a grammatical source as well (Scontras et al. 2019), merging more objective properties closer to the noun than the more subjective ones. Importantly, AORs may differ cross-linguistically both in their directionality and nature (Cornilescu & Nicolae 2016, Leivada & Westergaard 2019, Trainin & Shetreet 2021). For instances, adjectives in Romance, are predominantly postnominal. However, there is disagreement between various studies with respect to how they are ordered. According to Cinque (2010), adjectives in Romance reflect a mirror order of English (4a), a view which supports a syntactic account of AORs. However, according to Cornilescu & Giurgea (2013), Cornilescu & Nicolae (2017), Cornilescu & Cosma (2019), adjectives in Romance may be more freely ordered, a view which goes against the idea of a syntactic hierarchy. Luciu & Bleotu (forthcoming) and Truşcǎ & Bleotu (forthcoming), for instance, have shown experimentally that Romanian native speakers accept both Color-Size and SizeColor orders, unlike British English native speakers (see 6). (6) a. Color-Size un fluture albastru mititel (Romanian) a butterfly blue tiny ‘a tiny blue butterfly’ (English) b. Size-Color un fluture mititel a butterfly tiny ‘a blue tiny butterfly’ albastru (Romanian) blue (English) One possible explanation for this cross-linguistic difference could be that AORs in terms of semantic properties (such as Color and Size) are syntactic in English, but cognitive in Romanian. Such a parametric difference could explain why adjectives are more freely ordered in Romanian, but more strictly ordered in English. Interestingly though, Grohe & Schulz (2021) have recently brought evidence supporting the idea that, even in English, the adjectival hierarchy in terms of cognitive properties may be less rigid for certain types of adjectives. In particular, adults seem to be more flexible in their ordering of Color and Shape adjectives, either placing the Shape adjective closer to the noun than the Color one (7a) or placing the Color adjective closer to the noun than the Shape one (7b): (7) a. a blue square table b. a square blue table Such findings cast doubt upon the idea that AORs are syntactic in nature. If some orders are more rigid while others are more flexible, one would have to assume that some parts of the hierarchy are syntactic, while others are cognitive. A way out would be to simply assume that AORs are cognitive cross-linguistically, but, 16 BLEOTU, FOUCAULT AND ROEPER because of usage, certain orders are more fixed in English than in Romanian or other languages. While AORs in terms of semantic/cognitive dimensions have been dealt with extensively in the literature, ordering adjectives depending on the sets/subsets they identify has received little attention in the literature. 2.2 Set Subset Ordering Restrictions Investigating the recursive order of adjectives is important because adjectives are often used by speakers to identify subsets of sets of objects. For example, let us consider the adjective sequence in (8), combining the Color adjective green with the Size adjective long: (8) long green leaves This phrase is ambiguous between (a) a coordinative reading ([long and green] leaves) and (b) a recursive reading ([long [green leaves]]). While in the default case there may not be a referential difference between the coordinative and the recursive structure, in a situation where one wants to refer to the subset of long leaves out of the set of green leaves, the recursive structure is more specific. Note that one can easily refer to long leaves out of a set of green leaves ([long [green leaves]]) or to green leaves out of a set of long leaves ([green [long leaves]]) as ‘long and green leaves’. In other words, the coordinative structure is compatible with both referents. In contrast, the recursive reading is more specific. Our investigation of recursive uses and interpretations of multiple adjectives builds on the theoretical discussion on dynamic semantics in Scontras et al. (2019), as well as on insights from Roeper & Snyder’s (2004) and Roeper’s (2011) theory of recursion. Recursion is considered one of the most important properties of Universal Grammar (Roeper 2011), a property unique to human language (Hauser, Chomsky & Fitch 2002) which is responsible for the generation of an infinite number of combinations of phrases and sentences. 1 Recursion relies on the operation of Merge, combining two elements together and making hierarchical structure-building possible. According to Roeper (2011), direct recursion (see 9) can be defined as the Merge of a syntactic category with an identical syntactic category, which thus generates itself: X => X (X). Indirect recursion (see 10) can be defined as the embedding of a syntactic category through an intermediate identical category creating a potentially infinite loop of XP phrases inside XP phrases (Chomsky 1995): XP =>X (YP), and YP => Y (XP). (9) (10) 1 long and short leaves long short leaves But see Arsenijevic & Hinzen (2010) for arguments that recursion might be an interface property rather than a (purely) linguistic one. Long short leaves 17 Our investigation focuses on how adults hierarchically order and interpret recursive adjectives and RCs which represent a case of indirect category recursion. Previous experimental work on adjectives suggests that adults are able to produce and interpret recursive adjectives according to the Set-Subset distinction. Children younger than 6, however, face some challenges with both production and comprehension, often treating recursion as coordination. This study focuses on adult data in order to inform our future work with children’s acquisition of recursive adjectives and RCs. Bleotu & Roeper (2021 a, b) showed that Romanian-speaking adults are able to produce and interpret recursive sequences of two and three adjectives specifying the same dimension (Size) in an experiment involving multiple sets/subsets of objects (see Figure 1). They were able to link flori mari mici (11a) with the subsets of flowers labelled as 3 and 4 in Figure 1, as well as link flori mici mici mari (11b) with the subset of flowers labelled as 7 in Figure 1. (11) a. 2-Level Adj Recursion: flori mari mici flowers big small ‘small big flowers’ b. 3-Level Adj Recursion: flori mici mici mari flowers small small big ‘big small small flowers’ (Romanian) (English) (Romanian) (English) Figure 1: Picture employed in Bleotu & Roeper (2021a, b) In the foundational story-based experiment that looked at recursive set/subsets of adjectives and that we have based subsequent experiments and extensions, Foucault et al. (2022) showed that American English-speaking adults are also able to produce and interpret sequences of two and three adjectives specifying the same dimension (Size) (12) a. b. 2-Level Adj Recursion: big small mushrooms 3-Level Adj Recursion: small big small mushrooms 18 BLEOTU, FOUCAULT AND ROEPER Figure 2: Picture employed in Foucault et al. (2022) Regarding sequences of adjectives specifying different dimensions, Bleotu & Roeper (2022 a, b) conducted a forced choice task where participants heard two phrases and had to choose the one which best described a certain picture. They found that, in a situation where one wants to refer to the subset of long leaves from the set of green leaves (see Figure 3), Romanian adults prefer merging the Color adjective earlier than the Size adjective (13a). In a situation where one wants to refer to the subset of green leaves from the set of long leaves (see Figure 4), on the other hand, Romanian adults prefer merging the Size adjective earlier than the Color (13b): (13) a. N AColor ASize frunze verzi lungi leaves green long ‘long green leaves’ b. N ASize AColor frunze lungi verzi leaves long green ‘green long leaves’ (Romanian) (English) (Romanian) (English) Figure 3: Picture showing long green leaves from Bleotu & Roeper (2022 a, b) Figure 4: Picture showing green long leaves from Bleotu & Roeper (2022 a, b) Long short leaves 19 Based on the above findings, we conclude that adults observe the RSSO over AORs that consider cognitive dimensions and argue that the RSSO is a primary compositional principle at the syntax-semantic interface of UG. We here expand this line of investigation by asking whether RCs containing Adjs specifying the same dimension (Size) or different dimensions (Color, Size) observe the RSSO in a similar fashion to adjectives or not. 3 Experiment (N = 28) 3.1 Aim Our experiment investigates two questions: (i) how adult native American English speakers order RC and Adj, (ii) whether they observe RSSO for RC and Adj and interpret the closest modifier as set-defining. These questions have not been previously investigated (see a literature review from de Vries 2018), and, therefore, research in this area is needed to shed light on the syntax-semantic interface. RCs containing Adjs share with Adj the property of modification, and, interestingly, (some) adjectives have even been analyzed as reduced RCs (Cinque 2010). Nevertheless, there is an important difference which distinguishes Adjs from RCs in English: adjectives are prenominal, while RCs are postnominal. The positioning of Adjs and RCs containing Adjs with respect to the nominal heads is expected to affect their ordering. 3.2 Predictions In terms of ordering, if RSSO is a universal principle at the syntax-semantics interface, interpreting recursive syntax compositionally, then it should hold for RC as well, not just for Adj. Since Adj are prenominal, but RC are postnominal in English, we expect RCs to show the inverse order of prenominal adjectives (see 14, 15). For nouns modified by two RCs, if the RCs are ordered through symmetric coordination (N [RC1 & RC2] = N [RC2 & RC1]) or via extraposition (by moving the Set Modifier across the Subset Modifier- see 16), then we should see a difference in how participants order RC compared to Adj. When Adjs and RCs occur together, participants might preferentially merge the Adj first, creating a Set, and then the RC to create a Subset. However, some participants might choose to first merge the RC to create a Set and then the adjective to create a Subset. In terms of interpretation, modifiers closer to N should denote sets. This is expected both for sequences with Adjs specifying the same dimension (Size) and for sequences with Adjs specifying different dimensions (Color and Size). (14) (15) short long leaves = leaves that are long that are short = long leaves that are short long green leaves = leaves that are green that are long = green leaves that are long 20 BLEOTU, FOUCAULT AND ROEPER Move (16) N RCset RCsubset -> N tset RCsubset RCset leaves [that are long]set [that are short]subset -> leaves tset [that are short]subset [that are long]set 3.3 Procedure and materials Participants responded to two successive forced choice tasks (see Tables 1, 2): (i) an Ordering Task consisting of 16 questions in which participants had to choose which expression has the same meaning as an A2 A1 N sequence (A1 N RC2 or A2 N RC1; N RC1 RC2 or N RC2 RC1) (ii) an Interpretation Task consisting of 24 questions in which participants had to choose the set from which the character should pick certain items (e.g., can long leaves that are short be found among long leaves or short leaves) We organized the experiment such that all Ordering Task questions precede the Interpretation Task to avoid a possible priming effect of interpretation on ordering. In terms of categories, in the Ordering Task we tested A2 A1 N and A1 N RC2 (Adjective + Noun + Relative Clause) equivalences and A2 A1 N and N RC1 RC2 (Noun + Relative Clause + Relative Clause) equivalences. In the Interpretation Task, we tested three combinations containing Adjs (A2 A1 N), RCs (N RC1 RC2), and Adjs/RCs combined (A1 N RC2). In terms of the cognitive dimension expressed by the modifiers, we tested both structures containing: (i) Size-Size combinations (short-long, small-big, short-tall, thin-fat) and (ii) Size-Color combinations (long-green, big-red, thin-brown, tall-yellow) Ordering Task Which of the expressions below has the same meaning as short long leaves? A N RC i) short leaves that are long ii) long leaves that are short N RC RC i) leaves that are long that are short ii) leaves that are short that are long Interpretation Task Laura has a pile of short leaves and a pile of long leaves. Some of the short leaves are long, and some of the long leaves are short. Tim asks Laura for … A N RC …the long leaves that are short From which pile should Laura choose the long leaves that are short? (i) the pile of long leaves (ii) the pile of short leaves N RC RC …the leaves that are long that are short From which pile should Laura choose the leaves that are long that are short? (iii) the pile of long leaves (iv) the pile of short leaves AAN …the short long leaves From which pile should Laura choose the leaves? (i) the pile of long leaves (ii) the pile of short leaves Table 1: Examples of experimental items with Size &Size modifiers Long short leaves 21 Ordering Task Which of the expressions below has the same meaning as long green leaves? A N RC i) long leaves that are green ii) green leaves that are long N RC RC i) ii) leaves that are long that are green leaves that are green that are long Interpretation Task Laura has a pile of long leaves and a pile of green leaves. Some of the long leaves are green, and some of the green leaves are long. Tim asks Laura for … A N RC N RC RC AAN …the long leaves that are green. From which pile should Laura choose the long leaves that are green? (iii) the pile of long leaves (iv) the pile of green leaves …the leaves that are long that are green. From which pile should Laura choose the leaves that are long that are green? (i) the pile of long leaves (ii) the pile of green leaves …the long green leaves. From which pile should Laura choose the long green leaves? (i) the pile of long leaves (ii) the pile of green leaves Table 2: Examples of experimental items with Color & Size modifiers The two tasks differ in one important respect: the relevance of context. The Ordering Task probes into participants’ default ordering preferences in the absence of any context, whereas the Interpretation Task probes into participants’ interpretive preferences within a context. In the Interpretation Task we limited the choices to restrictive interpretations, leaving for future research an investigation of coordinative interpretations which are also possible (e.g., leaves that are long and green). 3.4 Results In the Ordering Task, participants mostly ordered [A1 N RC2] and [N RC1 RC2] as a mirror of [A2 A1 N] (see Figure 5). Figure 5: Proportion of Set-Subset answers in the Ordering Task 22 BLEOTU, FOUCAULT AND ROEPER A look at their individual answers reveals, however, that not all participants ordered the modifiers as we expected: There were 6 participants who gave the reverse order (<50%). A reverse ordering of RCs in [N RC RC] structures counts as a violation of RSSO. Interestingly, a different ordering of modifiers in [A N RC] structures could possibly be in line with the RSSO rather than violating it: it could be that they merged the RC first, associating it with a Set interpretation, and they merged the Adj later, associating it with a Subset interpretation. In this case, their ordering would still observe the RSSO, though it would not meet our expectation that Adjs should have priority in Merge over RCs. Using R (2021), we performed a logistic regression with answer as our dependent variable (coded as 1 if it is in line with the RSSO, and as 0 if it is not), and the fixed effects Structure ([A1 N RC2], [N RC1 RC2]), Adjective types (SizeSize, Size-Color) and the interaction between Structure and Adjective types, as well as random slopes per Participant and Item. We find no statistically significant differences between [A1 N RC2] and [N RC1 RC2] structures, or between Size-Size and Size-Color combinations. We see a significant interaction between Structure and Adjective type (ß = 11.221, SE = 5.729, Z = 1.959, p < .01): There were fewer recursive readings in [A2 A1 N] structures with Size-Color combinations than with Size-Size combinations but more recursive readings in [N RC1 RC2] structures with Size-Color combinations than with Size-Size combinations, whereas [A1 N RC2] structures had similar recursive readings for both combinations. In the Interpretation Task, in accordance with RSSO, participants mostly interpreted A1/RC1 as indicating a set and RC2 as indicating a subset (see Figure 6). Figure 6: Proportion of Set-Subset answers in the Ordering Task We conducted a parallel analysis to the one we ran in the Interpretation Task, but we used random effects per Participant and Item, since the model would not converge with random slopes. We found that [A1 N RC2] structures led to more recursive interpretations than [A1 A2 N] structures (ß = 3.964, SE = 1.816, Z = 2.183, p < .05), and that [N RC1 RC2] structures also led to more recursive interpretations than [A1 A2 N] structures, but marginally so (ß = 3.449, SE = 1.781, Long short leaves 23 Z = 1.937, p =.052). Moreover, there was a marginally significant interaction between [N RC1 RC2] structures and Adjective type, with more recursive interpretations of [[N RC1 RC2] structures with Size-Color adjectives than with Size-Size adjectives (ß = −3.879, SE = 2.266, Z = −1.712, p =.0869). Interestingly, a look at the individual answers in the Interpretation Task reveals that all participants observed the RSSO in the Interpretation Task, while not necessarily doing so in the Ordering Task. To gain more insight into the effect of the task, we conducted a logistic regression with answer as our dependent variable (coded as 1 if it is in line with the RSSO, and as 0 if it is not), and the fixed effects Task type (Ordering Task, Interpretation Task), Structure ([A1 N RC2] , [N RC1 RC2]), Adjective types (SizeSize, Size-Color) and all possible interactions, as well as random slopes per Participant and Item. We found a highly significant Task effect (ß = −3.327, SE = 1.508, Z = −2.206, p < .05) but no other significant effects. 4 Discussion Participants mostly ordered and interpreted both adjectival and RC modifiers in accordance with RSSO (see 17, 18). (17) (18) ✓ a. [[Aset N] RCsubset] [[ long leaves] that are short] ✓ b. [Aset [N RCsubset]] ✗ [ long [leaves that are short]] ✗ ✓ a. [[N RCset] RCsubset] [[ leaves that are long] that are short] ✓ b. [[N RCsubset ] RCset] ✗ [[ leaves that are short] that are long] ✗ The ordering of adjectival and RC modifiers in accordance with the RSSO can be accounted for either through adjunction or movement. According to an Adjunction Parameter (Branching Directionality) Account (Kremers 2003, Abels & Neeleman 2010), we can assume different parameter settings for Adjs and RCs: Adjs are adjoined to the left, while RCs to the right. Interestingly, when a noun is modified by an Adj and an RC, Adjs compose first, while RCs compose onto the Adj+N sequence. Adjs have priority possibly because of their ability to directly merge to bare nouns and create compound-like formations, unlike RCs. Another explanation could be that relative clauses are longer than Adjs. Importantly, we see that the RSSO cuts across categories, ordering different modifiers in the syntactic hierarchy producing a set-subset interpretation (see 19). 24 BLEOTU, FOUCAULT AND ROEPER (19) Asubset Aset N N RCset RCsubset NP 2 APsubset NP long 2 APset N short leaves NP 2 NP RCsubset 2 that are long N RCset leaves that are short Aset N RCsubset NP 2 NP RCsubset 2 that are long APset N short leaves According to a Movement Account, we derive the orders at issue from the basic order FPsubset FPset (see 20). To get the [[Aset N] RCsubset] order, we start from RCsubset Aset N, and we move [Aset+N] to SpecFPsubset. To get the [[N RCset] RCsubset]] order, we start from RCsubset RCset N and apply Roll-Up: we first move N to SpecFPset and then move [N+RCset] to SpecFPsubset. (20) FP 2 2 F FPsubset 2 RC 2 4 Fsubset FP 2 2 F FPset Step 2 2 Aset/ RCset2 4 Fset NP N Step 1 A Movement account would be in line with Cinque’s (2010) explanation of cross-linguistic variation (deriving a mirror order of English in Italian, Spanish, etc.). However, an Adjunction Parameter account in terms of Branching Directionality (Kremers 2003, Abels & Neeleman 2010) where Adjs/RCs are placed to the right of the noun and ordered according to RSSO can also capture the mirror effect. Moreover, the latter account is structurally simpler: setting the parameter for branching directionality differently across languages and even within the same language, provides a less costly representation for recursive modification. This is because a universal grammar relying on less movement is preferable to one relying on more movement. On these grounds, we believe that an Adjunction Parameter account in terms of Branching Directionality best captures our findings. Long short leaves 25 Apart from the general finding that Adj/RC modifiers are mostly ordered in accordance with the RSSO, we also notice slightly higher accuracy rates in the Interpretation Task compared to the Ordering Task. We argue that this result can be explained through additional contextual support in the Interpretation Task and no contextual cues in the Ordering Task. A future direction of investigation may involve an Ordering Task in the presence of contextual cues, as well as an Interpretation Task in the absence of any contextual cues. In this way, one could see whether participants’ ordering and interpretive preferences vary with contextual cues. The lower number of RSSO orders with A A N structures compared to A N RC structures is perhaps due to the ease of associating stacked adjectives with symmetric coordination (long (and) short = short (and) long, long (and) green = green (and) long), which is harder to do with A N RC structures (?? long leaves and that are short). The existence of a number of participants who did not order modifiers as expected may be explained in different ways. For A N RC structures, it may be that these participants simply associate the RC with the Set and the adjective with the Subset ([Asubset [N RCset]]), which would actually be in line with RSSO. For N RC RC structures, participants may adopt a conjunctive default in ordering, where the modifiers are symmetrically coordinated (N [RCset (&) RCsubset] = N [RCsubset (&) RCset]) and, consequently, they can appear in any order. Another possibility is that participants start from a N RCset RCsubset order but have a preference for extraposing the RCset to the right of RCsubset (see 15). Importantly, a linear order with a Subset modifier closer to the noun than the Set one is rare. Moreover, in the Interpretation task, no participant gave an interpretation not in line with the RSSO. Taken together, these findings suggest an overall preference for the RSSO both in ordering and interpretation, with the important caveat that ordering modifiers seems to allow other orders, e.g., on the basis of a conjunctive default. 5. Conclusion Our experimental evidence demonstrates that native adult speakers of American English prefer to order and interpret modifiers that are Adjs or RCs containing Adjs in accordance with the Recursive Set-Subset Ordering Principle (RSSO): they prefer to associate the closest modifier with a Set interpretation and the modifier which is higher up in the hierarchy with a Subset interpretation. This shows that the RSSO is not category-dependent but applies to a variety of categories of UG which function as modifiers of nouns. The Recursive Set-Subset Ordering Principle (RSSO) is thus an important automatic, innate principle, defining the syntax-semantic interface of UG. It derives from Merge and provides a Set-Subset interpretation for the recursive syntax of nouns modified by recursive Adjs, RCs or their combination. 26 BLEOTU, FOUCAULT AND ROEPER References Abels, K. & A. Neeleman. 2010. Linear Asymmetries and the LCA. Syntax 12(1). 25–74. https://doi.org/10.1111/j.1467-9612.2011.00163.x. Arsenijevic, B. & W. Hinzen. (2010). Recursion As a Human Universal and As a Primitive. Biolinguistics. 4. 165-173. http://doi.org/10.5964/bioling.8781. Bleotu, A. C. & T. Roeper. 2021a. Small big flowers or small and big flowers? Simple is Better and Roll-Up is Too Complex for Romanian 5-Year-Olds. In Proceedings of the 45th annual Boston University Conference on Language Development, ed. by D. Dionne & L. A.Vidal Covas, vol. 1. 66–79. Somerville, MA: Cascadilla Press. Bleotu, A. C. & T. Roeper. 2021b. Roll-Up is too complex for Romanian 5-year-olds. Evidence from recursive adjectives. In Proceedings of the Linguistic Society of America, v. 6, n. 1, 133143. Bleotu, A. C. & T. Roeper. 2022a. The Recursive Set-Subset Ordering Restriction overrides Adjective Ordering Restrictions. Evidence from Romanian 4-year-olds and adults. In BUCLD 46: Proceedings of the 46th Annual Boston University Conference on Language Development, ed. by Y. Gong & F. Kpogo, vol.1, 62–75. Somerville, MA: Cascadilla Press. Bleotu, A. C. & T. Roeper. 2022b. Children are more sensitive to the Recursive Set-Subset Ordering than to Adjective Ordering Restrictions. In Proceedings of the Linguistic Society of America, [S.I.], v.7, n.1, p. 5267. Chomsky, N. 1995. The Minimalist Program. Cambridge: MIT Press. Cinque, G. 1994. On the evidence for partial N-movement in the Romance DP. In Paths Towards Universal Grammar. Studies in Honor of Richard S. Kayne, ed. by G. Cinque, J. Koster, J.-Y. Pollock, L. Rizzi & R. Zanuttini, 85–110. Washington DC: Georgetown University Press. https://doi.org/10.1353/lan.2001.0097. Cinque, G. 2005. Deriving Greenberg’s Universal 20 and its exceptions. Linguistic Inquiry 36(3). 315–332. https://doi.org/10.1162/0024389054396917. Cinque, G. 2010. The Syntax of Adjectives. A Comparative Study. MIT Press. Cinque, G. 2020. The Syntax of Relative Clauses: A Unified Analysis. Cambridge: Cambridge University Press. https://doi.org/10.1017/9781108856195. Cornilescu, A. & I. Giurgea. Chapter 7. The Adjective. In A Reference Grammar of Romanian, ed. by C. Dobrovie-Sorin & I. Giurgea, vol. 1, 355–530. John Benjamins. Cornilescu, A. & R. Cosma. 2019. Revue Roumaine de Linguistique, LXIV, 3. 307–323. Bucureşti. Cornilescu, A. & A. Nicolae. 2016. Romanian adjectives at the syntax–semantics interface. Acta Linguistica Hungarica 63(2). 197-240. https://doi.org/10.1556/064.2016.63.2.3. de Vries, M. 2002. The Syntax of Relativization. LOT, Utrecht. Dixon, R. M. W. 1982. ‘Where Have All the Adjectives Gone?’ and other Essays in Semantics and Syntax. Berlin: Mouton de Gruyter. Foucault, D., A. C. Bleotu, U. Lakshmanan, E. Merritt, R. Sybing & T. Roeper. 2022. Relative gradable adjective recursion is more challenging for acquisition than possessive recursion. In Proceedings of the Linguistic Society of America, [S.I.], v.7, n.1, p. 5294. Grohe, L. & P. Schulz. 2021. Big round clocks and blue square tables: Adjective ordering preferences in elicited child and adult Production. In BUCLD 46: Proceedings of the 46th Annual Boston University Conference on Language Development, ed. by Y. Gong & F. Kpogo, vol.1, 242–255. Somerville, MA: Cascadilla Press. Hauser, M. D., Chomsky, N., & Fitch, W. T. 2002. The faculty of language: What is it, who has it, and how did it evolve? Science, 298(5598), 1569– 1579. https://doi.org/10.1126/science.298.5598.1569. Hollebrandse, B. & T. Roeper. 2014. Empirical results and formal approaches to recursion in acquisition. In Recursion: Complexity in Cognition, ed. by T. Roeper & M. Speas, 179–220. Dordrecht: Springer. Kremers, J. M. 2003. The Arabic Noun Phrase: A Minimalist Perspective. University of Nijmegen Doctoral dissertation. Long short leaves 27 Leivada, E. & M. Westergaard. 2019. Universal linguistic hierarchies are not innately wired. Evidence from multiple adjectives. PeerJ 7, e7438. https://doi.org/10.7717/peerj.7438. Luciu, A. & A. C. Bleotu. (to appear). How are Size, Age, Shape and Color adjectives ordered in English and Romanian? An experimental investigation. Bucharest Working Papers in Linguistics Roeper, T. 2011. The acquisition of recursion: How formalism articulates the child’s path. Biolinguistics 5(1-2). 57-86. Scontras, G., J. Degen & N. D. Goodman. 2017. Subjectivity predicts adjective ordering preferences. Open Mind 1. 53–65. Scontras, G., J. Degen & N. D. Goodman. 2019. On the grammatical source of adjective ordering preferences. Semantics and Pragmatics 12(7). 1–21. Scott, G.-J. 2002. Stacked adjectival modification and the structure of nominal phrases. Functional Structure in the DP and IP: The Cartography of Syntactic Structures, ed. by G. Cinque (ed.), vol. 1, 91–120. Oxford and New York: Oxford University Press. Sproat R. & C. Shih. 1991. The cross-linguistic distribution of Adjective Ordering Restrictions. In Interdisciplinary Approaches to Language, ed. by C. Georgopoulos & R. Ishihara, 565593. Dordrecht: Kluwer. https://doi.org/10.1007/978-94-011-3818-5. Trainin, N. & E. Shetreet. 2021. It’s a dotted blue big star: on adjective ordering in a post-nominal language. Language, Cognition and Neuroscience, 36(3). 320–341. Truşcǎ, D. & A. C. Bleotu (to appear). An experimental investigation of the ordering of Quality, Size and Color adjectives in English and Romanian. Bucharest Working Papers in Linguistics. Proceedings of CLS 58 (2022), 29-44 © Chicago Linguistic Society 2023. All rights reserved. 29 SE figure reflexives alternate with Voice in Romanian: Insights from a semi-artificial denominal paradigm Adina Camelia Bleotu1, and Rodica Ivan2 University of Bucharest1, UMass Amherst1,2, Acuity Insights2 1 Introduction We investigate SE figure reflexive constructions, i.e., structures where the se clitic occurs together with a verb and a body part PP (1a, b), by means of two experiments which rely on semi-nonce verbs (an Intentionality Experiment and a Change-ofState Experiment). Our results support Cornilescu & Nicolae (2017, 2021)’s proposal that there are two such types in Romanian (transitive and unaccusative), and that they are best analyzed as Voice alternation. The term figure reflexives has been adopted by Cornilescu & Nicolae (2017, 2021) from work by Manoliu-Manea (1996), Schäfer (2008, 2009) and Wood (2012, 2014) to refer to prepositional reflexive constructions such as those in (1), which involve a Figure-Ground relation (detailed in section 2). In Romanian, such constructions employ the clitic se and an obligatory (although not always overt) locative PP typically referring to an inalienably possessed object, e.g. față ‘face’, of the subject. Given the presence of the clitic se, we henceforth refer to Romanian figure reflexives like those in (1a) and (1b) as SE figure reflexives. (1) a. Maria s- a spălat pe față. Transitive Figure Reflexive Maria SE has washed PE face ‘Maria washed her face (lit. Maria washed herself on the face)’. b. Maria s-a luminat la față. Unaccusative Figure Reflexive Maria SE has brightened LA face ‘Maria SE has brightened LA face’. (example (1b) adapted from Cornilescu & Nicolae (2021), ex. (32b)) According to Cornilescu & Nicolae (2017, 2021), there are two types of SE figure reflexives: transitive constructions, which usually indicate some contact with a body part (a spăla ‘wash’, a tăia ‘cut’, a lovi ‘hit’, a scărpina ‘scratch’) and unaccusative constructions, which express change-of-state events (a se albi ‘whiten’, a se înroși ‘redden’, a se îngroșa ‘thicken’). Transitive SE figure reflexives like those in (1a), involve a thematic Voice head, which assigns the role of Agent to the subject (Maria). In contrast, unaccusative SE figure reflexives like those in (1b) have an expletive Voice head, which is unable to assign the Agent role to the subject non-agentive. One important question to address is whether the constructions in (1) exist independently from the lexical verbs they appear with (e.g. lumina ‘brighten’). An alternative to the existence of two distinct (transitive and unaccusative) SE reflexive 30 BLEOTU AND IVAN constructions would be deriving these interpretations from the argument structure of the lexical verbs themselves. To remove this alternative, we decided to test the syntactic and semantic distinction between these two constructions by employing semi-artificial denominal (SAD) verbs, i.e., non-existent verbs derived from existing nouns, such as a cireși ‘to cherry’ (derived from cireașă ‘cherry’). More specifically, we investigate the SAD paradigm in (2), whereby we take a comparative look at: (i) bare verb constructions like those in (2a) and simple SE clitic constructions like those in (2b), as well as (ii) transitive SE figure reflexives like those in (2c), which typically involve the preposition pe ‘on’, and unaccusative SE figure reflexives like those in (2d), which typically involve the preposition la ‘at’. By employing this novel SAD paradigm in our experiments, participants cannot rely on the argument structure of existing verbs as a basis for interpretation. (2) a. Maria Maria b. Maria Maria c. Maria Maria d. Maria Maria a has s-a SE has s-a SE has s-a SE has cireșit. cherried cireșit. cherried. cireșit cherried cireșit cherried Bare (SAD) Verb SE Clitic (SAD) Verb pe PE la LA mâini. SE (SAD) PE Figure Reflexive hands mâini. SE (SAD) LA Figure Reflexive hands In our investigation of how participants understand SE figure reflexive structures, we targeted two properties: intentionality, which is typical of agentive transitive constructions (in Experiment 1), and change-of-state interpretations, which are typical of unaccusatives (in Experiment 2). 2 Background Figure reflexives are a class of reflexive constructions which introduce a FigureGround relation (Svenonius 2003, Talmy 1978,1985) by means of a PP. The Figure has been defined as an object whose path or location is at issue (Svenonius 2003: 432), while the Ground represents a reference-point with respect to which the Figure’s path or location is described (Talmy 1985:61). In (3), for example, a Figure-Ground relation (on) characterizes the subject the book (Figure) and the PP object the bed (Ground). (3) The book is on the bed. The term figure reflexive has previously been used by Wood (2012, 2014) to refer to constructions such as (4a) in Icelandic, employing the morpheme -st, which, according to the author, is not reflexive itself, but rather an expletive element which contributes to a reflexive interpretation. Similarly to Romance SE, -st occurs in both SE figure reflexives 31 agentive (transitive) structures, such as (4a), as well as in non-agentive anticausative constructions, such as (4b): (4) a. Bjartur tróð-st gegnum mannþröngina. Agentive Figure Reflexive Bjartur.NOM squeezed-ST through the.crowd ‘Bjartur squeezed through the crowd.’ (ex. (13), p. 263, Wood & Marantz 2017) b. Hurðin opnaðist. door.the.NOM opened-ST ‘The door opened.’ Anticausative Construction (ex. (16b), p. 265, Wood & Marantz 2017) Wood (2012, 2014) assumes that figure reflexives in Icelandic are always agentive, and that the subject has two roles: a Figure role from the PP and an Agent role from Voice. In contrast to Icelandic, Romanian SE figure reflexives may be of two types: transitive (5a) and unaccusative (or anticausative) (5b). Additionally, although both Icelandic and Romanian figure reflexives involve a Figure-Ground relation, where the Ground is expressed by the PP complement, the Icelandic examples in (4) capture the movement of the Agent/Figure with respect to the location expressed by the Ground, whereas in Romanian examples like those in (5), there is a relation of inalienable possession (typically a body part) between the subject and PP object. (5) a. Maria s- a spălat pe față. Transitive Figure Reflexive Maria SE has washed PE face ‘Maria washed her face (lit. Maria washed herself on the face)’. b. Maria s-a luminat la față. Unaccusative Figure Reflexive Maria SE has brightened LA face ‘Maria SE has brightened LA face’. Cornilescu & Nicolae (2017, 2021) argue that, in transitive SE figure reflexives (5a), the Figure is SE, whereas in unaccusative figure reflexives (5b), the Figure is the lexical subject. Additionally, in transitive SE figure reflexives, the FigureGround relation is typically mediated via the preposition pe ‘on’, while in unaccusative SE figure reflexives (5b), the Figure-Ground relation is typically mediated by prepositions like la ‘at’. Both pe and la are semantically interpretable prepositions in these constructions. Thus, the pe preposition in transitive SE figure reflexives is different in nature from the pe preposition which represents a differential object marker for animate direct objects (Tigău 2010, Irimia 2020). Also, the la preposition in unaccusative SE figure reflexives is different in nature from the la preposition which marks the Dative case on animate objects. We focus on two main differences between the two two types of SE figure reflexives: Voice and agentivity (Cornilescu & Nicolae, 2017, 2021). We assume that Voice heads its own projection, which hosts external arguments (Marantz 1984, Rivero 1990, Kratzer 1994, 1996, a.o.). In transitive figure 32 BLEOTU AND IVAN reflexives, the Voice head is thematic and, consequently, it assigns an Agent role to the sentence subjects. In contrast, in unaccusative figure reflexives, the Voice head is expletive, and, consequently, it fails to assign any theta role to the subject. The derivations of the two constructions are illustrated and discussed below. Just like Cornilescu & Nicolae (2017, 2021), we adopt the idea presented in Dobrovie-Sorin (2017) that SE is a DP which must check its case before reaching the Person/Tense domain. The source and path of SE differ in the two constructions. In transitive SE figure reflexives like (6), SE starts out as an argument, a Figure in a Figure-Ground prepositional small clause. Evidence for this is that, given its argumental status, SE can be doubled with the strong pronoun pe sine ‘PE self’ in examples like (6) (Cornilescu & Nicolae, 2021). The function of SE is to reflexivize the predicate, and to establish coreference with the Agent subject (Maria). SE will afterwards move to the specifier of the aspectual position Spec TOP (Pesetsky and Torrego, 2004), in order to value its accusative feature. (6) Maria se spală pe Maria SE washes PE ‘Maria washes her face.’ față. face Transitive The reflexive clitic behaves differently in unaccusative configuration such as (7). Unlike in transitive constructions, SE is not an argument. This hypothesis is supported by the fact that SE in (7) cannot be doubled by a strong reflexive (* pe sine ‘PE self’). Since SE is not an argument, it does not receive a theta-role. It never occurs in SpecVoiceP. Instead, it merges in Spec TOP to value its accusative feature. SE figure reflexives (7) Maria se luminează la Maria SE brightens LA ‘Maria’s face brightens up.’ față. face 33 Unaccusative Apart from Voice, another important difference between the two types of SE figure reflexives involves lexical aspect. As shown by previous research (Alexiadou, Agnastopolou & Schäfer 2006, 2015, Harley 2012, Schäfer 2008, 2009), causatives and anticausautives/unaccusatives differ aspectually in terms of the predicates present in their lexical make-up. Causatives have a CAUSE projection, while anticausatives have a BECOME projection but lack CAUSE. Our study is centered on the experimental investigation of these two main differences teasing apart transitive and unaccusative SE figure reflexives: the agentive or expletive nature of Voice, and the presence or absence of CAUSE in the lexical make-up of the predicates at issue. In order to determine whether these two constructions exist independently of the lexical verbs that exemplify them, we choose to employ a SAD paradigm which makes it possible to investigate the same verb in different configurations. This is particularly important, since, if we investigated participants’ interpretations through sentences containing lexical verbs they are familiar with, their answers could potentially reflect a structural bias (SE PE/ SE LA) those verbs have in Romanian, rather than the interpretation associated with different types of figure reflexive constructions. While there are some verbs which can occur in both frames, such as albi ‘whiten’, most verbs cannot occur equally easily in both SE PE and SE 34 BLEOTU AND IVAN LA figure reflexive frames. For instance, the verb înroşi ‘redden’ occurs naturally with SE and body parts introduced by LA, but less so with SE and a PE body part (8). SAD verbs remove lexical bias, as they are compatible with both PE and LA. (8) a. Maria Maria b. Maria Maria s- a SE has s-a SE has înroşit reddened înroşit reddened la LA pe PE față. SE LA Figure Reflexive face buze cu un ruj. SE PE Figure Reflexive lips with a lipstick Another advantage of the SAD paradigm is that it sheds light on the interpretation of denominal verbs. Denominals are known to vary in their interpretation (Kelly 1998, Harley & Haugen 2007, Kiparsky 1997, Borer 2014, Bleotu 2017, 2019, Bleotu & Bloem 2020, 2021), giving rise either to (i) literal readings, derived by noun-incorporation into a light verb such as DO, or to (ii) figurative readings, derived by root-incorporation into BECOME/ACT (Kiparsky 1997). For instance, a SAD verb like a cireşi ‘to cherry’ can have either (i) literal readings, involving canonical actions related to the fruit cireaşă ‘cherry’ (e.g. eating, picking cherries), or (ii) figurative change-of-state interpretations (e.g. becoming red like a cherry/blushing). If there is a transitive-unaccusative dichotomy between the SE figure reflexives in the SAD paradigm in (2), we might expect participants to choose change-of-state BECOME LIKE readings to a greater extent for SAD unaccusative figure reflexives (2d) than for SAD transitive figure reflexives (2c). 3 Experiment 1 (The Intentionality Experiment) 3.1 Aim In Experiment 1, we investigate the agentivity of figure reflexives, looking into whether SAD transitive figure reflexives are interpreted as more agentive than SAD unaccusative figure reflexives. To this purpose, we ask participants which sentences (9a, 9b, or both) can be continued with the adverb intenţionat ‘intentionally’. If participants interpret SE PE constructions like (9a) as agentive transitive SE figure reflexives, we expect them to associate intentionality more often with SE PE figure reflexives than with SE LA figure reflexives. (9) a. b. Dumitru Dumitru Dumitru Dumitru s-a SE has s-a SE has piersicit peached piersicit peached pe PE la LA mâini. SE PE Figure Reflexive hands mâini. SE LA Figure Reflexive hands 3.2 Procedure, participants, and materials We tested 36 native Romanian undergraduate students from the University of Bucharest. Participants had to choose which of two sentences can be continued with intenţionat ‘intentionally’ (10). Their options included sentence (a), (b) or both. SE figure reflexives 35 (10) Care propoziție poate fi continuată cu ‘intenţionat’? Which sentence can be continued with intentionally? The experiment involved a training stage and a testing stage. In the training stage, participants were exposed to three trial items which contained existent lexical verbs (see Sample Training Item in Table 1). Participants were encouraged to follow their own intuitions: they were explained that some speakers might prefer a particular sentence (either (a) or (b)) or that they could like both options to an equal extent. In the testing stage (see Sample Critical Item in Table 1), participants were tested on SE PE and SE LA figure reflexives. They were given three possible options to choose from: a sentence containing a SE PE figure reflexive (a), a sentence containing a SE LA figure reflexive construction (b) or Ambele ‘Both’ (c). a. b. c. a. b. c. d. e. f. Sample Training Item a. Omul a căzut. man-the has fallen b. Omul a minţit. man-the has lied c. Ambele Both Sample Critical Item a. Dumitru s- a piersicit pe mâini Dumitru SE-has peached PE hands b. Dumitru s- a piersicit la mâini. Dumitru SE-has peached LA hands c. Ambele Both Sample Filler Items Passive-Active Intransitives with (in)animate subjects a. Sofia a fost prunitǎ. a. a. Rodica a lămȃit. Sofia has been plummed b. Rodica has lemoned b. Sofia a prunit. c. b. Copacul a lămȃit. Sofia has plummed d. The tree has lemoned c. Ambele e. c. Ambele Both f. Both Table 1: Sample items used in Experiment 1 (Intentionality Experiment) The experiment employed 9 critical items distributed in 9 Latin Square lists: each participant saw only one item per verb, and only one item per body part. The experiment also made use of 18 fillers: 9 fillers involving active-passive alternations, and 9 fillers involving intransitive verbs with animate and inanimate subjects (see Table 1). Based on a norming study we conducted to determine the dominant property associated with each noun root (see Bleotu & Ivan 2022 for more details), we employed 9 critical verbs: 3 FLAVOR verbs (a cireşi ‘to cherry’, a piersici ‘to peach’, a zmeuri ‘to strawberry’), 3 SHAPE verbs (a castraveţi ‘to cucumber’, a ciuperci ‘to mushroom’, a conopidi ‘to artichoke’), and 3 MIXED verbs, derived 36 BLEOTU AND IVAN from nouns defined in terms of several dominant property (a cartofi ‘to potato’, a pepeni ‘to watermelon’, a dovleci ‘to pumpkin’). We combined each of the 9 verbs with 9 body parts (cap ‘head’, faţă ‘face’, mȃini ‘hands’, picioare ‘feet’, gȃt ‘neck’, deget ‘finger’, spate ‘back’, burtă ‘stomach’) in order to avoid any bias due to a certain body part. 3.3 Analysis & Results The results were analyzed in R-4.0.5 (2021). No participants or items were excluded. We find that participants preferred PE sentences (58.95%) over LA sentences (18.52%) and the BOTH option (22.53%), as can be seen in Figure 1. For a comparison of these preference rates, we conducted a One-Way ANOVA analysis, with the number of intentionality continuations as the dependent variable and Answer Type (PE, LA, BOTH) as a fixed effect. We also ran Tukey tests for multiple comparisons of means. We find that (i) the rate of PE choices is significantly different from LA choices (β = 43.667, SE = 4.181, t = 10.444 , p < 0.001), (ii) the rate of LA choices is not significantly different from BOTH (β = 4.333, SE = 4.181, t = 1.036, p = 0.583), and (iii) the rate of PE choices is significantly different from BOTH (β = 39.333, SE = 4.181, t = 9.407, p < 0.001). Figure 1: Rate of ‘intentionally’ continuations in Experiment 1 3.4 Discussion Participants preferred to continue sentences containing PE figure reflexives with intentionally more than sentences containing LA figure reflexives. Our results support the hypothesis that PE figure reflexives are understood as agentive to a larger extent than LA figure reflexives. Nevertheless, we find that LA PPs can also be continued with intentionally, although to a lesser extent. This is somewhat expected, given that LA PPs can also occur in transitive figure reflexives (see 11). (11) Maria s-a tăiat la deget Maria SE has cut LA finger ‘Maria (intentionally) cut her finger.’ (intenționat). (intentionally) SE figure reflexives 37 Thus, we take our findings to confirm that, while both PE PPs and LA PPs can be interpreted as transitive, participants show a strong bias for interpreting SE PE figure reflexives as involving an [Agent] role. This bias is in accordance with the syntactic analysis proposed by Cornilescu & Nicolae (2017, 2021). 4. Experiment 2 (The Change-of-state Experiment) 4.1. Aim In Experiment 2, our main aim is to probe into the unaccusativity of SE LA constructions. In particular, if LA PPs are more likely to be associated with unaccusative syntactic structures, we expect participants to provide change-of-state interpretations more often for sentences containing SE LA figure reflexives (12d) than for sentences with SE PE figure reflexives (12c). A secondary aim of this experiment is to investigate whether reflexive SE verbs (12b) receive change-of-state interpretations more often than bare denominals (12a). Given that SE has been argued to be a marker of anticausativity (DobrovieSorin 1998, 2017; Dobrovie-Sorin & Giurgea, forthcoming), we expect that the presence of SE will lead to a higher rate of change-of-state interpretations for (12b). (12) SAD verb paradigm a. Dumitru a piersicit. Dumitru has peached b. Dumitru s-a piersicit. Dumitru SE has peached. c. Dumitru s-a piersicit Dumitru SE has peached d. Dumitru s-a piersicit Dumitru SE has peached Bare (SAD) Verb SE Clitic (SAD) Verb pe PE la LA mâini. SE (SAD) PE Figure Reflexive hands mâini. SE (SAD) LA Figure Reflexive hands 4.2. Procedure, participants, and materials 79 native speakers of Romanian (University of Bucharest students) took part in Experiment 2. Participants were asked which interpretation(s) they would accept for a given sentence from the SAD verb paradigm presented in (12). They could choose from (a) a BECOME paraphrase, (b) a DO paraphrase, or (c) both. Just like Experiment 1, Experiment 2 also involved a training stage and a testing stage (see Table 2). In the training stage, participants saw three items involving existing verbs. To make participants aware that they could choose any of the 3 options, one training item was biased towards a BECOME paraphrase, one training item was biased towards a DO paraphrases, and one was compatible with both paraphrases. In the testing stage, participants saw critical items belonging to 4 conditions: bare verb, SE clitic, SE PE Figure Reflexive, and SE LA Figure Reflexive. For 38 BLEOTU AND IVAN instance, they saw a sentence containing a SE clitic (see 13), and they could choose from the following three options: (13a) a BECOME paraphrase, (13b) a DO paraphrase, or (13c) both. Sample items are also listed in Table 2. (13) Sample Experiment 2 Item (SE Clitic condition) Pentru propoziţia “Dumitru s-a piersicit” ce interpretare aţi accepta? ‘For the sentence “Dumitru has peached” what interpretation would you accept?’ a. Dumitru a devenit ca o piersicǎ. Dumitru has become like a peach b. Dumitru a făcut ceva cu o piersicǎ. Dumitru has done something with a peach c. Ambele Both Target Sentence Sample Training Item Mihai a înroşit ouǎle. ‘Mihai reddened the eggs.’ Sample Critical Item by Condition Alexandru a dovlecit. Bare Verb a. Alexandru has pumpkined. b. SE Clitic Alexandru s- a dovlecit. Alexandru SE has pumpkined. c. SE PE Figure Reflexive Alexandru s- a dovlecit d. pe mȃini. e. Alexandru SE has pumpkinedf. on hands Alexandru s- a dovlecit g. la mȃini. Alexandru SE has pumpkined at hands SE LA Figure Reflexive Sample Filler Items Sofia a fost prunitǎ. Passive SAD ‘Sofia has been plummed.’ verbs Intransitive SAD verbs with inanimate subjects Copacul a lămȃit. ‘The tree has lemoned.’ Interpretative Options a. Ouǎle au devenit roşii. ‘The eggs became red.’ b. Mihai a fǎcut ceva cu ouǎle. ‘Mihai did something with the eggs.’ c. Ambele ‘Both’ BECOME DO BOTH a. Alexandru a devenit ca un dovleac. BECOME ‘Alexandru has become like a pumpkin.’ b. Alexandru a făcut ceva cu un dovleac. DO ‘Alexandru has done something with a pumpkin.’ c. Ambele BOTH Both a. Alexandru sau mâinile lui au devenit ca un dovleac. BECOME ‘Alexandru or his hands have become like a pumpkin.’ b. Alexandru a făcut ceva cu un dovleac. DO ‘Alexandru has done something with a pumpkin.’ c. Ambele BOTH Both a. Sofia a devenit ca o prunǎ. BECOME ‘Sofia became like a plum.’ b. Cineva i-a fǎcut ceva Sofiei cu o prunǎ. DO ‘Somebody did something to Sofia with a plum.’ c. Ambele BOTH Both a. Copacul a devenit plin cu lămȃi. BECOME ‘The tree became full of lemons.’ b. Copacul a fǎcut ceva cu nişte lămȃi. DO ‘The tree did something with some lemons.’ c. Ambele BOTH Both Table 2: Sample items used in Experiment 2 (Change-of-state Experiment) SE figure reflexives 39 The same 9 verbs from Experiment 1 were employed as critical stimuli in Experiment 2. However, for simplicity, all items featured a single body part, mȃini (“hands”). The experiment employed 9 critical items distributed in 4 Latin Square lists and 18 fillers (9 passive SAD sentences, and 9 sentences with intransitive SAD verbs and inanimate subjects). 4.3. Analysis & Results The results were analyzed in R-4.0.5 (2021). No participants or items were excluded. Overall, participants preferred BECOME paraphrases (60.47%- see Figure 2), which suggests a bias towards change-of-state readings. DO paraphrases were chosen only 25.87% of the total amount, and BOTH answers were chosen 13.64%. Figure 2: Interpretation by Condition in Experiment 2 To investigate the effect of Condition (Bare Verb, SE Clitic, SE PE Figure Reflexive, SE LA Figure Reflexive) on paraphrase rates (BECOME, DO, BOTH), we conducted a generalized linear mixed effects regression models with Condition as a fixed effect and random by-item and by-participant slopes. The results show that, with respect to the contrast between SE PE and SE LA figure reflexives, participants chose BECOME paraphrases to a lesser extent (β = –1.45, SE = 0.39, z = –3.75, p < 0.01) and DO paraphrases to a larger extent (β = 1.73, SE = 0.38, z = 4.51, p < 0.01) in the SE PE condition. Similarly, with respect to the contrast between the Bare Verb condition and the SE Clitic condition, participants chose BECOME paraphrases to a lesser extent (β = –1.32, SE = 0.5, z = –2.77, p < 0.01) and DO paraphrases to a larger extent (β = 1.61, SE = 0.56, z = 2.9, p < 0.01) in the Bare Verb condition. The rate of DO/BECOME paraphrases was not significantly different with respect to the contrast between the Bare Verb condition and SE PE, or the contrast between the SE Clitic condition and the SE LA. The rate of BOTH answers was the same across all four conditions. 40 BLEOTU AND IVAN 4.4 Discussion Our results reveal a general bias of our SAD paradigm towards change-of-state interpretations. Nevertheless, we found a significant contrast between SE PE figure reflexives and SE LA figure reflexives, in that participants chose BECOME paraphrases more often in the SE LA condition than in the SE PE condition. These results are compatible with the hypothesis that SE LA reflexives are associated with an unaccusative structure. Moreover, participants also chose DO paraphrases more often in the SE PE condition than they did in SE LA condition, which provides further support for an analysis of SE PE figure reflexives as transitive. 4.5. Exit poll To gain more insight into the way participants interpreted SE reflexives, we ran an informal exit-poll at the end of Experiment 2, where we asked participants to assess their interpretive preferences for 3 SAD verbs with SE clitics: a FLAVORdominant one, a SHAPE-dominant one and a MIXED one, (14) Să presupunem ca propoziția următoare înseamnă că Marina (sau o parte din Marina) a devenit ca o zmeură. ‘Let’s suppose that the following sentence means that Marina (or some part of Marina) became like a raspberry.’ Marina s-a zmeurit. Marina SE-has raspberried. Puteți să spuneți în ce sens a devenit Marina ca o zmeură? (forma, culoare, textură, mărime, aromă, altele) ‘Can you explain in what sense Marina became like a raspberry? (shape, color, texture, size, flavor, others)’ Our results suggest that, possibly because of its visual salience, color was the preferred change-of-state interpretation overall, followed by size and shape. Such a preference could potentially characterize figure reflexives as well. We leave this for future research. 5. General discussion The novel semi-artificial denominal verb (SAD) paradigm allowed us to test the existence of two types of figure reflexive syntactic structures in Romanian, transitive and unaccusative, as proposed by Cornilescu & Nicolae (2017, 2021). The SAD paradigm removes lexical verb bias, preventing verbs from biasing speakers towards a particular syntactic structure. Our experimental investigation provides evidence that unaccusative and transitive SE figure reflexive frames exist independently of the figure reflexive verb. Our experimental results show that although both SE PE and SE LA figure reflexives can be associated with intentionality (Experiment 1) and change-of-state interpretations (Experiment 2), there are significant differences between the two structures. Essentially, SE PE figure reflexives are ‘more agentive’ and SE LA SE figure reflexives 41 figure reflexives are ‘more unaccusative’. Thus, we interpret our experimental findings as supporting Cornilescu & Nicolae (2017, 2021)’s proposal that there are two types of SE figure reflexives in Romanian (transitive and unaccusatives). Another possible interpretation of the results could be that agentivity is actually encoded by the prepositions pe and la rather than the two structures. Nevertheless, we argue that this is not the case, given that the two prepositions can be used both with agentive and non-agentive predicates, as can be seen in (15a) and (15b). Consequently, we embrace the view that the prepositions are not associated with a particular Voice head: pe is not [+Agent] and la is not [−Agent]. Instead, it is the the syntactic structures that are [+/− agentive]. (15) a. Maria este pe scaun. Maria is PE chair ‘Mary is on the chair.’ b. Maria a mers la teatru. Maria has gone LA theater ‘Mary went to the theater.’ Finally, we would like to remark that the existence of two structures (SE PE figure reflexives and SE LA figure reflexives) can be argued to be a case of causative alternation (Levin & Rappaport Hovav, 1994, 1995) for Romanian figure reflexive constructions, such that unaccusative figure reflexives express internal causation and transitive figure reflexives express external causation. We leave this matter for consideration in future investigation. 6. Conclusion In the current paper, we have investigated Cornilescu & Nicolae (2017, 2021)’s proposal that there are two types of figure reflexive structures in Romanian (agentive and unaccusative) by means of a novel paradigm using semi-artificial denominal verbs (i.e., non-existent denominal verbs derived from existing nouns). We have brought evidence from two experiments that unaccusative and transitive figure reflexives differ in terms of at least two dimensions: agentivity and lexical aspect. We found that what Cornilescu & Nicolae (2017, 2021) call transitive figure reflexives are more easily associated with intentionality/agentivity and DO interpretations than unaccusative figure reflexives, while unaccusative figure reflexives are associated with intentionality/agentivity to a lesser extent, and they are more likely to express BECOME interpretations. Through the novel SAD paradigm, we have thus shown that the two figure reflexive structures exist independently of the lexical verbs that exemplify them, and that Cornilescu & Nicolae (2017, 2021) analysis accurately captures the empirical data. 42 BLEOTU AND IVAN Acknowledgments This research is part of a research project funded by UEFISCDI (PN-III-P1-1.1-PD-2019-0472) during the period 1 September 2020-28 February 2022: Experimental Insights into Denominals and Creativity (XIDenCre). We give thanks to the undergraduate students at the University of Bucharest who participated in our experiments, as well as to the audiences of AICED 2021, the UMass Amherst Syntax Workshop, GLOW 45 (Generative Linguistics in the Old World), and Chicago Linguistic Society 58. We also express our gratitude to Alina Tigǎu (the project mentor), Alexandra Cornilescu, Larisa Avram, Monica Irimia, Tom Roeper, Heidi Harley, Brian Dillon, Faruk Akkuş, Kyle Johnson, Shota Momma, and Rajesh Bhatt. References Alexiadou, A., E. Anagnostopoulou, & F. Schäfer. 2006. The properties of anticausatives crosslinguistically. In Phases of Interpretation, ed. by M. Frascarelli, 187-212. Berlin: Mouton. Alexiadou, A., E. Anagnostopoulou, & F. Schäfer. 2015. External Arguments in Transitivity Alternations: A Layering Approach. Oxford: Oxford University Press. Bleotu, A. C. 2017. What does ‘a cireşi’ (‘to cherry’) mean? Revisiting Kiparsky’s (1997) Canonical Use Principle from an experimental and theoretical standpoint. Ms. University of Bucharest. Available online: <lingbuzz/003743>. Bleotu, A. C. 2019. Towards a Theory of Denominal Verbs: A Look at Incorporation, Phrasal SpellOut and Spanning. (Series: Empirical Approaches to Linguistic Theory). Leiden, The Netherlands: Brill Publishing House. Bleotu, A.C., & J. Bloem. 2020. The distinction between true and pseudo denominals? It’s an illusion! In NELS 50: Proceedings of the Fiftieth Annual Meeting of the North East Linguistic Society (Vol. 1), ed. by M. Asatryan, Y. Song, & A. Whitmal, 39-48. Amherst: Graduate Linguistics Student Association, Department of Linguistics, University of Massachusetts. Bleotu, A.C., & J. Bloem. 2021. What’s the Meaning of a Nominal Root? Insights from Experiments into Denominals and Similarity. Poster presented at 39th Meeting of the West Coast Conference on Formal Linguistics (WCCFL). Available at: https://doi.org/10.25422/azu.data.c.5325401.v2. Bleotu, A. C., & R. Ivan. 2022. Voice Alternation in SE figure reflexives in Romanian: Evidence from a semi-artificial paradigm. In A Life in Linguistics: A Festschrift for Alexandra Cornilescu on her 75th Birthday, ed. by G. Alboiu, D. Isac, A. Nicolae, M. Tǎnase-Dogaru & A. Tigǎu, pp. 79–101. Bucureşti: Editura Universitǎţii Bucureşti. Borer, H. 2014. The category of roots. In The Roots of Syntax, the Syntax of Roots, ed. by A. Alexiadou, H. Borer, & F. Schäfer, 91–121. Oxford: Oxford University Press. Cornilescu. A., & A. Nicolae. 2017. On a class of figure reflexives in Romanian. Ms. Cornilescu A., & A. Nicolae. 2021. On a class of figure reflexives in Romanian: Ion se spală pe mâini ‘John washes his hands’. In Unraveling the complexity of SE. Studies in Natural Language and Linguistic Theory, ed. by G. Armstrong & J.E. MacDonald, vol 99, 187–218. Cham: Springer. https://doi.org/10.1007/978-3-030-57004-0_8. Dobrovie-Sorin, C. 1998. Impersonal se constructions in Romance and the passivization of unergatives. Linguistic Inquiry 29 (3). 399–437. Dobrovie-Sorin, C. 2017. Reflexive-marking in Romance: Voice and feature deficiency. In The Blackwell Companion to Syntax, 2nd revised edn., ed. by M. Everaert & H. van Riemsdijk. Oxford: Blackwell. Dobrovie-Sorin, C., & I. Giurgea. (forthcoming). Chapter 2. Voice. In A Reference Grammar of Romanian. Hale K., & S. J. Keyser. 2002. Prolegomenon to a Theory of Argument Structure. Cambridge: MIT Press. SE figure reflexives 43 Harley, H. 2012. External arguments and the Mirror Principle: On the distinctness of Voice and v. Lingua 125. 34–57. Harley, H., & J. Haugen. 2007. Are there really two classes of instrumentals?. Snippets 16. 6–7. Kelly, M. H.1998. Rule- and idiosyncratically derived denominal verbs: Effects on language production and comprehension. Memory & Cognition 26 (2): 369–381. Kiparsky, P. 1997. Remarks on denominal verbs. In Complex predicates, ed. by A. Alsina, J. Bresnan, & P. Sells. 473–499. Stanford: CSLI. Kratzer, A. 1994. The Event Argument and the Semantics of Voice. Ms, University of Massachusetts, Amherst. Kratzer, A. 1996. Severing the External Argument from its Verb. In Phrase Structure and the Lexicon, ed. by J. Rooryck, & L. Zaring, 109-137. Kluwer: Dordrecht. Irimia, Monica Alexandrina. 2020. Types of structural objects. Some remarks on differential object marking in Romanian. In Case, agreement and their interactions. New perspectives on differential argument marking, ed. by A. Bárány & L. Kalin. 76–126. Berlin: Mouton de Gruyter. Levin, B., & M. Rappaport Hovav. 1994. A Preliminary Analysis of Causative Verbs in English. Lingua 92: 35–77. Levin, B., & M. Rappaport Hovav. 1995. Unaccusativity: At the Syntax-Lexical Semantics Interface, Linguistic Inquiry Monograph 26. Cambridge, MA: MIT Press. Manoliu-Manea, M. 1996. Inalienability and topicality in Romanian: pragma-semantics of syntax. In The Grammar of Inalienability: A Typological Perspective on Body Part Terms and the Part–Whole Relation, ed. by H. Chappell & W. McGregor, 711–743. Berlin: Mouton de Gruyter. Marantz, A. 1984. On the Nature of Grammatical Relations. Cambridge, MA: MIT Press. Pesetsky, D., & E. Torrego. 2004. Tense, case, and the nature of syntactic categories. In The Syntax of Time, ed. by J. Guéron, & J. Lecarme. 495–538. Cambridge, MA: MIT Press. R Core Team. 2021. R: A language and environment for statistical computing. R Foundation for Statistical Computing. Vienna, Austria. Available online at https://www.R-project.org/. Rivero, M-L. 1990. The Location of Non-Active Voice in Albanian and Modern Greek. Linguistic Inquiry 21. 135-146. Schäfer, F. 2008. The Syntax of (Anti-)Causatives. External Arguments in Change-of-State Contexts. Amsterdam/Philadelphia: John Benjamins. Schäfer, F. 2009. The causative alternation. Language and Linguistics Compass 3 (2): 641–681. Svenonius, P. 2003. Limits on P: filling in holes vs. falling in holes. Nordlyd. Proceedings of the 19th Scandinavian Conference of Linguistics 3(2). 431–445. Svenonius, P. 2007. Adpositions, particles and the arguments they introduce. In Argument Structure, ed. by E. J. Reuland, T. Bhattacharya, G. Spathas, 63-103. Amsterdam / Philadelphia: John Benjamins. Talmy, L. 1978. Figure and ground in complex sentences. In Universals of Human Language, ed. by J. Greenberg, C. Ferguson, & H. Moravcsik, 627–649. Stanford, CA: Stanford University Press. Talmy, L. 1985. Lexicalization patterns: Semantic structure in lexical forms. In Language typology and syntactic description, ed. by T. Shopen, vol. 3: Grammatical categories and the lexicon, 57–149. New York: Cambridge University Press. Tigău, A. M. 2010. Towards an account of differential object marking in Romanian. Bucharest Working Papers in Linguistics 12 (1). 137–58. Wood, J. 2012. Icelandic Morphosyntax and Argument Structure. PhD dissertation, New York University. Wood, J. 2014. Reflexive -st verbs in Icelandic. Natural Language & Linguistic Theory 32 (4): 1387–1425. Wood, J., & A. Marantz. 2017. The interpretation of external arguments. In The Verbal Domain, ed. by R. D’Alessandro, I. Franco, & A. J. Gallego, 255–278. Oxford: Oxford University Press. Proceedings of CLS 58 (2022), 45-59 © Chicago Linguistic Society 2023. All rights reserved. 45 A new analysis of interpretive properties of the verb and subject in Spanish DOM Penelope Daniel University of Connecticut 1 Introduction Differential object marking (DOM) is a case pattern where only a subset of objects are case-marked, in contrast with a standard nominative-accusative case pattern where all objects are marked the same way. The exact subset of objects that are case-marked can vary across languages, but is typically defined according to the objects’ semantic properties. For example, Spanish has a well-known DOM pattern where the relevant semantic properties of the objects are specificity and animacy: in most cases, specific, animate objects are marked, while nonspecific and/or inanimate objects are not (1). The proper noun María, which can only have a specific, animate interpretation, must be marked with a, while the indefinite animate NP una mujer allows both possibilities, but is only interpreted as specific when marked with a. (1) a. Besó una mujer kissed a woman ‘He kissed a woman.’ (nonspecific) b. *Besó María kissed Mary ‘He kissed Mary.’ c. Besó a una mujer kissed DOM a woman ‘He kissed a woman.’ (specific) d. Besó a María kissed DOM Mary ‘He kissed Mary.’ (Rodríguez-Mondoñedo 2007) DOM has been widely studied within generative syntax, and has been the subject of numerous theoretical debates about the nature of case assignment. Many accounts of DOM focus on the interpretive properties of the object. However, I present oftenignored Spanish data that show that interpretive properties of the subject and the verb interact with DOM as well. In particular, Spanish DOM is always an option with accomplishments, it yields a bounded activity interpretation with activities and statives, and it is incompatible with inanimate, non-causer subjects. In order to address these interactions with the subject and verb, I propose an analysis in which different v-heads correspond with the availability of case on the 46 PENELOPE DANIEL object, as well as with the interpretive properties of the verb and subject. In section 2, I discuss the Spanish DOM data in more detail, in particular highlighting the way that DOM interacts with the interpretation of the subject and the verb. In section 3, I present my analysis, and in section 4, I argue that the data favor this type of approach to case assignment over a Dependent Case approach (in the domain of DOM). Section 5 concludes the paper. 2 Spanish DOM In most cases, specific, animate objects in Spanish are marked with the case marker a (1).1 Nonspecific and inanimate objects, on the other hand, are not a-marked. DOM is thus closely associated with the interpretation of the object, which is highlighted by the fact that it can yield a specific or animate interpretation for otherwise ambiguous NPs. As shown above, for an indefinite NP that may have a specific or nonspecific reading, the absence of a results in a nonspecific interpretation, as shown in (1a), while the presence of a results in a specific interpretation, as shown in (1c). Similarly, for an NP that can be either animate or inanimate, the presence or absence of a can help disambiguate between the two interpretations. As shown in (2), pueblo de Numancia can refer either to the people or the town of Numancia, and thus be either animate or inanimate: the presence of a yields the animate (‘people’) interpretation, while the absence of a yields the inanimate (‘town’) interpretation. (2) Buscan al/el pueblo de Numancia. looking.for DOM.the/the people/town of Numancia. ‘They are looking for the people/town of Numancia’ (Torrego 1998) Before moving forward, it is important to clarify some of the complexities of these properties that are often ignored. In particular, it has been observed that the notions of animacy and specificity are not quite fine-grained enough to capture the environments where DOM appears in Spanish. For example, RodríguezMondoñedo (2007) suggests that the notion of ‘autonomy of motion’ may be more accurate than ‘animacy,’ given that vehicles are compatible with DOM, as shown in (3). (3) ...atravesando el Pont Neuf, vi al barquito en cuestión crossing the Pont Neuf, saw DOM.the little.ship in question ‘Crossing the Pont Neuf, I saw the little ship in question.’ (Laca 1995; Rodríguez-Mondoñedo 2007) Furthermore, Rodríguez-Mondoñedo also proposes that the ‘possibility of having a referential interpretation,’ may be more accurate than the notion of ‘specificity.’ Importantly, the objects that cannot be marked are those that cannot be specific 1 While a is homophonous with the dative case marker/preposition ‘to,’ Bárány (2018) shows that this homophony is due to syncretism of accusative and dative case, based on the fact that a-marked direct objects pattern with non-a-marked direct objects in their syntactic behavior, rather than with dative indirect objects. I therefore follow Bárány (2018) in analyzing a as an accusative case marker. Interpretive properties of Spanish DOM 47 (e.g. objects of existential constructions, as shown in (4)), while the objects that must be marked can only be interpreted as specific (e.g. partitives, as shown in (5)) (Rodríguez-Mondoñedo 2007). However, the precise semantic characterizations of the properties are not crucial to my analysis, so I will continue to use ‘specificity’ and ‘animacy’ for the sake of demonstration. (4) Había (*a) un policía en el parque. there-was DOM a policeman in the park ‘There was a policeman in the park.’ (Rodríguez-Mondoñedo 2007) (5) Había besado *(a) varias de sus amigas. had kissed DOM several of his girlfriends ‘He kissed several of his girlfriends.’ (Rodríguez-Mondoñedo 2007) Importantly, in addition to these properties of the object, properties of the verb and subject, which have often been set aside in the literature, also interact with DOM in Spanish. For example, while animate, specific objects of accomplishments always have the option to be marked, a-marking has additional interpretive effects with non-telic predicates. Torrego (1998) shows that activities with marked objects are interpreted as bounded: according to Torrego, the difference between (6a) and (6b) is that while both have a repetitive reading, (6b) has an additional reading, where a single act of hiding is delimited. (6) a. Laura escondió un prisionero durante dos años. Laura hid a prisoner for two years b. Laura escondió a un prisionero durante dos años. Laura hid DOM a prisoner for two years ‘Laura hid a prisoner for two years’ (Torrego 1998) Torrego also shows that statives with marked objects are interpreted as bounded activities. Indeed, statives with marked objects show more activity-like properties, such as compatibility with the progressive. As shown in (7), the stative verb conocer (‘to know’) may be used in the progressive only if the object is marked. (7) Estoy conociendo *(a) un vecino. am knowing DOM a neighbor Intended: ‘I am getting to know a neighbor.’ (Torrego 1998) Finally, properties of the subject also interact with DOM. Torrego (1998) shows that DOM occurs when there is a causer subject (8) or animate subject (9), while inanimate, non-causer subjects are disallowed, as shown in (10). (8) El vino emborrachó *(a) varios invitados. the wine (made) drunk DOM several guests ‘The wine made several guests drunk.’ (9) El herido reclamaba a un médico the injured demanded DOM a doctor ‘The injured demanded a doctor.’ (Torrego 1998) (Torrego 1998) 48 (10) PENELOPE DANIEL *La situación reclamaba a un médico the situation demanded DOM a doctor Intended: ‘The situation demanded a doctor.’ (Torrego 1998) To summarize, Spanish DOM occurs with the following objects (11), verbs (12), and subjects (13). While only objects that are both animate and specific can be marked, they will only be marked in the context of a particular interpretation of the verb and subject, namely when the verb is interpreted as an accomplishment or bounded activity, and when the subject is either animate or an inanimate causer. Objects that are not both animate and specific are not marked, and no object, regardless of animacy or specificity, is marked when the verb is a non-bounded activity or stative, or when the subject is an inanimate non-causer. (11) Objects animate, specific animate, nonspecific inanimate, specific inanimate, nonspecific 3 7 7 7 (12) Verbs accomplishment 3 activity 7/3* stative 7/3* (*3 only when interpreted as a bounded activity) (13) Subjects animate 3 inanimate causer 3 inanimate, non-causer 7 3 Analysis While many approaches to Spanish DOM, such as Rodríguez-Mondoñedo (2007) and López (2012), do not directly address the interpretation of the verb and subject, Torrego (1998) argues that the interpretative properties of the verb and subject are a direct result of the same mechanism responsible for DOM. While I maintain the core of Torrego’s proposal that v is responsible for the interpretation of the subject, verb, and the case of the object, I provide an analysis that makes use of developments in the domains of case and v-heads that enable us to make this multifunctionality of v more precise. Specifically, I propose that different v-heads are responsible for the interpretation of the verb and subject, and that the v-head that values the case feature of specific, animate objects in Spanish is the same one that introduces animate or causer subjects, and yields an accomplishment or bounded activity interpretation. In the following subsections, I will elaborate in more detail on the relevant case feature on the object, and the relevant v-head that values it. Interpretive properties of Spanish DOM 49 3.1 On the nature of the case feature Recall that the difference between Spanish DOM and the basic NOM - ACC pattern is that only objects with certain semantic properties (animacy and specificity) are casemarked. I propose that the reason for this difference lies in the nature of the case features involved. Specifically, I propose that standard case assignment involves uninterpretable case features, while DOM involves interpretable case features. According to Bošković (2006, 2011), features are interpretable if they have semantic import. For example, Bošković (2006) argues that the case feature involved in the assignment of inherent case is an interpretable one, due to its semantic function of identifying the thematic role of an NP. Given the interaction between case-marking and the interpretation of the object in DOM, I propose that DOM should be analyzed as an additional phenomenon involving interpretable case features. Specifically, I propose that there is an interpretable case feature that is borne only by specific, animate objects in Spanish, and that this feature is what gives them their specific, animate interpretation.2 Importantly, this does not mean that nonspecific or inanimate objects are caseless: Bošković (2006) argues that interpretable case features do not participate in the traditional case filter, which is responsible for the licensing of arguments. I therefore assume that DOM driven by interpretable case features is distinct from that type of case assignment. I assume that the process of case assignment that is necessary for licensing will occur independently, but I will not discuss it in this paper, and I will show only the case features necessary for DOM in derivations. To differentiate this case feature from the one responsible for licensing of arguments, I will call it iDOM. In order to be morphologically realized as a-marking, the iDOM feature on an object needs to be valued. Note that this approach assumes that interpretability and valuation are independent from one another, following Bošković (2011), contra Chomsky (2000). In other words, it is possible for a feature to be valued and uninterpretable, or unvalued and interpretable. While interpretability depends on whether a feature has semantic import, whether a feature is valued or unvalued depends on whether its value is always the same, or whether its value depends on the context. According to Bošković (2011), the case features of NPs are unvalued, because the same NP may appear with one case in one context, and another case in another context. Case features on functional heads, however, are valued, because they always license the same case. Finite T, for example, always licenses nominative case. In Spanish, then, the iDOM feature of animate, specific objects is unvalued. Following previous approaches to Spanish DOM in which a-marked objects are located in a higher position than inanimate and nonspecific objects (Torrego 1998; Rodríguez-Mondoñedo 2007; López 2012), I assume that only animate, specific objects undergo object shift in Spanish. I adopt Bošković’s (2007, 2011) approach to movement, where movement is driven by a feature on the moving argument. 2 There are two possibilities for how all objects can either have this case feature or not. One possibility is that for each NP, there exist two versions in the lexicon: one with the case feature, and one without it. To avoid this burden on the lexicon, another option is that the case feature is located on the head of a KP, which has the option to be merged with any DP (see López 2012 for a KP analysis of Spanish DOM). I leave it to future research to tease these options apart. 50 PENELOPE DANIEL According to Bošković (2011), unvalued features function as probes, which must c-command their goal. Valuation is successful if the probe finds and enters an agree-relation with a matching, valued feature. This means that if an argument with an unvalued feature does not already c-command a matching valued feature, that argument must move to a position from which it does. I propose that this is what drives object shift in Spanish, as shown in (14): animate, specific objects have an unvalued iDOM feature, which they are unable to value in-situ, since they do not c-command a valued counterpart (14a). However, as will be discussed further in the following subsection, I propose that v carries a valued DOM feature (which will be uninterpretable (uDOM), since case is not interpreted on the verb). The object therefore raises to Spec,vP, where it successfully values its iDOM feature (14b). The interpretation, morphological realization of case, and movement of the object can therefore all be attributed to the unvalued iDOM feature. (14) a. vP VP v [val uDOM] V OBJ [ iDOM] b. vP OBJ v’ [val iDOM] VP v [val uDOM] V t 3.2 On the nature of the v-head I assume that the valued uDOM feature that is necessary for valuing the iDOM feature on animate, specific objects is located on a v-head, based on the traditional assumption that v is the locus of the feature responsible for accusative case (Chomsky 1995; Chomsky 2000; Chomsky 2001; Kratzer 1996). Crucially, a number of other functions besides case valuation have also been attributed to v-heads, such as encoding voice/aspect and introducing arguments. I propose that the interaction between DOM and the properties of the verb and subject are a consequence of this multifunctionality of v-heads. Specifically, I propose that in Spanish, there are different v-heads, and only one of them carries the valued iDOM feature. Note that it has been widely assumed that other case-assigning functional heads, Interpretive properties of Spanish DOM 51 such as T and Voice, can vary in whether they carry a case feature, according to their properties: finite T can assign nominative case but non-finite T cannot, and in ergative-absolutive systems, transitive Voice can assign ergative case while intransitive Voice cannot. Perhaps the closest parallel is with active and passive Voice/v, where the active Voice/v-head is standardly assumed to be able to assign accusative, while the passive cannot. I therefore assume that there are different v-heads (e.g., Folli & Harley 2005), and that these can also vary in whether they carry a case feature. In particular, I propose that the v-head that corresponds with accomplishments and bounded activities, and allows animate or causer subjects, carries the valued uDOM feature that values the iDOM feature on specific, animate objects in Spanish. The v-heads that correspond with non-bounded activities and statives, however, do not. I assume v-heads are the locus of these properties, following a number of authors who argue that different v-heads encode different event structure properties (Ramchand 2008; Rothmayr 2009; Cuervo 2015). In particular, I adopt Ramchand’s (2008) idea that v-like heads can directly instantiate aspectual semantic information such as initiation or causation, and dynamic processes, which I refer to as v[init] and v[proc], respectively. According to Ramchand, these heads can appear alone, or together, yielding different interpretations of the verb. For example, when v[proc] appears alone, the verb is interpreted as an activity that is not bounded. When v[init] appears alone, the verb is interpreted as stative. When v[init] and v[proc] appear together, the resulting interpretation of the verb is that of an accomplishment or bounded activity (which includes both activity and stative verbs with a bounded activity interpretation). For Ramchand (2008), when two heads co-occur, each one projects, with one vP above another. However, given that this combination of heads yields a unique interpretation, and, as I will discuss below, also has unique subject-introducing and case-assigning properties, I propose that when v[init] and v[proc] co-occur, the two heads are bundled (see Pylkkänen 2008; Harley 2017, who discuss the possibility of bundling v-heads with Voice), which enables them to act as a single head, which I call v[init+proc]. The inventory of v-heads will therefore include the distinct heads v[init], v[proc], and v[init+proc], which all encode different event structure information. In addition to introducing different predicate types, I adopt Ramchand’s idea that another function of these heads is licensing different types of arguments. In other words, each v-head has its own requirements for what kind of subject it can introduce. Note that in addition to their different case assigning properties, the active and passive Voice/v-heads also vary in their subject-introducing properties, with the passive prohibiting the introduction of overt subjects altogether (unless introduced in a by-phrase or with an oblique case). I propose that v[init+proc] is similar in the sense that it prohibits a particular kind of subject. Specifically, v[init+proc] is incompatible with inanimate, non-causer subjects. v[init] and v[proc], however, allow inanimate, non-causer subjects. I therefore propose that v[init+proc] is the head that is responsible for the accomplishment or bounded activity interpretation of the verb, introduces an animate or causer subject, and carries the uDOM feature in Spanish. v[init] and v[proc], on the other hand, which yield stative and non-bounded activity interpretations, and do not impose such restrictions on their subjects, both lack a uDOM feature. 52 PENELOPE DANIEL 3.3 Deriving Spanish DOM Putting everything together, the object will only be marked in Spanish when both a) the v-head is v[init+proc], and b) the object bears an unvalued iDOM feature. As shown in (15), modified from (14b), when both of these conditions are met, the unvalued iDOM feature on the object triggers object shift: the object raises to Spec,v[init+proc] and then probes the valued uDOM feature on v[init+proc]. As a result, the object successfully values its own iDOM feature, which is spelled out as a, and gives the object an animate and specific interpretation. In addition, v[init+proc] will ensure that the verb is interpreted as an accomplishment or bounded activity, and that the subject is animate, a causer, or both. (15) v[init+proc]P SBJ v[init+proc]’ OBJ v[init+proc]’ [val iDOM] VP v[init+proc] [val uDOM] V t On the other hand, DOM will not occur if either a) the object does not have an iDOM feature, or b) the v-head is v[init] or v[proc]. The former scenario is relatively straightforward: if the object has no unvalued iDOM feature, which is the case for nonspecific and inanimate objects, it cannot receive a value, so it will not be a-marked. As shown in (16), even when v[init+proc] is present with its valued uDOM feature, if the object has no unvalued iDOM feature, it will have no reason to shift, and remains in its base position where it will not be a-marked.3 3 Note that the subject c-commands the uDOM feature on v[init+proc]. Furthermore, since all NPs can exist with or without an iDOM feature, the subject could, in theory, be introduced with an unvalued iDOM feature, which could then be valued via agreement with v[init+proc]. This does not arise when the object has an iDOM feature, due to the timing of operations: when v[init+proc] enters the derivation, the object raises to its specifier and immediately values its iDOM feature, triggering the deletion of the uDOM feature on v[init+proc] before the subject is even introduced. The complication does arise, however, when the object does not have an iDOM feature, as in (16). While this is a more general complication within a system that assumes that an argument must ccommand its case assigner, I argue that it may be resolved through slight modifications, and I offer two possibilities for Spanish. One is that the subject is not actually introduced by the same v-head that carries the uDOM feature, but by a distinct head, Voice (see e.g., Legate 2014). Under this possibility, v[init+proc]’s restriction on inanimate non-causer subjects would be accomplished through a selectional relationship between Voice and v, where the Voice-head that selects for v[init+proc] can only introduce animate and/or causer subjects. In Daniel (in prep.), I argue that this type of selectional relationship between Voice and v is independently required to account for other differential argument marking phenomena cross-linguistically. The other possibility is that since Spanish is a Interpretive properties of Spanish DOM (16) 53 v[init+proc]P SBJ v[init+proc]’ VP v[init+proc] [val uDOM] V OBJ In the latter scenario, v[init] and v[proc] do not have a uDOM feature, so an object with an unvalued iDOM feature would not be able to receive a value, as shown in (17) for v[init] (the derivation for v[proc] would be identical). This prohibits the possibility of a-marking with statives and non-bounded activities.4 (17) v[init]P SBJ v[init]’ OBJ v[init]’ [ iDOM] VP v[init] 7 V t To summarize, v[init+proc] is responsible for the accomplishment or bounded activity interpretation of the verb, the prohibition of inanimate, non-causer subjects, and the ability to value the case feature of shifted objects. This correctly predicts pro-drop language, the subject NP is not introduced by v or Voice, and instead is introduced in a non-argument position, following Alexiadou & Anagnostopoulou (1998). For both of these options, I suggest that there is a phase boundary between the subject and the uDOM feature on v[init+proc], and that either the uDOM feature is deleted upon Spell-Out of the phase, or the subject cannot probe the uDOM feature across the phase boundary. 4 Note that since this approach assumes that unvalued features drive movement, the fact that the object cannot value its iDOM feature in Spec,v[init]P or Spec,v[proc]P might lead one to expect the object to keep raising without ever valuing its iDOM feature. However, recall that there is a separate mechanism of case assignment that is required for licensing. While that mechanism is outside the scope of this paper, I propose that the valuation of that case feature on the object has a freezing effect, preventing it from moving further. Whether this takes place before or after the unvalued iDOM causes it to raise to Spec,v[init]P or Spec,v[proc]P (in other words, whether the object is frozen in its base position or in Spec,v[init]P/v[proc]P), depends on whether animate, specific objects undergo object shift even when they are not a-marked. According to López (2012) they do, while according to Torrego (1998) and Rodríguez-Mondoñedo (2007) they do not. I do not commit to one position or the other here, and leave the matter to future research. 54 PENELOPE DANIEL the contexts where DOM does and does not occur, as shown in (18-20) below. Only animate, specific objects are marked, since these are the only objects with an iDOM feature. DOM also only occurs with accomplishments, as well as activities and statives that have a bounded activity interpretation, since these are the only verbs introduced by v[init+proc]. Finally, DOM only occurs with animate or inanimate causer subjects, since these are the only subjects that can be introduced by v[init+proc]. (18) (19) (20) Objects animate, specific animate, nonspecific inanimate, specific inanimate, nonspecific 3 7 7 7 Verbs accomplishment activity stative v[init+proc] v[init+proc]/v[proc] v[init+proc]/v[init] 3 3/7 3/7 Subjects animate inanimate causer inanimate, non-causer 3 3 7 iDOM feature no iDOM feature no iDOM feature no iDOM feature v[init+proc] v[init+proc] incompatible with v[init+proc] This analysis captures a number of different observations about Spanish DOM through a combination of insights from work on interpretable case and v-heads. The unvalued interpretable case features accounts for object shift and the interpretation of the object, and the v-heads account for the interpretation of the verb and subject. The valuation of the unvalued interpretable case feature via agreement with a particular v-head is what accounts for the morphological marking of the object. Since interpretable case features and the types of v-heads employed here have been independently motivated for other phenomena, Spanish DOM represents an important example of where all of these phenomena intersect and interact with one another. 4 Advantages over Dependent Case As mentioned in the introduction, DOM has been the subject of much debate in the literature on case assignment. While the proposed analysis assumes that case assignment is the result of a relationship between a functional head and the object, other approaches assume that case is assigned configurationally via Dependent Case assignment. Under the Dependent Case approach, case assignment is the result of two arguments co-occurring in the same domain relevant for case assignment (Marantz 2000; Baker & Vinokurova 2010; Baker 2015). Baker & Vinokurova (2010), for example, propose a Dependent Case analysis of Sakha DOM, where only specific objects are marked with accusative case, as shown in (21). Interpretive properties of Spanish DOM (21) a. Erel kinige-ni atyylas-ta. Erel book-ACC buy-PAST.3sS ‘Erel bought the book.’ b. Erel kinige atyylas-ta. Erel book buy-PAST.3sS ‘Erel bought a book/books.’ 55 (Baker & Vinokurova 2010) In the Dependent Case framework, a ‘dependent case’ is assigned when two noncase-marked arguments are located in the same case domain (which are commonly argued to correspond to phases; see Baker 2015). In a NOM - ACC language like Sakha, that case is accusative, assigned to the lower of the two arguments. For Sakha, Baker & Vinokurova (2010) propose that subjects are located in the phase above the object, as shown in (22), where bolding indicates the phase boundary. When the object is nonspecific, it remains within the lower phase, as shown in (22b). When the object is specific, it undergoes object shift into the next highest phase (22a). Specific objects therefore receive accusative case because they are in the same phase as the subject, but nonspecific objects will not since they are not in the same phase as another argument. (22) a. [vP Erel [VP book-ACC [VP t buy]]] b. [vP Erel [VP book buy]] (Baker & Vinokurova 2010) Since similar properties of the object are involved in both Sakha and Spanish DOM, one might want to adopt a similar approach to what (Baker & Vinokurova 2010) propose for Sakha: the specific, animate objects would undergo object shift into the same phase as the subject, while non-specific and inanimate objects would not. However, such an approach would have nothing to say about the interpretation of the subject and verb. An additional mechanism would therefore be required to account for these interactions in Spanish.5 In the current analysis, however, v-heads provide a natural way of correlating case with these kinds of interpretive effects. Furthermore, there is independent evidence that Spanish DOM occurs even when the object is the only argument in its phase. Villa-García (2012) shows that in Spanish ‘double-que’ constructions (23a,25), topics are base-generated in the left periphery. These double-que constructions fail to show reconstruction effects, such as disallowing bound variable readings (23a), unlike topic constructions with just the higher que, which do allow them (23b) (this in fact holds for all types of reconstruction effects, see Villa-García 2012). (23) 5 a. Dice que en su∗i/j hijo, que todo el mundoi tiene que confiar says that in his son that all the world has that trust One possibility is that the different v-heads discussed above vary not only in the interpretation of the verb and the subject, but also in their phasehood status. In other words, one might propose a Dependent Case analysis similar to the current analysis where, instead of differing in whether they have a uDOM feature, v-heads differ in whether they are phase heads: v[init]P and v[proc]P would be phases, so the object and subject would be in separate phases, and the object would not be assigned accusative. On the other hand, v[init+proc]P would not be a phase, so the object and subject would be in the same phase, and the object would receive accusative. However, see below for additional arguments against a Dependent Case approach to Spanish DOM. 56 PENELOPE DANIEL ‘S/he says that everybody has to trust his/their (=somebody else’s) son.’ [bound reading: 7] b. Dice que en sui/j hijo, todo el mundoi tiene que confiar says that in his son all the world has that trust ‘S/he says that everybody has to trust their son.’ [bound reading: 3] (Villa-García 2012) A Dependent Case account would therefore predict that the topicalized object in double-que constructions would not be a-marked, since it is not in the same phase as the matrix subject, and is higher than the embedded subject (meaning one might even expect the embedded subject to receive the a-marking under the Dependent Case approach), as shown in (24). However, this prediction is not borne out: the topicalized object is a-marked in the double-que example (25). (24) [vP SBJ [VP ... [CP que [XP OBJ que [vP SBJ [VP ...]]]]]] (25) Dicen que a su perro, que no lo reconocen say that DOM his dog that not cl. recognize ‘They say that they don’t recognize their dog.’ (Villa-García 2012) This problem does not arise within the current analysis: the object may simply probe v[init+proc] from topic position and value its iDOM feature, allowing it to receive accusative case (in line with Bošković’s 2007, 2011 approach to case, which Villa-García in fact argues for based on sentences like (25), where the basegenerated topic probes a lower functional head for case). (26) CP C XP que X’ OBJ [val iDOM] X ... que v[init+proc]P SBJ v[init+proc]’ v[init+proc] [val uDOM] VP Interpretive properties of Spanish DOM 57 As a result, not only does the Dependent Case approach have the conceptual disadvantage of requiring an additional mechanism to account for the interpretive effects of Spanish DOM, but it also makes the wrong predictions about DOM in doubleque constructions and cannot account for them, while the current analysis easily can. 5 Conclusion I have proposed an analysis of Spanish DOM in which a v-head, namely v[init+proc], bears the feature responsible for valuing the interpretable case feature of a shifted object. Crucially, I have shown that this analysis captures not only the case of the object, but also its interpretation, as well as the interpretation of the subject and verb. Note that this is especially important since the interpretive properties of the subject and verb have often not been directly addressed in recent analyses of Spanish DOM. This is accomplished through the multifunctionality of v-heads, which assign case, introduce subjects with certain semantic properties, and encode certain event structure properties: v[init+proc] yields an accomplishment or bounded activity reading and also licenses an animate or causer subject. In addition, the interpretable case feature on the object in DOM is responsible for its animate, specific interpretation, and for driving object shift to Spec,v[init+proc], which allows it to be valued and realized as a. In this approach to Spanish DOM, both the realization of case and all of its interpretive effects are derived through the unified mechanism of valuation by a v-head. I have also shown that this analysis is a better fit for Spanish DOM than a Dependent Case approach. On the conceptual side, the Dependent Case approach would require an additional mechanism to account for the interpretative properties of the subject and verb, while the proposed analysis derives them for free. More problematic, however, is empirical evidence that a Dependent Case approach would make incorrect predictions about DOM in double-que constructions, while I have shown that the proposed analysis yields the correct result. While I have only demonstrated how the type of analysis proposed here has advantages over Dependent Case for DOM in Spanish, I argue that this has broader implications for the theory of Dependent Case more generally. While one might argue that DOM could be assigned by Dependent Case in some languages, and by agreement with a functional head in others, I argue that it would be desirable for a theory of case to be able to account for all case phenomena. The Spanish DOM data is therefore problematic for Dependent Case, while agreement with a functional head appears to be a more promising direction for a unified theory of case assignment. This proposal also has implications for future work on DOM, as well as differential argument marking (DAM, which also includes differential subject marking (DSM)) more broadly. Since DAM is always closely tied with the interpretative properties, I argue in Daniel (in prep.) that interpretable features are always involved in DAM. However, since the precise interpretative effects vary across DAM patterns, it is possible that the head that values that interpretable feature, and the interpretation that results from its valuation, may vary across languages. 58 PENELOPE DANIEL References Alexiadou, A., & E. Anagnostopoulou. 1998. Parametrizing agr: Word order, v-movement and epp-checking. Natural Language & Linguistic Theory 16. 491–539. Baker, M. 2015. Case. Number 146. Cambridge University Press. Baker, M. C., & N. Vinokurova. 2010. Two modalities of case assignment: Case in Sakha. Natural Language & Linguistic Theory 28. 593–642. Bárány, A. 2018. Dom and dative case. Glossa: a journal of general linguistics 3. 1–40. Bošković, Ž. 2006. Case checking versus case assignment and the case of adverbial nps. Linguistic Inquiry 37. 522–533. Bošković, Ž. 2007. On the locality and motivation of Move and Agree: An even more minimal theory. Linguistic inquiry 38. 589–644. Bošković, Ž. 2011. On valued uninterpretable features. In Proceedings of Northeast Linguistic Society, volume 39, 109–120. Chomsky, N. 1995. The minimalist program. MIT press. Chomsky, N. 2000. Minimalist inquiries: The framework (mitopl 15). Step by step: Essays on minimalist syntax in honor of Howard Lasnik 89–155. Chomsky, N. 2001. Derivation by phase. In Ken Hale: A life in language, 1–52. MIT Press. Cuervo, M. C. 2015. Causation without a cause. Syntax 18. 388–424. Daniel, P. in prep. A unified analysis of differential argument marking. Ms., University of Connecticut. Folli, R., & H. Harley. 2005. Flavors of v. In Aspectual inquiries, 95–120. Springer. Harley, H. 2017. The “bundling” hypothesis and the disparate functions of little v. The verbal domain 3. 3–28. Kratzer, A. 1996. Severing the external argument from its verb. In Phrase structure and the lexicon, 109–137. Springer. Laca, B. 1995. Sobre el uso del acusativo preposicional en español. In El complemento directo preposicional, 61–92. Visor. Legate, J. A. 2014. Voice and v: Lessons from Acehnese, volume 69. MIT Press. López, L. 2012. Indefinite objects: Scrambling, choice functions, and differential marking. MIT press. Marantz, A. 2000. Case and licensing. Arguments and case: Explaining Burzio’s generalization 11–30. Pylkkänen, L. 2008. Introducing arguments, volume 49. MIT press. Ramchand, G. 2008. Verb meaning and the lexicon: A first-phase syntax, volume 116. Cambridge University Press Cambridge. Rodríguez-Mondoñedo, M. 2007. The syntax of objects: Agree and differential object marking. University of Connecticut. Rothmayr, A. 2009. The structure of stative verbs, volume 143. John Benjamins Publishing. Torrego, E. 1998. The dependencies of objects, volume 34. mit Press. Villa-García, J. 2012. The Spanish complementizer system: Consequences for the syntax of dislocations and subjects, locality of movement, and clausal structure. University of Connecticut. Proceedings of CLS 58 (2022), 59-79 © Chicago Linguistic Society 2023. All rights reserved. 59 Semiotic distinctions in compositional semantics Kathryn Davidson Harvard University 1 Introduction Within traditional areas of focus in linguistics, the long discussed arbitrariness of the linguistic sign (Saussure 1916) highlights the independence of many linguistic forms from their meaning, such as the fact that the form of the English word cat has no particular relationship to any concept of a cat. Put another way, the relationship between the word cat and the concept of a cat is entirely by convention, thus arbitary. Arbitariness is often contrasted with iconicity, in which forms bear some relation to their meanings. Examples of iconic linguistic forms come in all modalities including both spoken languages and sign languages, such as the Amer‘cat’1 which follows the shape of a cat’s ican Sign Language (ASL) word whiskers, English onomatopeia like meow that sound somewhat like the sounds a cat makes, or an English speaker’s or ASL signer’s gesture of patting a cat that co-occurs with either speaking or signing, all of which involve an expressive form used as part of language, having some non-arbitrary relationship to its meaning. Although there is recent interest in the extent of iconicity/non-arbitrariness among words across the world’s spoken languages (Blasi et al. 2016, Winter & Perlman 2021) and a surge of research on both iconic and noniconic aspects of sign languages and gesture (see Goldin-Meadow & Brentari 2017 and replies for overview), there is largely a lack of consideration for the role that iconicity plays in composing sentence meaning and interacting with the logical structure of language. For example, if a form A is iconic, and composes with form B, is the iconicity in the resemblance of a form to its meaning interpreted as part of the composite expression (A + B), or is it ignored in the interpretation of the composite? How is it affected by logical/functional operators in language like negation (not), disjunction (or), questions (does...?) etc? Within cognitive linguistic and psycholinguistic approaches, it is often assumed that iconicity is a mechanism for adding meaning/ content, for example via conceptual structure-mapping (Emmorey 2014). Work on iconicity in cognitive linguistics includes many rich observations on how iconicity can both support (Taub 2001) and constrain (Meir 2010) non-compositional meaning like metaphors, in general placing more focus on meaning via iconicity than on meaning via function composition. In contrast, formal semantic approaches foreground semantic compositionality, but mostly ignore iconicity. One line of formal semantic research that does engage with 1 This image and all other images of single lexical signs from American Sign Language are from the ASL Sign Bank, (Hochgesang et al. 2020) 60 KATHRYN DAVIDSON the question of iconic content and compositionality focuses on co-speech gestures, and finds that they differ in their compositional properties from (arbitrary) spoken words, such as their interaction with spoken language negation (Ebert & Ebert 2016, Tieu et al. 2017). Formal explanations have generally focused on the issue of gesture and speech existing in separate modalities (Schlenker et al. 2013, Schlenker 2018a, Tieu et al. 2019), attributing differences in composition to the mis-match between the spoken and visual language modality for co-speech gestures and not on the fact that the gestures are iconic. Esipova (2019), in contrast, argues that the distinction stems from structural considerations, not modality ones, and in fact is quite compatible with the approach taken here, but also does not take iconicity to be a driving factor in these different structural arrangements. Despite differences, all of these formal semantic approaches to gesture in spoken languages work under the assumption that the iconicity of the gestural content has the same potential for compositionality as non-iconic content, and in fact treat iconic meaning in gestures within the same kind of formal system used for non-iconic words in the language (Schlenker 2021). Outside of gesture, in related formal semantic work on sign languages the question of iconicity and compositionality has arisen in such varied areas as verb forms (Strickland et al. 2015, Kuhn & Aristodemo 2017), adjectives (Aristodemo & Geraci 2018), and anaphora (Kuhn 2020), and the approach has generally been to take iconicity as an advantageous property of the visual modality that allows us to see compositional structure that is present but might be hidden in spoken languages. The assumption again is that the structure in (more iconic) sign and (less iconic) speech is generally the same. Thus we see that in many different approaches to iconicity and sentence meaning, iconic content is often assumed to have similar compositional properties as non-iconic content, often without questioning the role that iconicity itself could play in constraining composition. The goal of this short paper is to provide some insights into the question of iconic compositionality from a semiotic perspective by dividing iconic content by what it requires to be interpreted, directly following work on depiction in semiotics and psychology (Clark 2016, Dingemanse et al. 2015, Dingemanse & Akita 2017, Ferrara & Hodge 2018, Kita 1997), consistent with a wide spectrum of philosophical approaches to depiction that separate its interpretation from the kind of interpretation required by symbolic compositional language (Fodor 2007, Burge 2018). Following in the footsteps of these psycho-semiotic foundations, we will divide iconic language into two categories. On the one hand, there are iconic form-meaning pairings which are symbols that derive meaning via convention: these include onomatopeia like chirp in English, expressive sensory ideophones like gelegele ‘shiny’ in Siwu (Dingemanse 2019) or gorogoro ’heavy rolling object’ in Japanese (Kita 1997), emblems that build on metaphors like the thumbs up gesture for ‘good’, and visually iconic signs like ‘cat’ in ASL. We can categorize all of these forms as having lexical/descriptive iconicity: the forms are not entirely arbitrary, since we can identify how aspects of the form relate to their meaning, but their meaning is not derived from the form but rather from this conventionalized form-meaning pairing (the stored lexicon), and so it has the potential to be represented by an entirely arbitrary symbol. For example, the ASL word and Semiotic distinctions in compositional semantics 61 Iconicity in the lexicon Depictive iconicity (Some) lexical signs: INFORM in ASL Classifier handshapes in ASL: (Some) size gestures: This huge tail Classifier movements, locations in ASL: ‘vehicle’ Ideophones in Siwu: gelegele ‘shiny’ Onomatopeia: Margola chirped in her cage. Categorical gestures: ‘vehicles like this’ Siwu ideophone w/depiction: gelegele-gelegelegelegelegele “Expressive" onomatopeia: Margola went “chiiirrrrp!" Quotations: Alexis was like, “No, David!” Table 1: Lexical/descriptive iconicity vs. depictive iconicity the English word for ‘cat’ each pick out the same concept; this mapping has to be stored conventionally, and could have been mapped differently, given the different words for ‘cat’ across both sign and spoken languages of the world, and their sometimes arbitrary nature (e.g. English cat). Similarly, gelegele and shiny might be iconic in having some non-arbitrary aspects of their form (perhaps they seem related to light, for example), but as symbols they contribute a basic meaning via a conventionalized form-meaning mapping. Iconicity within the symbolic lexicon contrasts with depictive iconicity that need not express stable or abstract concepts and thus could not be expressed equally as well by an arbitrary symbol, and instead must be interpreted not via the lexicon (form-meaning conventionalized mappings) but instead via a non-symbolic perceptual mapping of the same sort used to interpret images. Consider, for example, that a painting or a photograph of a cat contributes its meaning through resemblance of looking like a cat. Note that an image is necessarily of a particular cat (in a particular pose, etc) whereas a symbol need not commit to any particular instantiation of the cat kind or concept. In short, depiction involves showing, not telling. As Hodge & Ferrara (2022) note, depiction can accompany descriptive language in both spoken and signed languages, thus it seems possible to talk about “depictive language” as depictions occuring while using language, which can be more and less integrated into a linguistic utterance. Depictive iconicity includes modifications of onomoatopeia and ideophones to mimic a particular sound or perceptual experience, gestures that depict/show a particular size, depictive verbs in sign languages which illustrate a particular arrangement, and even quotation-like demonstrations that depict another’s actions, attitudes, or speech (Clark & Gerrig 1990, Davidson 2015), all of which are understood as providing meaning by showing. The interlocutor understands depictions through their own perceptual experiences of what a particular object or event might have looked like, felt like, sounded like, etc. Descriptive/lexical iconicity involves conventionalized forms mapping to discrete meanings (or an area/network of meetings, when we consider polysemy) in some kind of stored lexicon; depictive iconicity involves neither, and instead meaning derives from mapping a form directly to a particular pictoral/sensorial representation. In many areas of lingustics these have both been categorized as “iconicity", and this conflation seems to have been a source of confusion in both cognitive and 62 KATHRYN DAVIDSON formal approaches. Here we take this semiotic distinction as potentially crucial to understanding the role of iconicity in compositionality. Table 1 provides several contrastive examples: the first column involves forms that bear some relation to their meaning, i.e. they are not entirely arbitrary, but also must be stored in a lexicon and are interpreted as abstract symbols. In the right column are cases of iconic depiction in language, which are not interpreted merely as symbolic but necessarily involve understanding depictions. Notice that there are many cases where (descriptively iconic) symbols from the left column (e.g. classifier handshapes in sign languages, ideophones or onomatopeia in spoken languages) are especially natural starting bases for further using language depictively; see discussion of this point especially in Dingemanse 2015. 2 Depiction and compositionality We can build upon this distinction between lexical/descriptive iconicity and depictive iconicity to more carefully pose our original question: how should we think about the contributions of iconic language from the perspective of compositional semantics? For lexical/descriptive iconicity, the answer appears quite obvious: these are symbols that contribute their meaning just like non-iconic words do. For ‘cat’, or ‘inform’ (which is iconic, in that example, the ASL words it shows information coming out of the head) and the English words cat or inform (which do not have much obvious iconicity) are all symbols. The symbols and cat have the same compositional contribution: λx.cat(x) (a function that rehave turns true for entities that are cats).; similarly, the symbols inform and the same compositional contribution: λv.inform(v) (a function that returns true for informing events). Although some are motivated in their form, all of these signs are ultimately abstract symbols, and compose just like other abstract symbols. We see the same thing in onomatopeias like chirp or knock, which similarly are also abstract symbols (i.e. chirp contributes the function that returns true for chirping events, λv.chirp(v)), even if they are (descriptively/lexically) iconic in the sense that they are motivated in form and can be quite easily used to support depictions. While the compositional contributions of lexically/descriptively iconic words are straighforward, it is a wide open question how to think about depictive iconicity with regard to compositional semantics. On the one hand, why and how would a picture compose with language? If we were looking at a children’s storybook that contained text and pictures, we wouldn’t typically view the job of the formal semanticist (or syntactician, or any linguist) to understand how the pictures compose with the text. On the other hand, there has been a recent flourishing of interesting work on precisely this, how to understand pictures alone or pictures accompanying text, by cognitive scientists using the tools of formal semantics that were designed for Semiotic distinctions in compositional semantics 63 natural language. For example, Abusch (2012) and Abusch & Rooth (2022) apply a formal approach to the analysis of narrative picture sequences, Cohn (2012) similarly considers the semantic structure of comic sequences, and Greenberg (2021) applies the notion of entailment that we see in formal semantics to a broad range of pictures. These seem to counter the seemingly default view that pictures fall outside the domain of formal linguistic analysis; applying formal semantic tools to them can be thought of as part of a larger investigation into applying the tools of semantics to domains beyond language (Schlenker 2018b). Thus, how we handle depictive iconicity becomes an important question: do iconic depictions compose (or not) with symbolic structures in language, and if so, how? As we noted above, much work that investigates compositionality and iconicity takes place within gesture or sign languages, categories defined by their modality, not semiotic distinctions: iconic gestures can involve lexical iconicity and depictive iconicity, and similarly, sign languages involve both lexical iconicity and depictive iconicity. The goal of the current work is to gain some additional insight, then, beyond previous studies looking at linguistic iconicity and composition, by carefully focusing on depictive iconicity, where the stakes are most interesting. Section 3 presents two short quantitative studies investigating the composition of depictive language in written English, a language and modality not typically associated with iconicity. Section 4 presents a proposal based on this data and existing data in the literature, arguing that depictions cannot participate in the construction of alternatives, necessary for composition in many functional areas of the grammar like questions, negations, disjunction, etc. Section 5 applies the proposal to depictions in other domains including Japanese ideophones and depictive classifiers in American Sign Language, as well as English onomatopeia. Section 6 concludes. 3 Naturalness ratings 3.1 Methodology and design This section describes two short experiments in English to test the hypothesis that iconic language, especially iconic depictions, do not compose with logical operators in the same way as symbolic language. The goal is to understand compositional reflexes of this semiotic distinction. We will focus on the behavior of symbolic language vs. depictions under negation, since negation is in some sense the simplest logical operator, and every human language in the world seems to have a symbolic means of expressing negation. Negation has also been a domain in which compositional differences have been reported between (sometimes depictive) co-speech gestures and words with seemingly similar content in spoken languages. The goal will be to compare depictions and symbols in the context of positive and negative versions of the same logical structure, for example, contrasting the way that depictive vs. symbolic languge composes with positive adverbials (e.g. always) vs. adverbials that express negation (e.g. never), with the hypothesis that depictions may be less natural than symbolic content in a negative context. The methodological approach taken here in some sense “a little bit experimental” (Davidson 2020). Instead of hiding the contrast of interest among many fillers, we highlight the contrast by presenting participants with only ten total target sen- 64 KATHRYN DAVIDSON Figure 1: Screenshot of training trials tences simultaneously, an effective tactic in in previous work in experimental semantics (Marty et al. 2020). Participants, recruited via Prolific Academic, were asked to help teach an “alien from another planet” to blend in among speakers in the participant’s community. For the alien’s sake, participants were asked to rate sentences on a slider bar from “Very unnatural, I can’t imagine anyone saying it like this.” to “Very natural, I could imagine myself or someone else saying it like this.” The goal with this phrasing was to broaden the typical notion of acceptability, which might be difficulty to judge while incorporating depictions, to a task of “naturalness”, motivated by the kind of language teaching tasks often used to elicit acceptability judgments by children in language acquisition research. Each mini study involved training with three example sentences (the same for each study): one was natural, one intentionally unnatural (notably: ungrammatical), and one somewhat in between, as shown in the screenshot in Figure (1). Following the training trials, the ten target trials in each mini study were presented in a randomized order. Participants were reimbursed $1.50, an average of $27/hr as estimated by Prolific Academic, the site used to recruit all participants. Semiotic distinctions in compositional semantics 65 3.2 Depiction via vowel lengthening In the first mini-study, participants (n=52) were presented with a depictive contrast in vowel lengthing. The English adjective huge is a (mostly) arbitrary symbol, with little relation between the form and its meaning beyond perhaps some descriptive/lexical iconicity in the choice of low vowels for large things. It contributes its meaning by convention. However, vowel lengthening can be applied to the low vowel to metaphorically convey a larger size using a kind of depictive iconicity, wherein the length of the vowel depicts in a sense the size of a tree. Participants were asked to provide naturalness ratings of 10 total sentences. These were based on five sentence frames, in pairs that varied in whether the adjective involved a lengthened vowel, for a total of ten sentences, which were all presented on screen together in randomized (by participant) order. The sentence frames included a positive episodic sentence, one with a universal quantifier, a negative quantifier, a negative modal, and one intentionally constructed to be not well formed in the language (1). (1) a. The last time I felt so relaxed, I was sitting under a {huuuge, huge} palm tree in California, enjoying the moment. (positive episodic) b. My friend likes the shade so he always sits under a {huuuge, huge} palm tree next to the boardwalk. (universal quantifier: always) c. My friend likes the sunshine so he never sits under a {huuuge, huge} palm tree next to the boardwalk. (negative universal quantifier: never) d. There are small ones indoors, but I can’t find any {huuuge, huge} palm trees in Mass. I don’t think they exist. (negative modal: no exists) e. I love sitting under {huuuge, huge} in California, the shade feels great. (ungrammatical) The hypothesis was, first, that depiction would be most natural in a positive episodic sentence, since the sentence describes a particular event and one could, in part, metaphorically depict aspects of the huge tree through vowel lengthening. Depiction was hypothesized to be worst in negative contexts (like the negative universal quantifier or negative modal) since these involve no entailment that there is an event or object to depict. Thus, depiction should interact with sentence frame, such that sentences without depiction may vary (for example, sentences both with and without depiction are expected to be rated unnatural in the ungrammatical sentence frame), but depiction will be more unnatural under negation than non-depiction sentences, which should be natural in all but the ungrammatical sentence frame. Data were modeled in R as glm with the acceptability (response) as dependent variable and depiction (huge vs. huuuge) and sentence type/environment as independent variables (Response ∼ Depiction * SentenceType), with the resulting model plotted using ggplot and shown in Figure 2, and means given in Table 2. Results upheld the two main hypotheses involving depiction. First, depiction was rated highest in the case of the positive episodic (“particular”) sentence, in that 66 KATHRYN DAVIDSON Figure 2: Acceptability ratings of vowel lengthening depictions across five linguistic contexts, based on 52 participants recruited from Prolific Academic platform Sentence type Particular Always No Exist Never Ungrammatical Huge 86.36 62.6 75.8 64.58 7.7 Huuuge 80.8 58.98 49.64 38.84 6.26 Table 2: Means for Experiment 1 case not significantly different from the sentence without depiction. In the ungrammatical sentence frame, both were rated unnatural, and did not differ from each other. However, depiction diverged from non-depiction sentences in the negative sentence frames, where sentences without depiction were acceptable in the context of negation but sentences with depiction were rated significantly less natural. Note in particular that the positive universal quantifier and negative universal quantifier have exactly the same sentence structure and level of complexity, yet there was a significant interaction between negation and depiction (p < 0.01) such that negation decreased the acceptability ratings of (only) sentences involving depiction. 3.3 Depiction via demonstration The second mini-study also involves depictive iconicity using written language. In this experiment, participants were presented with a depictive contrast in quotation versus an embedded clauses; the first involves depictive iconicity via a demonstration (showing a particular saying/attitude event) while the other involves telling about the same kind of event. The structure followed the study in section 3.2: ten sentences were based on five sentence frames, again involving an episodic particular, a universal adverbial quantifier (always), a negative adverbial quantifier (never), a non-quantificational negative sentence, and an ungrammatical sentence. Both the depictive and non-depictive versions report the speaker’s request, but one does so via quotation, the other via an embedded clause. Semiotic distinctions in compositional semantics Sentence type Yesterday Always Didn’t Never Ungrammatical Demonstration 80.02 73.51 43.96 55.64 11.14 67 Embedded question 95.8 91.86 78.86 85.88 16.62 Table 3: Means for Experiment 2 (2) a. At the cafe yesterday, I {was like “could you open the windows?” /asked if they could open the windows.} (positive episodic/yesterday) b. At the cafe, I {am always like “could you open the windows?” / always ask if they can open the windows.} (universal quantifier: always) c. At the cafe, I {am never like “could you open the windows?” / never ask if they can open the windows.} (negative universal quantifier: never) d. At the cafe, I {didn’t go like “Can you open the windows?” /didn’t ask if they can open the windows.} (negative/didn’t) e. At the cafe, I them ask {like “could you open the windows?” /if they can open the windows.} (ungrammatical) Although the form of depiction was entirely different than the case of vowel lengthening, the predictions were the same: depictions should be most natural in positive sentences, especially episodic sentences, and diverge the most from non-depictive sentences under negation. Both depictive and non-depictive options should be unnatural in the ungrammatical sentence frame. The hypotheses were again mostly born out in the results, which were analyzed in parallel with the previous mini study. The same glm model (here, depiction being quote vs. embedding) plotted in ggplot is shown in Figure 3. The ungrammatical sentence frame was rated as highly unnatural both for depiction and non-depiction versions. Unlike vowel lengthening, there was an overall main effect of depiction, such that quotation/demonstration versions were generally rated as less natural than embedded clause versions. However, the key hypothesis, that there would be an interaction between depiction and negation, was upheld again: depiction diverged the most from non-depiction in the never and didn’t sentence frames. Focusing just on the adverbial quantifiers always and never, which have the same sentence structure (and thus provide the tightest control), we find a significant interaction between depiction and always vs. never (p < 0.05). Means are provided in Table 3. Full anonymized data for both studies are available on the project’s OSF website: https://osf.io/y4v9r/?view_only=6e2216967bb14c93bd0fd6f84553b6c2 Overall, it seems that iconic depiction shows a clear compositional sensitivity to negation. This has been reported for gestures (Ebert & Ebert 2016), but we show here that this is less about the modality, and instead about the depiction involved in these iconic forms. In the next section, we discuss why this might be so. 68 KATHRYN DAVIDSON 100 Acceptability 75 factor(Depiction) Demonstration 50 Embed 25 0 Always Didn't Never Ungrammatical Yesterday SentenceType Figure 3: Acceptability ratings of quotational depictions across five linguistic contexts, based on 50 participants recruited from Prolific Academic 4 Interim conclusions: The role of alternatives Although depictive iconicity is not typically discussed for written language (it is much more obvious in the case of depictive gestures or even spoken language vocal gestures, etc.), it does exist, as in the cases of vowel lengthening and quotation. Our interim conclusion from these two short quantitative studies is that even in written language, depictive iconicity is considered especially unnatural in negative contexts, and can be quite natural in other, positive (especially positive episodic) contexts. From a compositional standpoint, we might want to conclude that iconic depiction contributes differently to meaning in different semantic environments. Why might this be? One hypothesis about iconic content is that it requires existence. Kuhn (2020) argues that existence requirements on iconic content underly patterns in anaphora in sign languages. The difference between positive and negative versions of sentences in this task provide further support for this idea. In addition, we hypothesize that there may be a pragmatic component, since negation involves focus alternatives: depiction might be especially incompatible with the construction of pragmatic alternatives. Alternative semantics is type of semantic composition used to model sentence meaning in many different domains including questions, disjunction, negation, and focus. Consider, for example, that the formal semantic denotation of a question like Did Alex eat the pasta? is typically modeled as the set of possible answers to that question, e.g. {Alex ate the pasta, Alex didn’t eat the pasta}. The pragmatic/semantic contribution of focus is also modeled via alternatives: the focus semantic value of Alex ate the PASTA (with focus on pasta) is typically modeled as activating a set of alternatives, e.g. {Alex ate the pasta, Alex ate the bread, Alex ate the rice, ...}. Focus sensitive operators like only or negation interact with these alternatives, so that Alex only ate the PASTA negates everything in the set of alternatives that isn’t pasta (or isn’t entailed by eating pasta), and negation with focus as in Alex didn’t eat the PASTA seems to do the opposite, negating the alternative involving pasta and suggesting that one of the other alternatives must be true. Although merely a sketch of the way that alternative semantics is used in Semiotic distinctions in compositional semantics 69 compositional semantics, this should give a sense that many functional pieces of the grammar, like disjunction, questions, and negation, rely on composition with these kind of semantic/pragmatic alternatives (Rooth 2016). Why might depiction be incompatible with alternatives? Lupyan & Winter (2018) note that symbols generalize over details in a way that more (depictively) iconic language does not. For example, a picture of a guitar necessarily shows what kind of guitar something is: it has to commit to whether it is an electric or acoustic guitar. In contrast, a symbol like the English word guitar or the ASL sign ‘guitar’ has potential to generalize over that level of detail. Such is the nature of a symbol, versus a depiction. This makes symbols ideal for creating alternatives: we can imagine the alternatives for Alex didn’t play the GUITAR to be other kinds of instruments, e.g. {Alex played the piano, Alex played the french horn, Alex played the guitar, ...}. The key idea behind why depictions are incompatible with negation, questions, and other places we see alternatives is that iconic depictions are generally too specific to create convenient partitions for alternatives. Depiction being generally incompatible with building pragmatic alternatives is consistent with quite a lot of otherwise puzzling existing data in the literature. One place we see supporting data is in the study of co-speech gestures: Ebert & Ebert (2016) focus on the way that depictive gestures seem to be unable to be targeted by negation. They attribute this to the gestural content being essentially a supplement, “not at issue”, similar to a non-restrictive relative clause. Schlenker (2021) agrees with the not-at-issue status of gestures, and takes them to be a type of presupposition (“co-supposition”). Esipova (2019) also takes gesture to frequently be not-at-issue. Thus, there is a robust discusison of the difficulty of gesture to occur in the scope of negation. There are, however, two important distinctions between that work and this work that set it apart. First, that line of work generally makes the assumption that gestures are natural/acceptable in the contexts of negation, and investigates their meaning given that assumption. As we have seen, however, depictions can be rated quite unnatural in negative contexts, so we should first be verifying that they are well-formed before investigating what they mean. Second, previous work typically does not separate depictive gestures from more descriptive/lexically iconic gestures, since the gestural modality is taken to be the relevant feature, instead of the semiotic distinction (depiction vs. description). Given that the previous two sections showed lowered naturalness ratings for iconic depictions even in written language, we certainly want to keep in mind the possible variations in naturalness. Second, it seems that depiction matters, not necessarily modality, given that both of these studies were in written language, not visual/manual gesture. Another kind of data that supports the incompatiblity of depictions and pragmatic alternatives comes from the analysis of visual pictures. Esipova (2021) cleverly analyses the compositional interaction between pictures and negative modals by investigating a small corpus of prohibition signs. These generally involve a symbol for negation (e.g. a slash through a picture) and a depicted activity. Although the intended interpretation is typically clear from context, Esipova (2021) notes that a picture with such a prohibition typically has several potential interpretations, and concludes that pictures are information structurally ambiguous. What we might 70 KATHRYN DAVIDSON Figure 4: Proportion choosing the contrastive pair response, based on 79 participants (no exlusions) recruited from Amazon MT platform add to this observation is that whatever negation does target, that aspect of the picture ends up having to be interpreted as a symbol that generalizes over particulars (e.g. particular kind of activities). It’s unclear why this would be, except if depictions are incompatible with building pragmatic alternatives, and so any visual representation (e.g. a picture) that would naturally be interpreted as a depiction must be re-interpreted as a symbol in order to be targeted by negation, with this re-interpretation open to ambiguities. Finally, we see evidence in prior literature for incompatibility between depiction and alternatives by studying contrastive inferences involving gesture. A well known fact at the intersection of psycholinguistics and semantics is that the use of additional modifications like asking someone to pick up the tall glass, versus just asking them to pick up the glass, leads to contrastive inferences that the modifier is necessary, i.e. there are multiple glasses, of which one is tall (Sedivy 2007). This is typically shown through real time processing measures like eye-tracking, where participants who hear Point to the tall... will show more eye movements toward a medium sized glass than to an even taller pitcher if there is only one pitcher but two glasses, one medium sized and one small. In other words, the use of the modifier tall cues the listener to guess that whatever noun follows must not be unique (glass) since if there was only one item (e.g. pitcher) then the adjective tall wouldn’t be necessary. Alsop et al. (2018) tested contrastive inferences using a simple off-line forced choice behavioral paradigm, and showed that for spoken modifiers, participants showed robust contrastive inferences. In other words, they regularly assumed that the use of an adjective like short indicated that the intended referent was one of a pair of items (e.g. a short and tall glass) instead of a unique item in a pair (e.g. a short glass and a watering can, see Figure 4). Intruigingly, they did not find the same effect for depictive co-speech gestures: use of a modifying gesture had a small effect in the absence of any other modifier, yet overall led to no contrastive inferences (i.e. there was no preference for the unique item, and participants were ignoring the gesture for contrastive reasoning). Why might the depictive gestures not lead to contrastive inferences when the same information (e.g. short) in the spoken language did? Here again we might Semiotic distinctions in compositional semantics 71 attribute this to composition in alternative semantics: the gestural depiction may tell about the particular glass but it will not be taken into account in the composition of semantic alternatives. In general, it seems that depictive co-speech gestures in English provide quite good evidence for a lack of depiction in the composition of alternatives. It is consistent with the data from the previous section that focused on written English, and from the study of visual representatins/pictures. What about other languages? The next section, Section 5, describes qualitative judgements on the naturalness of negation and depictions in spoken Japanese, American Sign Language, and English that seem to provide further cross-linguistic evidence, before Section 6 concludes. 5 Cross-linguistic and cross-modality comparison As Dingemanse (2019) notes, depiction is perhaps easier to identify in other languages with rich arrays of depictive language. One class of languages known for rich depictive vocabulary are spoken languages with rich sets of ideophones, such as Japanese. Another class of languages known for rich depictive vocabulary are sign languages, such as American Sign Language. We will investigate each in turn here, before turning to an iconic corner of the English language. 5.1 Japanese ideophones An especially rich discussion of the semantic contribution of Japanese ideophones, relevant to the question of negation and depiction, comes from Kita (1997). Since we will not have space to cover the details here, the reader is referred to the paper for details between different classes of ideophones, for example, that seem to vary with their syntactic category and amount of depiction. What we focus on here are a category of ideophones which are used in spoken language to support depictions. They are somewhat sound-symbolic in that the form bear a relation to their meaning: in this sense they involve lexical/descriptive iconicity. However, they can also be used depictively: Kita (1997) shows that these ideophones are far more likely to accompany depictive co-speech gestures than non-ideophones, and they are especially natural when used to both tell and show how something feels, moves, sounds, etc. An especially relevant observation for our discussion here is that Kita (1997) shows that ideophones are resistant to use under negation. He contrasts the well formed positive sentence involving an ideophone (3) with a negative version of the same sentence (4) which is no longer acceptable. He further shows that a different (non-iconic, non-ideophone) modifier like sizukani ‘quietly’ is perfectly well formed under negation (5), illustrating that this is an incompatibility specifically between negation and ideophones. (3) Depiction, no negation tama ga gorogoro to korogat-ta no o mi-ta ball N OM Mimetic roll-Past Nominalizer ACC see-Past ‘(One) saw a ball rolled gorogoro.’ (gorogoro = movement of a heavy round object with continuous rotation) 72 KATHRYN DAVIDSON (4) Depiction, with negation (not acceptable) *tama ga gorogoro to korogat-ta no de wa na-i ball N OM Mimetic roll-Past Nominalizer C OP Focus Neg Pres ‘It was not the case that a ball rolled gorogoro.’ (5) Descriptive modifier, with negation tama ga sizukani korogat-ta no de wa na-i ball N OM quietly roll-Past Nominalizer C OP Focus Neg Pres ‘It was not the case that a ball rolled quietly.’ How might we view the Japanese pattern in light of the hypothesis that depictive material is incompatible with pragmatic alternatives? The idea is that a nonideophone such as sizukani ‘quietly’ is an abstract symbol that generalizes over all kinds of ways that balls can roll quietly, and as such contrasts with other abstract symbols for manners of rolling, such as ‘loudly’, etc. In contrast, the ideophone gorogoro is used precisely in contexts in which the speaker wants to not only describe but also depict for their interlocutor some aspect of the ball rolling, so the speaker can’t help but depict while saying the ideophone, and as such is incompatible with building alternatives. Note that this analysis assumes that a (symbolic) ideophone will generally be used in cases when one wants to (also) depict, and thus the incompatibility in the Japanese case is somewhat roundabout/indirect. In other words, there isn’t strictly a prohibition on using the ideophone in a strictly symbolic/lexical sense, only when it is used to help depict; it just happens that ideophones are best used for purposes of depiction. This might actually turn out to be a strength of the analysis, as we will see when we extend the discussion to English onomatopeia in the final section. First, though, we turn to a case of more obligatory depiction: classifiers in American Sign Language. 5.2 ASL depictive classifiers Sign languages around the world make use of a class of signs that involve conventionalized handshapes used to depict; in the formal linguistics literature the symbolic nature of these handshapes is emphasized and they are often referred to as classifier predicates (Zwitserlood 2012), since handshapes are chosen based on categories similar to the verbal classifier literature, e.g. humans, animals, vehicles, flat objects, etc. In cognitive linguistics literature, another key aspect of these signs is highlighted: the fact that their locations and movements are depictive, and thus they are known as depicting signs/depicting verbs (Dudis 2004). Like Japanese ideophones, they involve a symbolic base (the basic ideophone form in Japanese, the classifier handshape in ASL), which is most natural used to depict. In the case of ASL, the depiction is arguably even more necessary than in Japanese, as the entire purpose of this class of expressions is to depict objects, shapes, events, etc. Given the fundamentally depictive nature of depictive classifier predicates (to use an especially broad terminology), and our working hypothesis in this paper that depictions are incompatible with building alternatives, we might expect to see Semiotic distinctions in compositional semantics 73 a similar pattern to the Japanese ideophone for ASL classifiers. And indeed, the pattern is surprisingly similar, despite the different language and language modality. In (6), the depicting sign is used as a predicate to depict the way that the main character pulled a book off the shelf. This sign uses the “c” handshape, conventionalized for handling objects of the size and shape of a large book. The movement shows/depicts/demonstrates the way that the book was pulled off the shelf. In this positive sentence, the depictive classifier is considered natural/acceptable by signers of ASL. However, the same sign in (7) ‘pull down the book’ is no longer acceptable under the scope of negation; instead, to express the idea that it was not difficult (i.e. to target the pulling-down manner with negation), an entirely symbolic sign HARD has to be used, as in (8); to maintain a depicting classifier it has to be used outside the scope of negation (as in DC_c(hold book)), and cannot show what did not happen, but only depict what did happen. (6) Depiction, no negation ‘Of all the books in a row, it was difficult to pull one down’ (7) Depiction, with negation (not acceptable) *BOOK DSc (books lined up), NOT DS_c(pull down w/difficulty) ‘Of all the books in a row, it wasn’t difficult to pull one down’ (8) Descriptive modifier, with negation 74 KATHRYN DAVIDSON ‘Of all the books in a row, it wasn’t difficult to pull one down’ Our very preliminary finding for the interaction of depiction and negation in ASL, then, involving classifiers, is quite similar to what we find in Japanese involving ideophones. This is not to say that this exhausts depiction in either language, of course! Much research in sign linguistics focuses on the ways that depiction interacts with other aspects of the language; see especially relevant work by Ferrara & Hodge (2018) and Hodge & Ferrara (2022), who provide a detailed discussion of different semiotic aspects of language, with an especially informed view of sign languages, that are consistent with and an inspiration for the one I am adopting here. What I hope to emphasize as new is the way that we can test for and probe these different classes by means of the way that they interact with other parts of the grammar. For example, depictions in ASL seem to be resistant to contexts that involve negation, and also, it seems, less natural in questions. This might be precisely why we also need descriptions/symbolic language: because communication is not all about showing what happened but also about asking questions, answering in the negative, etc. 5.3 English onomatopeia In this section so far we have shown that robust depictive language, such as Japanese ideophones or ASL depictive classifiers, are notably less acceptable in the context of negation than in positive episodic contexts, and less acceptable than symbolic language in negative contexts. It’s worth noting before we leave this section that the same observations can be many in any corner of a language where we contrast depictive iconicity with symbolic/descriptive iconicity. Consider the case of English onomatopeia: in (9), a descriptively iconic symbol like the word chirp can quite easily support a further depiction, as in chirrrp-chirrping[expressed in a singsongy manner]. This is similar to the case of the Japanese ideophone being used to depict, or the ASL classifier handshape being used in a depiction. In the case of English onomatopeia, we also see that the sentence is much less natural/acceptable under negation (10, whereas a similar example with a non-depictive modifier under negation is fine (11). Note, moreover, that we find the same contexts with a polar question (12)-(13). Semiotic distinctions in compositional semantics 75 (9) Depiction, no negation The bird was chirrrp-chirrping[expressed in a sing-songy manner] on her perch. (10) Depiction, with negation (not acceptable) *The bird wasn’t chirrrp-chirrping[expressed in a sing-songy manner] on her perch. (11) Descriptive modifier, with negation The bird wasn’t chirping loudly on her perch. (12) Depiction, in question (not acceptable) *Was the bird chirrrp-chirrping[expressed in a sing-songy manner] on her perch? (13) Descriptive modifier, in question Was the bird chirping loudly on her perch? Thus we see again that the incompatibility of negatoin and depiction isn’t about iconicity in a broad sense (chirp is often considered iconic, from the perspective of non-arbitariness in the lexicon), or of gestures, but rather it is a property of iconic depiction in a narrow sense, including in spoken/written language. The extension of this data to questions further supports the hypothesis in Section 4 that this is about alternatives, and not, say, only about negation. 6 Discussion We began with the question of how iconicity in language relates to compositional semantics: how does the fact that a form is iconic affect the result of its composition with another form? Does this depend on what kind of composition is involved, or what other forms it combines with? We first proposed to distinguish iconicity in the lexicon from depictive iconicity, which we argued are too frequently conflated in literature on this topic. We then narrowed our question to how depictive iconicity composes with other aspects of the language, and presented a serious of short quantitative studies that showed that depictions in written English are rated by English speaking participants as less acceptable in contexts involving negation. The unacceptability of depiction in the context of negation was proposed to be due to the inability of depictions to build alternatives. This was further supported by existing data in co-speech gesture and picture semantics that find depictions behaving unlike symbolic language under negation and other alternative-invoking environments, such as contrastive inferences. Finally, we showed that these observations have cross-linguistic validity in rich forms of depiction in other languages such as ideophones in spoken Japanese and depictive classifiers in ASL, as well as onomatopeia in spoken English. There are three main takeaways from this short paper. First, the semiotic approach championed by many outside of formal semantics, such as Clark (2016), Kita (1997), Dingemanse et al. (2015), and Ferrara & Hodge (2018) seems to provide an important insight for formal compositional semantics, since we find this resistance to negation not necessarily in iconicity originating in the lexicon, but rather in depictively iconic language. In addition, we see possible positive influence 76 KATHRYN DAVIDSON in the other direction: alternative semantic frameworks have been extremely productive in modeling semantic composition in areas as varied as focus (Rooth 1992), questions (Hamblin 1976), disjunction (Alonso-Ovalle 2006), negation (Ramchand 1997), and broadly the organization of our entire discourse (Roberts 2012). When we look beyond positive episodic sentences that describe a particular event (as is commonly done in storytelling and narratives, and language elicited from picture prompts, etc.), we find that much of conversational language involves reasoning over pragmatic alternatives. Here we find different functions for different semiotic categories: depictions are excellent for showing the small details of particulars, but description is necessary for making generalizations of the sort involved in building alternatives for questions and their (focused) answers. The final takeaway more narrowly concerns the formal semantic analysis of iconic content like co-speech gestures, classifier predicates, and ideophones. The gesture literature has frequently note that gestures are largely incompatible with negation. Here, we extend this observation much more broadly to depictive content, irrespective of modality. Sometimes “gesture” is used to encompass everything here that we are calling iconic depictions (Goldin-Meadow & Brentari 2017), and so under that interpretation indeed, gestures seem to be unable to be targeted by negation. But we argue that this is due to their nature as depictive, not their modality. This connects with quite a lot of other literature on the not-at-issueness of gestural/depictive content more broadly, hopefully combining a large class of literature not by their language modality, but rather their semotic status. The ideas discussed here naturally extent to depictive content wherever it might be found. This paper covered written quotations, lengthened vowels (in writing), ideophones and onomatopeia in spoken and written language, and sign language depictive classifiers. There are as many ways to depict as there are to describe, and many of them appear together with symbolic language, so hopefully this will spur more work on the interaction of the formal semantics of functional/compositional language and iconic depictions. Acknowledgements Thanks to Mark Dingemanse, Masha Esipova, and the audience at the Chicago Linguistic Society 58 in Spring 2022, HSP (especially Sotoro Kita), and the joint meeting of the SPP/ESPP in Summer 2022, and members of the M&M Lab for comments on various stages of this project, and Jessica Tanner for recording ASL depictive classifier sentences. This work was supported by an NSF CAREER grant BCS-1844186. Semiotic distinctions in compositional semantics 77 References Abusch, D. 2012. Applying discourse semantics and pragmatics to co-reference in picture sequences. Abusch, D., & M. Rooth. 2022. Pictorial free perception. Alonso-Ovalle, L. 2006. Disjunction in alternative semantics. University of Massachusetts Amherst. Alsop, A., E. Stranahan, & K. Davidson. 2018. Testing contrastive inferences from suprasegmental features using offline measures. Proceedings of the Linguistic Society of America 3. 71–1. Aristodemo, V., & C. Geraci. 2018. Visible degrees in Italian sign language. Natural Language & Linguistic Theory 36. 685–699. Blasi, D. E., S. Wichmann, H. Hammarström, P. F. Stadler, & M. H. Christiansen. 2016. Sound– meaning association biases evidenced across thousands of languages. Proceedings of the National Academy of Sciences 113. 10818–10823. Burge, T. 2018. Iconic representation: Maps, pictures, and perception. In The map and the territory, 79–100. Springer. Clark, H. H. 2016. Depicting as a method of communication. Psychological review 123. 324. Clark, H. H., & R. J. Gerrig. 1990. Quotations as demonstrations. Language 764–805. Cohn, N. 2012. Comics, linguistics, and visual language: The past and future of a field. In Linguistics and the Study of Comics, 92–118. Springer. Davidson, K. 2015. Quotation, demonstration, and iconicity. Linguistics and Philosophy 38. 477–520. Davidson, K. 2020. Is “experimental” a gradable predicate. Proceedings of North East Linguistics Society (NELS) 50. Dingemanse, M. 2015. Ideophones and reduplication: Depiction, description, and the interpretation of repeated talk in discourse. Studies in Language. International Journal sponsored by the Foundation “Foundations of Language” 39. 946–970. Dingemanse, M. 2019. “Ideophone” as a comparative concept. Ideophones, mimetics, expressives 13–33. Dingemanse, M., & K. Akita. 2017. An inverse relation between expressiveness and grammatical integration: On the morphosyntactic typology of ideophones, with special reference to Japanese 1. Journal of Linguistics 53. 501–532. Dingemanse, M., D. E. Blasi, G. Lupyan, M. H. Christiansen, & P. Monaghan. 2015. Arbitrariness, iconicity, and systematicity in language. Trends in cognitive sciences 19. 603–615. Dudis, P. G. 2004. Depiction of events in ASL: Conceptual integration of temporal components. University of California, Berkeley. Ebert, C., & C. Ebert. 2016. The semantic behavior of co-speech gestures and their role in demonstrative reference. presentation, LSCP LANGUAGE seminar talk at the Institut JeanNicod, Département d’études Cognitives, école normale supérieure, Paris, Dec. 2016. Emmorey, K. 2014. Iconicity as structure mapping. Philosophical Transactions of the Royal Society B: Biological Sciences 369. 20130301. Esipova, M. 2019. Composition and projection in speech and gesture.. New York, NY: New York University dissertation. Esipova, M. 2021. On not-at-issueness in pictures. Glossa: a journal of general linguistics 6. 83. Ferrara, L., & G. Hodge. 2018. Language as description, indication, and depiction. Frontiers in Psychology 9. 716. Fodor, J. 2007. The revenge of the given. Contemporary debates in philosophy of mind 105–116. Goldin-Meadow, S., & D. Brentari. 2017. Gesture, sign, and language: The coming of age of sign language and gesture studies. Behavioral and Brain Sciences 40. Greenberg, G. 2021. Semantics of pictorial space. Review of Philosophy and Psychology 12. 847–887. Hamblin, C. L. 1976. Questions in montague english. In Montague grammar, 247–259. Elsevier. Hochgesang, J., O. Crasborn, & D. Lillo-Martin. 2020. ASL Signbank. https://ASLsignbank.haskins.yale.edu/. 78 KATHRYN DAVIDSON Hodge, G., & L. Ferrara. 2022. Iconicity as multimodal, polysemiotic, and plurifunctional. Frontiers in Psychology 13. Kita, S. 1997. Two-dimensional semantic analysis of japanese mimetics. Linguistics 379–415. Kuhn, J. 2020. Logical meaning in space: Iconic biases on quantification in sign languages. Language 96. e320–e343. Kuhn, J., & V. Aristodemo. 2017. Pluractionality, iconicity, and scope in French Sign Language. Semantics and Pragmatics 10. Lupyan, G., & B. Winter. 2018. Language is more abstract than you think, or, why aren’t languages more iconic? Philosophical Transactions of the Royal Society B: Biological Sciences 373. 20170137. Marty, P., E. Chemla, & J. Sprouse. 2020. The effect of three basic task features on the sensitivity of acceptability judgment tasks. Glossa: a journal of general linguistics (2016-2021) 5. 72. Meir, I. 2010. Iconicity and metaphor: Constraints on metaphorical extension of iconic forms. Language 86. 865–896. Ramchand, G. C. 1997. Questions, polarity and alternative semantics. In North East Linguistics Society, volume 27, p. 28. Roberts, C. 2012. Information structure: Towards an integrated formal theory of pragmatics. Semantics and pragmatics 5. 6–1. Rooth, M. 1992. A theory of focus interpretation. Natural language semantics 1. 75–116. Rooth, M. 2016. Alternative semantics. The Oxford handbook of information structure 19–40. Saussure, F. d. 1916. Course in general linguistics. Columbia University Press, 2011. Schlenker, P. 2018a. Gesture projection and cosuppositions. Linguistics and Philosophy 41. 295–365. Schlenker, P. 2018b. What is super semantics? Philosophical Perspectives 32. 365–453. Schlenker, P. 2021. Iconic presuppositions. Natural Language & Linguistic Theory 39. 215–289. Schlenker, P., J. Lamberton, & M. Santoro. 2013. Iconic variables. Linguistics and philosophy 36. 91–149. Sedivy, J. C. 2007. Implicature during real time conversation: A view from language processing research. Philosophy compass 2. 475–496. Strickland, B., C. Geraci, E. Chemla, P. Schlenker, M. Kelepir, & R. Pfau. 2015. Event representations constrain the structure of language: Sign language as a window into universally accessible linguistic biases. Proceedings of the National Academy of Sciences 112. 5968– 5973. Taub, S. F. 2001. Language from the body: Iconicity and metaphor in American Sign Language. Cambridge University Press. Tieu, L., R. Pasternak, P. Schlenker, & E. Chemla. 2017. Co-speech gesture projection: Evidence from truth-value judgment and picture selection tasks. Glossa: a journal of general linguistics 2. Tieu, L., P. Schlenker, & E. Chemla. 2019. Linguistic inferences without words. Proceedings of the National Academy of Sciences 116. 9796–9801. Winter, B., & M. Perlman. 2021. Iconicity ratings really do measure iconicity, and they open a new window onto the nature of language. Linguistics Vanguard 7. Zwitserlood, I. 2012. Classifiers. In Sign language: An international handbook. De Gruyter. Proceedings of CLS 58 (2022), 79-89 © Chicago Linguistic Society 2023. All rights reserved. 79 A pragmatic characterization of po in Camuno Matteo Fiorini University of Utah 1 Introduction The paper presents an overview of the basic pragmatic properties of the discourse particle po in Camuno, an endangered Gallo-Italic variety spoken in Vallecamonica, in the north of Italy. Cognate forms of po (po; pa; poi) are attested in numerous Northern Italian Dialects. See, for instance, Munaro and Poletto (2005), for Piedmont varieties; Hack (2014), and Dohi (2019) for Dolomitic varieties; Hinterhölzl and Munaro (2015) for Bellunese; Coniglio (2008), Cognola and Cruschina (2021), and Fiorini (2022) for Italian. All these forms derive from the Latin adverb post ‘afterward,’ which undergoes a process of grammaticalization from purely temporal adverb to functional particle (Hack, 2014). These elements are mainly attested in questions and introduce a number of pragmatic values like: (i) astonishment (Croatto, 1986); (ii) emphasis and focus (Poletto and Zanuttini, 2003); (iii) surprise and indignation (Hack, 2014); (iv) "inability to find an answer" or "concern or interest in the information being asked for" (Coniglio, 2008: 111). In the appropriate context, all these values can be expressed in Camuno, but, differently from most of the previous literature, I argue that they all stem from the properties of po. In particular, I claim that the particle adds a non-truth conditional (i.e., presuppositional) semantic import to a proposition p, conveying the speaker’s doxastic evaluation of p and the common ground. An analysis of Italian poi in presuppositional terms has been put forward by Cognola and Cruschina (2021). In their discussion, they focus on two key concepts: (i) Counter-expectationality; and (ii) Mirativity1 . Counter-expectuality is strictly related to mirativity (I use the term counter-factuality) and signals a mismatch between the expectation of at least one participant in the conversation and the actual state of affairs. The discussion of the semantic properties of poi in terms of expressive presupposition in Fiorini (2022) shows that pragmatics plays a bigger role in the interpretations above than suggested by C & C. This paper explores this intuition employing data from Camuno to provide a pragmatic analysis that is predicted to hold for at least some of the varieties where cognate forms of po are attested. In a nutshell, I propose that the central import of po to a proposition is uncontroversiality p: the speaker believes that p should be uncontroversial, but it is not in the utterance world. This characterization provides a common source for the numerous readings that, consequently, do not require any additional language-specific analysis. 1 Mirativity has been first identified as the grammatical category which marks the imprepareness of the speaker’s mind over a certain proposition (De Lancey, 1997). 80 MATTEO FIORINI As for the organization of the paper, section 2 introduces some basic background information regarding discourse particles and their analysis. Section 3 presents the discussion of the various uses of po and their properties. Finally, section 4 concludes the article by summarizing the main advantages of the line of analysis presented here and introducing some outstanding questions to be addressed by future works. 2 Preliminaries The analysis presented here requires to be framed appropriately by introducing some brief preliminary remarks regarding the main concepts referred to throughout the paper. 2.1 Discourse particles The discussion of the properties of discourse particles [henceforth: DiPs] raises various theoretical issues that cannot be exhaustively addressed here. What follows is a brief overview of the most crucial ones in order to situate the topic of this paper within the ongoing debate on the matter. First of all, I follow Grosz (2020) in employing the term "discourse particle" rather than the equally common "modal particles" to avoid confusion regarding different definitions of "modality" (see Thurmair, 1989; and Abraham, 2016). Secondly, I consider po to be a DiP, rather than a discourse marker, following the characterization in Diewald (2006), and Degand, Cornillie & Pietrandrea (2013) based on a distinction in scope properties. DiPs take scope over propositions or speech acts, while discourse markers over non-propositional discourse elements. Consider the following examples from Italian (1): (1) a. b. Poi magari non lo vuole nemmeno poi maybe NEG CL .3 SG want.3 SG . PRS even ‘Then maybe (s)he doesn’t even want it.’ Non lo vuole nemmeno poi NEG CL .3 SG want.3 SG . PRS even poi ‘(S)he doesn’t want it poi (obviously!)’. In (1-a), poi ‘then’ functions as a discourse marker since its role is to connect two parts of a larger discourse. In particular, its role is to signal that the topic described by p had already been discussed in a previous conversation by the participants to the present one. In (1-b), poi comments directly on the propositional content of p. From the theoretical point of view, we can identify three main approaches to the investigation of the way DiPs contribute to the overall meaning of a proposition: (i) syntactic-based, (ii) presuppositional, and (iii) use-conditional accounts. Since at least Jacobs (1991) and Abraham (1991), several accounts have proposed that DiPs interact with an illocutionary operator in the left periphery of the clause (Rizzi, 1997; Coniglio, 2011). The role of DiPs is to modify these operators by adding specific values to the proposition, i.e., modifying the illocutionary force of the utterance. Po in Camuno 81 Presuppositional accounts claim that DiPs act as truth-conditionally vacuous triggers of presupposition (Egg & Zimmermann, 2012; Kaufmann, 2012). Grosz (2020) argues for an analysis in terms of "expressive presupposition", i.e., a presupposition containing three indexical elements, which deictically refer to the speaker, the addressee, and the situation in the utterance context (see Sauerland, 2007; Schlenker, 2007; and Fiorini, 2022 for a similar treatment for Italian poi). One last set of works argues that the import of DiPs to the overall meaning of the proposition pertains to a separate dimension, labeled “expressive” (Potts, 2005; 2007) or “use-conditional” (Kratzer, 1999, in Grosz 2020; Gutzmann, 2015). Gutzmann (2015) introduces a new semantic type u which adds non-truth conditional meaning to a proposition. For the analysis of po, I follow Grosz’s (2020) proposal for a treatment of DiPs as triggers of expressive presuppositions. The use of the particle, in fact, is strictly connected to the beliefs of the speaker regarding both context and addressee. The presupposition is thus self-sufficient and verified by the speaker’s beliefs (and expectations) alone. 2.2 (Un)controversiality of a proposition As mentioned in the previous section, the non-truth conditional (i.e., presuppositional) semantic import of po conveys the speaker’s doxastic evaluation of the proposition. The presupposition that must hold for the interpretation to be felicitous stems from the two interrelated concepts of factuality and uncontroversiality, informally defined as in (2-a) and (2-b): (2) a. Factuality: a proposition p is factual iff p is true and “verifiable on the spot” (Gutzmann, 2015). b. Uncontroversiality: a proposition p is uncontroversial iff factual, and it is not challenged in the utterance context w, i.e., ¬p is not under consideration by any of the participants of a conversational event in w (Lidner, 1991; Zimmermann, 2011; Grosz, 2021)2 The presupposition licensing po is directly connected to the beliefs of the speakers regarding a proposition p. In particular, the speaker must believe that p should be controversial, regardless of the actual state of affairs. For this reason, po is thought to operate on the Common Ground Management level, which in Krika’s (2008: 246) definition, "contain[s] information about the manifest communicative interests and goals of the participants". This dimension of the common ground has been initially proposed to handle the interaction between questions, which are truth-conditionally vacuous objects, and common ground content. It can be thus considered to be operating over the doxastic, in addition to the epistemic, evaluation of the common ground by the speaker. By definition, the common ground management can be asymmetric, i.e., po refers to the way the common ground should develop according to at least one of the participants in a conversation. In this case, the assumption is for p to be uncontroversial. 2 In this definition, and thorough the paper, I follow Grosz (2010; 2021), in using the notation ¬p to indicate any alternative proposition to p, rather than the opposite of p. 82 MATTEO FIORINI 3 The import of po to the meaning of proposition From a pragmatic point of view, po introduces a counterfactual and mirative value. This reading is attested in all sentence types, however, the details of its import can vary. Both in declarative and in interrogative structures, the speaker believes that p is not (but should be) uncontroversial. However, in the former, this is because the addressee is believed to be considering ¬p, while in the latter, because p is believed to be non-factual. In both cases, the use of po identifies a unique appropriate proposition true in w. In declaratives, this makes p, in the speaker’s mind, verifiable on the spot. In interrogatives, p is false, but according to the speaker, so it is ¬p. The details of these patterns are discussed below. 3.1 Po in declarative structures The particle po signals that the speaker initially believes p, and that p is believed to be true by the addressee(s) as well, i.e., that p is factual, and ¬p is not under consideration. At the utterance time, p is no longer believed to be uncontroversial. Consider (3): (3) La ho hcèta la laura {po} ho à Brè {po} the her daughter CL .3 SG work.3 SG . PRS po up in Breno po ‘Her daughter works in Breno [and you should know that!].’ (3) is only felicitous if the speaker believes that the addressee is considering ¬p, i.e., that her daughter may work somewhere other than Breno. The speaker, in fact, believes that the addressee should know that p is factual and, consequently, uncontroversial in w. In certain contexts, po can introduce additional meaning by establishing a relationship between propositions. This affects the doxastic evaluation of a proposition r, which in this case involves a relation between propositions. Consider (4): (4) a. b. Foda la maja tutch i de le la ho htchèta? why CL.3SG eat.3SG.PRS every the days there the her daughter ‘Why does her daughter eat there every day?’ (Parké) la laura ho à Brè po because cl.3SG work.3SG up in Breno po ‘Because she works in Breno po [it’s the only explanation!].’ In (4), the speaker believes that p (she works in Breno) is factual, exactly like the one in (3). However, it is not p (or at least not only p) to be uncontroversial, but the inference r between p (she works in Breno and q (she eats in Breno everyday). According to the speaker, r = (q→p) is the only logical answer to the addressee’s question, so that any other alternative to p is believed to be false. In other words, knowing q, the addressee should be able to infer p. The two propositions forming the inference must be kept separated since, for instance, in (4), there is no evidence to suggest that the speaker assumes that the addressee is considering ¬p. For the way the context is set up, (4-b) merely suggests that the addressee is considering ¬(q→p)/¬r – e.g., there are many other reasons Po in Camuno 83 for her to be eating there. This is clearly visible in (6). The speaker’s question in (5) suggests that they do not consider p to be uncontroversial. Rather, (6-b) signals that it is q that should be uncontroversial: (5) a. b. (6) a. b. Foda la maja tutch i de le la ho htchèta? why CL .3 SG eat.3 SG . PRS every the days there the her daughter ‘Why does her daughter eat there every day?’ Te he ke la ho htchèta la laura ho à Brè? you know that the her daughter CL .3 SG work.3 SG . PRS up in Breno ‘Do you know that her daughter works in Breno?’ He al ho yes CL .3 SG know.1 SG . PRS ‘Yes, I Know.’ Pota, l’ è per kel ke la maja fo fo pota CL .3 SG be.3 SG . PRS for that that CL .3 SG eat.3 SG . PRS out out le po there po ‘Well, that’s why she eats there po [why else?!].’ What (5) and (6) show is the property of a factual proposition of being self-evident in w. By establishing the appropriate premises, the speaker makes it possible for the addressee to infer p, even when not previously known. To summarize, the use of po in declarative structures signals that the speaker believes that a proposition p should be uncontroversial (U). However, this does not hold at the utterance time, resulting in the general mirative reading described in the literature for cognate forms of po. The typical conversational progression where po is licensed is articulated as follows (Bs = Speaker’s beliefs; Ba = Addressee’s beliefs), (7): (7) T-1. Bs(U(p)) T 0. Bs(Ba(¬p)) = PRESUPPOSITION FOR po Once again, such a presupposition is an expressive one, i.e., there is no need for the speaker to have any concrete reason to believe that (Ba(¬p)). 3.2 Po in interrogative structures Following the traditional Hamblian semantics (Hamblin, 1973), questions are modeled after their possible answer(s). A wh-phrase activates a set of alternatives and, in the answer, the correct one is selected. Speakers can introduce further presuppositional content by manipulating the set of alternatives (see, e.g., Gunlogson, 2001; Krifka, 2004). This is the type of function that po carries out, resulting in a biased reading for the hosting interrogative. The use of po signals that there is a single proposition p that should be true in the utterance context w. However, the speaker believes that p is false in w, i.e., it is not factual. 84 MATTEO FIORINI Take the wh-interrogative in (8): (8) a-la fat-ho kè po la Marina? has=CL .3 SG prepare what po the Marina ‘What did Marina prepare po?’ A question like (8) is only licensed in contexts where the speaker holds previous beliefs regarding what Carla prepared. The difference between p and PO-p lies in the expectations of the speaker regarding the answer. An apparent paradox arises with po-interrogative structures since the speaker knows that p is false, however, they also believe that only p could be true in the utterance context. From a pragmatic point of view, by uttering (8), the speaker is seeking additional information to re-interpret the context and validate their beliefs. Let us set up a more articulated context for (8) by making use of a complex proposition r (9): (9) a. b. C ONTEXT: Marina invited the speaker for a barbeque night. However, she is a strict vegetarian with no interest in meat substitutes. The speaker encounters the addressee in Marina’s apartment building’s lobby and ask (8) q = M. invited us to a barbeque p = M. prepared meat. B( S ): (q –> p) & (p = 0 & p) in w. In (9), the speaker knows that it cannot be the case that M. prepared meat. However, according to their beliefs, the only possible type of food to be consumed at a barbeque is meat, i.e., in the context (9-a), M. could only be preparing meat. An important difference between interrogative and declarative po-structures is that in the former, the beliefs of the addressee (in the speaker’s mind) are irrelevant. The conversational progression licensing the particle mimics, in fact, the one in (7), with the difference that the presupposition is satisfied by the non-factuality of p, rather than by the attitude of the speaker towards the beliefs of the addressee (10): (10) T-1. Bs(U(p)) T 0. Bs(¬p) = PRESUPPOSITION FOR po 3.3 Not-a-big-deal po A third use of po can be labeled "non-a-big-deal" reading since it adds a special type of contrastive value related to the (assumed) discrepancy between the speaker’s beliefs and the addressee’s. Consider the example in (11): (11) a. bahta! mandarò al me fiol al colegio. enough send.1 SG . FUT CL .3 SG my son to-the boarding-school ’Enough, I’ll send him to boarding school!’ Po in Camuno b. 85 l’ a po biit na bira CL .3 SG have.3 SG . PRS po drink. PRT a beer ‘He merely drank a beer [nothing too serious].’ None of the propositions in (11) is controversial, i.e., po-p (He drunk a beer) is both factual and known by the participants to the conversation. The use of (11-b), however evokes a proposition q that can be paraphrased as he deserves being sent to military school. What the speaker rejects is thus an inference that they assumed to be believed by the addressee Bs(Ba(p→q)). The speaker believes that the addressee considers the beer-drinking event described by p to warrant sending their son to boarding school. The speaker rejects this inference while introducing p to the discourse and marking it as uncontroversial since they believe that p is the only event that could motivate the addressee’s utterance. A number of pragmatic considerations can be made for (11) considering, for instance, the de facto establishment of a threshold that may differ between speaker and addressee, i.e., one of the events justifying q. The presence of such a threshold is fundamental for the interpretation of interrogative "not-a-big-deal" structures like (12) that we can represent as r = (x → p = a. will attend boarding school): (12) a. b. Foda al l’ a mandat al why CL .3 SG CL .3 SG have.3 SG . PRS send.PRT to=the po? colegio boarding=school po ‘Why did he send his son to boarding school?’ kè al ho fiol po? fat A-l have.3 SG . PRS do.PRT what the his son po ‘What did his son do?’ The use of po signals that not only the event resulting in q is not part of the speaker’s knowledge (the unmarked meaning of p), but also that the speaker is not aware of, nor can imagine, an event serious enough to justify q. The particle’s contribution to the meaning of the complex proposition r is again a mirative one, since the speaker also manifests their surprise over the fact that such an even exist3 . 3.4 Summary and shared properties The three context of use for po described above share a number of properties typical of discourse particles. First of all, given their doxastic nature, they all operate on the common ground management level since they add information to the utterance regarding the state of mind of the speaker with respect to the necessary or expected way the common ground should develop. A direct consequence of this property is the classification 3 Additional considerations that must be left to future study can be made regarding the state of mind of the speaker and their beliefs regarding the addressee’s moral values. For instance, the speaker may expect an answer adding material to the premise, i.e., A’s son indeed did something justifying q. It can also be the case, however, that the addressee is expected to justify r 86 MATTEO FIORINI of po in the domain of modality, i.e., it deals with possible worlds rather than the actual state of affairs. Secondly, they all entail a certain degree of mirativity and counter-factuality, in that by selecting a single proposition that can (or could) hold in the utterance context, the speaker expresses their surprise over its not uncontroversiality. Thirdly, depending on the yet-to-be-investigated assumptions regarding the addressee made by the speaker, they introduce a variable level of negotiation of the truth values of the proposition (Abraham, 2016). In other words, the speaker is more or less open to modifying their initial beliefs once the context is somehow changed. In the contexts employed here, this is obtained by adding information to the common ground like "Carla was transferred", "It is possible to cook vegetables and cheese on a barbeque", and "A’s son actually vandalized the principal’s car." Finally, their common role can be described, from a semantic point of view, as one of manipulation of sets of alternatives (Rooth, 1985, et seq.) activated by po4 4 Conclusions and outstanding questions The goal of this paper was to identify and discuss the basic, underlying properties of the discourse particle po. The data I presented suggest an analysis based on the idea of (non-)uncontroversiality of a proposition that can successfully accommodate the licensing of the particle both for declarative and interrogative structures. Consequently, its distribution and its numerous interpretations can be reduced to pragmatic implicatures without the need to propose additional properties for specific cases (cf. Cognola and Cruschina, 2021). The analysis also provides additional evidence for a treatment of discourse particles from a presuppositional perspective rather than a syntactic one. The acceptability of po in virtually all sentence types, in fact, makes the direct connection with the illocutionary force of the sentence assumed in these accounts unnecessary. Considering the common etymological origin and the comparable readings described in the literature, it is plausible to expect the analysis to be able to capture the properties of the cognate forms attested in other Northern Italian Dialects. A number of questions arise both from the analysis and the data introduced here. First of all, the "no-big-deal" reading described here requires further investigation to identify the possible underlying contrastive value that can unify the uses of po (as suggested for Italian in Fiorini, 2022). The distribution of po seems also to affect its interpretation. The particle can occupy a number of positions within TP, and have different prosodic properties. While I do not think that positing the presence of two allomorphic particles is plausible, it is possible to have more than one po, each with its own scope, in a proposition. The issue requires specific data and analysis that must be left to future works. Finally, the relation between the speaker, the addressee, and the context must be investigated further to identify possible additional implications for a conversation, including a po-statement. 4 This property is crucial for the investigation of the varieties described by Hack (2014) where the cognate form pa is used as a pure focus- or obligatory question- particle. Po in Camuno 87 To conclude, this brief paper offered new data and a different take on the topic of discourse particles. These elements have been widely investigated in works discussing Germanic languages and have been receiving increasing attention in the Romance literature. The paper contributes to this trend by adding empirical coverage and theoretical insights to this body of work. References Abraham, W. 2020. Modal Particles between Context, Conversation, and Convention. Cambridge: Cambridge University Press. Abraham, W., 2016. Discourse marker= discourse particle= thetical= modal particle? A futile comparison. In Discourse particles: Formal approaches to their syntax and semantics, ed. by Bayer, J. & V. Struckmeier, 241-280. Berlin: Walter de Gruyter. Bonan, C. 2019. "On clause-internally moved wh-phrases: wh-to-foc, nominative clitics, and the theory of Northern Italian wh-in situ." PhD thesis. Genève, CH: Université de Genève. Cognola, F. & Cruschina, S. 2021. Between time and discourse: A syntactic analysis of Italian poi. In Annali di Ca’ Foscari. Serie orientale, 55. 87-116. Coniglio, M. 2008. Modal particles in Italian. In University of Venice working papers in linguistics 18:1. Ed. by, L. Brugè. 91–130. Coniglio, M. 2011. Die Syntax der deutschen Modalpartikeln: ihre Distribution und Lizenzierung in Haupt- und Nebensatzen. Berlin: Akademie Verlag. Croatto, E. 1986. Vocabolario Ampezzano(Ampezzano dictionary). Cortina D’Ampezzo, ITA: Cassa rurale ed artigiana. Degand, L., B. Cornille & P. Pietrandrea. 2013. Discourse Markers and Modal Particles: Categorization and description. Amsterdam: John Benjamins. Diewald, G. 2006. Discourse Particles and Modal Particles as Grammatical Elements. In Approaches to Discourse Particles, ed. by, K. Fischer. 403-425. Amsterdam: Elsevier. Delancey, S. 1997. Mirativity: The grammatical marking of unexpected information. Linguistic Typology 1(1). De Gruyter Mouton. Dohi, A. 2019. La particella pa nelle varietà del ladino dolomitico con particolare attenzione al fassano. [The particle pa in Dolomitic Ladin varieties with special attention to Fassano]. PhD thesis, Trento: University of Trento. Egg, M & M. Zimmermann. 2012. Stressed Out! Accented Discourse Particles: The Case of ’doch’. Proceedings of Sinn Und Bedeutung, 16:1, 225-238. Fiorini, M. 2022. Italian poi as a trigger of expressive presupposition. In: Proceedings of the Linguistic Society of America 7(1). 52-69. Grosz, P. G. 2020. Discourse Particles. In: The Wiley Blackwell Companion to Semantics, 1–34. John Wiley & Sons, Ltd. Gunlogson, C. 2001. True to form: Rising and falling declaratives as questions in english. PhD thesis, Santa Cruz, CA: UC Santa Cruz. Gutzmann, D. 2015. Use-Conditional Meaning: Studies in Multidimensional Semantics. Oxford: Oxford University Press. Hack, F. M. 2014. The Particle Po In The Varieties Of Dolomitic Ladin – Grammaticalisation From A Temporal Adverb Into An Interrogative Marker. Studia Linguistica 68(1). 49–76. Hamblin, Charles L. 1973. Questions in Montague English. Foundations of Language 10(1). 41–53. Hinterhölzl, R. & N. Munaro. 2015. On the interpretation of modal particles in non-assertive speech acts in German and Bellunese. In: Discourse-oriented Syntax edited by Josef Bayer, Roland Hinterhölzl and Andreas Trotzke. John Benjamins Publishing Company. Jacobs, J. 1991. On the semantics of modal particles. In Discourse Particles, ed. by, W. Abraham. 141-162. Amsterdam: John Benjamins. Kaufmann, M. 2012. Epistemic particles and performativity, Proceedings of SALT 22, 208-225. Krifka, M. 2004. The Semantics of Questions and the Focusation of Answers. In: Topic and Focus: A Cross-Linguistic Perspective, 139–151. Dordrecht: Kluwer Academic Publishers. Krifka, M. 2008. Basic Notions of Information Structure. Acta Linguistica Hungarica, 55(3-4). 243–276. 88 MATTEO FIORINI Lindner, K. 1991. ‘Wir sind ja doch alte bekannte’. The use of German ja and doch als modal particles. In: Discourse Particles: Descriptive and theoretical investigations on the logical, syntactic and pragmatic properties of discourse particles in German. Ed. by, W. Abraham. Amsterdam: John Benjamins. Munaro, N. & C. Poletto. 2005. On the diachronic origin of sentential particles in North-Eastern Italian dialects. Nordic Journal of Linguistics 28(2). 247–267. Poletto, C. & R. Zanuttini. 2003. Making Imperatives: Evidence from Central Rhaetoromance. In The syntax of Italian Dialects, ed. by, C. Tortora, 175-206. Oxford: Oxford University Press. Potts, C. 2005. The Logic of Conventional Implicatures. Oxford: Oxford University Press. Potts, C. 2007. The expressive dimension. Theoretical Linguistics, 33, 165-198. Rizzi, L. 1997. The Fine Structure of the Left Periphery. In Elements of Grammar: Handbook in Generative Syntax. Ed. by, L. Haegeman. 281–337. Springer Netherlands. Rooth, M. E. 1985. Association with Focus (Montague Grammar, Semantics, Only, Even). PhD thesis, Amherst, MA: University of Massachusetts Amherst. Sauerland, U. 2007. Beyond unpluggability. Theoretical Linguistics, 33. 231-236. Schlenker, P. 2007. Expressive presupposition. Theoretical Linguistics, 33. 237-245. Thurmair, M. 1989. Modalpartikeln und ihre Kombinationen. Berlin, New York: Max Niemeyer Verlag. Zimmermann, M. 2011. Discourse particles. In An international Handbook of Natural Language Meaning. Ed. by, K. von Heusinger, C. Maienborn & P. Portner. Berlin, Boston: De Gruyter Mouton. 2012-2038. Proceedings of CLS 58 (2022), 89-103 © Chicago Linguistic Society 2023. All rights reserved. 89 UR Underlearning of Mandarin Chinese Tone 3 Sandhi Words1 Boer Fu Massachusetts Institute of Technology 1 Introduction This paper addresses a long-standing question in the field of phonological learning: can speakers learn UNDERLYING REPRESENTATIONS (UR) that are different from their SURFACE REPRESENTATIONS (SR) in the absence of morphological alternation (see Yip 1996, McCarthy 2005, Rasin et al. 2020, Richter 2021)? In languages that are rich in morphology, alternations offer help for speakers to learn non-identical UR-SR mappings. For example, in German final devoicing, a word-final voiced consonant in the UR is mapped to a voiceless consonant in the SR. Speakers of German have no trouble learning the non-identical UR-SR mapping, since they can rely on morphological alternations between singular and plural nouns (e.g. [taːk] ‘day.SG’ and [taːgə] ‘day.PL’). What happens when there is no morphological alternation? Can speakers still learn non-identical UR-SR mappings? I set out to address this question by exploring Mandarin Chinese. The language’s shortage of inflectional morphology makes it an interesting test case for UR learning. I focus on the famous process of TONE 3 SANDHI, which results in a tonal mismatch between the UR /T3 T3/ and the SR [T2 T3]. Can Mandarin speakers learn the non-identical tonal UR-SR mapping in the absence of morphological alternations? Previous literature on the phonological learning of tone 3 sandhi has mainly focused on how productive the process is, examining whether Mandarin speakers can consistently apply the sandhi rule in nonce words (see J. Zhang & Lai 2010, C. Zhang & Peng 2013). There has not been much work to investigate whether Mandarin speakers can learn the non-identical UR. In fact, it is usually assumed that the speaker URs of sandhi words are identical to their dictionary entry, T3 T3. I challenge the assumption that all Mandarin disyllabic words that are purported to have undergone tone 3 sandhi are consistently learned with a /T3 T3/ UR by speakers. I show that many sandhi words are in fact stored with a surface-identical UR /T2 T3/ by some Mandarin speakers, contrary to the dictionary entry. This is demonstrated via a novel AABB reduplication diagnostic that I have designed. I conducted a pilot survey of 6 Mandarin speakers, making use of the AABB diagnostic. The results indicate that many sandhi words are learned with identical UR-SR mapping, or UNDERLEARNED. The term refers to the underapplication of tone 3 sandhi in the speaker’s grammar. I argue that underlearning only takes place when the target sandhi word is compositionally Many thanks to Adam Albright, Fulang Chen, Edward Flemming, Filipe Kobayashi, Michael Kenstowicz, Patrick Niedzielski, Donca Steriade, Stanislao Zompì for valuable discussions. The project also benefited from comments at the MIT reading groups Morphun and Phonology Circle, as well as audience feedback at CLS 58. All remaining mistakes are my own. 1 90 BOER FU opaque, because speakers cannot establish morphological alternations between the target word and related words elsewhere in the lexicon that might reveal the non-identical UR. Therefore, I conclude that morphological alternation is crucial for the speaker to learn non-identical UR-SR mappings. The rest of the paper is organized as follows. Section 2 introduces Mandarin tone 3 sandhi and the resulting incomplete neutralization. Section 3 discusses why the UR learning of sandhi words is a problem for both learners and linguists alike. In Section 4, I delineate the AABB reduplication diagnostic. Results from a pilot survey are reported in Section 5. Section 6 and 7 discuss the implication in phonological learning and Chinese orthography. Section 8 concludes. In the paper, italics are used for dictionary entries, and slashes denote true UR. 2 Tone 3 sandhi and incomplete neutralization Tone 3 sandhi is perhaps the most well-studied phonological phenomenon in Mandarin Chinese. It involves two tones: tone 2, or the rising tone, and tone 3, the low dipping tone that is often accompanied with creaky voice. The sandhi rule states that whenever a tone 3 is followed by another tone 3, it surfaces as a tone 2. Formally, the rule is written as T3 ฀ T2 / __ T3. The tonal process leads to a case of neutralization in disyllabic words. Since all words with /T3 T3/ as their UR surface as [T2 T3], they become indistinguishable from words that have /T2 T3/ in UR. (1) shows an extreme case of neutralization. The minimal pair (1a) /wu3 pi3/ ‘five stroke input system’ and (1b) /wu2 pi3/ ‘unparalleled’ are neutralized into [wu2 pi3] after the application of tone 3 sandhi. (1) Neutralization in disyllabic words a. /wu3 pi3/ [wu2 pi3] 五笔 b. /wu2 pi3/ [wu2 pi3] 无比 ‘five stroke input system’ ‘unparalleled’ Mandarin tone 3 sandhi is reported to be a case of INCOMPLETE NEUTRALIZATION (see W. Wang & K. Li 1967, M. Lin et al. 1980, Liu 2013). The term is used to describe a phenomenon in which a contrast is maintained in production but lost in perception (see Port et al. 1981). Some Mandarin speakers are shown to be able to produce a small acoustic difference between /T3 T3/ and /T2 T3/ words, but it has been demonstrated that speakers cannot reliably perceive the difference. The acoustic difference is manifested in the pitch range of the rising tone on the initial syllable. Specifically, the rising tone [T2] derived from an underlying /T3/ begins at a lower pitch and ends at a lower pitch, compared to the rising tone that is underlyingly /T2/ (Zee 1980, Liu 2013, Yuan & Chen 2014). Mandarin listeners fail to perform better than chance at identifying the correct UR (W. Wang & K. Li 1967, M. Lin et al. 1980, Liu 2013). UR underlearning of tone 3 sandhi words 91 3 UR learning problem The incomplete neutralization of tone 3 sandhi leads to a learning problem. How do children acquiring Mandarin learn two distinct URs for a pair of lexical items that share the same SR? In fact, can they learn non-identical UR at all? As we have seen, however little pitch difference there is between /T3 T3/ and /T2 T3/ words, the acoustic information is of no use to Mandarin learners, since it cannot be reliably perceived. As to morphological alternations, Mandarin does not have much to offer. There is no productive word-internal morphological process, segmental or suprasegmental, that triggers change in the SR of sandhi words. Nevertheless, there is morphological alternation in Mandarin, in the form of compounding. If a sandhi word is composed of two morphemes that appear elsewhere in the lexicon, the learner has a chance to observe the tonal UR of the individual morphemes. This can be said for the neutralized pair in (1). (1a) [wu2 pi3] ‘five stroke input system’ is made up of the individual morphemes [wu3] ‘five’ and [pi3] ‘stroke’. Mandarin learners can establish a [wu2~wu3] alternation for the morpheme ‘five’ and infer that the UR for ‘five stroke’ is /wu3 pi3/. (1b) [wu2 pi3] ‘unparalleled’ is composed of [wu2] ‘neg’ and [pi3] ‘to compare’. The learner can see that the negation morpheme appears as [wu2] everywhere, concluding that ‘unparalleled’ is /wu2 pi3/ in UR. The above examples show that Compositional Transparency in compounds can help Mandarin learners acquire non-identical UR-SR mappings by providing morphological alternations. We should expect a learning asymmetry between compositionally transparent sandhi words and compositionally opaque sandhi words. If the target sandhi word is compositionally transparent, Mandarin learners ought to have no trouble learning a non-identical /T3 T3/ UR. However, if the target sandhi word is compositionally opaque, then not all learners will be able to posit a /T3 T3/ UR in accordance with the dictionary entry T3 T3. It is expected that a number of learners will store a /T2 T3/ UR, identical to the SR. In order to test the hypothesis on the learning asymmetry between words that are compositionally transparent and opaque, we need to find out, as a first step, the tonal UR in the mental lexicon of adult Mandarin speakers. As phonologists, we have no opportunity to see the tonal UR of words that are purported to have undergone tone 3 sandhi, for all the reasons listed above that makes the same task difficult for the Mandarin learners. To tackle this challenge, I have designed a novel AABB reduplication diagnostic, which can reveal the tonal UR Mandarin speakers have posited for words with [T2 T3] SR. 4 AABB reduplication diagnostic AABB reduplication is a semi-productive morphological process in Mandarin. It takes a disyllabic AB base, reduplicates each syllable, and forms a quadrisyllabic AABB word. The AB base can be an adjective, a verb, or a noun. The resulting AABB form expresses quantification of the AB base (N. Zhang 2015), as in (2). 92 (2) a. b. c. BOER FU AABB reduplication [kan1 tɕiŋ4] ‘clean’ [jow2 ɥy4] ‘to hesitate’ [ʂɨ2 kʰə4] ‘moment’ e. [kan1 kan1 tɕiŋ4 tɕiŋ4] f. [jow2 jow2 ɥy4 ɥy4] g. [ʂɨ2 ʂɨ2 kʰə4 kʰə4] ‘spotless’ ‘indecisive’ ‘every moment’ Only adjective-to-adjective AABB reduplication in (2a) is relatively productive. As a preview, the AABB reduplication diagnostic for tonal UR will make use of the unproductive noun-to-noun AABB reduplication, as seen in (2c). 4.1 AABB reduplication as a diagnostic for tonal UR The tonal sequence in an AABB reduplicated form can help disambiguate between /T3 T3/ bases and /T2 T3/ bases, as exemplified by (3). (3) AABB reduplication with /T3 T3/ vs. /T2 T3/ a. /two3 ʂan3/ ‘to evade’ c. [two2 two2 ʂan2 ʂan3] ‘evasive’ d. [two2 two3 ʂan2 ʂan3] b. /xʊŋ2 xwo3/ ‘flourishing’ e. [xʊŋ2 xʊŋ2 xwo2 xwo3] ‘very f. *[xʊŋ2 xʊŋ3 xwo2 xwo3] flourishing' (3a) /two3 ʂan3/ ‘to evade’ has two variant AABB forms in (3c&d) (Feng 2003). The second syllable [two2~3] (in boldface) in the quadrisyllabic adjective is the site of tonal variation. (3b) /xʊŋ2 xwo3/ ‘flourishing’ only has one AABB reduplicated form, which is (3e) [xʊŋ2 xʊŋ2 xwo2 xwo3]. The alternative tonal pattern of (3f), where the second syllable surfaces as [xʊŋ3], is illicit. A /T3 T3/ base has two variant AABB forms, because the reduplicated form can either be derived from the SR or the UR of the base. As seen in Table 1. The SR route takes the base [two2 ʂan3] ‘to evade’ to [two2 two2 ʂan2 ʂan3]. The UR route in Table 2, on the other hand, takes /two3 ʂan3/ and changes it into [two2 two3 ʂan2 ʂan3]. We can also see why a /T2 T3/ base can only have one AABB form. Its SR and UR are identical, leading to the same reduplicated form. AB Base SR AABB Reduplication Tone 3 Sandhi AABB reduplicated SR [two2 ʂan3] ‘to evade’ two2 two2 ʂan3 ʂan3 two2 two2 ʂan2 ʂan3 [two2 two2 ʂan2 ʂan3] [xʊŋ2 xwo3] ‘flourishing’ xʊŋ2 xʊŋ2 xwo3 xwo3 xʊŋ2 xʊŋ2 xwo2 xwo3 [xʊŋ2 xʊŋ2 xwo2 xwo3] Table 1: AABB reduplication via SR AB Base UR AABB Reduplication Tone 3 Sandhi AABB reduplicated SR /two3 ʂan3/ ‘to evade’ two3 two3 ʂan3 ʂan3 two2 two3 ʂan2 ʂan3 [two2 two3 ʂan2 ʂan3] /xʊŋ2 xwo3/ ‘flourishing’ xʊŋ2 xʊŋ2 xwo3 xwo3 xʊŋ2 xʊŋ2 xwo2 xwo3 [xʊŋ2 xʊŋ2 xwo2 xwo3] Table 2: AABB reduplication via UR UR underlearning of tone 3 sandhi words 93 The different behavior of /T3 T3/ and /T2 T3/ base in AABB reduplication means that we can utilize the morphological process as a diagnostic for tonal UR of the AB base. Given that the AABB reduplication is a semi-productive process, with the appropriate instruction, Mandarin speakers can apply it to any existing disyllabic word, forming a novel AABB reduplicated form. If the AB base word is one that is purported to have undergone tone 3 sandhi, what the speaker accepts as the AABB form can inform us of how they store the lexical item in UR. Specifically, if the sandhi word is truly stored as /T3 T3/ in the UR, then [T2 T2 T2 T3] and [T2 T3 T2 T3] are both possible as the reduplicated form. However, if the sandhi word is learned with a /T2 T3/ UR, then only [T2 T2 T2 T3] is licit as the AABB form. The alternative AABB form [T2 T3 T2 T3] will be rejected by the speaker, as schematized in Table 3. The diagnostic is stated in (4). Base UR: /T3 T3/ Base UR: /T2 T3/ AABB: [T2 T2 T2 T3] AABB: [T2 T3 T2 T3] Accept (via SR) Accept (via UR) Accept (via UR or SR) Reject! Table 3: AABB form judgement table (4) The AABB reduplication diagnostic: For a word with surface [T2 T3], if the speaker rejects [T2 T3 T2 T3] as an AABB reduplicated form, then they have posited a /T2 T3/ UR for the lexical item. Otherwise, the UR is /T3 T3/. 5 Speaker judgement survey I conducted a survey on Mandarin speakers’ judgement on the tones of AABB reduplicated forms. The survey serves two purposes. Firstly, it is a proof-of-concept trial run for the novel AABB reduplication diagnostic. The aim is to find out if the diagnostic can reliably reveal the tonal UR of sandhi words. Secondly, it is designed to detect if there is an asymmetry in the phonological learning of sandhi words, depending on the compositional transparency of the target word. 5.1 Methods The speaker judgement survey makes use of the AABB reduplication diagnostic. Mandarin speakers are given a list of disyllabic words with surface [T2 T3] and asked to judge the tones of the AABB reduplicated form. The survey is framed as a language game. At the beginning of the survey, the speakers are given the instruction that there is a novel AABB reduplication process that takes in any noun of the shape AB and turns it into an AABB noun, which now has the meaning ‘every AB’. The speakers are also provided with examples sentences. In addition to regular nouns, there are place names in the survey. They are given a slightly different novel meaning of ‘everywhere in AB’. For each word, the speaker is prompted with a question: “The word ____, after AABB reduplication, is pronounced as?” They are asked to choose between 94 BOER FU three options: (a) [T2 T2 T2 T3], (b) [T2 T3 T2 T3], and (c) Both forms are fine. The question prompt is written in Simplified Chinese. Each option has the AABB form written in PINYIN, the romanization script used in China, which transcribes lexical tones with diacritics. The pinyin transcription is also accompanied by the AABB form written in Chinese characters. (5) is a sample question, with English translation in italics. (5)“蚂蚁”一词，组成AABB的叠词后，读音是： The word “ant”, after AABB reduplication, is pronounced as? a. má má yí yǐ 蚂蚂蚁蚁 [ma2 ma2 ji2 ji3] b. má mǎ yí yǐ 蚂蚂蚁蚁 [ma2 ma3 ji2 ji3] c. 两种形式都可以。 Both forms are fine. If the surveyed speaker chooses option (a), it suggests that they have stored the word for ‘ant’ as /ma2 ji3/ in their mental lexicon. Option (b) maps to a /ma3 ji3/ UR. Option (c) “Both forms are fine” also points to a speaker UR of /ma3 ji3/. Recall that the AABB reduplication diagnostic states that it is only a rejection of the form [T2 T3 T2 T3] that counts as evidence for a /T2 T3/ UR. Given the availability of option (c) “Both forms are fine”, any speaker who chooses option (a) [ma2 ma2 ji2 ji3] does so as a rejection to the alternative (b) [ma2 ma3 ji2 ji3]. This means that their stored UR for the lexical item is truly /ma2 ji3/. If a dictionary T3 T3 word is shown to have been learned as /T2 T3/ by a speaker, a case of underlearning is identified. The term “underlearning” is named after underapplication, not underperformance. To learn a T3 T3 target word as /T2 T3/, the speaker is acquiring a grammar in which the sandhi rule is underapplied. It is important to note that whenever a case of underlearning is identified, the speaker is never in the wrong. The /T2 T3/ UR they have acquired is a valid and robust representation of the lexical item. 5.2 Participants 6 native speakers of Mandarin answered the survey, presented on Google Forms. 5.3 Materials There are 40 disyllabic words with [T2 T3] SR included in this survey. 26 have the citation tones of T3 T3. 14 are labeled as T2 T3. Among the disyllabic words across the two tonal categories, some are compositionally transparent, others are opaque. The 14 non-sandhi T2 T3 words are included, in order to make sure that selecting the AABB tonal SR is not a trivial task for the speakers surveyed. The first 2 lexical items the speakers see in the survey are the compositionally transparent ɥy3 san3 ‘umbrella’ and jɑŋ2 san3 ‘parasol’, serving as a baseline. 5.4 Results and discussion UR underlearning of tone 3 sandhi words 95 The speakers surveyed are shown to have robust judgement on the tones of the AABB reduplicated forms, even though they are not attested. This can be seen in their responses to the first two AB bases in the survey, ɥy3 san3 ‘umbrella’ and jɑŋ2 san3 ‘parasol’. These two words are both compositionally transparent, therefore the Mandarin speaker URs for them are expected to align with the dictionary entries. This is indeed what we observe, as shown in Table 4. The AABB reduplication diagnostic is a reliable measure to gauge speaker UR. Dictionary tones ɥy3 san3 ‘umbrella’ jɑŋ2 san3 ‘parasol’ [T2 T2 T2 T3] [T2 T3 T2 T3] Both are fine Speaker UR 0% speakers 33.3% speakers 66.7% speakers /T3 T3/ 100% speakers 0% speakers 0% speakers /T2 T3/ Table 4: Speakers have robust judgement on AABB tones Among the 26 sandhi words listed as T3 T3 in the dictionary, 9 have been underlearned as /T2 T3/ by one or more speakers surveyed. The 9 sandhi words subject to underlearning can be described as (i) compositionally opaque and (ii) containing an initial syllable that has an available tone 2 alternative. These two aspects of UR underlearning are discussed in 5.4.1 and 5.4.2 respectively. 5.4.1 Compositional transparency Recall that the hypothesis the survey aims to test is a learning asymmetry between sandhi words that are compositionally transparent and those that are opaque. It is largely borne out. 8 out of the 9 sandhi words that have been underlearned by a few speakers are opaque. The words in Table 5 all contain an initial syllable that does not contribute to the overall meaning of the word in a transparent manner. Sandhi word (dictionary tone) a. law3 老 ʂu3 鼠 ‘old’ ‘rodent’ b. ma3 蚂 ji3 蚁 ? ‘ant, termite’ c. ma3 马 tʰʊŋ3 桶 ‘horse’ ‘bucket’ d. mɤŋ3 猛 ma3 犸 ‘ferocious’ ? English UR: /T3 T3/ UR: /T2 T3/ ‘rat’ 33.3% speakers 66.7% speakers ‘ant’ 50% speakers 50% speakers ‘toilet’ 66.7% speakers 33.3% speakers ‘mammoth’ 66.7% speakers 33.3% speakers Table 5: Compositionally opaque words are prone to /T2 T3/ underlearning. In Table 5, (a) law3 ʂu3 ‘rat’ is an animal name. It is ostensibly composed of law3 ‘old’ and ʂu3 ‘rodent’. However, there is nothing ‘old’ about rats. The initial syllable law3 is not contributing transparently to the overall meaning of the word, which makes it prone to underlearning. In (b) ma3 ji3 ‘ant’, the initial syllable ma3 does not have a meaning other than ‘ant’. The word formation of ma3 ji3 is akin to that of cranberry in English, where berry clearly denotes a category, while cran does not have an independent meaning. The syllable ma3 appears again in 96 BOER FU (c) ma3 tʰʊŋ3 ‘toilet’, albeit transcribed with a different character. This time, it is written with the famous ‘horse’ character. If the meanings of the character components of ma3 tʰʊŋ3 were used for word composition, one would arrive at ‘horse bucket’, not ‘toilet’. The comical result could be due to ma3 tʰʊŋ3 being a historical compound, where the etymology is unclear. For Mandarin learners, the initial syllable in ‘toilet’ does not have the meaning of ‘horse’. Therefore, the surface [ma2] does not necessarily map onto /ma3/, resulting in underlearning. (d) mɤŋ3 ma3 ‘mammoth’ is a loanword, which is naturally compositionally opaque. In contrast, the compositionally transparent sandhi words in Table 6 are never underlearned. (a) ɥy3 san3 ‘umbrella’ begins with the morpheme ɥy3 ‘rain’, which transparently points to the context in which the physical object is used. (b) tʰu3 ɻɑŋ3 ‘soil’ is a coordinative compound made out of two nouns describing very similar concepts, tʰu3 ‘earth’ and ɻɑŋ3 ‘soft soil’. In (c) ʂwej3 mu3 ‘jellyfish’ is an animal that dwells in ʂwej3 ‘water’. As to (d) li3 xaj3 ‘Caspian Sea’, one only need to refer to a map to see that it is a landlocked sea ‘inside’ a continent. Sandhi word (dictionary tone) a. ɥy3雨 san3伞 ‘rain’ ‘umbrella’ b. tʰu3土 ɻɑŋ3壤 ‘earth’ ‘soft soil’ c. ʂwej3水 mu3母 ‘water’ ‘mother’ d. li3里 xaj3海 ‘inside’ ‘sea’ English UR: /T3 T3/ UR: /T2 T3/ ‘umbrella’ 100% speakers 0% speakers ‘soil’ 100% speakers 0% speakers ‘jellyfish’ 100% speakers 0% speakers ‘Caspian Sea’ 100% speakers 0% speakers Table 6: Compositionally transparent words are not underlearned. 5.4.2 Availability of tone 2 alternative Sandhi word (dictionary tone) a. kan3橄 lan3榄 ? ? b. kow3枸 tɕʰi3杞 ? ? English UR: /T3 T3/ UR: /T2 T3/ ‘olive’ 100% speakers 0% speakers ‘goji berry’ 100% speakers 0% speakers Table 7: Exceptional opaque words that are not underlearned. Unexpectedly, a few sandhi words that are compositionally opaque are not underlearned by any of the surveyed speakers, as shown in Table 7. (a) kan3 lan3 ‘olive’ and (b) kow3 tɕʰi3 ‘goji berry’ are monomorphemic plant names. I argue that the compositionally opaque sandhi words in Table 7 all lack a crucial ingredient to trigger underlearning — an available tone 2 alternative for the initial syllable. In other words, their initial syllables have a TONAL GAP at tone 2. A Mandarin syllable is said to have a tonal gap when one of the 4 available lexical tonal configurations is unattested in the lexicon. This can be observed in the comparison between ma, the linguists’ favorite syllable for Mandarin tone UR underlearning of tone 3 sandhi words 97 demonstrations, and pʰan. In Table 8, ma enjoys the full paradigm, appearing in tone 1, 2, 3, and 4, as ‘mother’, ‘hemp’, ‘horse’, and ‘scold’. In contrast, pʰan can only appear in tone 1, 2, and 4. There is no *pʰan3 anywhere in the lexicon. Tone 1 ma1‘mother’ pʰan1 ‘to climb’ Tone 2 ma2 ‘hemp’ pʰan2 ‘plate’ Tone 3 ma3 ‘horse’ *pʰan3 Tonal Gap Tone 4 ma4 ‘scold’ pʰan4 ‘to judge’ Table 8: Tonal gap Tonal gap offers a clue to speakers’ divergent learning pattern for the two sets of compositionally opaque sandhi words in Table 5 and 7. In a frequently underlearned opaque sandhi word, the initial syllable, which is the site of tonal UR ambiguity, does not have a tonal gap at tone 2. This is shown in Table 9. In other words, the underlearned sandhi word’s initial syllable has an available tone 2 alternative. In contrast, the initial syllable in an opaque sandhi word that is never underlearned has a tonal gap at tone 2, or that it does not have an available tone 2 alternative, as seen in Table 10. Sandhi Word (Dictionary Tone) a. law3 ʂu3 ‘rat’ b. ma3 ji3 ‘ant’ Initial Syllable Tone 2 alternative law2 ‘labor’ ma2 ‘hemp’ Tone 3 alternative law3 ‘old’ ma3 ‘horse’ Rate of /T2 T3/ Underlearning 66.7% speakers 50% speakers Table 9: Available tone 2 alternative for initial syllable leads to underlearning. Sandhi Word (Dictionary Tone) a. kan3 lan3 ‘olive’ b. kow3 tɕʰi3 ‘goji berry’ Initial Syllable Tone 2 alternative *kan2 Tonal Gap *kow2 Tonal Gap Tone 3 alternative kan3 ‘to dare’ kow3 ‘dog’ Rate of /T2 T3/ Underlearning 66.7% speakers 50% speakers Table 10: Unavailable tone 2 alternative for initial syllable blocks underlearning. The Mandarin learner hears the initial syllable of a sandhi word in its surface [T2]. Their task is to map it onto an underlying tone. When the syllable can appear in tone 2 or tone 3 in the wider lexicon, the learner needs to select either /T2/ or /T3/ as its UR. Sometimes they make a decision that disagrees with the dictionary entry, which results in underlearning of the sandhi word. However, when the initial syllable has a tonal gap at tone 2, as is the case for the opaque sandhi words in Table 10. the Mandarin learner knows that the syllable cannot possibly be /T2/ in UR. Therefore, they always map the initial syllable to /T3/, landing on a /T3 T3/ UR for the entire word. No underlearning takes place. I have demonstrated that a Mandarin sandhi word is prone to underlearning if (i) it is compositionally opaque; (ii) its initial syllable has an available tone 2 98 BOER FU alternative. On the flip side, a sandhi word is almost never underlearned if it is (iii) compositionally transparent; or (iv) compositionally opaque but has an initial syllable with a tonal gap at tone 2. 5.4.3 Overlearning In the previous two subsections, I have discussed the factors contributing to underlearning of sandhi words transcribed as T3 T3 in the dictionary. But what about overlearning? Do words listed as T2 T3 ever get learned as having a /T3 T3/ UR, in which the Mandarin learner’s mental lexicon includes an overapplication of tone 3 sandhi for these words? The answer is almost never, with exceptions. 14 of the 40 words surveyed are listed as T2 T3 in the dictionary. 12 are learned consistently by the surveyed speakers as having /T2 T3/ in UR. The 2 non-sandhi items in Table 11 show some degree of overlearning. Non-Sandhi Word (Dictionary Tone) a. li2厘 mi3米 ‘ancient unit’ ‘meter’ b. xu2湖 pej3北 ‘lake’ ‘north’ English ‘centimeter ’ ‘Hubei’ UR: /T3 T3/ 33.3% speakers 16.7% speakers UR: /T2 T3/ 66.7% speakers 83.3% speakers Table 11: Non-sandhi T2 T3 words overlearned as /T3 T3/ These results are unexpected, because we usually expect a learning bias towards faithful mapping between the UR and the SR, as a result of lexicon optimization (see Prince & Smolensky 1993). When a Mandarin-learning child hears a word with [T2 T3] SR, they are expected to map it to an identical UR /T2 T3/, unless given evidence that suggests otherwise. As we have seen in the discussion of sandhi words underlearning, evidence can come in the form of transparent morphological alternation involving the initial syllable or a tonal gap at tone 2. The 2 non-sandhi words in Table 11 do not provide such evidence. They include an initial syllable that is transcribed as tone 2 in the dictionary, therefore they have no reason to participate in morphological alternation with any lexical item that contains the initial syllable in tone 3, at least prescriptively speaking. It is possible that some learners have reanalyzed the initial syllable as a different morpheme. For example, li2 ‘ancient unit’ in (a) li2 mi3 ‘centimeter’ could have been reanalyzed as li3 ‘inside’ or li3 ‘mile’. Similarly, xu2 ‘lake’ in (b) xu2 pej3 ‘Hubei’ could be taken to mean xu3 ‘tiger’ or xu3 ‘amber’. An alternative explanation for the unexpected overlearning is a confound that results from the question sequence in the survey. Both items in Table 11 are located at the very end of the survey (item number 39 and 40). A certain level of speaker fatigue is expected. For future studies, a randomized stimulus sequence for each participant is necessary to identify real cases of overlearning. UR underlearning of tone 3 sandhi words 99 6 Implications for phonological learning The investigation on the UR learning of Mandarin sandhi words aims to answer the research question: can speakers learn non-identical UR-SR mapping in the absence of morphological alternation? The answer appears to be no. Sandhi words listed as T3 T3 in the dictionary are a case of non-identical UR-SR mapping, at least prescriptively speaking. In Mandarin, morphological alternation comes in the form of transparent compounding. If a sandhi word is compositionally transparent, it offers evidence of morphological alternation with individual morphemes that appear elsewhere in the lexicon. However, if a sandhi word is compositionally opaque, its individual syllables are not recurring morphemes in the lexicon, therefore no morphological alternation can be found. The results from the AABB reduplication survey show that the morphological alternations provided by compositionally transparent sandhi words make the UR learning task straightforward for the Mandarin speaker. For these words, the speakers are almost guaranteed to acquire the non-identical mapping between /T3 T3/ and [T2 T3] consistently. When there is a true absence of morphological alternation, as in compositionally opaque sandhi words, Mandarin speakers do not always acquire the prescriptive non-identical UR-SR mapping. Instead, they frequently acquire identical UR-SR mappings for these opaque sandhi words. The widespread speaker variation we observe in the UR learning of opaque sandhi words raises the possibility that compositional transparency, and therefore what count as morphology alternations, might differ from speaker to speaker. In the borrowed animal name mɤŋ3 ma3 ‘mammoth’, the initial syllable mɤŋ3 might have been reanalyzed as the morpheme ‘ferocious’ by some Mandarin learners. This can account for the 66.7% of surveyed speakers who have acquired the dictionary UR /mɤŋ3 ma3/, as opposed to underlearning the lexical item. It can be said that these speakers have identified a morphological alternation between [mɤŋ2 ma3] ‘mammoth’ and [mɤŋ3] ‘ferocious’, that the other 33.3% have not. Other opaque sandhi words might be subject to the same type of reanalysis, or learner-initiated search for morphological alternations. This might account for why there is no opaque sandhi word in the survey that is underlearned by 100% of the speakers. For example, ma3 ji3 ‘ant’ might be reanalyzed as ma3 ‘horse’ ji3 ‘ant’, given that both ants and horses are animals that carry weight. The results from the survey on Mandarin speakers’ tonal UR show that the phonological learning of non-identical UR-SR mapping in the absence of morphological alternation is, if not impossible, difficult. In addition, it is the child language learner who decides on what counts as a morphological alternation, not the lexicographer or the adult literate speaker. This is likely true for the UR learning of other languages as well. For a child acquiring German final devoicing, the task of determining what counts as morphological alternation is a trivial one. It is evidently not trivial for a child acquiring Mandarin tonal UR. In order to confirm the key role of speaker-specific morphological alternations in UR learning, some concrete evidence of such speaker variation is 100 BOER FU needed. A future study will benefit from a lexical association experiment, which can firmly link a speaker’s UR to their morphological representation of sandhi words. 7 Implication for Chinese orthography Throughout the discussion above, I have assumed that phonological learning generally precedes the learning of the writing system. Therefore, Chinese orthography is not expected to affect the learning of Mandarin tonal URs. In this section, I consider the possibility that Mandarin speaker’s URs for sandhi words are influenced by the written Chinese characters. Help from written Chinese is only expected to be required in sandhi words that are not considered to be compositionally transparent by all speakers. In such a scenario, the orthography might be able to aid tonal UR learning directly or indirectly. If the target word contains common Chinese characters, the learner can directly identify the tonal UR. ma3 tʰʊŋ3 马桶 ‘toilet’ is such an example. The first syllable is written with a common character 马. In isolation, 马 means ‘horse’ and is pronounced as [ma3]. The child learner can make use of their knowledge of the character to posit /ma3 tʰʊŋ3/ as the UR. Sometimes, the target word is written with specialized characters that rarely appear in the environment. Nevertheless, many of these specialized characters often contain a common phonetic component, which offers an indirect clue for the learner. Most Chinese characters are made up of two parts: a SEMANTIC COMPONENT and a PHONETIC COMPONENT. The semantic component, or the radical, indicates the semantic category of the character. The phonetic component provides some rough clues to its pronunciation. For example, the character 橄 in kan3 lan3 橄榄 ‘olive’ has a phonetic component敢, which is a common character pronounced as [kan3]. A Mandarin-speaking child encountering the word 橄榄 ‘olive’ for the first time, might be able to infer that 橄 ought to be pronounced as敢 [kan3]. However, the pronunciation of the phonetic component has been subject to sound change over time. A Mandarin speaker might find the phonetic component as helpful as the English spelling ough, which is used in rough, ought, through, drought, each with a different pronunciation. For instance, the character 枸 in kow3 tɕʰi3 枸杞 ‘goji berry’ has the phonetic component 句, which is pronounced as [tɕy4] on its own. It shows up as [tɕy1] in 驹, [ɕy4] in 煦, [kow3] in 狗, [kow4] in 够, to name a few common characters. The tone varies, as well as the segments of the syllable. A child learning the word 枸杞 ‘goji berry’ will not be able to pinpoint the UR of 枸 with accuracy. Despite the one-to-many mapping between the phonetic component 句 and its pronunciation in related characters, the child might still be able to locate 狗 [kow3] as the indicator to the sound of 枸. A probabilistic learning process is likely to be involved, in which the learner searches through all the Chinese characters sharing the same phonetic component in their inventory, making a prediction on the most likely pronunciation of the UR underlearning of tone 3 sandhi words 101 target character. A computational model of the character learning process can be an interesting extension of this project. If it is indeed the case that knowledge on Chinese orthography has helped Mandarin learners in acquiring tonal URs, then we should expect an age of acquisition effect. Specifically, sandhi words that are acquired before literacy ought to not be affected by orthography, therefore displaying a higher rate of underlearning. In contrast, sandhi words acquired after the child has learned the writing system are expected to show a lower rate of underlearning. For any word, its age of acquisition surely differs from speaker to speaker, but its lexical frequency might serve as a good enough approximation of the age of acquisition within a population. Words with higher lexical frequency are more likely to be acquired at an earlier age. This is yet another avenue for future research. Regardless of the role that Chinese orthography plays in phonological learning, the results from the speaker survey challenge the long-held view that Chinese characters are logographic, namely that it transcribes meanings, as opposed to sounds in an alphabetical writing system. Instead, there are multiple examples of meaningless characters in the survey. There are characters that only appear in certain animal and plant names like 蚂 in 蚂蚁 ‘ant’ and 橄 in 橄榄 ‘olive’, with no independent meaning on their own. Even characters that enjoy a high token frequency in the written language can appear meaningless in a compositionally opaque word. The best example is the character 马. On its own, it means ‘horse’, but when it is found inside马桶 ‘toilet’, it has no meaning. In the synchronic Mandarin grammar, its only function is to transcribe the sound of [ma2] in [ma2 tʰʊŋ] ‘toilet’. The Chinese writing system includes many more characters like马, used to transcribe sound, not meaning. 8 Conclusion This project investigates the tonal UR learning in Mandarin words that are purported to have undergone tone 3 sandhi. It poses two questions: what UR have the Mandarin speakers learned for sandhi words? What does Mandarin speakers’ behavior in UR learning tell us about phonological learning in general? To address the first question, I have designed a novel AABB reduplication diagnostic, which is shown to be an effective tool at gauging individual speaker’s UR for sandhi words. For the second question, I incorporated the diagnostic in a speaker judgement survey, in order to see if morphological alternations, found only in compositionally transparent compounds, are a requisite for establishing non-identical UR-SR mappings. A key finding of the survey is that Mandarin speakers frequently learn identical UR-SR mappings for compositionally opaque sandhi words that have non-identical UR-SR mappings prescriptively, in a process termed underlearning. A few compositionally opaque sandhi words that are never underlearned appear to be exceptions at first sight, but closer inspection shows that tonal gaps can serve as a supplementary tool for the learning of non-identical UR-SR mapping. The survey results confirm the hypothesis that there is a 102 BOER FU learning asymmetry between compositionally transparent and opaque sandhi words. They also strongly support the claim that morphological alternations are crucial to the learning of non-identical UR-SR mappings. At the same time, the between-speaker variation in UR learning points to the possibility that morphological alternations are speaker-specific. For a morphological alternation to be useful, it needs to be identified by the child language learner as such. References Feng, G. B. 2003. Lexical category specific constraints: Mandarin verb versus adjective reduplication. USC Working Papers in Linguistics 1. 1-12. Lin, M., L. Lin, G. Xia, & Y. Cao. 1980. Putonghua erzici biandiao de shiyan yanjiu. [Experimental study of Mandarin disyllabic word tone sandhi.] Zhongguo Yuwen [Studies of the Chinese Language] 1980(1). Liu, X. 2013. Shangsheng biandiao de shengxue yu ganzhi shiyan yanjiu. [Acoustic and perceptual experimental study of third tone sandhi.] Wenjiao Ziliao [Data of Culture and Education] 625. 139-142. McCarthy, J. J. 2005. Taking a free ride in morphophonemic learning. Catalan Journal of Linguistics 4(19). 19-55. Port, R., F. Mitleb, & M. O’Dell. 1981. Neutralization of obstruent voicing in German is incomplete. J. Acoust. Soc. Am. 70(S1). S13-S13. Prince, A., & P. Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Rasin, E., I. Shefi, & R. Katzir. 2020. A unified approach to several learning challenges in phonology. In Proc. NELS 50, 73-87. Richter, C. L. 2021. Alternation-sensitive Phoneme Learning: Implications for Children’s Development and Language Change. University of Pennsylvania dissertation. Wang, W. S., & K. P. Li. 1967. Tone 3 in Pekinese. Journal of Speech and Hearing Research. 10(3). 629-636. Yip, M. 1996. Lexicon Optimization in Languages without Alternations. Yuan, J., & Y. Chen. 2014. 3rd tone sandhi in Standard Chinese: a corpus approach. Journal of Chinese Linguistics. 42(1). 218-237. Zee, E. 1980. A spectrographic investigation of Mandarin tone sandhi. UCLA WPP 49. 98-116. Zhang, C., & G. Peng. 2013. Productivity of Mandarin third tone sandhi: a wug test. Eastward Flows the Great River: Festschrift in Honor of Prof. William S-Y. Wang on his 80th Birthday. 256-282. Zhang, J., & Y. Lai. 2010. Testing the role of phonetic knowledge in Mandarin tone sandhi. Phonology 27(1). 153-201. Zhang, N. N. 2015. The morphological expression of plurality and pluractionality in Mandarin. Lingua 165. 1-27. Proceedings of CLS 58 (2022), 103-118 © Chicago Linguistic Society 2023. All rights reserved. 103 Question Answer Pairs in Hong Kong Sign Language (HKSL) Linghui Eva Gan University of Connecticut 1 Introduction This paper discusses the syntactic and semantic properties of Question Answer Pairs (QAPs) in Hong Kong Sign Language (hereafter HKSL). The construction is shown in (1). The first clause adopts the form similar to an interrogative, the second clause provides the answer, and the whole construction denotes a statement.1 Despite the fact that different terms have been used to refer to the construction (to be introduced shortly), in this paper, I refer to it as a Question Answer Pair (QAP). The first clause is referred to as the Q(uestion)-clause, and the second clause is the A(nswer)-clause. br AARON LIKE WHAT, CAT. 2 (1) ‘What Aaron likes is cats.’ QAPs are commonly used across sign languages (Wilbur 1996; Caponigro & Davidson 2011; Stickles 2012; Branchini 2014; Kimmelman & Vink 2017; Hauser 2018; Khristoforova & Kimmelman 2021; Huddlestone 2021). Cross-linguistic variations are attested, including whether or not the wh-phrase must appear at the final position of the Q-clause; whether the non-manual marking of the Q-clause is exclusively a brow raise (br); and whether the A-clause always expresses exhaustivity, etc. There are also controversies regarding the nature of the construction. This paper aims to explore the syntactic and semantic properties of QAPs in HKSL, which aims to contributes to the understanding of QAPs in sign languages in general and the information packaging strategies in HKSL. The paper is organized as follows: In the rest of Section 1, I briefly summarize the theoretical backgrounds regarding the syntactic and the semantics properties of QAP across sign languages. In Section 2, I discuss the syntactic properties of QAPs in HKSL. I show that QAPs in HKSL display intermediate clausal status between a single syntactic unit and a discourse-level question and answer combination (i.e., two sentences). In Section 3, I discuss a few semantic properties of QAPs in HKSL. I show that QAPs share some similarities with topic-comment constructions but are different in nature. In Section 4, I briefly discuss the use of the final position in sign languages and conclude the paper. 1 The interrogative in the first clause can either be a wh-question or a polar question, and it typically co-occurs with a brow raise (br). 2 Glossing convention: Throughout this paper, examples from languages other than HKSL will be specified with the source of the language source in square brackets. For sign language examples, manual signs are glossed in small caps. IX stands for pointing signs. The number 1, 2, or 3 following it refers to first person, second person, and third person, respectively, e.g. IX -1 ‘I’, IX -2 ‘you’. The non-manuals are indicated by the line and the abbreviation above the manual glossing (br: brow raise; bf : brow furrow). 104 LINGHUI EVA GAN 1.1 Theoretical background In the literature, there are three main milestone-like proposals regarding the nature of QAPs. They are all base on American Sign Language (ASL) data. In the early discussions, the construction is regarded as a rhetorical question (BakerShenk 1985). Baker & Cokely (1980); Neidle et al. (2000); Valli & Lucas (2000) adopt this term. Wilbur (1996) later argues against this view. She proposes that the construction is the equivalence of a wh-cleft (pseudocleft) in spoken languages like (2). In particular, the Q-clause is a wh-XP that acts as the predicate, and the Aclause is the subject. The wh-XP moves leftward to Spec CP, and the subject moves to Spec IP. Semantically, A-clause makes use of the clause-final position to express focus. For QAPs in Italian Sign Language (LIS), Branchini (2014) also regard them as wh-clefts. (2) [English] What I like is apple. However, Caponigro & Davidson (2011) argue that a QAP is different from a wh-cleft. They propose that it is a Clausal Question-Answer Pair (see also Davidson et al. (2008a), Davidson et al. (2008b)). Syntactically, it contains a silent copula eBE taking an interrogative CP (the Q-clause) as its subject, and an IP with possible elided material as its complement (i.e., the A-clause). Semantically, the Q-clause conveys a question and the A-clause delivers a true answer to that question. Moreover, they argue that the Q-clause and the A-clause in a QAP always assert two identical propositions, and that QAPs in ASL exhibit exhaustivity. Several following studies adopt a similar term Question Answer Pair, including Kimmelman & Vink (2017) on Sign language of the Netherlands (NGT), Hauser (2018) on French Sign Language (LSF), Herrmann et al. (2019) on German Sign Language (DGS), Khristoforova & Kimmelman (2021) on Russian Sign Language (RSL), and Huddlestone (2021) on South African Sign Language (SASL). Now I turn to discuss properties of QAPs in HKSL. The data were collected from three native HKSL signers (average age 40) through Zoom. ‘Native’ is defined as being born to deaf parents and having been exposed to/using HKSL as the primary language of communication since birth. I followed fieldwork consultation method of elicitation and acceptability judgment tests. For the consistency of presenting the contexts and sentences to be judged, I set up Google Forms for the data collection. The consultants shared their screens with the researcher (the author) while answering the Google Forms and were recorded with consent. The judgment tests adopt a three-point scale (good, so-so, bad). 2 Clausal properties of QAPs in HKSL In this section, I will show that QAPs in HKSL display intermediate clausal status between a single syntactic unit and a discourse-level question and answer combination. I will first discuss properties from the two ends and then turn to in-between properties. Accordingly, I argue that HKSL locates somewhere similar to NGT in the grammaticalization trajectory (Hauser 2018). Question answer pairs in HKSL 105 2.1 QAP as a single syntactic unit QAPs in HKSL show characteristics of a single syntactic unit. The first piece of evidence is that a scope ambiguity is attested in a QAP in HKSL. In general, quantifier scope relation is evoked when the same clause has more than one quantifier. For instance, the English sentence in (3) is ambiguous. Despite the rigid scope (every > one) that is consistent with the structure hierarchy between the two quantifier NPs (every man and one woman), the structural lower quantifier NP one woman can take the wide scope through Quantifier Raising (QR), yielding the inverse scope of the sentence (one > every). (3) [English] Every man likes one woman. (a) (b) (ambiguous) Reading a: For all man x, there is a woman y such that x loves y. (every > one) Reading b: There is a woman y and for all man x, x likes y. (one > every) By contrast, in a discourse-level question and answer combination like (4), the question and the answer are independent sentences. Therefore, even though each sentence has a quantifier NP, there is no scope relation between the quantifier NPs, and (4) is not ambiguous. (4) [English] A: B: Who does every man like? One woman. (unambiguous) In HKSL, scope ambiguity is attested in certain conditions (Gan 2022a). For instance, when the structurally lower NP phrase contains the sign ONE and the higher NP phrase contains the sign ALL, an inverse scope is possible (ONE > ALL) yielding ambiguity (5). (5) MAN ALL LIKE WOMAN ONE . ‘Every man likes one woman.’ (ambiguous) (a) (b) (ALL > ONE) (ONE > ALL) For every man x there is a woman y such that x likes y. There is a woman y such that every man x likes y. In a QAP in HKSL, when the Q-clause contains an ALL-phrase and the A-clause contains a ONE-phrase, the sentence is also ambiguous (6). This indicates that the A-clause is in the same syntactic structure as the Q-clause, so that the scope relation of the two quantifiers is evoked. In other words, (6) is more similar to (3) than (4). br (6) MAN ALL LIKE WHO , WOMAN ONE . ‘Every man likes a woman.’ (a) (b) (?ambiguous) ?For every man x there is a woman y such that x likes y. (ALL > ONE) There is a woman y such that every man x likes y. (ONE > ALL) 106 LINGHUI EVA GAN Another potential evidence is that wh-phrases in HKSL always locate at the right periphery of the Q-clause. In regular wh-questions in HKSL, argument whphrases appear at the right periphery of the question. One exception is the subject WHO . It can either be clause-initial or clause-final (Gan 2022b), as in (7). bf (7) (a) EAT FISH NOT WHO ? ‘Who does not eat fish?’ bf (b) WHO EAT FISH NOT ? (Gan 2022b:6,7) Nevertheless, in a QAP, the subject WHO in the Q-clause has to appear clause-finally and cannot appear clause-initially (8). br (8) (a) EAT FISH NOT WHO , AARON . ‘The person who does not eat fish is Aaron.’ br (b) * WHO EAT FISH NOT, AARON . Such a constraint on the position of the wh-phrase in a Q-clause could be related to the clausal connectivity between the Q-clause and the A-clause, so the wh-phrase has to be immediately adjacent to the A-clause. Wilbur (1996:217) has a similar observation for ASL. While wh-phrases in ASL can remain in-situ, moves to the right, or to the left (Petronio & Lillo-Martin 1997; Neidle et al. 1998b; Neidle et al. 1998a; Neidle et al. 1996), they can only appear to the right in the Q-clause in a QAP. Wilbur argues that QAP is thus a single unit instead of a looser language pair. Admittedly, more tests are needed, such as the dependency between the wh-phrase and elements in the A-clause. 2.2 QAP as a discourse-level question answer combination Nevertheless, some QAPs in HKSL do show properties similar to a discourse-level question and answer combination, i.e., the question and the answer are two independent sentences. In particular, QAPs with an indirect answer in the A-clause cannot be embedded. In HKSL QAPs, in addition to a specific answer like in (8 a) that responds to the question raised in the Q-clause, the A-clause can also be an indirect answer, like ‘I don’t know’ (9) (the set of possible answers to WHERE is a set of places, although ‘I don’t know’ is not a member of this set, it is a valid answer). br (9) GLADYS GO WHERE , IX -1 DON ’ T - KNOW Lit. ‘Gladys went where, I don’t know.’ ‘I don’t know where Gladys went.’ Recall that a QAP expresses a statement (i.e., it is declarative), if (9) is a single syntactic unit, it should be able to be embedded as a clausal complement of a matrix verb that takes [-wh] complements, similar to (10) and the English example (11). (10) KENNY THINK [CP GLADYS GO SHATIN ] . ‘Kenny thinks that Gladys went to Shatin.’ Question answer pairs in HKSL 107 (11) [English] (a) (b) I don’t know where Gladys went. Kenny thinks that [CP I don’t know where Gladys went]. However, this is not the case. From the available interpretation, one can tell the syntactic configuration of the sentence in (12). The reading in (12 a) is not available. For this reading, the QAP structurally acts as the complement of the matrix verb THINK . Thus, (12 a) shows that (9) cannot be the complement of THINK as a whole. br (12) KENNY THINK GLADYS GO WHERE , IX -1 DON ’ T KNOW. (a) (b) # ‘Kenny thinks that I don’t know where Gladys went.’ [CP KENNY THINK [QAP [Q-clause GLADYS GO WHERE ] [A-clause IX -1 DON ’ T - KNOW ]]] ‘Where does Kenny think that Gladys went? I don’t know. [QAP [Q-clause KENNY THINK [GLADYS GO WHERE ]], [A-clause IX -1 DON ’ T - KNOW. ]] By contrast, the sentence as a whole can be interpreted as a QAP. The Q-clause contains a long-distance wh-question ‘Where does Kenny think that Gladys went?’, and the A-clause gives an indirect answer to it (12 b). Thus, the contrast between the two interpretations discussed above can be taken to imply that the Q-clause and the A-clause in (9) are not in a single syntactic unit. 2.3 In-between properties Some QAPs in HKSL got different acceptability judgements from the consultants. For some signers, the two clauses form a single unit, whereas for others, it is more like a discourse-level combination. The first set of data is about the binding of SELF. In HKSL, the sign SELF shows some logophor-like characteristics. It can appear independently as the subject within an embedded clause, being long-distance bound by the matrix subject (13 a). However, it cannot appear independently in a matrix clause (13 b). (13) (a) (b) (c) MOM SAY SELF BUSY. ‘Mom said that she (herself) is busy.’ * SELF BUSY. Intended: ‘I am busy.’ IX -1 BUSY. ‘I am busy.’ Thus, for a QAP that has SELF in the A-clause and its antecedent in the Q-clause, the grammaticality of the sentence can reflect whether or not the A-clause is independent from the Q-clause (14 b). (14) ‘Momi said that shei is busy.’ br (a) MOM i SAY WHAT , IX -3 i BUSY. (b) % MOMi SAY WHAT, SELFi BUSY. (%: inconsistent judgements: one signer judges as acceptable and two judge as bad.) br 108 LINGHUI EVA GAN Interestingly, compared with (14 a) that is acceptable for all three consultants, (14 b) got different judgements. One signer judges the sentence as acceptable, implying that the sentences is a single syntactic unit, so that SELF in the A-clause is bound by MOM in the Q-clause. By contrast, two signers judge the sentence as ungrammatical, which implies that for them, the Q-clause and the A-clause in (14 b) are more of a discourse-level question and answer combination. Thus, SELF being unbound as the subject in an independent sentence (the A-clause) is ungrammatical. Another piece of evidence comes from conditionals. In a conditional construction in English, a wh-cleft can either act as the antecedent or the consequence clause, forming one sentence (15). However, when involving a (discourse-level) wh-question and its answer, the sequence contains two sentences (16). (15) (a) If what Aaron brings back is a gold medal, Mom will cook shrimp. (b) If Aaron brings back a gold medal, what Mom will cook is shrimp. (16) If Aaron brings back a gold medal, what will Mom cook? Shrimp. Therefore, whether or not QAPs can be embedded in a conditional could inform us about its clausal status. In conditionals in HKSL, the conditional antecedent has to co-occur with a brow raise (br), and the manual sign IF is optional. Opposite judgements are again attested. For two signers, a QAP can either be the conditional antecedent or the consequence in a conditional (ok for both 17 b and 17 c). These judgements imply a tighter connectedness between the clauses, i.e., the sentences are more similar to (15). The other signer judges the sentences as bad, implying a looser connectedness, i.e., the sentences are more similar to a discourse-level question and answer sequence like (16). br (17) (a) ( IF ) AARON TEST SCORE ONE - HUNDRED, MOM COOK SHRIMP. ‘If Aaron gets a full score in the exam, mom will cook shrimps.’ br AARON TEST SCORE HOW - MANY, ONE - HUNDRED ], MOM COOK SHRIMP. (b) %[QAP (c) %AARON SCORE ONE - HUNDRED, [QAP MOM COOK WHAT, SHRIMP ]. (%: inconsistent judgments: two signers judge as ok and one judges as bad.) br br Comparing the judgements by each signer in (14) and (17) (Table 1), sign A and signer B seem to have consistent but opposite judgements regarding the connectedness between the Q-clause and the A-clause. And Signer C’s judgments are not consistent. Thus, the table shows that the intermediate status of QAPs appear not signer A signer B signer C SELF ok (1 clause) * (2 clauses) * (2 clauses) conditional ok (1 clause) * (2 clauses) ok (1 clause) Table 1: Summary of judgments on (14) and (17) from different consultants only across signers. Within the same signer, some QAPs seem to be more singleunit-like, while others are closer to discourse-level question and answer combinations. More scrutiny is needed to explore the possible constraints. For now, I leave it to future study. Question answer pairs in HKSL 109 Above all, QAPs in HKSL show intermediate status between a single syntactic unit and a discourse-level question and answer combination. Some QAPs are clearly single-unit-like, others are more similar to a discourse-level question and answer combination, whereas some show unstable status between the two among the three consultants. Kimmelman & Vink (2017) have similar observations for NGT. According to them, QAPs in NGT have properties that are wh-cleft like, but there are also QAPs that are rhetorical question-like. They thus propose that QAPs in NGT is undergoing grammaticalizaiton. Partially adopting Kimmelman & Vink (2017), Hauser (2018) provides a unified account and proposes the distribution of QAPs in different sign languages in a grammaticalization trajectory. Hauser proposes that the grammaticalization starts from being two independent sentences that are related at the discourse level, and ends up being a complex construction that is relative-clause like. There are five steps (p. 53), Step 1 is when a rhetorical question and its answer are used as a focus strategy. In Step 2, the two clauses are linked at the discourse level. In Step 3, the two clauses are conventionalized into a new construction as a single syntactic unit. In Step 4, the construction becomes a pseudo-free-relative. At this stage, the construction has a hybrid status CP and DP, and it lacks truly nominal properties to be used referentially (thus, it is not a real free-relative). In Step 5, the wh-constituent is normalized into a real free relative. Based on the previous literature and her own data, Hauser proposes that NGT locates between Step 2 and 3. ASL locates in Step 3, LSF locates in Step 4, and LIS locates in Step 5. Adopting Hauser (2018)’s grammaticalization trajectory, HKSL seems to locate at a stage similar to NGT, which is in between Step 2 and Step 3. Admittedly, more consistent cross-linguistic comparisons should be done for further verification. Now I move on to discuss the semantic properties of QAPs in HKSL. 3 Semantic properties of QAPs in HKSL In this section, I first compare QAPs with topic-comment construction, and then discuss the exhausitivity in QAPs. I propose that a QAP is a focus construction, but it is different from a wh-cleft. 3.1 Comparing Q-clauses with topics Q-clauses and a topic in HKSL share similarities. First, the Q-clause of a QAP in HKSL is typically marked with a brow raise (br), which is also used in marking topic-like information in the language, including aboutness topic (18 a) (Sze 2008; Sze 2011), topicalized constituent (18 b), and conditional antecedents (18 c). br (18) (a) VEGETABLE , AARON LIKE CORN . ‘As for vegetables, Aaron likes corn.’ (aboutness topics) br (b) JESSICA FAVORITE FIVE , I KNOW. ‘I know that Jessica favorites the number five.’ br (c) AARON EXAM SCORE 100, MOM COOK SHRIMP. (topicalized CP) 110 LINGHUI EVA GAN ‘If Aaron gets 100 in the exam, Mom will cook shrimps.’ (conditional antecedent) Moreover, similarities in information packaging is attested when comparing a QAP with a topic-comment construction. A topic is what the utterance in which it occurs is about, and the comment is the rest of the sentence. It seems that the Q-clause conveys old information like a topic does, and the A-clauses expresses new information, similar to the comment. This is reflected by the fact that the A-clause cannot appear in a topic. Recall that in a regular topic-comment construction like (19 a), the topic phrase (marked by br) conveys old information, and the comment clause IX -1 KNOW conveys new information. When a topic phrase contains a whole QAP, the sentence is ungrammatical (19 b). I argue that this is because the A-clause carries new information, so it cannot be marked by the brow raise signaling topic-like old information, yielding the ungrammaticality. br (19) (a) (b) [TopicP JESSICA FAVORITE FIVE], [CommentP IX -1 KNOW. ] ‘I know that Jessica likes the number five the most.’ * [TopicP [Q-clause JESSICA FAVORITE NUMBER br HOW - MANY ], br [A-clause FIVE]], [CommentP IX -1 KNOW. ] ‘I know that Jessica likes the number five the most.’ Nevertheless, Q-clauses possess properties that are different from a topic. Specifically, some types of comment is not allowed in A-clause. The comment of a topic can be either directly or indirectly relevant to the topic. For instance, in (20 a), the comment is specifically about the topic, i.e., the property of the O(rientation)-camp. In (20 b), however, the comment is about the weather on the day that the O-camp takes place, which is only indirectly relevant to the topic. (20) (a) As for the Orientation camp, it was fun. (b) As for the Orientation camp, the weather was nice. If a QAP is a topic-comment construction, both types of comment should be possible in the A-clause. This is not the case. It seems that in a QAP, the A-clause has to directly address the question raised by the wh-phrase in the Q-clause, and In an answer similar to the comment in (20 b) is not allowed. For instance, in (21), the wh-phrase HOW in both Q-clauses asks for a set of properties of the O-camp. If the A-clause answers the question with a property, the QAP is acceptable (21 a). But if the A-clause is a proposition about the weather on the day of the O-camp, the QAP is not acceptable (21 b). br (21) (a) O - CAMP HOW , FUN . ‘As for the O(rientation)-camp, (it was) fun.’ br (b) *O - CAMP HOW, WEATHER NICE . Intended: ‘As for the O(rientation)-camp, the weather was nice.’ Similarly, in (22 a), the A-clause KIND denotes a property, thus the appropriate wh-phrase in the Q-clause is HOW instead of WHAT (that denotes a set of nominal Question answer pairs in HKSL 111 entities). And in (22 b), the wh-phrase in the Q-clause is WHO, which denotes a set of individual. Therefore, the A-clause has to be an individual instead of a property like SMART. br (22) (a) TEACHER IX -3 HOW /* WHAT, KIND . ‘How’s that teacher, (she is) kind.’ br (b) ACTIVITY IX RESPONSIBLE ARRANGE WHO , AARON /* SMART. ‘Who is in charge of organizing the activity? Aaron/*Smart.’ In this sense, as the A-clause provides a possible answer that is a member of the set raised by the wh-phrase in the Q-clause, QAPs behave similarly to a regular wh-question and its corresponding answer (23). (23) A: What about the Orientation camp? (a) (b) #B: The weather was nice. B: It was fun. Recall that content-wise, what an A-clause and a Q-clause denote can be nonidentical (9), repeated below as (24). This is also attested in other sign languages, such as Russian Sign Language (RSL) (25). But why is a sentence like (24) acceptable but not (21 b)? br (24) GLADYS GO WHERE , IX -1 DON ’ T - KNOW Lit. ‘Gladys went where, I don’t know.’ ‘I don’t know where Gladys went.’ (25) [RSL](Khristoforova & Kimmelman 2021:ex. 8, 9) (a) (b) DO WHAT, NEE RESCUE ‘What to do? One should be rescued.’ RECTANGLE WHICH , KNOW - NOT ‘What is this rectangular thing? (I) don’t know.’ I suggest that this is due to a discourse requirement on QAPs. Specifically, the A-clause must be congruent to the Question Under Discussion (QUD) raised by the Q-clause. A QUD is a question that constitute a ‘live issue’ that have been introduced for discussion at a given point in the conversation, and has not yet been downdated (Hamblin 1973; Roberts 1996). I suggest that the A-clause has to be either a semantically congruent answer, which is the case of (21 a), or a pragmatically congruent answer, like ‘I don’t know’ in (24). For the latter, although the A-clause is not a member of the set raised by the wh-phrase in the Q-clause, it responds to the question intention raised by the Q-clause (Groenendijk & Stokhof 1984). By contrast, in sentences like (21 b), the A-clause is unacceptable because it is neither a semantically congruent answer nor a pragmatically congruent one. 112 LINGHUI EVA GAN 3.2 QAP as a focus construction In ASL, there are controversial data on whether the A-clause provides exhaustive information. On the one hand, Caponigro & Davidson (2011); Stickles (2012) show that the A-clause expresses exhausitivity. For instance, in (26), if the A-clause of a QAP only mentions parts of the contextually salient exhaustive set, the QAP is infelicitous (26 a). If both of the contextually salient item are mentioned in the A-clause, the QAP is acceptable (26 b). (26) [ASL] (Caponigro & Davidson 2011:ex. 75)] Context: Meg bought the contextually salient book and CD. (a) (b) # MEG BUY WHAT, BOOK . ‘The thing/What Meg bought is the book.’ MEG BUY WHAT, BOOK , CD . ‘The thing/What Meg bought are/is the book and the CD.’ In addition, one can reject the A-clause by giving supplementations of the list provided. In a it-cleft in English like (27 a), Italy is the only place that Mary went, and a rejection of the exhausitivity is felicitous (27 b). In ASL, a similar rejection is possible, implying that the list given by the A-clause is exhaustive (28). (27) [English](E. Kiss 1998:ex.11) (a) (b) It was Italy where Mary went. No, Mary went to Spain, too. (28) [ASL] (Stickles 2012:ex.4) br (a) (b) A: MARY BUY WHAT, JACKET. ‘What Mary bought is a jacket.’ B: NO , MARY BUY BACKPACK TOO . ‘No, Mary also bought a backpack.’ However, not all QAPs express exhaustivity in the A-clause. There are ‘mentionsome’ QAPs in ASL. For instance, in (29), the A-clause only conveys a partial answer to the question expressed by the Q-clause. (29) [ASL] (Caponigro & Davidson 2011:ex. 80) Context: Starbucks is one of the three places where one can find coffee in the area considered to be relevant. CAN FIND COFFEE WHERE , STARBUCKS . ‘You can find coffee at Starbucks.’ In HKSL, mixed judgments regarding the exhausitivity of the A-clause are also attested. One the one hand, in (30), the QAP can be rejected, indicating that the A-clause provides an exhaustive set. Also, (31) is felicitous only under the context in which Aaron knows what Mom bought exactly (i.e., it is no more than vegetables and meat, in other words, the A-clause is an exhaustive list). (30) Context: Aaron bought chocolate, tea and a banana: Question answer pairs in HKSL 113 br (a) (b) A: AARON BUY WHAT, CHOCOLATE , TEA . ‘What aaron bought was chocolate and tea.’ B: NO , IX -2 MISTAKE , AARON BUY BANANA HAVE . ‘No, you are wrong. Aaron also bought a banana.’ br (31) AARON KNOW MOM BUY WHAT, VEGETABLE , MEAT. (a) (b) Context 1: Aaron knows that Mom only bought vegetable and meat. # Context 2: Aaron knows that Mom bought vegetable, meat, and also some fruit. However, in (32), when judging whether what Brenda repeated is true, the signers gave mixed judgements. One signer judges (32 b) as infelicitous, which indicates that the A-clause in (32 b) denotes exhausitivity. However, this sentence is felicitous for two signers, indicating that for them, the A-clause in (32 b) does not denote exhausitivity. (32) Context: In a retelling game, Brenda must truthfully repeat all/part of the content of what Connie said. (a) Connie: LISA BUY HAT, BOOK , UMBRELLA , FINISH . ‘Lisa bought a hat, a book, and an umbrella, and nothing else.’ (b) Brenda: % LISA BUY WHAT, HAT. ‘Lisa bought a hat.’ (%: inconsistent judgements: signer A: infelicitous; signer B& signer C: felicitous) br I propose that the controversies in the exhausitivity expressed by the A-clause is due to focusing. In particular, in QAPs in HKSL, the A-clause can convey both information focus and contrastive focus. Only when it is the latter, the A-clause must denote exhausitivity. Take (33) as an illustration, signer A asked an open question, and signer B answers it with a list of what Aaron bought. In this case, the A-clause conveys information focus. It can either be an exhaustive list (33 a) or a non-exhausitive one (33 b). (33) Context: Yesterday Aaron went shopping. He bought some bananas, chocolates, tea, and nothing else. Signer A and signer B are talking about it. Signer B knows what Aaron bought. br A: YESTERDAY AARON GO BUY - BUY YES - NOT - YES? ‘Aaron went shopping yesterday, right?’ (a) (b) B: YES . AARON BUY WHAT, BANANA , CHOCOLATE , TEA . ‘Yes. Aaron bought bananas, chocolate, and tea.’ (exhaustive) B: ? YES . AARON BUY WHAT, BANANA . ‘Yes. Aaron bought bananas.’ (non-exhaustive) By contrast, in (34), signer B in rejects what signer A said, thus it conveys contrastive focus, and the QAP is only felicitous when the A-clause is exhaustive (34 a) but not non-exhaustive (34 b). Following this, because the repetition in (32 b) conveys information focus, the A-clause does not have to be exhaustive. 114 LINGHUI EVA GAN (34) Context: Yesterday Aaron went shopping. He bought some bananas, apples, tea, and nothing else. A and B are talking about it. B knows what Aaron bought. br A: YESTERDAY AARON BUY WHAT APPLE , TEA , FINISH . ‘Yesterday Aaron bought apple, tea, and nothing else.’ br (a) B: NO , IX -2 WRONG . AARON BUY WHAT APPLE , TEA , BANANA . (exhaustive) ‘No, you are wrong. Aaron bought apples, tea, and bananas.’ br (b) B: # NO , IX -2 WRONG . AARON BUY WHAT TEA , BANANA . ‘No, you are wrong. Aaron bought tea and bananas.’ (non-exhaustive) Thus, I agree with Wilbur (1996) that QAP is a focus construction. Nevertheless, a QAP in HKSL is different from a wh-cleft. In particular, while QAPs can only be specificational, wh-clefts can either be specificational or predicational (Higgins 1973). Specificational wh-clefts specifies a value of a variable, and predicational wh-clefts merely says something more about the variable without specifying its value. To illustrate, in (35 a), the content after the copula specifies the variable of ‘the person in charge’; and in (35 b), although we know that the person in charge is smart, who the person is is still unknown. (35) [English] (a) (b) The person who is in charge is Aaron. The person who is in charge is smart. (specificational) (predicational) By contrast, in a QAP in HKSL, the A-clause can only be specificational but not predicational, as in (22 b), repeated below as (36), and (37). This is similar to ASL (Hoza et al. 1997) and LSF (Hauser 2018). br (36) ACTIVITY IX RESPONSIBLE ARRANGE WHO , AARON /* SMART. ‘Who is in charge of organizing the activity? Aaron/*Smart.’ br (37) (a) IX -3 EAT MEAT NOT WHY , WANT ENVIRONMENTALLY - FRIENDLY. ‘The reason why he does not eat meat is (that he) wants to (be) environmentally-friendly.’ br (b) #IX -3 EAT MEAT NOT WHY, AFFECT MOM TROUBLE - MINDED . ‘The reason that he doesn’t eat meant bothers his mom.’ There are more tests supporting the point that QAPs are not wh-clefts. For instance, similar to LIS, NGT, LSF, and RSL (Branchini 2014; Kimmelman & Vink 2017; Hauser 2018; Khristoforova & Kimmelman 2021), different wh-phrases can be used in Q-clause. This is not the case for wh-clefts (38). (38) (a) [English] * Which book he bought was The Fountainhead. Question answer pairs in HKSL (b) 115 AARON LIKE BLUE RED WHICH , BLUE . ‘The color that Aaron prefers between blue and red is blue.’ Also, the Q-clause and the A-clause are non-reversible (39), which is different from wh-clefts. Moreover, unlike wh-clefts, the Q-clause in a QAP in HKSL can be a polar question (similar to LSF, NGT, RSL, and SASL) (41). br (39) (a) IX -1 DISLIKE WHAT , POSS -3 TIE . ‘What I dislike is his tie.’ br (b) *POSS -3 TIE , IX -1 DISLIKE WHAT. (c) *POSS -3 TIE , IX -1 DISLIKE WHAT. br (40) [English] (a) (b) (41) (a) His tie is what I dislike. What I dislike is his tie. [English] *Whether Eva went to school is she did/yes. br (b) EVA GO SCHOOL , YES . ‘Eva went to school.’ 4 Concluding remarks In this study, I discuss syntactic and semantic properties of Question Answer Pairs (QAPs) in Hong Kong Sign Language (HKSL). Syntactically, I show that QAPs in HKSL display mixed clausal properties, suggesting an intermediate status between a single syntactic unit and a discourse-level question and answer combination. Semantically, I show that the Q-clause in a QAP shares certain similarities with a topic, but they are different in nature. In a QAP, an A-clause has to be a congruent response to the question raised by the Q-clause. This can be either semantically congruent or pragmatically congruent, which I suggest is similar to regular interrogatives and their corresponding answers. QAPs with an indirect but congruent A-clause also show that the Q-clause and its corresponding A-clause do not always denote identical propositions as Caponigro & Davidson (2011) propose. I also show that QAP is a focus construction. I argue that only when it expresses constractive focus, must the A-clause be exhaustive. Studying the properties of QAPs helps us understand the information packaging strategies in sign languages. The clause final position in sign languages has been argued to be phonologically salient (Crasborn & Van Der Kooij 2012; Wilbur 1996; Wilbur 1994). It also carries focus, and quantificational elements like wh-phrases and negator (Lee 2006; Neidle et al. 2000; Lillo-Martin & de Quadros 2005; Kimmelman 2012; Petronio & Lillo-Martin 1997). In QAPs, placing the A-clause to the right-periphery of a QAP, and locating the wh-phrase at the right-periphery of the Q-clause could be relevant to using the right-periphery position to express salient information. 116 LINGHUI EVA GAN In spoken languages, constructions that play a similar pragmatic function role as a QAP is occasionally attested. As an anecdotal evidence, in a cooking tutorial I watched, to raise the audience’s attention, the Cantonese speaking chef frequently used rhetorical questions with an answer when introducing each cooking step. The question arises: why are QAPs so commonly used across sign languages in a rather wide range of discourse contexts, whereas in spoken languages, the use of rhetorical questions seems to be less wide-spread? As a preliminary stipulation, this could be related to the lack of plasticity of sign languages that Wilbur (1996) proposes. She argues that ASL and many sign languages are [-plasticity], and the intonational prominence is fixed in the clause-final position. As a result, a syntactic means must be used to make the salient information fall under the prominence. Following this, sign languages commonly resort to QAPs as one of the syntactic constructions to convey salient information. By contrast, spoken languages vary in being [± plasticity], and there are other ways in expressing salient information than adopting a particular construction. This is relevant to the fact that sign languages are in general relatively young. Thus, unlike many spoken languages, many sign languages have not yet develop complex constructions like clefts or free relatives, so they adopt a combination of small clauses instead (Progovac 2015). References Baker, C., & D. Cokely. 1980. American Sign Language. MD: Silver Spring: T.J. Publishers. Baker-Shenk, C. 1985. The facial behavior of deaf signers: evidence of a complex language. American annals of the deaf 130. 297–304. Branchini, C. 2014. On Relativization and Clefting: An Analysis of Italian Sign Language. Berlin: De Gruyter Mouton & Ishara Press. Caponigro, I., & K. Davidson. 2011. Ask, and tell as well: QuestionAnswer Clauses in American Sign Language. Natural Language Semantics 19. 323–371. Crasborn, O., & E. Van Der Kooij. 2012. On the weight of phrase-final prosodic words in a sign language. Sign Language & Linguistics 15. 11–38. Davidson, K., I. Caponigro, & R. Mayberry. 2008a. Clausal Question-Answer Pairs: evidence from ASL. In Proceedings of the 27th West Coast Conference on Formal Linguistics, ed. by N. Abner & J. Bishop, 108–115, Somerville, MA. Cascadilla Proceedings Project. Davidson, K., I. Caponigro, & R. Mayberry. 2008b. The semantics and pragmatics of Clausal Question-Answer Pairs in ASL. Semantics and Linguistic Theory 18. 212. E. Kiss, K. 1998. Identificational focus versus information focus. Language 74. 245–273. Publisher: Linguistic Society of America. Gan, L. E. 2022a. Scope relation and structure hierarchy in Hong Kong Sign Language (HKSL): exploring ditransitives. In Proceedings of the 30th Conference of the Student Organization of Linguistics in Europe, ed. by A. Holtz, I. Kova, & R. Puggaard-Rode, 201–219, Leiden. Published by Leiden University Centre for Linguistics. Gan, L. E., 2022b. Syntactic structure of argument wh-questions in Hong Kong Sign Language. Groenendijk, J., & M. Stokhof. 1984. Studies on the Semantics of Questions and the Pragmatics of Answers. Amsterdam: University of Amsterdam dissertation. Hamblin, C. L. 1973. Questions in Montague English. Foundations of Language 10. 41–53. Hauser, C. 2018. Question-answer pairs: the help of LSF. In Proceedings of the Formal and Experimental Advances in Sign Language Theory Conference, ed. by V. Kimmelman, volume 2, 44–55. Question answer pairs in HKSL 117 Herrmann, A., S. Proske, & E. Volk. 2019. Question-Answer Pairs in sign languages. In Questions in Discourse, 96–131. Brill. Higgins, F. R. 1973. The pseudo-cleft construction in English. Massachusetts Institute of Technology dissertation. Hoza, J., C. Neidle, D. MacLaughlin, J. A. Kegl, & B. Bahan. 1997. A unified syntactic account of rhetorical questions in American Sign Language. In Syntactic Structure and Discourse Function: An Examination of Two Constructions in American Sign Language, ed. by C. Neidle, D. MacLaughlin, & R. G. Lee, 1–23. Boston: Boston University. Huddlestone, K. 2021. Negation and polar questionanswer clauses in South African Sign Language. Sign Language & Linguistics 24. 63–86. Publisher: John Benjamins. Khristoforova, E., & V. Kimmelman. 2021. Question-answer pairs in Russian Sign Language: a corpus study. In The FEAST journal, volume 4, 101–112. Kimmelman, V. 2012. Doubling in RSL and NGT: Towards a unified explanation. MIT Working Papers in 2012 65. 57–81. Kimmelman, V., & L. Vink. 2017. Question-Answer Pairs in Sign Language of the Netherlands. Sign Language Studies 17. 417–449. Lee, Y. F. J., 2006. Negation in Hong Kong Sign Language. Lillo-Martin, D., & R. M. de Quadros. 2005. Focus Constructions in American Sign Language and Língua de Sinais Brasileira. In TISLR 8 Proceedings. Neidle, C., J. Kegl, B. Bahan, D. MacLaughlin, & R. G. Lee. 2000. The Syntax of American Sign Language: Functional Categories and Hierarchical Structure. Cambridge, MA: The MIT Press. Neidle, C., J. A. Kegl, B. Bahan, D. Aarons, & D. MacLaughlin. 1996. Rightward wh-movement in American Sign Language. In Rightward Movement, ed. by D. Beerman, D. LeBlanc, & H. van Riemsdijk. Philadelphia: John Benjamins. Neidle, C., D. MacLaughlin, & R. G. Lee. 1998a. Wh-Questions in ASL: A Case for Rightward Movement. Technical report. Neidle, C., D. MacLaughlin, R. G. Lee, B. Bahan, & J. Kegl. 1998b. The Rightward Analysis of wh-Movement in ASL: A Reply to Petronio and Lillo-Martin. Language 74. 819–831. Petronio, K., & D. Lillo-Martin. 1997. Wh-movement and the position of Spec-CP: Evidence from American Sign Language. Language 73. 18–57. Progovac, L. 2015. Evolutionary Syntax. Oxford Studies in the Evolution of Language. Oxford, New York: Oxford University Press. Roberts, C. 1996. Information Structure: Towards an integrated formal theory of pragmatics. Semantics and Pragmatics 5. 1–69. Stickles, E. 2012. Focus constructions in ASL: evidence from pseudoclefting and doubling. University of Pennsylvania Working Papers in Linguistics 19. 1–9. Sze, F. 2008. Topic Constructions in Hong Kong Sign Language. Bristol University dissertation. Sze, F. 2011. Nonmanual markings for topic constructions in Hong Kong Sign Language. Sign Language & Linguistics 14. 115–147. Valli, C., & C. Lucas. 2000. Linguistics of American Sign Language: An Introduction. Gallaudet University Press. Wilbur, R. B. 1994. Foregrounding structures in American Sign Language. Journal of Pragmatics 22. 647–672. Wilbur, R. B. 1996. Evidence for the function and structure of wh-clefts in American Sign Language. International review of sign linguistics 1. 209–256. Proceedings of CLS 58 (2022), 119-139 © Chicago Linguistic Society 2023. All rights reserved. 119 VRK Mutation in West Papua: Phonological Variation Across Time and Space Emily Gasser Swarthmore College 1 Introduction A general expectation of productive phonological rules is that they will target natural classes of sounds, apply in a unified set of environments, and have some phonetic motivation for the change they incur. These expectations motivate the proposal of theories which posit naturalness or groundedness in phonology (Ohala 1974, Ohala 1983, Donegan & Stampe 1979, Browman & Goldstein 1992, Archangeli & Pulleyblank 1994, Hayes 1999, Hayes et al. 2004). Regardless of whether ‘rules’ are defined as SPE-style ordered processes or the descriptive generalizations which underlie an Optimality Theoretic analysis, this expectation holds: in a particular environment, we see a change in uvulars, or high vowels, or voiceless stops, in a way that follows logically from their environment and is repeated in its broad strokes in language after language. Whether we see phonetic naturalness as intrinsic to the phonological grammar or arising from acquisition or other pathways of diachronic change, the rules we describe overwhelmingly have some level of observable phonetic motivation in the class which they target, the environments they describe, and the change which they effect. In some cases, however, these expectations go unmet. This paper describes VRK mutation, an unnatural (or ‘crazy’) rule in the sense of Bach & Harms (1972), which is found in at least 20 of the 43 or so languages of the South Halmahera-West New Guinea (SHWNG) branch of Austronesian. In canonical VRK mutation, the segments /B/ (orthographic v), /r/, and /k/ behave differently from the rest of the phoneme inventory when they appear as the second member in a cluster, leading to surface [mb], [nd], or [Ng], respectively. /B, r, k/ do not form a natural class in these languages, and in at least some cases a phonetic justification is also difficult to find, as in the /rr/ → [nd] change found in Wamesa [wad]. This pattern differs significantly from superficially similar ones found in e.g. Southern Paiute (Uto-Aztecan, Sapir 1930), Fore (Gorokan, Scott 1978) and the Sestu Campidanian dialect of Sardinian (Romance, Bolognesi 1998), as the voiceless stops /p/ and /t/ are not involved, nor is [G] present in most SHWNG varieties. The individual changes involved in VRK mutation are not cross-linguistically unusual — voicing of /k/ to [g] after a nasal is as standard a phonological rule as one might hope to find — but taken together in the larger phonological context of these languages they form a noteworthy confluence of effects. That sounds can pattern phonologically in ways that fail to align with phonetic classification has long been recognized (e.g. Sapir 1925). A number of unnatural processes such as Breton mid-vowel height/length distinction (Anderson 1981), 120 EMILY GASSER Icelandic velar fronting (ibid.), Zuni hyperpalatalization (Buckley 2000), and Polish back vowel alternations (Sanders 2003) have been described, and the processes that allow them to arise and be learned are well-studied (Wilson 2006, Hayes & White 2015, White 2017). SHWNG VRK mutation is not novel in type, but in extent: it is more widespread and more persistent than other documented unnatural rules, appearing in 20 or more languages whose common ancestor was spoken around 3000 years ago. These sorts of unnatural patterns are of interest beyond the confines of the languages in which they appear. They show us the limits of what phonological systems can do, and in doing so challenge theories of (morpho-)phonology which posit grounded and/or innate rules. Because they are not grounded in (synchronic) phonetic pressures, they give the lie to any theories which demand that phonology arise strictly out of such a basis. On the other hand, they are clearly not universal — few phonologists would posit a universal constraint motivating palatalization of velars before /a/, or grouping /B, r, k/ as a natural class to the exclusion of /s/, /m/, /p/, etc. They also shed light on what learners of a language may acquire: if there is indeed a bias towards learning natural processes, these processes show that it cannot be absolute. The VRK case, given its widespread and persistent nature, shows that these rules are not as diachronically fragile as we might expect, and can persist through time and multiple language splits more or less intact, successfully acquired and maintained over the generations. This paper looks at VRK mutation from three angles. §2 describes its distribution within SHWNG, and describes its instantiation in a few of the betterdocumented languages. Outside of Wamesa, VRK mutation has not previously been recognized as a pattern either within or across languages; where it is described at all it is as a series of individual rules among others, despite the unity of outcomes for these segments in contrast to others. §3 gives an overview of the origins of the pattern and its diachronic development from earlier *b, *d, and *g or *k. §4 concludes. 2 Typology and distribution The South Halmahera-West New Guinea languages, spoken on and around the southern part of Halmahera and northwestern New Guinea in far eastern Indonesia, are highly under-documented, especially as compared to some other subgroups of Austronesian, and most are endangered. SHWNG is a sister to the better-known Oceanic subgroup within Eastern Malayo-Polynesian, and includes 40–45 languages (see Gasser et al. 2023, Forthcoming). A map of the SHWNG languages of northwest New Guinea is given in Fig. 1. VRK mutation appears only in languages spoken in and around Cenderawasih Bay1 and the Mamberamo Delta – that is, those languages which do not belong to the Raja Ampat-South Halmahera (RASH) branch of SHWNG, which I will refer to collectively as Eastern SHWNG (ESHWNG). This is for now a geographical, not genetic, grouping, though there may well be a subgrouping argument to be made. 1 Biak is spoken in some villages in the Raja Ampat islands, but its largest speaker populations are on Biak Island in Cenderawasih Bay, and it clearly does not subgroup with the Raja Ampat languages. VRK mutation in West Papua 121 Figure 1: The SHWNG languages of northwest New Guinea. The languages discussed here are: Wamesa (1), Wooi (2), Ambai (3), Biak (4), Waropen (5), Warembori (6). Adapted from an unpublished SIL map, with additional information from Arnold (2018), Eberhard et al. (2022), Kamholz (2014), Remijsen (2001) and Sawaki (2017). While some version of VRK mutation is attested in at least 20 languages within ESHWNG — the lack of documentation of many varieties precludes a more exact count — its productivity and instantiation varies widely. In some languages, such as Wamesa (Gasser 2014), VRK mutation is synchronically active and applies in a wide range of environments. In others its application is more limited, only occurring following nasal segments, in compound words, or as a fossilized morpheme structure constraint. The exact set of sounds which trigger mutation also varies somewhat: while /B/ and /r/ participate in all cases, some languages drop /k/ or add triggering segments such as /w/, /F/, or /G/. While a full accounting of VRK mutation’s realizations in all of the ESHWNG languages is beyond the scope of this paper, the remainder of this section describes how it appears in five of the most well documented SHWNG languages: Wamesa, Ambai, and Wooi from the Yapen subgroup, the Biakic language Biak, and Waropen, considered a primary subgroup of SHWNG. A SHWNG family tree, showing our current best understanding of the genetic relationship of these languages, is given in Figure 2. 2.1 VRK in Wamesa Wamesa is spoken along the southwest coast of Cenderawasih Bay and the southern Bird’s Head by about 6000 speakers. Data from Wamesa comes from Henning et al. (1991), Gasser (2014, 2015), and my ongoing fieldwork on the language. The Wamesa phoneme inventory is given in Figure 3. In most positions, /b/ and /B/ are contrastive, as are /r/ and /d/, while [k] and [g] are not, outside of loans from Dutch and Indonesian. All Wamesa phonemes may appear word-initially and 122 EMILY GASSER Figure 2: The internal structure of SHWNG, with representative languages from each branch. The languages discussed here are given in bold. VRK mutation is found in most languages outside of the Raja Ampat-South Halmahera (RASH) branch. Adapted from Kamholz 2023. intervocalically; words are mostly vowel-final with some exceptions. The Wamesa inventory is fairly typical for eastern SHWNG (see Gasser et al. 2023). Bilabial Alveo-dental Velar Nasal m n N Plosive pb td k Fricative B s Tap/Trill r Figure 3: The Wamesa phoneme inventory Only two types of CC clusters are allowed in the language. The first type consist of a nasal [m, n, N] followed by a homorganic voiced stop [b, d, g]. These may appear morpheme-internally, as in the examples in (1), or across morpheme boundaries. These clusters always span a syllable boundary. (1) Wamesa morpheme-internal ND clusters a. [sembie] ‘moon’ b. [anda] ‘crown of thorns sea star’ c. [aNgadi] ‘coconut’ These ND clusters are in fact the only environment in which [g] surfaces in Wamesa (excepting loans); [N] is always followed by a [g] except when root-initial. The second type of cluster consists of any consonant directly preceded or followed by a glide, [j] or [w]. These glides are derived from the underlying high vowels /i/ and /u/, which may undergo post-lexical reduction when consonant-adjacent in a sequence of vowels — that is, when reduction does not remove the syllable nucleus. These will not be further discussed here. VRK mutation in West Papua 123 Other types of clusters do not appear morpheme-internally in Wamesa. They do arise across a morpheme boundary, most often via affixation of verbs, predicative adjectives, and inalienably possessed nouns. When this occurs, the general process is deletion of the first member of the cluster (C1 ), as in the examples in (2). C1 in these clusters is always /r/, /t/, or /n/, as these are the only affix-final consonants in the language: /r/ in the dual number subject agreement markers, /t/ in the plural subject markers and the applicative, and /n/ in the (exceedingly rare) causative. (2) Wamesa Cluster Simplification a. /tur-masoi/ → [tumasoi] ‘1DU . INCL buy’ b. /met-samuai/ → [mesamuai] ‘2PL collect’ c. /bu-it-pera/ → [buipera] ‘2SG uses (it) to cut’ There are two exceptions to the normal deletion process in clusters. First, when a root begins with a voiced stop /b/ or /d/, the preceding C will surface as a nasal homorganic to the root-initial segment (3). Voiced stops are of low frequency in word-initial position in general; within the word classes that may take agreement affixes, my data contains only five which are /b/-initial and one which is /d/-initial.2 (3) Affixation of voiced stop-initial words a. /tur-baba/ → [tumbaba] ‘1DU . INCL are big’ b. /met-batojar/ → [membatojar] ‘2PL are thick’ c. /set-deriasi/ → [senderiasi] ‘3PL . HUM3 are near’ It is the second exception to the deletion rule, VRK mutation, which is of primary interest here. In a C1 C2 cluster, if C2 is one of the set /B, r, k/, then the cluster surfaces as as a homorganic ND cluster at the place of articulation of C2 . This occurs regardless of the identity of C1 (though as mentioned above, C1 in practice is always a /t/, /r/, or a nasal); even a sequence of two identical /r/s will surface as [nd], as in (4b). This can take place at the prefix-root boundary (4) or between prefixes (5), though not over a stem-clitic or word boundary. At the juncture between constituents of a compound word, VRK mutation gives evidence for a semantically empty linking morpheme which triggers the process (6), historically an unspecified nasal. (4) Affix-root boundary a. /met-Bori/ → [membori] ‘2PL buy’ b. /tur-rama/ → [tundama] ‘1DU . INCL come’ c. /bu-it-kore/ → [buiNgore] ‘2SG use (it) to hold’ d. /j-on-raBesie/ → [jondaBesie] ‘1SG cause (them) to separate’ (5) Affix-affix boundary a. /met-Be-marisiani/ → [membemarisiani] ‘2PL are spicy’ 2 Compare this for example to 22 conjugating /p/-initial words and 38 /t/-initial ones, out of a total sample of 1187 lexical roots. 3 HUM = ‘human’. 124 EMILY GASSER b. /sur-Be-marisiani/ → [sumbemarisiani] ‘3DU are spicy’ c. /i-it-Be-kuas-e/ → [jimbekuase] ‘1SG use (it) to paint’ (6) Within compounds a. /a+C+Besie/ → [ambesie] ‘delicious’ (lit. ‘eat well’) b. /ru+na+C+rau/ → [runandau] ‘hair’ (lit. ‘head-LOC-leaf’?) c. /dia+C-kariria/ → [diaNgariria] ‘crocodile’ (lit. ‘nasty fish’) Mutation of /B/ and /r/ is exceptionless, but the behavior of /k/ is variable. When C2 is /k/, normal C1 deletion will sometimes apply in place of VRK mutation with variation between speakers, lexical items, and utterances. Thus a form like /setkubira/ ‘3PL . HUM-bathe’ is attested in my data as both [seNgubira], with VRK mutation, and [sekubira] with normal C1 deletion. This may be a shift towards regularization of forms with /k/, perhaps as a symptom of language endangerment; as Wamesa is spoken less and less in daily life its less-regular patterns are falling out of use. The fact that older sources (Saggers 1979, Ramar et al. 1983, Henning et al. 1991) do not describe variability in the behavior of /k/ suggests that his is the case. Mutation of /k/ may be particularly vulnerable compared to the other segments, as the resulting [Ng] cluster is far more limited in its distribution than the analogous [mb] and [nd] clusters. If /k/ is removed from the list of mutating consonants, the remaining two do form a natural class, comprised of all of the Wamesa voiced continuants, perhaps contributing some additional pressure in support of the change. The behavior of /k/ in this environment is in some other related languages as well. Viewed as a collection of individual rules, Wamesa VRK mutation seems perfectly phonologically natural. It is cross-linguistically common for segments like /B/, /r/, and /k/ to surface as voiced stops after nasals, and place assimilation within clusters is similarly frequent. Nasalization of another consonant before /b/ and /d/ is perhaps less ubiquitous, but still not particularly noteworthy. In this view, there is nothing remarkable about the VRK pattern except perhaps its appearance across so many SHWNG languages. I argue, however, that this view misses the larger generalization. When taken together in the context of the full phonological system, and particularly within a phonological framework such as Optimality Theory which does not incorporate ordering, VRK mutation is indeed unexpected. /B/, /r/, and /k/ do not form a natural class, and incorporating /b/ and /d/, arguably also part of the pattern, does not create one. It is possible to break VRK mutation down into component rules: one turning the voiced continuants /B/ and /r/ into stops, one voicing /k/, and one turning C1 into a nasal and assimilating its place to that of C2 . However, this overlooks the larger generalization that the triggering C2 segments act exceptionally compared to the rest of the phoneme inventory, including for example all voiceless segments other than /k/, where cluster simplification occurs through C1 deletion. Splitting the pattern in two treats this unity of behavior as a mere coincidence. Similarly, calling Wamesa VRK mutation a purely morphological process also seems to miss the generalization, for two reasons. First, VRK mutation in Wamesa holds morpheme-internally; no sequences surface anywhere in the language which violate it. Second, it’s not the case that a certain synchronically arbitrary set of prefixes or roots trigger mutation; rather, it takes place in all cases where /B/ or /r/ VRK mutation in West Papua 125 follows another C. Even with /k/, where the application of mutation is variable, its appearance does not depend on the particular morphemes involved, but appears to be random, with the same word unpredictably participating or not within utterances by the same speaker. This is the signature of a slowly unraveling phonological rule, not a morphological one. Looking for a phonetic motivation introduces other problems. If we assume that nasalization of C1 precedes stopping/voicing of C1 then it is natural and common for /B/, /r/, and /k/ to become stops after a nasal, but it is unclear why C1 should nasalize before those segments and not any other fricative or voiceless stop (i.e. /s/, /p/, /t/). If the order of operations is reversed, we can posit that /B, r, k/ become [b, d, g] first, followed by nasalization of C1 before a voiced stop, a reasonable rule. However, this does not explain why only these three segments become voiced stops following another (not yet nasalized) C: /k/ → [g] / N __ is a natural rule, but /k/ → [g] / t __, as in (4c), is not. If we forgo rule ordering and use a simultaneous framework such as Optimality Theory, we lose the chicken and egg problem just described, but end up with something unsatisfying and stipulative, as the constraints must target /B, r, k, b, d/ to create an ND output from an illicit cluster, while all other segments trigger C1 deletion. The specifics of the application also pose difficulties for phonetic grounding; cf. the /rr/ → [nd] change in (4b). One motivation which has been suggested is the Syllable Contact Law (Venneman 1988), which prefers falling sonority over a syllable boundary. This fails, however, in cases where the underlying segments are /r/ or a nasal followed by /B/ or /k/ (cf. example 5b); here, the difference in sonority is actually reduced by VRK mutation. 2.2 Ambai Ambai is spoken along the southeastern coast of Yapen Island and on a few small nearby islands. Ambai data comes from Silzer’s (1983) dissertation and Grace (1955-56) as cited in the ABVD (Greenhill et al. 2008). Ambai’s phoneme inventory differs minimally from Wamesa’s: it adds the voiceless bilabial fricative /F/, and removes /N/. Surface [N] appears only as an allophone of /m/ and /n/. Determining the exact set of VRK phones in Ambai is fraught. The sounds /r/, /k/, and /F/ certainly take part in mutation, surfacing as [d], [g], and [p], respectively. Silzer’s orthographic conventions conflate with /B/ with [w] in most cases, often making it impossible to tell which sound is being represented. Silzer’s featurebased formulation of the VRK mutation rule should include /B/ in the VRK set, but all of his examples are spelled with w. Given cognate forms found in other Yapen languages, it appears that both /B/ and /w/ are part of the Ambai VRK set, and merge to [b] in this environment. Because of uncertainty of the underlying sounds involved in any given form, I follow Silzer’s representation and use /w/ where he does. I also disregard his (somewhat questionable) assertion that all [w] comes from underlying /u/, and treat the /w/ as underlying, though all examples could be straightforwardly re-written with the /w/ as intermediate in the derivation. Ambai is somewhat more permissive than Wamesa with regards to its surface clusters: where Wamesa only allows homorganic nasal plus voiced stop clusters, the voiceless obstruents [p], [t], and [s] may also surface as C2 after a homorganic nasal in Ambai (7). 126 (7) EMILY GASSER a. [ampa] ‘shell armband’ b. [embai] ‘moon’ c. [antu] ‘child’ d. [rando] ‘banana’ e. [ansun] ‘clothing’ f. [waNgori] ‘crocodile’ When illegal clusters arise over a morpheme boundary, Silzer lists two sets of processes which may apply to them. Morpheme-final segments in Ambai are limited to the set /p, t, s, r, m, n/. When C1 is non-nasal, it will simply delete, leaving C2 intact (8). (8) /tat-madu/ → [tamadu] ‘1PL . INCL-speak’ Unlike in Wamesa, this default deletion pattern also applies when C2 is a VRK phone if C1 is non-nasal, shown in (9) for a subset of the verbal agreement paradigm. (9) Ambai verbal subject agreement: ‘to recline’ ‘to walk’ ‘to bite’ ‘to wash’ wata ra kiri feram Prefix /wata/ /ra/ /kiri/ /Feram/ 1DU . INCL /tur-/ [tuwata] [tura] [tukiri] - 3PL /et-/ [ewata] [era] [ekiri] [eFeram] Morpheme-final nasals behave differently from the non-nasal phonemes in the same position when introduced into cluster. A nasal C1 will assimilate to the place of articulation of the following C2 . If that C2 is one of the VRK phones, mutation occurs. As described above, /F, w, B, r, k/ in this environment will surface as [p, b, b, d, g], respectively. Mutation can be seen to apply across a prefix-root boundary (10), most clitic-root boundaries (11), and a root-root boundary within a compound (12). (10) a. /man-waru/ → [mambaru] ‘across’ b. /an-FeFe/ → [ampeFe] ‘cause’ c. /man-ru/ → [mandu] ‘two (animate)’ (11) a. /uanan=Fo/ → [wanampo] ‘the wind’ b. /rotan=wa/ → [rotamba] ‘the bag’ (12) /man/ ‘male’ + /rirau/ ‘married’ → [mandirau] ‘married man’ Note that similar post-nasal stopping and voicing does not occur with voiceless stops other than /k/, or with the voiceless fricative /s/, which might be expected to pattern with /F/. This is seen in surface nasal + voiceless obstruent clusters in words such as those in (7). Unlike in Wamesa, the behavior of /k/ is exceptionless in Ambai. VRK mutation in West Papua 127 2.3 Wooi Wooi (Sawaki 2017, p.c.; Nikolaus Himmelmann p.c.), spoken on the far western tip of Yapen and on nearby Num Island, takes the Wamesa phoneme inventory and adds /ñ/, /c/, and /h/. The VRK phones are /B/ and /r/, without /k/. Wooi allows not just homorganic nasal+stop clusters but also heterorganic ones, as in (13a), though homorganic NC are the most common (Sawaki 2017). Other types of clusters also occur more rarely, as in (13b)–(13c). Despite this phonotactic permissiveness, [B] never appears following another C in a cluster, and [r] does so only in (rare) [mbr] clusters (13d). Root-internally, [k] does appear following [N] (13e), unlike in the other languages described here. (13) a. [ramdempe] ‘tomorrow’ b. [naknak] ‘breadfruit’ c. [karpiapia] ‘rotten’ d. [kambrei] ‘hole’ e. [aNkati] ‘coconut’. Once again, the behavior of clusters is most easily observable in subject-verb agreement. In Wooi, a prefix-final C in will become nasal and assimilate to the place of articulation of a following voiceless stop, or delete before another nasal.4 On /h/-initial verbs, the prefix-final C deletes and the /h/ surfaces as [s]. (14) Wooi verbal agreement I: ‘to eat’ ‘to cut’ ‘to sit’ ‘to call’ ang perang mahoy ha prefix /aN/ /peraN/ /mahoi/ /ha/ 1SG /j-/ [jaN] [peraN] [mahoj] [ha] 3DU /hur-/ [huraN] [humperaN] [humahoj] [husa] 1PL . INCL /tat-/ [tataN] [tamperaN] [tamahoj] [tasa] Where the root begins with /B/ or /r/, however, mutation occurs. Wooi takes the pattern of reduced /k/ participation to its logical endpoint and does not include it in the mutation pattern at all; instead /k/ has fully joined the default pattern and acts as any other voiceless stop, triggering nasalization and place assimilation in the preceding C but remaining itself unchanged, like /p/ and /t/. (15) Wooi verbal agreement II: ‘to be hungry’ 4 ‘to go’ ‘to speak’ vavihni ra kavio prefix /BaBihni/ /ra/ /kaBio/ 1SG /j-/ [BaBihni] [ra] [kaBio] 3DU /hur-/ [humbaBihni] [hunda] [huNkaBio] 1PL . INCL /tat-/ [tambaBihni] [tanda] [taNkaBio] No verb stems beginning with voiced stops are included in the available data. 128 EMILY GASSER Mutation also does not occur in compounds in Wooi, except where it is fossilized in forms like [riuandauN] ‘hair’. Wooi is different from Ambai and Wamesa, as the targets of VRK mutation do form a natural class: the voiced continuants. In that sense, Wooi VRK mutation is a less unnatural process than that the other languages, though again the /r-r/ → [nd] change, as in [hunda] ‘3DU-go’ in (15), remains difficult to motivate phonetically. 2.4 Biak Biak, the largest and most geographically widespread of the eastern SWHNG languages, is the subject of van den Heuvel’s (2006) dissertation, which describes the dialect spoken in and around the village of Wardo in the western part of Biak Island. Wordlists from other dialects come from Gasser (2016), van Hasselt & van Hasselt (1947), Than et al. (2011), and Xavier Bach (p.c.). The Biak consonant inventory is given in Figure 4. It is very similar to the inventories found in the Yapen languages, most notably differing in its inclusion of labio-dental /f/. Biak /t/ is rare, appearing mostly in loans, due to a *t > /k/ sound change. The resulting /k/ is realized variably as [k] or [P] depending on dialect, speech rate, and other variables. An acute accent in transcriptions denotes a long vowel. The set of VRK phones in Biak is /B, r, k/. Bilabial Labio-dental Nasal m n Plosive pb Fricative B Trill Approximant Alveolar Palatal Velar (t) d f k s r w j Figure 4: The Biak consonant inventory (van den Heuvel 2006) Biak is far more permissive in its consonant cluster phonotactics than any of the Yapen languages. Earlier in the history of the Biakic family, unstressed vowels were lost from many initial syllables (Blust 1978, Kamholz 2014), leading to a wide range of acceptable surface clusters. van den Heuvel (2006: 39) lists 53 clusters which may surface in word-initial position and 38 which appear word-finally. These include both homo- and heterorganic pairs, as well as many which violate the Sonority Sequencing Principle (Kenstowicz 1994). The contrast with Yapen cluster phonotactics is striking. Where many of those languages entirely rule out tautosyllabic clusters without glides, the Biak lexicon includes monomorphemic surface forms including those in (16), in addition to forms with the fossilized prefixes /k/, /m-/, and /r-/, as in (17), and the productive inflectional affixes, as in (18). A similarly wide range of intervocalic clusters also appear. VRK mutation in West Papua (16) a. [pdef] ‘onwards’ b. [pokm] ‘able’ c. [rmomn] ‘angry’ d. [fnder] ‘forget’ (17) a. [kpéf] ‘burst’ b. [msór] ‘angry’ (18) a. [n-babo] ‘3PL . INAN-new’5 b. [s-pon] ‘3PL . AN-first’ 129 The mutating phones may appear as C2 in Biak clusters, as in (19). (19) a. [wekurB] ‘heel’ b. [aBrui] ‘bean’ c. [srepk] ‘short’ With the huge range of clusters found in Biak, it is striking which clusters never appear: nasals followed by [B], [r], or [k]. With the possible exception of wordfinal [Nk], these do not surface morpheme-internally, and of all the clusters which may arise over a morpheme boundary, they are the only ones which undergo phonological change in the surface form. The only C-final verbal agreement prefixes in Biak are /s-/ ‘3PL . AN’, which does not trigger mutation, and /n-/ ‘3PL . INAN’, which does; examples therefore come largely from other morphological juxtapositions. VRK mutation is more complex and segment-specific in Biak than in the Yapen languages. Biak /B/ is realized as [b] after a nasal, with optional assimilation of preceding /n/ to [m]. This mutation is obligatory within a word, as in (20a), and “preferred but optional” across a word boundary within an Intonational Phrase (van den Heuvel 2006: 55), as in (20b)–(20c). (20) a. /n-Bark/ → [nbark] ‘3PL . INAN-live’ b. /rofan=ja bakn B<j>e=d-ja/ → [rofanja bakm bjedja] ‘the dog’s body’ c. /Be Bar Bar/ → [Be Ban Bar] ‘to both sides’ The realization of /Nr/ sequences vary depending on the place of the nasal and the class of the word it appears in. Word-internally, the /r/ surfaces, but an additional voiced stop homorganic to the nasal is inserted into the cluster, as in (21). Across a word boundary but within an IP, if both elements are verbs or nouns, a nasal + /r/ cluster will surface as [ndr], otherwise the result is [nd], as in (22). (21) (22) a. /i-mráne/ → [imbráne] ‘3SG-walk’ b. /n-ra/ → [ndra] ‘3PL . INAN-go’ /i-mrán ro di-ne/ → [imbrán do dine]6 ‘3SG-walk LOC place-this’ A similar split is seen with /r+r/ sequences across word boundaries within the same IP. Between verbs and nouns, this sequence is “optionally” realized as [rd], while between other classes of words there is “a very strong tendency” for /r+r/ to surface as [rd] (van den Heuvel 2006: 56-57). 5 AN = ‘animate’, INAN = ‘inanimate’. Focusing on the /N+r/ interaction, van den Heuvel transcribes this instance of the /mr/ cluster as surface [mr], without the intrusive [b]; given his earlier use of the same form in isolation to illustrate the /mr/ → [mbr] rule, I have adjusted the transcription here to reflect that alternation. 6 130 (23) EMILY GASSER a. /mniBr/ ‘bee’ + /rum/ ‘nest’ → [mniBrrum] ∼ [mniBrdum] ‘bee’s nest’ b. /su-frar rao/ → [sufrar dao] ‘3DU-run until’ Finally, Biak /k/ is realized as [g] after a nasal except when it is word-final; preceding /n/, but not /m/, assimilates to [N]. This process applies within words, including across morpheme boundaries, but not across word boundaries (24). The velars [N] and [g] only surface in Biak as allophones of /n/ and /k/, respectively. (24) a. /ánkar/ → [áNgar] ‘deceive’ b. /m-kák/ → [mgák] ‘afraid’ (fossilized prefix) c. /nk-ún kapírare/ → [Ngún kapirare] ‘1PL . EXCL take a baby with us’7 Mutation of /B/ and /r/ is visible in compounds such as [mambri] ‘hero’ from /man/ ‘male’ plus /Bri/ ‘angry’ and [Bandum] ‘inside’ from /Bar/ ‘side’ plus /rum/ ‘inner part’. It is unclear from the available data whether mutation of /k/ applies within a compound. The processes described here — post-nasal voicing, nasal place assimilation, stop insertion — are not in themselves particularly notable, but in the context of Biak in particular and SHWNG more broadly they stand out. First, it is precisely the mutating set found elsewhere in the subgroup, /B, r, k/, which trigger these changes. Second, no other sounds do. Within Biak, /n/ only undergoes place assimilation before /B/ and /k/, not other sounds (cf. [nfo] ‘3PL . INAN-full). Only /k/ voices after a nasal, and only /B/ and /r/ trigger stop insertion. The other fricatives and voiceless stops, /f, s, p, t/, fail to undergo any similar processes. As in e.g. Ambai, mutation only applies when the VRK phone appears after a nasal (or, in the case of /r/, another /r/). These rules involve the same elements that appear in the Yapen versions of VRK mutation, as they originated as a single pattern in an ancestor language, but Biak has complicated the process since the languages diverged. 2.5 Waropen Examples of VRK mutation abound in Waropen. Data in this section comes from Held (1942) and Carolyn Rowse’s draft English translation thereof, as well as Blust & Trussel (2010). The Waropen phoneme inventory unusual only in its inclusion of /G/. The phoneme transcribed as w by Held is in fact the bilabial fricative [B] (1942: 15). The full set of Waropen consonant contrasts is given in Figure 5. The mutating segments in Waropen are /B, r, k, G/. The only clusters in the language are homorganic ND, as in (25). Waropen [N] only surfaces before [g]. (25) 7 a. [Gamburu] ‘bamboo lance’ b. [mandana] ‘big’ c. [aNgoro] ‘crocodile’ Van den Heuvel transcribes the verbal agreement prefix /nk/ as surface [ng] rather than [Ng]. Given his earlier statement that such sequences “always” surface as [Ng] (2006: 55), it seems likely that this is a typo rather than a failure of place assimilation, but this is not crucial here. VRK mutation in West Papua 131 Bilabial Alveolar Nasal m n Plosive pb td kg Fricative B s G Tap/trill Velar r Figure 5: The Waropen consonant inventory (Held 1942)) Waropen has a range of strategies to deal with clusters which arise over a morpheme boundary. In some cases, C1 is simply deleted (26), much as in the default Wamesa pattern. In other instances different patterns of consonant change or coalescence take place (27). In all of these, the underlying cluster surfaces as a single C. This can take place between the elements of a compound word or at a stem-affix boundary. Only nasals and [r] may appear word-finally in Waropen; morphemefinal “supporting vowels” often drop when no longer word-final, creating a cluster. (26) (27) a. b. c. a. b. c. /iamoko + dere/ → [iamodere] ‘to flee inland’ /mun + tata/ → [mutata] ‘to cut up small’ /kik-usa-ki/ → [kikuki] ‘they cast them’ /fam/ + /fero/ → [fapero] ‘to strike dead, extinguish’ /isa + koda/ → [itoda] ‘to prick against’ /gogoko + Be/ → [gogope] ‘bad’ However, when the first element is a nasal and the second is /B/, /r/, /k/, or /G/, instead of a single C the result is a homorganic ND cluster (28). (28) a. b. c. d. /mun + Buara/ → [mumbuara] ‘to strike away’ /fam + riGara/ → [fandiGara] ‘to knock down’ /fama + ki/ → [faNgi] ‘he strikes them’ /masin + Ga/ → [masiNga] ‘the water’ The outcome of a particular cluster is not always predictable. While /sk/ surfaces as [k] in (26c) above, it can also surface as [t], as in (27b). Further, mutation to ND is evidently not the only possible outcome of a VRK phone after a nasal: Held gives one example where the result is [t], in [ianitu] ‘to look upwards’ from /ianim + ru/. Two other environments can give rise to surface ND clusters in Waropen. First, a preceding nasal will assimilate to the place of articulation of following /d/; /nd/ clusters surface intact (29).8 Second, in some instances an ND cluster will form over a morpheme boundary where no apparent additional C appears synchronically; likely due to a historically present segment whose effect has since fossilized (30). 8 Held gives one example where nasal + /b/ surfaces somewhat unexpectedly as [p]: [bipaBa] ‘old woman’, from /bin + baBa/, but does not otherwise discuss clusters with /b/ as C2 . 132 (29) (30) EMILY GASSER /fam + doma/ → [fandoma] ‘to drive in’ a. /ni + buino/ → [nimbuino] ‘middle, side’ b. /ano + ora/ → [aNgora] ‘to obstruct by shooting arrows’ All three methods of cluster resolution are present in the verbal agreement paradigm. Representative portions of the paradigm are given in (31) for a V-initial verb root, one beginning with a ‘normal’ consonant, and one with each of /B/ and /r/. Held gives no paradigms for /k/- or /G/-initial verb roots. We see C1 deletion in the 2SG and 3PL forms of all C-initial verbs and all plural forms of the /m/-initial verb, coalescence in the 1PL . INCL with /r/ and /B/, and VRK mutation in the 1PL . EXCL and 2PL forms. (31) Waropen verbal subject agreement: ‘to eat’ ‘to kill’ ‘to row’ ‘to go’ ano muna Bo ra prefix /ano/ /muna/ /Bo/ /ra/ 1SG /ra-/ [rano] [ramuna] [raBo] [rara] 1PL . INCL /k-/ [kano] [muna] [bo] [da] 1PL . EXCL /aN-/ [aNgano] [amuna] [ambo] [anda] 2PL /miN-/ [miNgano] [mimuna] [mimbo] [minda] 3PL /kik-/ [kikano] [kimuna] [kiBo] [kira] There is some evidence that mutation may occur in some cases across a word boundary. Held gives two examples with the irrealis particle [bo] ∼ [Bo]. When the preceding word ends in a vowel [Bo] surfaces, but after a nasal we see [bo] instead. (32) a. [nanimba Bo raBo na Momi] ‘If I were adult I would go to Momi.’ b. [ranienanam bo tokio] ‘If I were slack, we could well waste them.’ Though the picture is complicated by historical processes which have merged it with the mutation of other consonants, VRK mutation after nasals is nonetheless strongly present in Waropen. The unpredictable behavior of other clusters, and the occasional exceptions to the regular VRK pattern, are assumed to be lexicalized and not part of the productive phonology. 3 Diachrony This section gives a brief overview of the diachronic roots of VRK mutation. The most recent well-reconstructed ancestor of the SHWNG languages discussed here is Proto-Malayo-Polynesian (PMP), spoken about 5000 years ago (Blust 1999, 2009). Proto-Malayo-Polynesian is the ancestor language of all 1200 or so Austronesian languages spoken outside of Taiwan. VRK mutation in West Papua 133 The /B/ phoneme in the modern eastern SHWNG languages can be straightforwardly traced back to PMP *b, as in PMP *babaq > Wamesa [BaBa], Biak [BaB] ‘below’ (Kamholz 2014). (The reverse is not always true: *b sometimes also appears in ESHWNG as /b/.) ESHWNG /r/ derives from a merger of PMP *d, *l, and *R, as in PMP *dahun > Wamesa [rau], Biak [rám], Waropen [rana] ‘leaf’; PMP *lima > Wamesa [rim], Biak [rim], Waropen [rimo] ‘five’; PMP *Rumaq > Biak [rum], Waropen [ruma] ‘house’ (van den Heuvel 2006, Kamholz 2014). Modern /k/ is once again the odd man out: only three possible reflexes of PMP *g persist in ESHWNG, and it deletes in all of these. Likewise, *k was independently lost multiple times in various SHWNG languages. That leaves modern /k/ with no clear antecedent in PMP. Proto-SHWNG has yet to be reconstructed, but I suggest that after the deletion of PMP *g, new words with *g and *k entered the lexicon through borrowing or innovation. It is the reflexes of these new forms give us modern /k/. The RASH languages show /p/ and /b/ from *b and /l/, /r/, or ∅ from *d, *l, *r, and *R (Kamholz 2014), explaining the lack of VRK mutation in those varieties. A key intermediate stage in the development of VRK mutation is preserved in Warembori (Donohue 1999), spoken in the Mamberamo River delta to the east of most other SHWNG languages. Warembori differs from the other varieties discussed here in that [b]/[B] and [d]/[r] are not contrastive: the voiced stops appear word-initially and post-nasally, and lenite to [B] and [r] intervocalically, as in (33). The nasals undergo place assimilation. (33) a. /e-bodo-do/ → [eBororo] ‘(it is) my tongue’ b. /waden+bo-do/9 → [waremboro] ‘(it is a) ‘mouth of a river’ c. /dodo+dan-do/ → [dororando] ‘(it is) rainwater’ Warembori /k/ also voices to [g] after a nasal, as in (34). This is unique to /k/; other voiceless segments remain so after a nasal. Neither [N] nor [g] surfaces in Warembori except as the realization of an /Nk/ cluster. (34) /waden-kaindu/ → [wareNgaindu] ‘two rivers’ The only clusters attested outside of loan words are homorganic nasal + obstruent, where C2 may be a fricative, as in (35a), or a stop of either voicing value, as in (35b) and the previous examples. All morphemes in Warembori are either nasal- or vowel-final, so other types of clusters have no opportunity to arise. (35) a. [nsora] ‘be full (of food)’ b. [anta] ‘fish’ The first step towards the historical development of VRK mutation would have looked like the Warembori pattern, where /b/ and /d/ lenite to [B] and [r] intervocalically. Step two involves generalization of lenition to word-initial position, at least in some cases. Vowel-final prefixes may have helped this change along, making lenition of root-initial segments more common in surface forms. Once this happens, [b] and [d] are limited to post-nasal environments, and a new generation of 9 The underlying form here is given by Donohue as /waren-bo/, with medial /r/ and no indicative /-do/ morpheme, though he claims no underlying /r/ and includes the [-ro] suffix in the output. 134 EMILY GASSER learners can reanalyze [B] and [r] as underlying. (Why all of this failed to happen with some reflexes of *b, which appear in all environments as [b], is a topic for future research.) Some instances of [r] would not originally have alternated with [d], having descended from *r, *R, or *l. Analogical pressure, along with a preference for unmarked ND clusters over [nr], would have generalized this alternation to /r/s from other etymological sources. Meanwhile, as individual languages and branches lost their inherited *k, ProtoSHWNG *g, may have devoiced to fill the gap in the inventory, particularly in initial position. The *g would have remained voiced post-nasally. This led to the modern pattern in which /k/ voices after a nasal but the other voiceless stops, which descend from voiceless segments in the proto-language, do not. From here, individual languages have added their own quirks and twists to VRK mutation as historical change continues apace, complexifying the pattern, maintaining it, or dropping it in whole or in part. In Proto-Yapen the alternation extended to apply after all C1 s, Biak vowel syncope created new clusters throughout the language, and elsewhere subsequent changes applied to create the range of synchronic alternations we see today. In some varieties such as Dusner (Dalrymple & Mofu 2012), VRK mutation ceased to be synchronically active, and its remaining effects as a phonotactic tendency in inherited roots will be obscured as lexical replacement continues over time. In all modern ESHWNG languages other than Warembori (and perhaps neighboring Yoke, which appears to have a similar pattern), further change has led to a contrastive distribution of /b/ vs. /B/ and /d/ vs. /r/, so that mutation is a neutralizing change, though this is not the case in most languages for [g] and [k], which in most cases are still in an allophonic relationship. Each of these historical changes is in itself perfectly natural; it is the combination of their cumulative outcomes that creates the unnatural result in the synchronic grammars. The historical context is crucial for understanding the pattern as it stands, though of course speakers are unaware of mutation’s diachronic source and must be able to generate it independently in their grammars. Typologically, we expect any unnatural rule to have a similarly prosaic historical source, and Buckley (2000) and Hayes & White (2015) demonstrate this to be the case for the patterns they describe. As Ohala (1992) points out, sound change is atelic, and individual changes are blind to the outcomes they create. It is this that allows unnatural rules to emerge, and in fact predicts that, occasionally, they should. 4 Conclusions It is clear that VRK mutation began in an ancestor of the ESHWNG languages spoken in the neighborhood of 3000 years ago, making it a highly persistent process despite its status as an unnatural rule, as learners have continued to successfully acquire and pass it down over time. The set of segments that it targets, usually /B, r, k/, is not a natural class, and is not unifiable by any feature set that does not also include other phonemes of the languages in question. In cases where the /k/ no longer takes part (Wooi, among others), the scenario is somewhat better, as /B/ and VRK mutation in West Papua 135 /r/ comprise the full set of voiced continuants10 in the languages in question. This may in fact be the motivation for dropping /k/ from the pattern. Waropen, one of only two ESHWNG languages which includes the voiced velar fricative /G/ in its inventory, is also one of the few in which /k/ appears to fully participate in mutation. Excluding /k/ in this case would not yield a natural class, reducing the pressure to do so. It may be that in some cases VRK mutation has been split into two processes by speakers, with one targeting /B/ and /r/ and the other targeting /k/. I argued in §2.1 that I do not believe that this is the best analysis in Wamesa, and that argument can be extended to many of the other varieties — the VRK phones specifically are exceptional within the language’s treatment of clusters, and considering them separately misses that generalization. In other languages, where the behavior of /k/ is particularly erratic or where it has been dropped from the pattern entirely, there is a much stronger case to be made for a non-unified account. That an unnatural pattern can be successfully learned shouldn’t be too surprising. Recent work in computational and experimental phonology has shown that while unnatural patterns are disprefered in language learning, they can sometimes still be acquired (Newport & Aslin 2004, Wilson 2006, Moreton 2008, Hayes et al. 2009, Becker et al. 2011, White 2017). These studies speak to the ongoing debate about the origins of phonological typology, and whether they rest in channel bias, which attributes typology to the availability of paths of diachronic change leading to a given surface pattern, analytic bias, in which the language faculty itself disprefers some patterns, or some combination of the two. The overall consensus seems to be that both factors play a role. Wilson, Moreton, White, et al. argue that we should take phonetically substantive analytic bias seriously as a bias, not a prohibition, in that it will make an unnatural pattern less likely to be learned but not ban it outright. VRK mutation fits in well with these conclusions. Hayes & White (2015: 270271), argue that the saltation pattern they survey only ever arises out of a series of unrelated ‘historical accidents’, never as a single change itself; that “phonetic change is ‘blind enough’ to create synchronic conundrums as changes accumulate”. This is exactly what we see in VRK mutation: the changes to each individual sound are phonetically motivated, but their confluence is a phonological conundrum. They go on to point out, as others have also done (c.f. White 2017), that given sufficient evidence, learners will overcome the substantive bias of the language faculty and successfully acquire an unnatural pattern. What counts as ‘sufficient’ is an open question, but many of the SHWNG languages seem clearly to meet it with regards to VRK mutation. In many cases, mutation-eligible clusters are formed whenever a verb has a non-singular subject or an inalienable noun has a non-singular possessor. Given that several highly frequent roots, including ‘go’, ‘come’, ‘paddle’11 , ‘bathe’, ‘head’, and ‘hand’ are VRK-initial, examples of mutation would be frequent enough, and the change incurred salient enough, that learners would have plenty of evidence to draw from. A laboratory study of the learnability of VRK mutation and the directions in which learners tend to over- or undergeneralize is an 10 where ‘continuant’ is defined to include the tapped/trilled /r/ despite its brief closure and exclude nasals, neither of which is entirely uncontroversial, see e.g. Ohala (1990) and Cohn (1993). 11 These are seaward-oriented societies, where canoes have long been the primary mode of transport and a part of daily life. 136 EMILY GASSER avenue for future work. Overall, VRK mutation is highly characteristic of the SHWNG languages, a group which are only just starting to receive attention in the literature. Unusual patterns such as this deserve our attention; as Blust (2017: 325) points out, “There are good reasons to argue that nonconforming observations in science should be highlighted... since it is precisely nonconforming observations that present the greatest challenges to our assumptions about the world.” SHWNG continues to be a source of typologically interesting observations, and further documentation of these languages will continue to enrich our understanding of what languages can do. References Anderson, S. R. 1981. Why phonology isn’t “natural”. Linguistic Inquiry 12. 493–539. Archangeli, D., & D. Pulleyblank. 1994. Grounded Phonology. Cambridge, MA: MIT Press. Bach, E., & R. T. Harms. 1972. How do languages get crazy rules? In Linguistic Change and Generative Theory, ed. by R. Stockwell & R. Macaulay, 1–21. Bloomington: Indiana University Press. Becker, M., N. Ketrez, & A. Nevins. 2011. The surfeit of the stimulus: Analytic biases filter lexical statistics in Turkish laryngeal alternations. Language 87. 84–125. Blust, R. 1978. Eastern Malayo-Polynesian: A subgrouping argument. In Second International Conference on Austronesian Linguistics: Proceedings, ed. by S. Wurm & L. Carrington, 181– 254. Canberra: Department of Linguistics, Research School of Pacific Studies, Australian National University. Blust, R. 1999. Subgrouping, circularity and extinction: some issues in Austronesian comparative linguistics. In Selected papers from the 8th International Conference on Austronesian Linguistics, volume 1, 31–94. Blust, R. 2009. The Austronesian Languages. Canberra: Pacific Linguistics. Blust, R. 2017. Odd conditions: Context-sensitive sound change in unexpected contexts. Journal of Historical Linguistics 7. 322–371. Blust, R., & S. Trussel, 2010. Austronesian comparative dictionary. Online database. http://www.trussel2.com/acd/. Bolognesi, R. 1998. The phonology of Campidanian Sardinian: a unitary account of a selforganizing structure. The Hague: Holland Institute of Generative Linguistics. Browman, C. P., & L. Goldstein. 1992. Articulatory Phonology: An overview. Phonetica 49. 155–180. Buckley, E. 2000. On the naturalness of unnatural rules. In Proceedings from the Second Workshop on American Indigenous Languages, volume 9 of UCSB Working Papers in Linguistics, 1–14. Cohn, A. C. 1993. The status of nasalized continuants. In Nasals, Nasalization, and the Velum, ed. by M. K. Huffman & R. A. Krakow, volume 5 of Phonetics and Phonology, 329–367. New York: Academic Press. Dalrymple, M., & S. Mofu. 2012. Dusner. Number 4887 in Languages of the World/Materials. Munich: LINCOM GmbH. Donegan, P. J., & D. Stampe. 1979. The study of natural phonology. In Current Approaches to Phonological Theory, ed. by D. Dinnsen, 126–173. Bloomington: Indiana University Press. Donohue, M. 1999. Warembori, volume 341 of Languages of the World/Materials. Munich: Lincom Europa. Gasser, E. 2014. Windesi Wamesa Morphophonology. Yale University dissertation. Gasser, E. 2015. Wamesa Talking Dictionary. Living Tongues Institute for Endangered Languages. www.talkingdictionary.org/wamesa. Gasser, E. 2016. Biak Talking Dictionary. Living Tongues Institute for Endangered Languages. www.talkingdictionary.org/biak. VRK mutation in West Papua 137 Gasser, E., L. Arnold, & D. Kamholz. 2023. Languages of South Halmahera and West New Guinea. In The Oxford Guide to the Malayo-Polynesian Languages of Southeast Asia, ed. by A. Adelaar & A. Schapper. Oxford: Oxford University Press. Gasser, E., L. Arnold, & L. I. S. Rødvand. Forthcoming. The languages of south Halmahera and northwest New Guinea. In Languages and Linguistics of Island Southeast Asia and the Pacific, ed. by B. Palmer & E. Smith-Dennis. Berlin, Boston: de Gruyter Mouton. Grace, G., 1955-56. Field notebooks. http://digicoll.manoa.hawaii.edu/grace/Pages/PDFlist.html. Greenhill, S. J., R. Blust, & R. D. Gray. 2008. The Austronesian Basic Vocabulary Database: From bioinformatics to lexomics. Evolutionary Bioinformatics 4. 271–283. Hayes, B. 1999. Phonetically driven phonology, the role of optimality theory and inductive grounding. In Functionalism and Formalism in Phonology, ed. by M. Darnell, E. Moravcsik, F. Newmeyer, M. Noonan, & K. Wheatley, volume I, 243–285. Amsterdam: John Benjamins. Hayes, B., R. Kirchner, & D. Steriade (eds.) 2004. The Phonetic Bases of Phonological Markedness. Cambridge: Cambridge University Press. Hayes, B., & J. White. 2015. Saltation and the p-map. Phonology 32. 267–302. Hayes, B., K. Zuraw, P. Siptár, & Z. Londe. 2009. Natural and unnatural constraints in Hungarian vowel harmony. Language 85. Held, G. J. 1942. Grammatica van het Waropensch (Nederlandsch Noord Nieuw-Guinea). Bandung: Nix. Henning, J. C., T. A. Henning, M. Sawaki, O. N. Mananian, T. Yoteni, & D. Webori-Mamori. 1991. Sane pai ve pir kavavo nana kavo Wondama/Perbendahara’an kata bahasa Wandamen/Wandamen vocabulary. Number 8 in Series B, Publikasi Khusus Bahasa-Bahasa Daerah. Irian Jaya, Indonesia: Universitas Cenderawasih and Summer Institute of Linguistics. Kamholz, D. 2014. Austronesians in Papua: Diversification and change in South Halmahera–West New Guinea. University of California at Berkeley dissertation. Kenstowicz, M. 1994. Phonology in Generative Grammar. Oxford: Blackwell Publishers. Moreton, E. 2008. Analytic bias and phonological typology. Phonology 25. 83–127. Newport, E. L., & R. N. Aslin. 2004. Learning at a distance i. statistical learning of non-adjacent dependencies. Cognitive psychology 48. 127–162. Ohala, J. J. 1974. Phonetic explanation in phonology. In Papers from the Parasession on Natural Phonology, 251–74. Chicago: Chicago Linguistic Society. Ohala, J. J. 1983. The origin of sound patterns in vocal tract constraints. In The Production of Speech, ed. by P. MacNeilage, 189–216. New York: Springer. Ohala, J. J. 1990. There is no interface between phonology and phonetics: A personal view. Journal of Phonetics 18. 153–172. Ohala, J. J. 1992. What’s cognitive, what’s not, in sound change. In Diachrony within synchrony: language history and cognition, ed. by G. Kellermann & M. Morrissey, 309–355. Frankfurt am Main: Peter Lang. Ramar, O., N. Ongkodharma, & R. Flaming. 1983. Vekavavo nana kavo Wondama/Percakapanpercakapan dalam bahasa Wandamen/Wandamen conversations. Number 6 in Series A, Publikasi Khusus Bahasa-Bahasa Daerah. Irian Jaya, Indonesia: Universitas Cenderawasih and Summer Institute of Linguistics. Saggers, N. 1979. A sketch grammar of Wandamen. Master’s thesis. Sanders, R. N. 2003. Opacity and Sound Change in the Polish Lexicon. UC Santa Cruz dissertation. Sapir, E. 1925. Sound patterns in language. Language 1. 37–51. Sapir, E. 1930. Southern Paiute, a Shoshonean language. Proceedings of the American Academy of Arts and Sciences 65. 1–296. Sawaki, Y. 2017. A Grammar of Wooi: An Austronesian Language of Yapen Island, Western New Guinea. The Australian National University dissertation. Scott, G. 1978. The Fore Language of Papua New Guinea. Number 47 in Series B. Canberra: Pacific Linguistics. Silzer, P. J. 1983. Ambai: An Austronesian language of Irian Jaya, Indonesia. Australian National University dissertation. Than, J., A. Kafiar, K. Waimuri, S. Korwa, J. Krey, L. Rumsarwir, & A. Rumaropen. 2011. Kamus 138 EMILY GASSER Bahasa Daerah Biak. Pemerintah Kabupaten Biak Numfor Dinas Pariwisata dan Kebudayaan. van den Heuvel, W. 2006. Biak: Description of an Austronesian Language of Papua. Vrije Universiteit Amsterdam dissertation. van Hasselt, J. L., & F. J. F. van Hasselt. 1947. Noemfoorsch Woordenboek. Amsterdam: N.V. Drukkerij en Uitgeverij J.H. de Bussy. Venneman, T. 1988. Preference Laws for Syllable Structure. Berlin: Mouton de Gruyter. White, J. 2017. Accounting for the learnability of saltation in phonological theory: A maximum entropy model with a P-map bias. Language 93. 1–36. Wilson, C. 2006. Learning phonology with substantive bias: An experimental and computational study of velar palatalization. Cognitive science 30. 945–982. Proceedings of CLS 58 (2022), 139-153 © Chicago Linguistic Society 2023. All rights reserved. 139 Polarity, homogeneity, and authority Julie Goncharov and Gurmeet Kaur University of Göttingen 1 Introduction This paper discusses Completive Compound Verbs (CCVs) in Hindi-Urdu (HU), which have been argued to have a distribution similar to Positive Polarity Items (PPIs).1 In particular, CCVs have been shown to be unacceptable under negation in simple declarative sentences, such as in (1), where the completive compound complex paRh lii is used (e.g., Hook 1973; Singh 1998; Homer & Bhatt 2019). (1) mira-ne kitaab (*nahiiN) paRh lii Mira-ERG book.F. SG NEG read take.PFV. F. SG ‘Mira did (not) read the book.’ However, the empirical picture is more complex than it has been previously assumed. We demonstrate that there are (at least) two important aspects in which the ban on negated CCVs is dissimilar from the polarity phenomena. First, CCVs are disallowed under negation in some speech acts (declaratives and immediate imperatives), but not in others (deferred imperative), as shown by the contrast between (1)/(2a) and (2b) (also see Hook 1973). (2) a. ye kitaab (*mat) paRh l-o this book NEG read take-IMM ‘(Don’t) read this book!’ b. ye kitaab (mat) paRh le-naa this book NEG read take-DFR ‘(Don’t) read this book!’ Second, the ban on CCVs under negation can be lifted in declaratives and immediate imperatives in certain configurations, which only partly overlap with the configurations that affect the polarity phenomena. For instance, (1) improves with the presence of the quantifier saare/saarii ‘all’ in the object DP, but not with the presence of the quantifier har ‘every’ in the same position, as shown in (3a) and (3b). By contrast, the presence of either quantifier can affect polarity sensitive phenomena in the language. 1 We would like to thank Muskaan, Neha Kulshreshtha, Parvinder Kaur, Usha Udaar and Yash Sinha for discussion of the Hindi-Urdu data. This research has been supported by the German Research Foundation (DFG) grants: ZE 1040/13-1 to Julie Goncharov and ZE 1040/11-1 to Gurmeet Kaur. 140 (3) GONCHAROV AND KAUR a. mira-ne saarii kitaabeiN nahiiN paRh liiN Mira-ERG all.F. PL book.F. PL NEG read take.PFV. F. PL ‘Mira didn’t read all the books.’ nahiiN paRh lii b. * mira-ne har kitaab Mira-ERG every book.F. SG NEG read take.PFV. F. SG Intended: ‘Mira didn’t read every book.’ These observations have received little to no attention in the literature, despite the fact that they undermine not only the polarity-based account of CCVs (e.g., Homer & Bhatt 2019), but also competing accounts in terms of completion (e.g., Singh 1998), as we will show. For the analysis, we focus on the first observation, that is, the interaction between negated CCVs and speech acts. We propose that CCVs have what we call a homogeneity requirement on the completion of an action expressed by the main verb. According to this requirement, either in all speaker’s doxastic possibilities the action is completed within a contextually set time span or in all speaker’s doxastic possibilities the action is not completed within that time span. We show that this requirement interacts with different speech acts in a different way which explains the grammaticality of negated CCVs in deferred imperatives, but not in declaratives or immediate imperatives. We put the investigation of the ameliorating contexts (our second cluster of observations) outside of the scope of this paper. Here, we use them primarily to argue against the existing accounts of CCVs in HU and to set the desiderata for any account of CCVs. The paper is organized as follows: in section 2, we present the distribution of negated CCVs across declarative and imperative sentences. Section 3 provides our main proposal explaining this distribution. Section 4 presents ameliorating contexts for the occurrence of negated CCVs in declaratives and immediate imperatives. In Section 5, we argue that the two existing accounts of CCVs that we are aware of cannot accommodate the facts from HU presented in this paper. Section 6 offers some concluding remarks. 2 CCVs across speech acts There are two ways to express the perfective in Hindi-Urdu (henceforth HU). The first way is by addition of perfective (PFV) morphology on a simple verb, see (4a). The second way is via a completive compound verb (CCV). CCVs are verbal structures that consist of the stem form of the main verb and an inflecting light verb such as de ‘give’, le ‘take’, Daal ‘put’, maar ‘hit’ (Hook 1973; Butt 1995; Singh 1998; Butt 2003; Homer & Bhatt 2019). An illustration is given in (4b). (4) a. mira-ne kitaab paRhii Mira- ERG book.F. SG read.PFV. F. SG ‘Mira read the book.’ b. mira-ne kitaab paRh lii Mira- ERG book.F. SG read take.PFV. F. SG ‘Mira read the book.’ Polarity, homogeneity, and authority 141 It is well-known in literature (Hook 1973; Butt 1995; Singh 1998; Homer & Bhatt 2019, among others) that negation of CCVs is disallowed in declaratives, as shown in (5a). The simple perfective, by contrast, can be negated; see (5b). (5) a. b. * mira-ne kitaab nahiiN paRh lii Mira- ERG book.F. SG NEG read take.PFV. F. SG ‘Mira did not read the book.’ mira-ne kitaab nahiiN paRhii Mira- ERG book.F. SG NEG read.PFV. F. SG ‘Mira did not read the book.’ What is less discussed is that the above restriction also extends to (a subset of) imperatives in the language. HU makes a morphological distinction between imperatives that require the addressee to act immediately after the imperative is issued (immediate imperatives/IMM) and imperatives that require the addressee to act later (deferred imperatives/DFR) (Sharma 1999, Banerjee & Kaur 2022). See (6a) for an example of the IMM imperative and (6b) for an example of the DFR imperative. (6) a. kitaab paRh-o book read-IMM ‘Read this book.’ b. kitaab paRh-naa book read-DFR ‘Read this book (later).’ Like in declaratives, CCVs cannot be negated in IMM imperatives, (7a). The DFR imperative, in contrast, shows no such restriction, (7b) (Hook 1973). (7) a. ye kitaab (*mat) paRh l-o this book NEG read take-IMM ‘(Don’t) read this book.’ b. ye kitaab (mat) paRh le-naa this book NEG read take-DFR ‘(Don’t) read this book.’ The simple perfective does not obtain in imperatives, regardless of negation. This is an independent restriction in the language, which we do not discuss in this paper, but see Kaur & Goncharov (2022). (8) a. kitaab paRh-(*aa)-o book read-PFV- IMM ‘Read the book!’ b. kitaab paRh-(*aa)-naa book read-PFV- DFR ‘Read the book!’ 142 GONCHAROV AND KAUR The above observations, as summarized in Table (1), lead to the following puzzle: why are CCVs disallowed under negation in declaratives and IMM imperatives but not in DFR imperatives? Negated CCV Negated PFV DECL IMM DFR N Y N NA Y NA Table 1: Distribution of PFV under negation across speech acts 3 Accounting for CCVs across different speech acts 3.1 The meaning of PFV and CCVs in HU In order to capture the polarity sensitivity of CCVs, we first need to understand the semantic contribution of CCVs. We begin with the established observation that sentences with PFV in HU have a weak meaning, which requires the event to be discontinued, but does not require the event to reach its natural end-point (e.g., Singh 1998; Arunachalam & Kothari 2011; Altshuler 2013). This is illustrated in (9). (9) a. mira-ne kitaab paRhii #aur use ab-tak paRh rahii hai Mira-ERG book read.PFV and it.ACC still read PROG be.PRS lit: ‘Mira read a book and is still reading it.’ b. mira-ne kitaab paRhii lekin khatam nahiiN kii Mira-ERG book read.PFV but finish NEG do.PFV ‘Mira read a book but did not finish it.’ For this reason, we assign the weak truth-conditions for a sentence with PFV that the event is discontinued at an arbitrary end-point, which we represent as the state FINsome , see (10). In (10), ⇤K is the declarative speech act operator (Meyer 2013; Krifka 2015), tδ is a pronominal time set by tense (ignored here), I c is some non-zero, non-infinite contextually set time interval (which is a parameter for ⇤K ), and FINsome is an abbreviation for the event in question being in a non-ongoing state with an arbitrary end-point (Ramchand & Svenonius 2014; Altshuler 2016). (10) Truth-conditions for (9a): ⇤K [ FINsome (book-reading(m)) at tδ ] tδ ∈ I c Throughout the presentation of our proposal, we assume event semantics and a system similar to Ramchand & Svenonius, in which verbs denote an event description and the aspect embeds this event description into a situation of a particular kind. For example, the truth-conditional meaning of PFV in HU will be as in (11), where FINsome is a situation where the event is discontinued at some arbitrary point. The semantic meaning of a sentence with PFV, such as in (9a), will be roughly a set of situations where book-reading events whose agent is Mira are discontinued at an arbitrary end-point, see (12). However, as our main point concerns the interaction with the speech act rather than event semantics, for the rest of the paper, we will use abbreviations like in (10). Polarity, homogeneity, and authority 143 (11) PFV λP λs.∃e[P (e) ∧ F IN some (s, e)] (12) λs.∃e[book − reading(e) ∧ Ag(e) = m ∧ F IN some (s, e)] In addition to the truth-conditional meaning, PFV has a non-truth conditional component, which ensures that the speaker holds it possible that the event can reach its natural end-point, see (13). This non-truth-conditional component of PFV is formulated in (14) where FINall = finish at a natural end-point state. The meaning of a sentence with PFV can be pragmatically strengthened either to mean that the event has reached its natural end-point, as in (13), via the inference in (14) or to mean that the event didn’t reach its natural end-point, as in (9b), via a mechanism similar to some but not all strengthening. (13) mira-ne kitaab paRhii aur puurii paRhii Mira-ERG book read.PFV and full.F. SG read.PFV ‘Mira read a book completely.’ (14) Non-truth-conditional contribution of PFV: if ⇤K [ ∃t ∈ I c : FINsome (book-reading(m)) at t ], then ⌃K [ ∃t ∈ I c : FINall (book-reading(m)) at t ] The sentences with CCVs are different from sentences with simple PFV-verbs in that in the former, the inference that the event has reached its natural end-point (FINall ) is obligatory, see (15). (15) a. mira-ne kitaab paRh lii #aur use ab-tak paRh rahii Mira-ERG book read take.PFV and it.ACC still read PROG hai be.PRS lit: ‘Mira read a book and is still reading it.’ b. mira-ne kitaab paRh lii #lekin khatam nahiiN kii Mira-ERG book read take.PFV but finish NEG do.PFV lit: ‘Mira read a book but did not finish it.’ Thus, we propose that a CCV makes a non-truth-conditional contribution, which, in positive sentences, strengthens speaker’s certainty about completion of the event. In particular, we propose that a CCV contributes what we call a homogeneity requirement on completion of the event, which we capture as an excluded middle inference here.2 According to this requirement, a CCV is felicitous only if either in all of speaker’s K-worlds the action is completed within I c or in all her K-worlds the action is not completed within I c , see (16). (16) 2 Non-truth-conditional contribution of CCV: ⇤K [ ∃t ∈ I c : FINall (book-reading(m)) at t ] ∨ ⇤K [ ¬ ∃t ∈ I c : FINall (book-reading(m)) at t ] The interaction between polarity and homogeneity has interested linguists since at least Fodor’s (1970) observation about definite plurals, whose recent accounts make a straightforward connection to the derivation of polarity sensitive items (e.g., Bar-Lev 2018). Our use of homogeneity is different from the technical point of view, but the same in spirit. 144 GONCHAROV AND KAUR Taking stock, a sentence like in (15a) has the weak truth-conditions in (17a) where the speaker commits herself to the statement that the event has discontinued at some arbitrary end-point (FINsome ). PFV adds the requirement that the speaker holds it possible that the event will be completed, see (17b). Finally, the addition of a CCV strengthens the inference brought in by PFV to the inference that in all speaker’s K-worlds the event has reached its natural end-point (FINall ), see (17c). (17) a. Truth-conditions for 15a: ⇤K [ FINsome (book-reading(m)) at tδ ] b. Contribution of PFV: ⌃K [ ∃t ∈ I c : FINall (book-reading(m)) at t ] c. Contribution of a CCV (strengthening): ⇤K [ ∃t ∈ I c : FINall (book-reading(m)) at t ] tδ ∈ I c 3.2 Under negation Let us now turn to what happens when negation is present, as in (18). As we show shortly, in negative sentences, the homogeneity requirement of CCVs in (16) cannot be satisfied. (18) mira-ne kitaab (*nahiiN) paRh lii Mira-ERG book.F. SG NEG read take.PFV ‘Mira did (not) read the book.’ This is because there are only two ways to fulfil the requirement in (16) and both of them require the speaker to be omniscient about how the event in question develops throughout I c . As we assume that the speaker is not omniscient, (16) is unsatisfiable and thus, CCVs are out in negative sentences. To make things concrete, let us look at the representation of (18) in (19). The truth-conditions in (19a) say that the event is not in the finish-at-some-arbitrary-point state. Given that the negated PFV expresses a non-ongoing event, see (20), (19a) is true only in case the event has not begun by tδ . (19) a. Truth-conditions for 18: ⇤K [ ¬ FINsome (book-reading(m)) at tδ ] tδ ∈ I c b. Non-truth-conditional contribution of PFV: if ⇤K [ ∃t ∈ I c : FINsome (book-reading(m)) at t ], then ⌃K [ ∃t ∈ I c : FINall (book-reading(m)) at t ] (vacuous) c. Non-truth-conditional contribution of CCV: ⇤K [ ∃t ∈ I c : FINall (book-reading(m)) at t ] ∨ ⇤K [ ¬ ∃t ∈ I c : FINall (book-reading(m)) at t ] (unsatisfiable) (20) mira-ne kitaab nahiiN paRhii #aur use ab-tak paRh rahii hai Mira-ERG book NEG read.PFV and it.ACC still read PROG be.PRS ‘Mira did not read a book (#and is still reading it).’ To satisfy (19c), the speaker must be omniscient and either (i) know that the event will begin at some later time t ∈ I c and then reach its natural end-point within Polarity, homogeneity, and authority 145 I c (satisfying the first disjunct), or (ii) know that the event remains non-started throughout I c (satisfying the second disjunct). As we said above, we standardly assume that the speaker is not omniscient. Therefore, the condition in (19c) is unsatisfiable in negative sentences which leads to CCVs being unacceptable under sentential negation. 3.3 CCVs in imperatives Let us recall that in imperatives the situation is more complex: IMM-imperatives pattern with declaratives in that CCVs cannot be negated there, (21a), whereas in DFR-imperatives, CCVs are fully acceptable under negation, (21b). (21) a. ye kitaab (*mat) paRh l-o this book NEG read take-IMM ‘(Don’t) read this book!’ b. ye kitaab (mat) paRh le-naa this book NEG read take-DFR ‘(Don’t) read this book!’ To explain (21), we build on the idea that in contrast to declaratives, in imperatives the speaker has Authority, i.e., perfect knowledge about the necessity of ordered propositions, see (22) (Kaufmann 2012). (22) When ‘p!’ is uttered, ⇤K ⇤R p ↔ ⇤R p modality) (where R is some prioritizing In a nutshell, our proposal for imperatives is that Authority ensures that nonstarted events remain non-started throughout I c (which, in a sense, implements the second omniscient strategy above). In this case, the homogeneity requirement on CCVs is satisfied since if in all speaker’s K-worlds it is necessary that the event is non-started, it is entailed that in all speaker’s K-worlds the event is not completed. The difference between IMM-imperatives, which do not allow CCVs under negation, and DFR-imperatives, which do, is in how long the Authority lasts. It has been noted that Authority in DFR-imperatives spans a longer period of time, whereas Authority in IMM-imperatives is limited to a short time interval overlapping with the utterance time. This difference is illustrated in (23). (23) a. issii waqt baazaar jaa-o this.foc time market go-IMM ‘Go to the market right this moment!’ b. * issii waqt baazaar jaa-naa this.foc time market go-DFR Intended: ‘Go to the market right this moment!’ In our system we can capture this difference by specifying Authority duration relative to I c . In IMM-imperatives, Authority spans a time interval that is properly included in I c , (24a), whereas in DFR-imperatives, Authority is longer than I c , (24b). 146 (24) GONCHAROV AND KAUR a. When ‘p!-IMM’ is uttered, ⇤K [ ∀t ∈ J : ⇤R pt ] ↔ ⇤R p b. When ‘p!-DFR’ is uttered, ⇤K [ ∀t ∈ J : ⇤R pt ] ↔ ⇤R p J ⊂ Ic Ic ⊆ J Let us look at the meaning of a negated imperative with a CCV, assuming for simplicity the modal view on imperatives (Kaufmann 2012). (25) a. Truth-conditions for 21: ⇤R [ ¬ BEG(book-reading(you)) ] b. Contribution of the Authority condition: ⇤K [ ∀t ∈ J : ⇤R ¬ BEG(book-reading(you)) at t ] ↔ ⇤R ¬ BEG(bookreading(you)) c. Contribution of CCV: ⇤K [ ∃t ∈ I c : FINall (book-reading(you)) at t ] ∨ ⇤K [ ¬ ∃t ∈ I c : FINall (book-reading(you)) at t ] Now, if J ⊂ I c , as is the case with IMM-imperatives, there will still be speaker’s K-worlds within I c , in which it is uncertain whether your reading the book begins or not. Therefore, none of the disjuncts of (25c) can be satisfied. If, on the other hand, I c ⊆ J, as is the case with DFR-imperatives, in all speaker’s K-worlds within I c , your reading of the book remains non-started (which entails that the event does not reach its natural end-point within I c ). Thus, the second disjunct of the homogeneity condition in (25c) is satisfied, which explains why negated CCVs are acceptable in DFR-imperatives. In summary, we have proposed that a homogeneity requirement is at the heart of the ban on CCVs under negation. This homogeneity requirement interacts with speech act operators in declarative sentences and imperatives, which explains different distribution of CCVs in these environments. 4 Ameliorating contexts for negated CCVs It has been noted in the literature that the ban on CCVs under negation can be lifted in declaratives in certain configurations (Homer & Bhatt 2019, Hook 1973). This section provides a more comprehensive list of ameliorating contexts for negated CCVs in HU by presenting three additional contexts that improve the occurrence of CCVs under negation in declaratives and immediate imperatives.3 Based on Hook (1973), Homer & Bhatt (2019) show that although CCVs are ungrammatical in the immediate scope of negation in simplex declaratives, they can occur under negation in at least three contexts. In particular, negated CCVs can occur in the antecedent of a counterfactual conditional, as shown in (26a), in subjunctive relative clauses, as shown in (26b), and in jab tak ‘as long as’-clauses shown in (26c). 3 The speakers we consulted agree that the following contexts improve the occurrence of negated CCVs in declaratives and immediate imperatives - however, there is inter-speaker variation in the degree of acceptability. Polarity, homogeneity, and authority (26) 147 a. agar maiN-ne ram-ko laDDuu nahiiN khilaa diye if I-ERG Ram-DAT laddu.M . PL NEG feed give.PFV. M . PL ho gayaa hotaa behosh to vo zarur hote, be.HAB . M . PL then he definitely unconscious be go.PFV be.HAB . M . SG ‘If I hadn’t fed Ram laddus, he would have definitely lost consciousness.’ b. yahaaN aisaa koi bhii nahiiN [jis-ne Sita-ke liye ye kaam here such some ever NEG REL - ERG Sita-GEN for this work naa kar diyaa ho] NEG do give. PFV. M . SG be. SBJV ‘There is no one here who hasn’t done this work for Sita.’ c. [jab tak mira-ne kitaab nahiiN paRh lii] [tab tak when till Mira-ERG book NEG read take.PFV. F. SG then till karan intezaar kartaa rahaa] Karan wait do.HAB . M . SG stay.PFV. M . SG ’Karan kept waiting until Mira finished reading the book.’ We provide three additional contexts that significantly improve the occurrence of CCVs under negation both in declaratives and immediate imperatives. The first ameliorating context is provided by the quantifier ‘all’. Consider the following example - when the internal argument is a quantified phrase containing ‘all’, it becomes possible to negate the CCV. This is shown for the declarative in (27a), and for the immediate imperative in (27b) - these examples are more natural with stress on ‘all’. (27) a. mira-ne saarii kitaabeiN nahiiN paRh liiN Mira-ERG all.F. PL book.F. PL NEG read take.PFV. F. PL ‘Mira did not read all the books.’ b. saarii kitaabeiN mat paRh l-o all.F. PL book.F. PL NEG read take-IMM ‘Don’t read all the books.’ Crucially, the universal quantifier ‘every’ and numerals, even with stress placed on them, do not cause this amelioration. This is shown in (28) with declaratives. Immediate imperatives pattern alike. (28) a. b. * mira-ne har kitaab nahiiN paRh lii Mira-ERG every book.F. SG NEG read take.PFV. F. SG Intended: ‘Mira did not read every book.’ * mira-ne do kitaabeiN nahiiN paRh liiN Mira-ERG two book.F. PL NEG read take.PFV. F. PL Intended: ‘Mira did not read the two books.’ The second ameliorating context for negated CCVs in declaratives and immediate imperatives is provided by in-adverbials, which are known to occur with telic but not atelic predicates. 148 GONCHAROV AND KAUR (29) a. John built the house in a year. b. #John loved Beth in a year. CCVs in HU are telic predicates which allow in-adverbials, as shown in (30). (30) mira-ne ye kitaab ek raat meiN/(#tak) paRh lii Mira-ERG this book one night in/for read take.PFV. F. SG ’Mira read this book in a night.’ The in-adverbial has no bearing on the grammaticality of a positive CCV. However, consider the declarative and the immediate imperative in (31) - a phonologically stressed in-phrase adverbial significantly improves the presence of a CCV under negation in these sentences. (31) a. mira-ne kitaab ek (hii) raat meiN nahiiN paRh lii Mira-ERG book one FOC night in NEG read take.PFV. F. SG ’Mira did not read the book in a night.’ b. ek (hii) raat meiN kitaab mat paRh lo one FOC night in book NEG read take.IMM ’Don’t read the book within one night.’ The third ameliorating context comes from contrastively focused main verb. In HU, the occurrence of negation before the verb is its default position. This is illustrated in (32). However, negation can also occur between the verb and the auxiliary, in which case, a contrastive reading of the verb becomes easily available. This is shown in (33) from Homer & Bhatt (2020:3). (32) ram-ne seb nahiiN khaayaa thaa Ram-ERG apple NEG eat.PFV. M . SG be.PST. M . SG ‘Ram had not eaten the apple.’ (33) ram-ne seb khaayaa nahiiN thaa (sirf suNghaa Ram-ERG apple eat.PFV. M . SG NEG be.PST. M . SG only smell.PFV. M . SG thaa) be.PST. M . SG ‘Ram hadn’t eaten the apple (he merely smelled it).’ As seen already, placement of negation in its default position in a CCV leads to ungrammaticality in perfective declaratives and immediate imperatives. However, these structures are significantly improved when negation occurs between the main and the light verb. Consider the examples in (34) - the position of negation after the main verb causes this verb to be contrastively focused, and ameliorates the presence of the CCV under negation. (34) a. mira-ne phuul toR nahiiN diyaa thaa (sirf mira-ERG flower break NEG give.PFV. M . SG be.PST. M . SG only suuNghaa thaa) smell.PFV. M . SG be.PST. M . SG ‘Mira did not pluck the flower, she only smelled it.’ Polarity, homogeneity, and authority 149 b. phuul toR mat do (sirf suuNgho) flower break NEG give.IMM only smell.IMM ’Don’t pluck the flower, only smell it!’ We now have a more comprehensive list of contexts that ameliorate negated CCVs. These are: (i) antecedent of a counterfactual conditional, (ii) a subjunctive relative clause, (iii) a jab tak-clause, (iv) a quantified object DP containing ‘all’, (v) an in-adverbial, and (vi) contrastive focus on the main verb. 5 Existing accounts On the empirical front, this paper has presented two findings: first, apart from positive declaratives, CCVs in HU are also licensed in negated DFR imperatives. Secondly, there are at least six contexts that improve the occurrence of CCVs under negation in declaratives and IMM imperatives. In view of these findings, this section evaluates two existing approaches to negated CCVs in HU. We demonstrate that both accounts lack empirical coverage - neither of them can explain the licensing of CCVs in negated DFR imperatives. Furthermore, under both approaches, only a subset of ameliorating contexts can be accounted for. There are two approaches to explaining the ban on negated CCVs in declaratives in HU. These are (i) the PPI-based approach (Homer & Bhatt 2019), and (ii) the completion-based approach (Singh 1998). According to the PPI-based approach, CCVs (but not simple perfective verbs) are positive polarity items, akin to English would rather, which restricts them from occurring in the scope of negation. (35) a. b. John would rather read the book. * John wouldn’t rather read the book. According to the completion approach (Singh 1998), use of negation is pragmatically infelicitous with a CCV, which is a marker of completion. For illustration, let us revisit the example in (1), repeated here as (36). Singh claims that ‘not read a book’ expressed with a CCV, as in said example, can have either of the following two meanings: (i) the agent did not start reading the book, or (ii) she started reading the book but did not complete it. Since neither of the meanings suggest completion of the event, the CCV is blocked and the simple perfective suffices. (36) mira-ne kitaab (*nahiiN) paRh lii Mira-ERG book.F. SG NEG read take.PFV. F. SG ‘Mira did (not) read the book.’ We evaluate these approaches against the empirical findings of the paper in the rest of this section: CCVs across speech acts: Let us start with the distribution of CCVs across speech acts. Recall that CCVs are disallowed under negation in some speech acts (declaratives and immediate imperatives), but not in others (deferred imperative). The PPI-based approach does not make any prediction about the interaction between negated CCVs and the type of speech act they occur in. In particular, if CCVs are 150 GONCHAROV AND KAUR regular PPIs, then their grammatical status under negation in deferred imperatives but not in immediate imperatives or declaratives, remains unexplained. Like the PPI-based approach, the completion based approach also fails to explain the grammatical occurrence of CCVs in deferred imperatives. As for ameliorating contexts, each of the approaches can explain some of the contexts but not all. Ameliorating contexts (i)-(iii): The PPI-based approach can account for the first three ameliorating contexts, listed as follows: (i) antecedent of a counterfactual conditional, (ii) a subjunctive relative clause, (iii) a jab tak-clause. In particular, the three contexts can be treated as rescuing environments, where rescuing is understood as the effect of making an otherwise ungrammatical occurrence of a PPI grammatical by placing it in the scope of a second Downward entailing expression. For example, the negation of would rather, which is typically ungrammatical, becomes grammatical in the antecedent of a conditional, and under the DE quantifier few, see (37). (37) a. If you wouldn’t rather be here, you are lying. b. Very few people wouldn’t rather sit somewhere else. Thus, one might conclude that CCVs are regular PPIs with some caveats that have to do with their lexical category and possibly language-specific properties of HU. However, the old and new data we have presented in this paper raises important challenges for the PPI-based treatment of CCVs in HU. As stated already, the PPIbased approach does not make any prediction about the interaction between negated CCVs and the type of speech act they occur in. Furthermore, this approach cannot explain all ameliorating contexts. Ameliorating context (iv): According to this context, a negated CCV can occur in the presence of a quantified object containing ‘all’ but not ‘every’. Prima facie, the PPI-based account seems capable of explaining amelioration context (iv) as an instance of ‘shielding’, defined as follows: a PPI becomes acceptable in the scope of a clausemate negation if a universal item intervenes at logical form (Nicolae 2017, Kroch 1974). For instance, some can be licensed in the scope of a clausemate negation if a strong scalar term such as always intervenes (38). (38) Mary doesn’t always understand something. (neg > always > some) Similarly, under the PPI-based approach to CCVs, their amelioration under negation in the presence of the universal quantifier ‘all’ can be treated as an instance of shielding. However, the PPI based approach cannot explain as to why this ‘shielding’ effect does not extend to other universal items such as ‘every’. Upon further development, the completion based approach may fare better for context (iv). Under the completion based account, the difference between ‘all’ and ‘every’ can be obtained by capitalizing on the collective versus distributive distinction in meaning between the two quantifiers - har ‘every’ and saaraa ‘all’ in HU behave differently when they occur in the subject position of an obligatorily collective predicate like ‘gather’. In such cases, only collective saaraa is grammatical; har that has (predominantly) a distributive reading is not possible with ‘gather’. Polarity, homogeneity, and authority (39) a. b. 151 saare laRke kamre meiN ikaTThaa ho gaye all boys room in gather be go.PFV. PL ‘All boys gathered in the room.’ * har laRkaa kamre meiN ikaTThaa ho gayaa every boy room in gather be go.PFV. SG lit: ‘Every boy gathered in the room.’ Ameliorating context (v): Since the completion based account is based on the event semantics of the CCVs, it may be possible to develop the account to explain the amelioration of negated CCVs with in-adverbials, which apply to quantized (telic or event) predicates (context v). It is not clear how to apply the PPI-based approach to this context. Ameliorating context (vi): An application of either of these approaches to explain ameliorating context (vi) with contrastive focus on the main verb, is unclear. The above discussion clearly demonstrates that existing approaches to negated CCVs in HU have limited empirical coverage. A novel account to explain the old and new data discussed in this paper is therefore highly desired. 6 Concluding remarks In this paper, we revisited the ban on negated CCVs by expanding the empirical domain to include imperatives - negated CCVs are grammatical in deferred imperatives but not in immediate imperatives and declaratives. We argued that a homogeneity requirement about the completion of an event is at the heart of this distribution. This homogeneity requirement has an epistemic component that spans a contextually set time interval. This component interacts with the speech act operator in declaratives and imperatives. Under negation, this interaction results in imposing an impossible requirement on the speaker’s epistemic state in declaratives and immediate imperatives (requiring her to be omniscient about the development of the event). In deferred-imperatives, this omniscience requirement is fulfilled by Authority (speaker’s perfect knowledge about the command) that spans longer than the contextually set interval for CCVs. We also provided ameliorating contexts that improve the occurrence of negated CCVs in declaratives and immediate imperatives. These amelioration facts and the interaction of negated CCVs with speech acts demonstrates that existing accounts of CCVs under negation do not suffice, and a distinct analysis is needed. Our future work will focus on providing a unified account of all ameliorating contexts for negated CCVs, including deferred imperatives where negated CCVs occur freely. References Altshuler, D. 2013. There is no neutral aspect. In Proceedings of Semantics and Linguistic Theory 23, 40–62. 152 GONCHAROV AND KAUR Altshuler, D. 2016. Events, States and Times. An essay on narrative discourse in English. Berlin, Boston: De Gruyter. Arunachalam, S., & A. Kothari. 2011. An experimental study of Hindi and English perfective interpretation. Journal of South Asian Linguistics 4. 27–42. Banerjee, N., & G. Kaur. 2022. Deferred imperatives across Indo-Aryan. In Proceedings of Formal Approaches to South Asian Languages (F)ASAL-10. Bar-Lev, M. 2018. Free choice, homogeneity, and innocent inclusion. Hebrew University of Jerusalem dissertation. Butt, M. 1995. The structure of complex predicates in Urdu. Center for the Study of Language (CSLI). Butt, M. 2003. The light verb jungle. In Harvard Working Papers in Linguistics 9, 1–49. Fodor, J. 1970. The linguistic description of opaque contexts. Hebrew University of Jerusalem dissertation. Homer, V., & R. Bhatt. 2019. Licensing of PPI indefinites: Movement or pseudoscope? Natural Language Semantics 27. 279–321. Homer, V., & R. Bhatt. 2020. Restructuring and the scope of negation in Hindi-Urdu. Glossa: a journal of general linguistics 5. Hook, P. 1973. The compound verb in Hindi. University of Pennsylvania dissertation. Kaufmann, M. 2012. Interpreting Imperatives. Studies in Linguistics and Philosophy. Springer Netherlands. Kaur, G., & J. Goncharov. 2022. Aspect in Hindi-Urdu and transition between syntactic domains. In Proceedings of The 13th Generative Linguistics in the Old World in Asia conference, 133– 148. Krifka, M. 2015. Bias in commitment space semantics: Declarative questions, negated questions, and question tags. In Proceedings of Semantics and Linguistic Theory 25, 328–345. Kroch, A. S. 1974. The semantics of scope in English. Massachusetts Institute of Technology dissertation. Meyer, M.-C. 2013. Ignorance and Grammar. Massachusetts Institute of Technology dissertation. Nicolae, A. C. 2017. A new perspective on the shielding property of positive polarity. In Proceedings of Semantics and Linguistic Theory 27, 266–281. Ramchand, G., & P. Svenonius. 2014. Deriving the functional hierarchy. Language Sciences 46. 152–174. Sharma, G. 1999. A pragmatic survey of Hindi imperatives. Annali di Ca’ Foscari, Serie orientale 38(3). 245–316. Singh, M. 1998. On the semantics of the perfective aspect. Natural Language Semantics 6. 171–199. Proceedings of CLS 58 (2022), 153-167 © Chicago Linguistic Society 2023. All rights reserved. 153 Maximization of the concord domain Anna Grabovac University College London 1 Introduction The term ‘concord’ refers to the agreement phenomena typically found in the nominal domain. In (1), for example, the demonstrative and adjective show concord for case, gender, and number with the head noun1 . (1) de-m neu-en Student-en the-DAT. M . SG new-DAT. M . SG student-DAT. M . SG ‘to the new student’ (German) Since concord involves the sharing of features among elements, it has been popular to simply define concord as a form of agreement. To this end, many have attempted to unify analyses of nominal concord with more familiar examples of argumentpredicate agreement, such as (2). (2) [The student]3SG reads3SG a book. Despite apparent similarities, others have argued that the differences between concord and this kind of agreement are not trivial. Norris (2014) outlines several differences, including the number of loci of expression, locality, and sensitivity to case. Regarding the first, Norris points out that the features involved in concord typically have multiple loci of expression in various syntactic positions, such as heads, adjuncts, and specifiers. Agreement features, by contrast, are often limited to a single locus of expression—usually a head. In terms of locality, concord is typically restricted to a single extended projection, whereas agreement can occur between elements in different extended projections. Finally, agreement appears to be conditioned by case (often nominative), while concord displays no such restriction. In light of these differences, it has been argued that analyses of concord should not be subsumed under agreement (Ackema & Neeleman 2020; Giusti 2008; Norris 2012, 2014; Polinsky 2016). Norris (2014) proposes that concord results from the realization of features from dominating nodes on available terminals (see also Ackema & Neeleman 2020). This general idea is demonstrated in (3), where the [f ] feature located on XP is realized on terminals 1–3 (slashes denote the result of concord). 1 Glossing abbreviations: 3 = third person, SG = singular, PL = plural, M= masculine, NOM = nominative, GEN = genitive, DAT = dative 154 ANNA GRABOVAC (3) XP [f ] 1 /f / 2 3 /f / /f / While concord patterns can be—and indeed have been—modeled under agreementbased approaches, the forced unification of concord and agreement comes at a cost: the resulting analysis is unnecessarily complicated. This complexity stems from the fact that in generative literature, standard agreement is a simple one-to-one mapping of features between nodes under c-command (Chomsky 2000, 2001). For example, person and number features are mapped from the subject to the verb in (2). By contrast, concord appears to be derived via a many-to-one mapping since the features involved in concord, such as case, gender, and number, originate in different places within the extended projection (and in fact may be introduced in their own projections) (Ackema & Neeleman 2020). In concord, these features are then individually mapped to their location of expression, such as the adjective in (1)2 . If, instead, a strict one-to-one mapping is maintained in concord, then the relation becomes one of domination rather than c-command; assuming that features percolate through the extended projection, only the topmost node contains all of the relevant features (Ackema & Neeleman 2020). Subsequent sections will return to this idea and the complications that arise in a purely agreement-based approach to concord. Expanding on Norris’s (2014) theory of concord as the spellout of features from dominating nodes, I emphasize the importance of domain maximization throughout the derivation (see Grabovac 2022). In the syntax, domain maximization consists of feature percolation as high as possible, and in concord (post-syntax), features are realized as low as possible. Normally, domain maximization respects the boundaries imposed by extended projections, but subsequent sections will reflect on feature percolation and impoverishment as possible methods of domain extension. The remainder of this paper is organized as follows: in the next section, I discuss examples of basic concord, comparing a purely agreement-based analysis and the concord-as-spellout approach. In §3, I use Slavic numeral constructions to demonstrate how the concord-as-spellout approach easily derives a number of complex concord patterns. This section returns to an evaluation of agreement-based analyses in light of these complex patterns and discusses the outcomes of domain extension. §4 concludes with a reflection on some of the major implications of the concord-as-spellout approach. 2 Although I take agreement to be a one-to-one mapping, this does not mean that agreement features cannot appear on multiple elements. For example, Norris (2014) points out that in the Bantu languages, agreement features appear on both the auxiliary and the verb. I analyze this as two instances of a one-to-one agreement mapping. Maximization of the concord domain 155 2 Evaluating Agree The precise formulation of Agree, the Minimalist operation underlying agreement (Chomsky 2000, 2001), has been subject to revision over the years. However, there is general consensus that the probe, which bears uninterpretable features, and the goal, which bears interpretable features, are in a c-command relation3 . When agreement occurs, the features of the goal value the uninterpretable features of the probe (a one-to-one mapping). In the remainder of this section, I evaluate Agree against typical argument-predicate agreement and nominal concord. 2.1 Argument-predicate agreement The general concept of Agree involving c-command and a mapping of features from goal to probe is simple enough to apply to canonical argument-predicate agreement. Let us consider how this plays out with the example in (4), where the verb agrees with the subject in person and number features. (4) [The student]3SG reads3SG a book. In the simplified structure in (5), the DP subject clearly c-commands the verb, and the 3 SG features located on the DP can be copied to the verb in agreement. Thus, subject-verb agreement results from a simple one-to-one mapping under ccommand. (5) ... DP ... [3 SG ] V ... [3 SG ] 2.2 Nominal concord 2.2.1 The application of Agree in concord Let us now consider the application of standard Agree mechanisms—c-command and a one-to-one mapping—against the distribution of features in (6). I will assume the structure in (7), where gender (represented by γ) originates on the noun (see Kramer 2016), number is introduced in #P just above NP (Ritter 1992), and case is introduced in KP (Bittner & Hale 1996). 3 There is some discussion in the literature on whether the probe should c-command the goal or vice versa (see Ackema & Neeleman 2018; Carstens 2016). For simplicity, I will merely assume that Agree requires a c-command relation between probe and goal, but I remain agnostic as to the directionality of agreement. 156 ANNA GRABOVAC (6) de-m neu-en Student-en the-DAT. M . SG new-DAT. M . SG student-DAT. M . SG ‘to the new student’ (7) KP K (German) [ DAT ] DP [ DAT ] D #P [ SG ] [K: ___, #: ___, γ: ___ ] # [ SG ] AP N [K: ___, #: ___, γ: ___ ] [K: ___, #: ___, γ: M ] Considering first the AP in (6), it is necessary to account for its dative, masculine, and singular features. Assuming agreement results from a one-to-one mapping under c-command, then one logical possibility is that AP agrees with N. However, N itself requires values for case and number. Thus, the one-to-one nature of agreement is compromised, since the individual gender, number, and case features originate in distinct projections. In order to derive the desired features on AP, we would have to assume that it probes upward for number and case but downward for gender4 . Similar issues arise with D, which would probe downward for number and gender but upward for case. If, instead, a strict one-to-one mapping is forced, c-command must be abandoned. Since features percolate through an extended projection (see (Anderson 1992; Cole et al. 1993; Grimshaw 2000), KP is the only node that contains all relevant features for a single mapping to AP and D. It could be possible to maintain a one-to-one mapping by introducing an ordering of valuation. More concretely, AP and D could agree with N once N’s features have been valued. Alternatively, we could adopt the feature-sharing version of Agree (Danon 2012; Frampton & Gutmann 2006; Pesetsky & Torrego 2007) which does not require the goal to contain valued instances of the features needed by the probe. Thus, elements could enter into an agreement relation with N even before N’s number and case features have been valued. Through feature-sharing, values for case and number are later supplied. While it is possible to make modifications to Agree to reconcile its definition with the distribution of features in simple examples like (6), §3 will show that not all examples are this straightforward. In particular, we will see that a one-to-one mapping from N to one of its modifiers is not always a viable analysis, as N may not realize the same features as the modifier. On the other hand, both basic and more complex examples of concord follow easily from the concord-as-spellout analysis. 4 The fact that AP functions as a probe at all is somewhat questionable, as probes are typically heads. Abney (1987) takes adjectives to be heads, but this is not a common assumption in current work. Regardless, it is clear that the theory of phrase structure one adopts plays a role in agreement (see Norris 2014). Maximization of the concord domain 157 2.2.2 Concord as spellout Following work by Norris (2014) and Ackema and Neeleman (2020), I take concord to result from the spell-out of features from dominating nodes on available terminals. Using this theory to derive the distribution of features in (6), we have the following derivation, adopting a broadly Minimalist, Distributed Morphology (DM) architecture of the grammar. Beginning in the syntax, features percolate through the extended projection and collect on KP, as shown in (8)5 . (8) KP [ DAT SG M ] K DP [ SG M ] [ DAT ] D #P [ SG M ] [K: ___, #: ___, γ: ___ ] # [M] [ SG ] AP N [K: ___, #: ___, γ: ___ ] [K: ___, #: ___, γ: M ] Moving out of the syntax, we arrive at the concord stage of the derivation. In this stage, the features located on the dominating KP are spelled out on available terminals. Case, gender, and number features are realized throughout the entire construction. As mentioned in §1, I use slashes to denote the result of concord. (9) KP [ DAT ] K DP [ DAT ] D #P [ SG ] / DAT M SG / # [ SG ] AP N / DAT M SG / / DAT SG / [ M ] The next section applies these stages of the concord-as-spellout analysis to more complex examples of concord. 5 I assume upward percolation of all features in contrast to Norris (2014), who assumes upward percolation of φ-features but downward percolation of case. Upward percolation avoids violations of Inclusiveness (Chomsky 1995), which roughly states that the properties of nodes must be recoverable from dominated structure (see Neeleman & van de Koot 2002). 158 ANNA GRABOVAC 3 Domain extension: Evidence from Slavic numeral constructions Slavic numeral constructions such as the Russian and Bosnian/Croatian/Serbian (BCS) examples in (10)–(12) have long been a topic of contention in the literature6 . This continued interest in numeral constructions largely stems from the complexity of their associated concord patterns. Slight variations in patterns across the Slavic languages pose difficulty for the formulation of a cohesive analysis. This section will demonstrate how concord as spellout straightforwardly accounts for these complex patterns. Note that (10)–(12) are merely a sampling of the Slavic patterns, but see Grabovac (2022) for a broader set of data. (10) <èt-i> pjat’ <èt-ix> nov-yx student-ov this-NOM . PL five.NOM this-NOM . PL new-GEN . PL student-GEN . M . PL ‘these five new students’/ ‘five of these new students’ (Russian) (11) <èt-im> pjat-i <èt-im> nov-ym student-am this-DAT. PL five.DAT this-DAT. PL new-DAT. PL student-DAT. M . PL ‘to these five new students’/ ‘to five of these new students’ (Russian) (12) <ov-ih> pet <ov-ih> nov-ih studen-a-ta this-GEN . PL five this-GEN . PL new-GEN . PL student-PL - GEN . M ‘these five new students’/ ‘five of these new students’ (BCS) Before discussing the derivations of the patterns in (10)–(12) it is first necessary to clarify the syntactic structure of these examples and establish some terminology in order to better evaluate the patterns. In each of the examples, a demonstrative is allowed above or below the numeral with no effect on the concord pattern (though the lower demonstrative typically gives rise to a partitive reading). Given Cinque’s (2005) analysis of Universal 20 (Greenberg 1963), which predicts the possible orders of demonstrative, numeral, adjective, and noun within the extended nominal projection, I analyze the numeral constructions with the structure in (13). Crucially, (13) consists of two extended projections—two syntactic domains—since a demonstrative below the numeral gives rise to unattested orders within a single extended projection7 . For simplicity, I will disregard #P in this discussion, but this has no effect on the analysis. 6 Given space constraints, I restrict the focus to three patterns displayed by higher numeral constructions, which contain numerals ‘five’ and above, but this is by no means the full range of patterns. The concord patterns displayed by Slavic numeral constructions vary according to language, numeral class, and case environment. 7 The structure in (13) predicts the possibility of examples containing two demonstratives simultaneously. Though rare, these types of examples are available for some speakers under a focus reading. Maximization of the concord domain (13) KP 159 [ EXTERNAL CASE ] K [ EXTERNAL CASE ] Dem/AP Num KP [ GENITIVE ] K [ GENITIVE ] Dem/AP N I analyze the lower domain with a genitive case phrase given the common observation that Slavic numerals impose genitive case on their complements (see e.g., Bošković 2006; Franks 1995; Ionin & Matushansky 2018; Klockmann 2017). The higher domain reflects the external case environment considering examples like (10) in which two distinct cases are realized. However, it is also necessary to address (11) and (12) in which the distribution of demonstratives suggests two syntactic domains, but the case distribution—only a single case realized throughout—suggests one concord domain. Now that the structure has been motivated, let us establish names for the patterns, all of which are based on Babby’s (1985) heterogeneous-homogeneous distinction. I will refer to (10) as aligned heterogeneous. The example is heterogeneous in the sense that two concord domains are distinguishable on the basis of case, and aligned because the break in these domains aligns with the break in the underlying syntactic domains. (11), then, is downward homogeneous since only one concord domain is apparent, and the externally assigned case is realized downward throughout the construction. (12) is upward homogeneous since the internally assigned genitive case percolates upward to be realized throughout the construction in a single domain of concord. The derivations also rely on three key hypotheses (see Grabovac 2022): (i) obligatory head-head agreement between N and the lower K; (ii) potential headhead agreement between Num and the higher K; (iii) impoverishment restricted to heads and the dominating nodes to which features have percolated. Further motivation for each of these will be provided as they apply in the following derivations. We will also see that impoverishment and feature percolation are two methods of domain extension. While impoverishment can extend the domain lower, percolation extends the domain higher. This is critical in the derivations of homogeneous patterns. 3.1 Deriving the aligned heterogeneous pattern The aligned heterogeneous pattern is repeated in (14). While this example is from Russian, the pattern is prevalent more generally for Slavic numeral constructions in structural case environments (see e.g., Franks 1995). Here, the externally assigned case (nominative) is realized on the numeral and above, while genitive is realized on the elements below the numeral. 160 ANNA GRABOVAC (14) èt-i pjat’ nov-yx student-ov this-NOM . PL five.NOM new-GEN . PL student-GEN . M . PL ‘these five new students’ (Russian) In this derivation, I assume that the numeral does participate in an agreement relation with the higher K, following hypothesis (ii). This assumption stems from the observation that Russian numerals are declinable; by contrast, we will later see that BCS numerals, which are indeclinable, do not participate in agreement. This potential for agreement is partly conditioned by the numeral’s semi-lexical status, which translates to a variability in feature specification (see Corver & van Riemsdijk 2001; Klockmann 2017; Vos 1999). This variability of the numeral contrasts with the fully lexical noun, which I assume always agrees with the lower K. In (15), the numeral and noun both agree for case. Feature percolation also occurs in this stage, as we saw earlier in (8). Here, the φ-features are able to percolate from the lower domain through the higher domain. On the other hand, the numeral’s participation in agreement for case blocks percolation of the genitive assigned to the lower domain. This trade off in percolation takes insight from the concept of relativized heads as discussed in Di Sciullo and Williams (1987). It is also important to highlight the numeral’s semi-lexical status. §1 mentioned that one of Norris’s (2014) criteria for distinguishing agreement and concord is that concord is normally limited to features within a single extended projection. The structure in (15) is composed of two extended projections, but I assume that the boundary between them is more transparent than usual because of the numeral’s semi-lexicality (see Vos (1999) for a discussion of transparency and semi-lexicality). (15) KP [ NOM M PL ] [ NOM M PL ] K [ NOM ] [ NOM M PL ] Dem/AP Num KP [ GEN M PL ] [ NOM ] [ GEN M PL ] K [ GEN ] Dem/AP N [ GEN M PL ] Moving out of the syntax, (16) depicts the result of concord. Here, the features from dominating nodes are spelled out on available terminals. This results in NOM.PL in the higher domain and GEN.PL in the lower domain. The break in concord domains corresponds to the break in syntactic domains—the aligned heterogeneous pattern. Maximization of the concord domain (16) KP 161 [ NOM M PL ] K [ NOM ] Dem/AP / NOM PL / Num KP [ GEN M PL ] [ NOM ] K [ GEN ] Dem/AP N / GEN PL / [ GEN M PL ] Recalling the discussion of the purely agreement-based analysis in §2, let us suppose that the modifiers agree with the head noun for a one-to-one mapping established under c-command. In the aligned heterogeneous pattern, this works well enough to derive the GEN.PL features of the lower modifier. However, this does not derive the NOM.PL features of the demonstrative in the higher domain, since nominative case is never present on the noun. Similarly, there is no empirical motivation to assume that plural is present on the numeral, so it is also not possible for a oneto-one mapping between the numeral and the demonstrative. It may be possible to assume a kind of roll-up movement whereby the required features are collected, allowing for a one-to-one mapping to be established. For example, Klockmann (2017) suggests that the numeral is initially merged lower in structure before moving to derive the correct word order. As the numeral moves, features are collected. However, this kind of approach to agreement seems to be making the essentially the same assumptions as the concord-as-spellout analysis in which features are collected on KP. Moreover, this additional movement is uncommon for theories of agreement and is problematic in light of Universal 20 (see Cinque 2005). Thus, the agreement-based derivation of the aligned heterogeneous pattern seems unnecessarily complicated when compared to the concord-as-spellout approach in which the pattern follows naturally. The following sections further illustrate possible outcomes of the concord-asspellout approach when domain maximization is leveraged. 3.2 Downward extension of the concord domain We have seen that the system attempts to maximize the ultimate concord domain first by percolating features as high as possible in the syntax and then by realizing features as low as possible in concord. While domain maximization is typically restricted to extended projection boundaries, the next two derivations demonstrate that the concord domain can be extended downward or upward. The downward homogeneous pattern is exemplified with Russian in (17), where the externally assigned case is realized throughout the construction in a single domain of concord. This pattern is commonly displayed by Slavic numeral constructions in lexical case environments. 162 ANNA GRABOVAC (17) èt-im pjat-i nov-ym student-am this-DAT. PL five.DAT new-DAT. PL student-DAT. M . PL ‘to these five new students’ (Russian) As before, the derivation begins in the syntax. Since (17) is a Russian example, I assume that the numeral agrees for case as we saw in (15). Feature percolation occurs in line with relativized heads—the φ-features are able to percolate through the higher domain since the numeral contributes no φ-features of its own, but the case of the numeral blocks percolation of genitive from the lower domain. (18) KP [ DAT M PL ] [ DAT M PL ] K [ DAT ] [ DAT M PL ] Dem/AP Num KP [ GEN M PL ] [ DAT ] [ GEN M PL ] K [ GEN ] Dem/AP N [ GEN M PL ] Moving out of the syntax, the impoverishment rule in (19) applies to delete genitive in the lower domain. A couple of points are worth noting: for one, I assume that impoverishment rules can refer to the content of dominating nodes. This follows from the setup of the system since features are located on heads and then percolate. Additionally, the semi-lexical status of the numeral plays an important role. While the cross-domain effects of (19) may at first be concerning on the basis of locality, I assume that the semi-lexicality of the numeral creates a more transparent boundary between the two domains than would a fully lexical noun (see Vos 1999). (19) [GEN] cannot occur on nodes dominated by [DAT] Maximization of the concord domain (20) KP 163 [ DAT M PL ] [ DAT M PL ] K [ DAT ] [ DAT M PL ] Dem/AP Num KP [ GEN M PL ] [ DAT ] [ GEN M PL ] K [ GEN ] Dem/AP N [ GEN M PL ] Finally, the content of dominating nodes is spelled out on available terminals in concord. Since case has been eliminated altogether from the lower domain, the dative of the higher domain is spelled out throughout the construction as the most local case. This produces an extended domain of concord for the externally assigned case. (21) KP [ DAT M PL ] K [ DAT ] Dem/AP / DAT PL / Num KP [ DAT ] K Dem/AP N / DAT PL / [ DAT M PL ] 3.3 Upward extension of the concord domain This section will demonstrate how upward extension of the concord domain is obtained in examples such as (22). Here, we see the internally assigned genitive case realized both above and below the numeral. (22) ov-ih pet nov-ih studen-a-ta this-GEN . PL five new-GEN . PL student-PL - GEN . M ‘these five new students’ (BCS) Example (22) comes from BCS. In BCS, numerals are indeclinable which is often taken as evidence for caselessness (Bošković 2006; Stjepanović 2012; Wechsler 164 ANNA GRABOVAC & Zlatić 2003)8 . Thus, I assume that the numeral does not participate in agreement with K. In (23), we see that this lack of agreement allows the internally assigned genitive to percolate upward into the higher domain, along with the φ-features. This is allowed because the numeral contributes no features of its own. Here, we see the semi-lexicality of the numeral and its lack of participation in agreement allowing for upward extension of the concord domain, through percolation higher than usual. (23) KP [ NOM M PL ] [ GEN M PL ] K [ NOM ] [ GEN M PL ] Dem/AP Num KP [ GEN M PL ] [ GEN M PL ] K [ GEN ] Dem/AP N [ GEN M PL ] The result of concord is very straightforward. GEN.PL spells out throughout the construction as the most local set of dominating features9 . (24) KP [ NOM M PL ] [ GEN M PL ] K [ NOM ] Dem/AP / GEN PL / Num KP K [ GEN ] 8 Dem/AP N / GEN PL / [ GEN M PL ] Technically speaking, only the class of higher numerals (5+) is indeclinable in BCS. Lower numerals (2, 3, 4) are subject to speaker variation, but some speakers do decline them. 9 It is logical to wonder whether the genitive case realized by the higher modifier/demonstrative is really the result of upward percolation. Alternatively, some analyses have proposed that this modifier originates below the numeral where it is assigned genitive case, and then subsequently moves (e.g., Caruso 2011). This violates Universal 20, as a demonstrative below the numeral gives rise to an unattested order of elements within a single extended nominal projection. Moreover, moving the demonstrative alone is problematic—Cinque (2005) argues that all movements to derive Universal 20 must include N (see also Abels & Neeleman 2012). Maximization of the concord domain 165 4 Conclusions This paper has argued that concord results from the spell out of features from dominating nodes on available terminals. This concept of concord as spellout was inspired by Norris’s (2014) theory but expanded through the concept of domain maximization. To this end, I argued that the system attempts to maximize the concord domain first by percolating features as high as possible and then by realizing features as low as possible. Domain maximization can also be extended through impoverishment (as in the downward homogeneous pattern) or higher percolation due to the feature specification of the heads involved (as in the upward homogeneous pattern). The resulting system is able to derive a variety of patterns, some of which involve a mismatch between the concord domains and underlying syntactic domains. More broadly, I have argued that concord as spellout provides a simpler alternative to analyses of nominal concord than those based solely on standard agreement mechanisms. While it would be desirable to unify concord and agreement under a single analysis, I have shown that unnecessary complications arise from agreementbased analyses which are avoidable under the concord-as-spellout account. References Abels, K., & A. Neeleman. 2012. Linear asymmetries and the LCA. Syntax 15. 25–74. Abney, S. P. 1987. The English Noun Phrase in its Sentential Aspect. Massachusetts Institute of Technology dissertation. Ackema, P., & A. Neeleman. 2018. Features of person: From the Inventory of Persons to Their Morphological Realization. Linguistic Inquiry Monograph Seventy-Eight. Cambridge, Massachusetts: The MIT Press. Ackema, P., & A. Neeleman. 2020. Unifying nominal and verbal inflection: Agreement and feature realization. In Nominalization: 50 Years on from Chomsky’s Remarks, ed. by A. Alexiadou & H. Borer, 29–52. United Kingdom: Oxford University Press. Anderson, S. R. 1992. A-morphous morphology. Cambridge: Cambridge University Press. Babby, L. H. 1985. Noun phrase internal case agreement in Russian. Russian Linguistics 9. 1–15. Bittner, M., & K. Hale. 1996. The structural determination of case and agreement. Linguistic Inquiry 27. 1–68. Bošković, Ž. 2006. Case checking versus case assignment and the case of adverbial NPs. Linguistic Inquiry 37. 522–533. Carstens, V. 2016. Delayed valuation: A reanalysis of goal features,“upward” complementizer agreement, and the mechanics of case. Syntax 19. 1–42. Caruso, Ð. Ž. 2011. Nominal phrases in Croatian as DPs. In Online proceedings of GLOW in Asia Workshop for Young Scholars, 16–30. Chomsky, N. 1995. The Minimalist Program. Cambridge, Massachusetts: The MIT Press. Chomsky, N. 2000. Minimalist inquiries: The framework. In Step by Step: Essays on Minimalist Syntax in Honor of Howard Lasnik, ed. by R. Martin, D. Michaels, & J. Uriagereka, 89–155. Cambridge, Massachusetts: The MIT Press. Chomsky, N. 2001. Derivation by phase. In Ken Hale: A Life in Language, ed. by K. Hale, 1–52. Cambridge, Massachusetts: The MIT Press. Cinque, G. 2005. Deriving Greenberg’s Universal 20 and its exceptions. Linguistic Inquiry 36. 315–332. Cole, P., G. Hermon, & L.-M. Sung. 1993. Feature percolation. Journal of East Asian Linguistics 2. 91–118. 166 ANNA GRABOVAC Corver, N., & H. van Riemsdijk. 2001. Semi-lexical categories. In Semi-lexical Categories, ed. by N. Corver & H. van Riemsdijk, 1–19. Berlin: Mouton de Gruyter. Danon, G. 2012. Two structures for numeral-noun constructions. Lingua 122. 1282–1307. Frampton, J., & S. Gutmann. 2006. How sentences grow in the mind: Agreement and selection in an efficient minimalist syntax. In Agreement systems, ed. by C. Boeckx, 121–157. Amsterdam: John Benjamins. Franks, S. 1995. Parameters of Slavic morphosyntax. New York: Oxford University Press. Giusti, G. 2008. Agreement and concord in nominal expressions. In The Bantu-Romance Connection: A comparative investigation of verbal agreement, DPs, and information structure, ed. by C. D. Cat & K. Demuth, volume 131 of Linguistik Aktuell/Linguistics Today, 201–237. Amsterdam: John Benjamins. Grabovac, A. 2022. Maximizing the concord domain: Concord as spellout in Slavic. University College London dissertation. Greenberg, J. H. (ed.) 1963. Universals of Language. Oxford: The MIT Press. Grimshaw, J. 2000. Locality and extended projection. In Lexical Specification and Insertion, ed. by P. Coopmans, J. Grimshaw, & M. B. Everaert, 115–134. Amsterdam: John Benjamins Publishing Company. Ionin, T., & O. Matushansky. 2018. Cardinals: The Syntax and Semantics of Cardinal-Containing Expressions. The MIT Press. Klockmann, H. E. 2017. The design of semi-lexicality: Evidence from case and agreement in the nominal domain. Utrecht: LOT. Kramer, R. 2016. The location of gender features in the syntax. Language and Linguistics Compass 10. 661–677. Neeleman, A., & H. van de Koot. 2002. The configurational matrix. Linguistic Inquiry 33. 529– 574. Norris, M. 2012. Towards an analysis of concord (in Icelandic). In Proceedings of the 29th West Coast Conference on Formal Linguistics, 205–213. Cascadilla Proceedings Project. Norris, M. 2014. A Theory of Nominal Concord. University of California, Santa Cruz dissertation. Pesetsky, D., & E. Torrego. 2007. The syntax of valuation and the interpretability of features. In Phrasal and clausal architecture: Syntactic derivation and interpretation. In honor of Joseph E. Emonds, ed. by S. Karimi, V. Samiian, & W. K. Wilkins, Linguistik Aktuell/Linguistics Today, 262–294. Amsterdam: John Benjamins. Polinsky, M. 2016. Agreement in Archi from a minimalist perspective. In Archi: Complexities of agreement in cross-theoretical perspective, ed. by O. Bond, G. G. Corbett, M. Chumakina, & D. Brown, 184–232. Oxford: Oxford University Press. Ritter, E. 1992. Cross-linguistic evidence for number phrase. Canadian Journal of Linguistics/Revue canadienne de linguistique 37. 197–218. Sciullo, A.-M. D., & E. Williams. 1987. On the Definition of Word. Cambridge, Massachusetts: The MIT Press. Stjepanović, S. 2012. Two cases of violation repair under sluicing. In Sluicing: Cross-Linguistic Perspectives, ed. by J. Merchant & A. Simpson, 68–82. Oxford: Oxford University Press. Vos, R. 1999. The grammar of direct and indirect partitive constructions. Tilburg University, The Netherlands dissertation. Wechsler, S., & L. Zlatić. 2003. The Many Faces of Agreement. Stanford, California: CSLI Publications. Proceedings of CLS 58 (2022), 167-182 © Chicago Linguistic Society 2023. All rights reserved. 167 Negative concord in the acquisition of English and German: Some results from a corpus study* Johannes Hein1 , Cory Bill2 , Imke Driemel1 , Aurore Gonzalez3 , Ivona Ilić1 and Paloma Jeretič2 1 Humboldt University of Berlin, 2 Leibniz Centre for General Linguistics (ZAS) Berlin, 3 University of Milan-Bicocca 1 Introduction This paper is concerned with the acquisition of negative indefinites (NIs) by children who are acquiring English or German. In particular, it focuses on children’s non-adult like productions of those indefinites where a sentential negator (not/n’t or nicht) is realized in the NI’s clause. In (standard) English and (standard) German, a NI can express the negation of a sentence on its own. For English, the NI no in (1a) indicates that it is not the case that Emma ate apples. Likewise for German, the NI nichts ‘nothing’ conveys that it is not the case that Milan sees something (2a). When the same sentence contains a sentential negation in addition to the NI, it yields a double negation reading (1b; 2b). Commonly, these sentences require a specific context to be felicitous. (1) a. Emma ate no apples. (= ¬[Emma ate apples]) b. Emma didn’t eat no apples. (= [Emma ate some apples]) (2) a. Milan sieht nichts. (= ¬[Milan sees something]) Milan sees nothing ‘Milan doesn’t see anything.’ German b. Milan sieht nicht nichts. (= [Milan sees something]) Milan sees not nothing ‘Milan sees something.’ In negative concord (NC) languages such as Italian, Hungarian, or Bosnian/Serbian/Croatian (BCS), similar constructions with a sentential negator and a negatively marked indefinite convey only a single semantic negation. Thus, while the sentence in (3) from BCS contains both a negative marker ne and a negative indefinite (a so-called negative concord item, NCI) ništa, the proposition is only negated once. In fact, the sentential negative marker is obligatory in those languages. * This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (LeibnizDream, grant agreement No 856421). 168 HEIN ET AL. Milan ne vidi ništa. Milan not sees nothing ‘Milan doesn’t see anything.’ (3) BCS (Progovac 1994, 40) During language acquisition, children are faced with the task of determining which type of language they are acquiring, a double negation (DN) language like German or English1 , or a NC language like Italian or BCS. Previous work on this matter investigating children’s comprehension and adult artificial language learning indicate a bias towards NC. Specifically, children (3;6–6;5) acquiring English or German strongly favour a single negation interpretation of sentences like (1b) and (2b), respectively (Thornton et al. 2016, Nicolae & Yatsushiro 2020, see also Moscati 2020, Tagliani et al. 2022 on Italian). Likewise, learners acquire an artificial language with NC more easily than one with DN (Maldonado & Culbertson 2021). While the observed bias may result from NC being encoded in the grammar that children entertain at this point in their language development, it could equally well be explained by extra-grammatical factors. That is, as DN arguably has higher processing demands and is pragmatically restricted, children might favour NC readings on account of their immature processing and pragmatic abilities. Investigating children’s production is likely to provide further insights into this bias and may even allow us to adjudicate between these two explanations. If children who acquire a DN language like English or German produce sentences that show erroneous NC, we can more confidently claim that the observed bias is due to a phase in which NC is a proper part of the learner’s grammar. Nicolae & Yatsushiro report two German examples from a search of the Leo corpus on the CHILDES database (MacWhinney 1991, Behrens 2006), and Miller provides ample instances in English from one child, Sarah from the Brown corpus (Brown 1973). However, both studies have limited validity due to their size and the fact that Sarah was exposed to a lot of NC productions in parental speech. We present results of an in-depth corpus study on several English- and Germanacquiring children corroborating Nicolae & Yatsushiro’s and Miller’s promising but restricted findings. Our results show that children produce a substantial number of NC-like constructions during acquisition in both languages. However, there are considerable quantitative and distributional differences between the two languages. The paper is structured as follows. In section 2, we present previous work on the acquisition of NIs in some more detail. In section 3, we lay out our methodology and procedure before presenting the results in section 4. Section 5 contains a discussion of the results and emerging differences between English and German with regard to the initial research question, and section 6 concludes the paper. 1 The status of English as a DN language is notoriously difficult to assess. While there are certainly a number of English varieties that clearly show NC the matter is still under debate for Standard English (Blanchette 2017). For the purposes of this paper, we take standard English to be a nonnegative concord language, as indicated by the adult data in Thornton et al.. Negative concord in the acquisition of English and German 169 2 Previous work on negative indefinites in acquisition As mentioned in section 1, there is some work on the acquisition of NIs in English and German from the perspective of comprehension. Thornton et al. conducted a study on 3;6 to 5;8 years old English-acquiring children’s interpretations of sentences with a NI and a sentential negator like (4). They also tested a control group of English-speaking adults, which were expected to interpret them as conveying the double negation meaning in (4a). The research question was whether children would behave like adults and assign the DN interpretation in (4a), or whether they rather assign the NC interpretation in (4b). (4) The girl who skipped didn’t buy nothing. a. The girl who skipped bought something. b. The girl who skipped bought nothing. Thornton et al. found that children chose a NC interpretation (i.e., (4b)) over a DN interpretation (i.e., (4a)) of such sentences 75% of the time. Adults, on the other hand, preferred such an interpretation in only 18% of cases. Nicolae & Yatsushiro ran a similar study in German, investigating whether children (4;2–6;5 years old) interpret sentence like the one in (5) as conveying a single semantic negation or a double negation. Their results show that the NC interpretation was preferred by the children approximately 95% of the time. The adult controls, however, did so in only about 15% of cases. (5) Der Hase hat kein Gemüse nicht gegessen. The rabbit has no vegetable not eaten. ‘The rabbit ate not vegetables.’ German Supplementing the comprehension experiment, Nicolae & Yatsushiro also conducted a search of the Leo corpus through the CHILDES database (MacWhinney 1991; Behrens 2006). They found that the two-year-old child had produced several NC utterances recorded in this transcript, including (6). (6) keine Glocken nicht da! keine bells not there ‘no bell there!’ German (Leo 2;02, Behrens 2006) We therefore find a clear bias for NC in both English- and German-acquiring children, at least in comprehension. This preference can even be observed in children acquiring NC languages when looking at DN corners of those languages (Moscati 2020, Tagliani et al. 2022). Take Italian as an example. Despite Italian being a (non-strict) NC language it shows DN interpretations of two negative elements in a sentence in certain contexts. One such context is a configuration whereby a preverbal NCI co-occurs with a negative marker as in (7). A second one is fragment answers to negative questions, as in (8). 170 HEIN ET AL. (7) NESSUNO non ha mangiato. nobody not has eaten ‘Nobody didn’t eat’ = ‘Everybody ate.’ (8) Q: Chi non è venuto? who NEG is come ‘Who didn’t come?’ Italian (Penka 2011, 19) Italian A: Nessuno. nobody ‘Nobody.’=‘Nobody didn’t come./Everybody came.’ (Moscati 2020, 170) Moscati conducted an experiment on the latter using a Truth-Value Judgement Task. He found that fragment answers to negative questions were assigned the (expected) DN reading by children (4;6–6;3 years old) only 37% of the time. This contrasts with adults, whose DN responses to the same items almost reached 60%. Given this bias for NC in children, we ask whether children, in particular those who are acquiring a DN language like (standard) English or German, produce a substantial amount of sentences that contain both a sentential negator and a NI, but that are intended to convey only a single semantic negation. A partial yes-answer can already be given for German on the basis of Nicolae & Yatsushiro’s findings. To answer the question for English and for a broader set of German data, we conducted an in-depth corpus search, as detailed in the following section. Indeed, we found a number of NC-like utterances both for English and German but with quantitative and distributional differences between both languages, which will be discussed in section 4. 3 Corpus study We selected corpora from the CHILDES database (MacWhinney 1991) of German and English aiming at a roughly equal total number of utterances in order to keep the two languages comparable. As there are fewer corpora for German than for English, we first gathered data from all available German corpora (as of December 2021) with typically developing children of the right age (0–7). We collected data from 43 German-speaking children aged 0–14;10 across 8 corpora (Caroline, MacWhinney 1991; Grimm, Grimm 2007; Leo, Behrens 2006; Manuela, Wagner 2006; Miller, Miller 1979; Rigol, Lieven & Stoll 2013; Stuttgart, Lintfert 2009; Wagner, Wagner 1985), and from 7 English-acquiring children (5 American English, 2 British English) aged 0;7–7;10 across three corpora (Brown, Brown 1973; MacWhinney, MacWhinney 1991; MPI-EVA-Manchester, Lieven et al. 2009). We checked for both languages whether any of the caregivers’ speech recorded in the corpus contained NC utterances which we took to be indicative of the child acquiring a NC dialect of the language. This led to the exclusion of Sarah from the Brown corpus which is incidentally also the child that the only previous corpus study on the acquisition of NIs in English is based on (Miller 2012). The total number of utterances amounts to 363,028 for German and 328,972 for English. As shown in Negative concord in the acquisition of English and German 171 Fig. 1, the distribution of the utterances across ages is similar in German and English with most data being available between 25 and 38 months of age. Since we did not have any English data beyond the age of 7;10, we only consider German utterances up to that age and ignore utterances produced by older children in this paper. This reduces the amount of relevant utterances in German to 338,407. Figure 1: Distribution of utterances across age From our data we extracted all child utterances that contained at least one NI (no, nobody/no-one, nothing, never for English; kein, niemand, nichts, niemals and relevant inflected forms for German). We found 2,529 such utterances in English and 3,370 in German. Each utterance was coded for the type of NI and whether it cooccurred with sentence negation. We also tagged the position of the NI as either preverbal or postverbal, excluding for German those that appeared in a verb-final (embedded) or verb-initial (polar question) clause and for English those that were found in clauses with Subject-Auxiliary-Inversion (i.e. polar questions). For English, we further coded whether the sentential negation was affixal n’t or phrasal not. Fragment utterances that did not contain a verb, participle, or predicational element and mistaggings (e.g. English no tagged as a response particle) were annotated and excluded from further analysis. This left us with 909 utterances in English and 2,665 in German. All annotations were done by native speakers of the respective language. 4 Results For English, we found a total of 184 NC utterances involving multiple negative elements. This amounts to 20.2 % of the 909 utterances that contain at least one NI. The vast majority, i.e. 178, are utterances where, in addition to a single NI, there also appears a sentence negator like not or n’t (9). In four cases, NC is established 172 HEIN ET AL. between a single NI and never (10). In the remaining two cases, there are two NIs and an additional sentential negator (11). (9) a. We don’t want no gas. (Adam 3;11, Brown 1973) b. No tigers don’t bit you? (Mark 2;08, MacWhinney 1991) c. I don’t care about nothing. (Ross 5;04, MacWhinney 1991) d. He won’t hurt his head never. (Eleanor 2;11, Lieven et al. 2009) e. No one’s not drying him, mum. (Fraser 3;00, Lieven et al. 2009) (10) a. I never have no cookies. (Adam 4;00, Brown 1973) b. I never heard of no flying dinosaur. (Adam 4;06, Brown 1973) c. I promise I’ll never hurt nobody again. (Ross 3;11, MacWhinney 1991) d. I’ll never care about nothin(g). (Ross 5;04, MacWhinney 1991) (11) a. I can’t do nothing with no string. (Adam 4;02, Brown 1973) b. She didn’t use no nothing of paper. (Adam 4;05, Brown 1973) For German, we found a considerably smaller amount of NC utterances, i.e. 45, which amount to 1.7 % of the 2,665 utterances that contain at least one NI. Except for one utterance, where NC is between a NI and nie ‘never’ (13), all other NC utterances contained the sentential negator nicht ‘not’ in addition to a NI (12). (12) a. Kein Gewitter kommt nicht heute. no thunderstorm comes not today ‘There’s no thunderstorm coming today.’ b. Wir haben noch keine Zudecke nich. we have yet no duvet not ‘We don’t have a duvet yet.’ c. Kein Teller kann s net sein. no plate can it not be ‘It can’t be a plate.’ (13) Geht ja nie niemand hin. goes PRT never nobody there ‘Nobody ever goes there.’ German (Leo 2;03, Behrens 2006) (Simone 3;07, Miller 1979) (Sebastian 5;04, Lieven & Stoll 2013) (Sebastian 5;04, Lieven & Stoll 2013) The results are summarized in Table 1. Figs. 2 and 3 show the proportion of NC utterances across age for English and German, respectively. We find that in both languages there is a contiguous span of about ten months (14 in English, 10 in German) where the error proportion is not zero. However, this span lies between the 44th and 58th age month for English but Negative concord in the acquisition of English and German 173 Table 1: Counts of different utterances. Utterance count Language total with NI with NC Proportion (NC/NI) English German 909 2665 184 45 328,972 338,407 20.2 % 1.7 % between the 25th and 35th age month for German. Thus, the error rate in English is highest at around 50 months reaching 35 % while in German the peak appears much earlier at about 27 months and only reaches ca. 3.5 %. Figure 2: Proportion of utterances with NC over age (English) Figure 3: Proportion of utterances with NC over age (German) 174 HEIN ET AL. Interestingly, a similar difference between English and German is found when looking at the overall use of NIs. Fig. 4 shows the proportion of NI-containing utterances across age. While the proportion of NI-utterances starts to sharply rise between 27 and 37 months for the German-acquiring children, it only substantially increases between 40 to 50 months for the English-acquiring children. The rate of NI-utterances then reaches a phase of relative stability at around 1.5 % for German and 1 % for English. The observation that the proportion of NI-utterances is overall lower in English fits with the fact that out of a roughly similar number of utterances (328,972 for English, 338,407 for German), only 909 contained non-fragment NIs (0.28 %) in English while for German there were 2,665 (0.79 %). Figure 4: Proportion of utterances containing an NI over age Turning to the different types of NIs, we distinguished five in English (no, nothing, nobody, noone, never) and four in German (kein, nichts, niemand, niemals). The proportion of each type within the total of NI-containing utterances is given in Fig. 5. For English, no occurs most often (46.2 %), followed by never (26.7 %). Nobody and nothing appear in 13.4 % and 13.1 % of cases while noone is used least frequently (0.5 %). The situation is slightly different for German. As in English, the negative indefinite determiner kein ‘no’ is the most frequent, but at 70 % it is substantially more frequent than its English counterpart. Nichts ‘nothing’ occurs 21.7 % of the time, while niemals ‘never’ takes third place at 6.4 %. Niemand ‘nobody’ appears in only 1.8 % of all utterances that contain a NI. For each type of NI Fig. 6 shows the proportion of NC utterances within all utterances that contain the respective type of NI. For English, we observe that the error proportion is highest with utterances containing nothing (35.8 %) followed by those containing no (28.1 %), noone (20 %), nobody (10.6 %) and never (4.9 %). These proportions are considerably larger than those for German, where the highest error proportion is observed at 2.2 % with kein ‘no’, followed by niemand ‘nobody’ (2 %), niemals ‘never’ (1.2 %) and nichts ‘nothing’ (0.2 %). Selecting only those utterances that exhibit NC, Fig. 7 shows the proportion of Negative concord in the acquisition of English and German 175 Figure 5: Proportion of each NI type in all NI-utterances Figure 6: NC error proportion by type of NI each type of NI present within them. For English, we observe that most of the 184 NC utterances contain the NI no (63.3 %). 22.9 % contain nothing, 6.9 % contain nobody, 6.4 % contain never and 0.5 % contain noone. In German, the respective counterpart of no, i.e. kein, is involved in an even larger proportion of all NC utterances, namely in 91.1 %, meaning that NC almost exclusively occurs with this type of NI. 4.4 % of NC errors (2 utterances) contain niemals ‘never’ and only 2.2 % each contain nichts ‘nothing’ and niemand ‘nobody’ (1 utterance each). Turning to the position of the NI, Fig. 8 shows the overall count of NIs by position with respect to the finite verb, excluding cases where it cannot appear preor postverbally for independent reasons (i.e. verb-initial clauses in German and 176 HEIN ET AL. Figure 7: Proportion of each NI type in all NC errors English, and verb-final ones in German). For English, 393 NIs appear preverbally, 495 occupy the postverbal position. The distribution in German strongly skews to the postverbal position where 2,246 NIs appear. Only 91 occur preverbally. Figure 8: Number of NIs by position Fig. 9 shows the proportion of NC by position of the NI. For English, of all preverbal NIs 5.6 % appear in a NC relation, while postverbal NIs appear in NC expressions 32.5 % of the time. The error proportion for postverbal NIs is therefore about 6 times higher than that for preverbal ones. This difference is statistically significant (p < 10−5 , χ2 -test). This distribution contrasts with what we found in German. While the error proportion within preverbal NIs is 5.5 % – almost exactly Negative concord in the acquisition of English and German 177 as high as that of preverbal NIs in English – it is five times higher than that for postverbal NIs, which is only 1.1 %. Again, this asymmetry is significant (p = 0.0043, Fisher’s exact test). The data are summarised in Table 2. Table 2: Counts by NI position count concord prop. English German preverbal postverbal preverbal postverbal 393 22 495 161 91 5 2,246 24 5.6% 32.5% 5.5% 1.1 % Figure 9: Proportion of NC-errors by NI position A final result concerns the type of negation in English. In the three English corpora we investigated, we found 15,669 utterances negated with the affixal/contracted negator n’t and 6,200 utterances containing the phrasal negator not. Within the utterances that show NC, 156 contain affixal negation n’t whereas 24 are negated using the phrasal negation not. The proportion of NC is higher in utterances with affixal n’t (1 %) than in those with phrasal not (0.4 %). Also, the ratio of affixal to phrasal negation is higher for NC-utterances (6.50) than for the overall corpora (2.53). This difference is significant (p < .00001, χ2 test). At this point, it is worth mentioning some of the things we do not find in the data. Recall that there were a handful of NC-utterances in which the NI cooccured with never (10) or its German counterpart nie(mals) (13) and there were two cases where two NIs appeared in the same clause as an additional sentential negation (11). However, apart from these utterances, there were no utterances that contained two NIs in the same clause, that is, we did not find sentences like (14) or (15). 178 HEIN ET AL. Table 3: Counts and proportions of different negation types in English. all utterances NC utterances % NC (14) a. Nobody drew no crocodile. n’t not n’t/not 15,669 156 6,200 24 2.53 6.50 1% 0.4% b. I gave no picture to no friend. (15) Kein Hase hat keine Karotte gefressen. no rabbit has no carrot eaten ‘No rabbit ate no carrot.’ In fact, for German there was no single utterance with multiple NIs and sentential negation (recall that we searched the corpora for single instances of the NIs; the search would have returned such examples, if they existed). 5 Discussion With regard to the initial question of whether children acquiring English or German produce sentences with NC, the results clearly show that they do so. We find a substantial amount of NC-type errors for both English (184) and German (45). Nonetheless, those errors occur in only a minority of the utterances that contain a NI (20.2 % in English; 1.7 % in German). In fact, most utterances with a NI are adult-like in the sense that there is no additional sentential negative marker present. This suggests that the children never pass through a phase of acquisition where their grammar is equivalent to that of a NC language like Italian or BCS. While NC-errors constitute a minority of NI utterances in both languages, we find two somewhat unexpected differences between English- and German-acquiring children. First, English-speaking children make many more NC-errors than their German counterparts (cf. Table 1). And second, they make them at an older age than the latter (cf. Figs. 2 and 3). English-acquiring children’s production of more NC-utterances is unlikely to be the result of them being exposed to more NC input by their parents/caregivers or by media and pop culture for the following reasons. First, we did not find a single instance of NC in the recorded speech of any of the adults in the respective English corpora. Second, the four American children were recorded in the early 1960s and 1970s when NC variants of English were not as ubiquitously represented in the media and culture as in present times.2 One might hypothesize that the higher error proportion is a consequence of English not being a DN language but rather a NC language or some hybrid type (Blanchette 2017,Robinson & Thoms 2021). Under this view, NC utterances are no longer errors and their greater proportion would corroborate these claims. We would, however, like to suggest an alternative explanation based on two main factors. First, it is much more natural in German to use a NI to negate a sentence than in English. 2 Of course we cannot exclude the possibility that the children were exposed to NC in the speech of their peers or extended family who do not feature in the recordings. Negative concord in the acquisition of English and German 179 The most natural way to express the proposition that it is not the case that Emma ate an apple is (16a) in German whereas (16b) is dispreferred. In contrast, English speakers prefer the equivalent of (16b), i.e. (17b), over the equivalent of (16a), i.e. (17a). (16) a. Emma hat keinen Apfel gegessen. Emma has no apple eaten ‘Emma ate no apple.’ b. Emma hat nicht einen Apfel gegessen. Emma has not an apple eaten ‘Emma didn’t eat an apple.’ (17) a. Emma ate no apple. b. Emma didn’t eat an apple. As a consequence, it is likely that German children receive more input that contains NIs, which enables them to figure out their correct meaning and usage earlier than English children. A second factor is the much richer and more regularized system of negative polarity items (NPIs) in English as opposed to German. In negated sentences, these NPIs most often occur with the overt sentential negator not/n’t. Since there is some competition between NIs and NPIs, at least in object position, children might struggle to distinguish the two. This is corroborated by the observation that at early ages children use NPIs as if they were NIs, that is, as expressing the negation of a proposition (Davidson 2020,Illingworth et al. 2022). Taken together, these factors are not only able to account for the higher error proportion and delayed error peak in English, but also for the observation that English children only start to properly use NIs about 13 months later than German children (cf. Fig. 4). Turning to the intriguing difference in the positioning of NIs between the two languages, we make three key observations: (i) the majority of NIs appear postverbally in German while the distribution is much more balanced in English (cf. Fig. 8); (ii) the error proportion among the preverbal NIs is almost equal between English (5.6 %) and German (5.5 %) (cf. Fig. 9); and (iii) the error proportion is much higher with postverbal NIs (32.5 %) as compared to preverbal ones in English whereas it is much lower (1.1 %) in German (cf. Fig. 9). Observation (ii) suggests that the difficulty with preverbal NIs has the same source in both languages. Observation (i) might be explained if we assume that there is something difficult about preverbal NIs and that, as a consequence, children try to avoid them. That is, as a V2 language, German readily allows word order changes such that a NI, even if it is the subject, can appear postverbally. Thus, the majority of NIs appears in postverbal position. As word order is stricter in English, children simply cannot avoid producing a preverbal NI, if that NI is the subject. That children have difficulty with preverbal NIs and tend to avoid them if possible dovetails with findings by Bill et al.. In an elicitation study based on a picture description task, they observed that adults almost exclusively produced a description with a preverbal NI (18a) whereas children strongly preferred the meaning-equivalent sentence where the NI occurs postverbally (18b). We leave for future research here what exactly makes preverbal NIs difficult for children. 180 HEIN ET AL. (18) a. Keine Katze trägt einen Hut. no cat wears a hat ‘No cat is wearing a hat.’ b. Alle Katzen tragen keinen Hut. all cats wear no hat ‘All cats are not wearing a hat.’ As for observation (iii), the near absence of errors with postverbal NIs in German suggests that whatever causes the preverbal NIs to be difficult for children does not likewise affect the postverbal position. That English children nonetheless show a higher error rate for postverbal NIs as compared to German can be attributed to their abovementioned struggle to distinguish NIs from NPIs. Since NPIs are not licensed in preverbal position in negated sentences they do not compete with preverbal NIs. This explains the asymmetry between error proportions with preverbal and postverbal NIs in English. Finally, the observation that NC-errors occur significantly more often with contracted negation n’t, which is commonly taken to be a syntactic head, than with phrasal negation not could be taken to support the view that there is a link between the head-status of negation and the presence of NC (Zeijlstra 2004,2021; pace Maldonado & Culbertson 2021). 6 Conclusion We presented a thorough corpus study investigating whether and to what extent English- and German-acquiring children produce sentences with apparent NC as suggested by results of previous comprehension and learning studies. While we found a substantial number of such sentences the majority of children’s NI-utterances are adult-like in the sense that they did not show NC. It is therefore unlikely that there is a phase during acquisition where English or German children entertain a NC grammar. The observed differences in quantity and distribution of NIs and NC between English and German can be explained by independent properties that differ between them such as the presence vs. absence of a rich and regular system of NPIs or strict vs. relatively free word order. They therefore do not stringently support claims that English is underlyingly a NC language (Miller 2012, Blanchette 2017, Robinson 2022). References Behrens, H. 2006. The input-output relationship in first language acquisition. Language and Cognitive Processes 21. 2–24. Bill, C., K. Yatsushiro, & U. Sauerland, 2019. Asymmetries in children’s negative determiner production. Poster, Boston University Conference on Language Development (BUCLD) 44. Blanchette, F. 2017. Micro-syntactic variation in american english negative concord. Glossa: a journal of general linguistics 2. 1–32. Brown, R. 1973. A first language: The early stages. Cambridge, MA: Harvard University Press. Davidson, K., 2020. A negative concord stage in negative polarity acquisition. Poster, Boston University Conference on Language Development (BUCLD) 45. Grimm, A. 2007. The Development of Early Prosodic Word Structure in Child German: Simplex Words and Compounds. Potsdam, Germany: University of Potsdam dissertation. Negative concord in the acquisition of English and German 181 Illingworth, C. H., J. W. D. Kang, H. Gibbs, K. Davidson, & R. Feiman, 2022. Negative polarity or negative concord? some children think ‘any’ means ‘no’. Poster, BUCLD 47. Lieven, E., D. Salomo, & M. Tomasello. 2009. Two-year-old children’s production of multiword utterances : A usage-based analysis. Cognitive Linguistics 30. 481–508. Lieven, E., & S. Stoll. 2013. Early communicative development in two cultures. Human Development 56. 178–206. Lintfert, B. 2009. Phonetic and Phonological Development of Stress in German. Stuttgart, Germany: University of Stuttgart dissertation. MacWhinney, B. 1991. The CHILDES project: Tools for analyzing talk. Hillsdale, NJ: Erlbaum. Maldonado, M., & J. Culbertson. 2021. Nobody Doesn’t Like Negative Concord. Journal of Psycholinguistic Research 50. 1401–1416. Miller, K. 2012. Sociolinguistic variation in Brown’s Sarah corpus. In Proceedings of the 36th annual Boston University Conference on Language Development, ed. by A. K. Biller, E. Y. Chung, & A. E. Kimball, 339–348, Somerville, MA. Cascadilla Press. Miller, M. 1979. The Logic of Language Development in Early Childhood. Berlin: SpringerVerlag. Moscati, V. 2020. Children (and some adults) overgeneralize negative concord: The case of fragment answers to negative questions in Italian. University of Pennsylvania Working Papers in Linguistics 26. 169–178. Nicolae, A. C., & K. Yatsushiro, 2020. Not eating kein veggies: negative concord in child German. Linguistic Evidence 2020, Proceedings. Penka, D. 2011. Negative Indefinites. Oxford: Oxford University Press. Progovac, L. 1994. Negative and positive polarity: a binding approach. Cambridge: Cambridge University Press. Robinson, M. 2022. Negative concord as a window into social perception of morphological and syntactic variables. New York: New York University dissertation. Robinson, M., & G. Thoms. 2021. On the Syntax of English Variable Negative Concord. In Proceedings of the 44th Annual Penn Linguistics Conference, p. 195–204. ScholarlyCommons. Tagliani, M., M. Vender, & C. Melloni. 2022. The acquisition of negation in Italian. Languages 7. 116.1–23. Thornton, R., A. Notley, V. Moscati, & S. Crain. 2016. Two negations for the price of one. Glossa: a journal of general linguistics 1. 1–30. Wagner, K. R. 1985. How much do children say in a day. Journal of Child Language 12. 475–487. Wagner, M. 2006. First Steps to Communication: A Pragmatic Analysis. Tübingen: Narr Verlag. Zeijlstra, H. 2004. Sentential negation and negative concord. University of Amsterdam dissertation. Zeijlstra, H., 2021. Types of Negative Concord. DGfS 2021, slides. Proceedings of CLS 58 (2022), 183-198 © Chicago Linguistic Society 2023. All rights reserved. 183 Japanese clausal argument ellipsis and embedded clause periphery Shiori Ikawa 1 , Akitaka Yamada 2 and Yoichi Miyamoto 3 1 Rutgers University and 2,3 Osaka University 1 Introduction An overt extraction from an elided clause is known to yield an ungrammatical sentence, as shown in (1)–(2) (Shinohara 2006; Saito 2007; Sakamoto 2018). (1) a. John-wa [CP Mary-ga hon-o katta-to] omotta-ga John-TOP Mary-NOM book-ACC bought-C thought-but ‘John thought [Mary bought a book], but’ b. Ken-wa [CP ∆ ] omowa-nakat-ta. Ken-TOP think-NEG-PST ‘Ziro didn’t think [CP ∆]’ (Shinohara 2006:2 (2a)) (2) a. Hon-oi John-wa [CP Mary-ga ei katta-to] book-ACC John-TOP Mary-NOM bought-C omotta-si thought-but ‘As for the booki , Taro said that Hanako bought ti but ... .’ ∗ Ken-wa [CP Mary-ga b. zassi-oi ei katta-to] omotta. magazine-ACC Ken-TOP Mary-NOM bought-C thought ‘as for the magazinei , Ziro thought [that Hanako bought ei ].’ (Shinohara 2006:2 (2b)–(2c)) In both examples, the clausal argument (i.e., the to-clause) is elided in the sentences in (b). While such clausal ellipsis itself is generally possible in Japanese as illustrated in (1), the unacceptability of (2b) indicates that movement of a phrase, zassi ‘magazine’ here, out of an ellipsis site is not allowed. This generalization, however, has been challenged by recent studies such as Takahashi (2020) and Otani & Tatsumi (2021) (O&T, henceforth). The primary aim of the current paper is to show that examples from these previous studies are, in fact, not genuine instances of extraction from an elided clause, thus defending the commonly-accepted view that overt extraction is prohibited from an elided clause; we argue that what looks like a counterexample involves movement from an embedded clause periphery, that is, a position outside the ellipsis site, not from inside the ellipsis site. As a secondary goal, we also argue for the need for long-distance licensing of ellipses, à la Aelbrecht (2010). Previous ellipsis literature entertained two opposing viewpoints on the licensing conditions for ellipsis operations. One view is that 184 IKAWA, YAMADA AND MIYAMOTO ellipsis is licensed in a local context to the licensing head (Lobeck et al. 1995; Merchant 2001; Merchant 2004; Saito & Murasugi 1990; Saito et al. 2008). For example, Merchant (2001:60) proposes that the E-feature in a head “issues an instruction to the PF system to skip its complement for purposes of parsing and production,” as schematically shown below. (3) [XP X [YP | licensor ... ] ] ⇒ [XP X [YP ... ] ] The other school claims that ellipsis can be licensed long-distance (Aelbrecht 2010; Takahashi 2020). Thus, YP can be elided by being licensed by the distance head X, as shown in (4).1 (4) [XP X [ZP [WP [YP | licensor ... ] ] ] ] ⇒ [XP X [ZP [WP [YP ... ] ] ] ] We show that the data in question provide cross-linguistic support for the longdistance approach to ellipsis licensing. 2 Data 2.1 Background on extraction out of an elided clausal argument While the data like (1)–(2) suggest that extraction from an ellipsis site is prohibited (Shinohara 2006; Saito 2007; Sakamoto 2018), recent studies have shown that overt extractions from an ellipsis site are possible under some conditions. For example, Otani & Tatsumi (2021), based on the observation from Takahashi (2020:fn.20), have noted that fronting of a phrase out of an ellipsis site is possible if (i) the fronted phrase receives contrastive prosody and (ii) the elided clause is headed by elements such as tokoro and koto. Consider the example below. (5) 1 a. Kono biru kara-wai Taroo-ga [Hanako-ga ei this building from-FOC Taro-NOM Hanako-NOM detekita-tokoro]-o mikaketa-si exited-C-ACC saw-and ‘From this buildingi , Taro saw [Hanako come out ei ] and ... .’ b. ano biru kara-wai Ziroo-ga [Hanako-ga ei that building from-FOC Ziro-NOM Hanako-NOM detekita-tokoro]-o mikaketa. exited-C-ACC saw ‘From that buildingi , Ziro saw Hanako [come out ei ].’ (O&T:3) According to Aelbrecht (2010), long-distance ellipsis licensing is mediated by a head that selects the elided phrase, the head of WP in (4): The licensor agrees with the mediating head, and the mediating head triggers the ellipsis of its complement. As this mediation does not play a crucial role in our analysis, we abstract away from this detail of the mechanism. Japanese clausal argument ellipsis 185 In the preceding clause (see (5a)), the phrase kono biru kara-wa is fronted out of the overt clause headed by tokoro. In (5b), a comparable phrase, ano biru kara-wa, is fronted despite the ellipsis of the clausal argument, provided that the boldface elements receive a contrastive pitch contour.2 At first glance, this appears to be a counterexample to the aforementioned generalization that extraction from an elided clause is categorically disallowed. The contrastive prosody alone is not enough to license this extraction; the choice of the complementizer also matters. This is clear from the unacceptability of example (6) from O&T. The general syntactic pattern in (5b) resembles that in (6), but it contains a different complementizer, to. That is, the extraction in question is not available when the embedded clause is headed by to. (6) a. Kono-biru-kara-wai Taroo-ga [Hanako-ga ei detekita-to] this-building-from-TOP Taro-NOM Hanako-NOM exited-C omotta thought ‘From this buildingi , Taro thought [Hanako came outei ].’ ∗ b. ano-biru-kara-wai Ziroo-ga [Hanako-ga ei detekita-to] that-building-from-TOP Ziroo-NOM Hanako-NOM exited-C omotta. thought ‘From that buildingi , Ziro thought [Hanako came out ei ].’ (O&T: 6 (28)) Note that other heads than to, including the interrogative clausal head ka (see (7)) and the nominalizing C head no (see (8)), allow extraction from an elided clause under the contrastive prosody.3 (7) a. Kono biru-kara-wai Taroo-ga [Hanako-ga ei detekita]-ka this building-from-FOC Taro-NOM Hanako-NOM exited-C kinisiteiru-si, wonder-and ‘From this buildingi , Taro wonders if [Hanako came out ei ], and ...’ Ziroo-ga [Hanako-ga ei detekita]-ka b. ano biru-kara-wai that building-from-FOC Ziro-NOM Hanako-NOM exited-C kinisiteiru. wonder ‘From this housei , Ziro wonders if [Hanako came out i ].’ 2 While Takahashi (2020) and Otani & Tatsumi (2021) refer to this movement as “fronting topicalization,” we have some reservations for their treatment, given the fronted phrase’s focus prosody and the fact that it forms an alternative semantic set. We set aside the question of where the exact final landing site is and refer to this movement simply as ‘fronting.’ 3 Otani & Tatsumi (2021) list some additional complementizers that allow the kind of extraction in question. We assume that they are all treated in a similar manner to tokoro or no. See Section 3.1 for a more detailed discussion about the status of these complementizers. 186 (8) IKAWA, YAMADA AND MIYAMOTO a. Kono biru kara-wai Taroo-ga [Hanako-ga ei detekita-no]-o this building from-FOC Taro-NOM Hanako-NOM exited-C-ACC mikaketa-si, saw-and ‘From this buildingi , Taro saw [Hanako come out ti ] and ... .’ b. ano biru kara-wai Ziroo-ga [Hanako-ga ei detekita-no]-o this building from-FOC Ziro-NOM Hanako-NOM exited-C-ACC mikaketa. saw ‘From that buildingi , Ziro saw Hanako [come out ti ].’ These data, according to Otani & Tatsumi (2021) and Takahashi (2020), suggest that the ban on extraction from an ellipsis site can be lifted, and they base their analysis on them. However, we show in the next section that what look like examples of extraction in (5b), (7b), and (8b) are, in fact, not genuine examples of extraction from an ellipsis site. 2.2 Against the extraction analysis: the binding data To empirically show our point that the fronted phrase is not extracted out of the clausal complement, let us examine the interaction between extraction from an ellipsis site and anaphor binding. If the fronted phrase has moved out of the elided position, it is base-generated inside the embedded clause, as indicated by t in (9). (9) The extraction structure: (To be rejected) [ Fronted Phrasei [ Subj [CP [TP Subj ti V T ] C ] V ] ] Notice that the original position of the fronted phrase is c-commanded by the embedded subject. This structure predicts that reconstruction of the fronted phrase below the embedded subject should be possible. Thus, supposing the reciprocal anaphor ‘otagai’ must be c-commanded by its antecedent (Saito 1992), it is predicted that a reciprocal anaphor inside the fronted phrase should be able to be bound by the embedded subject even when its original position is in the ellipsis site.4 This prediction is not borne out, however, as shown in (10). (10) 4 a. Otagaii -no biru-kara-waj Taroo-ga [Hana-to Mary-gai each other-GEN building-from-FOC Taro-NOM Hana-and Mary-NOM ej detekuru-tokoro]-o mikaketa-si, saw-and exit-C-ACC ‘From each other’si buildingj , Taro saw [Hana and Maryi come out ej ], and ...’ But see Hoji (2006) for the view that otagai does not necessarily need c-command from their antecedents. Note, however, that the pattern in (10)-(11) remains mysterious under such a view. Japanese clausal argument ellipsis b. 187 Ziroo-ga [Hana-to Mary-gai * otagaii -no ie-kara-waj each other-GEN house-from-FOC Ziro-NOM Hana-and Mary-NOM ej detekuru-tokoro]-o mikaketa. exit-C-ACC saw ‘From each other’si housej , Ziro saw [Hana and Maryi come out ej ].’ Here the fronted phrase includes the reciprocal anaphor otagai ‘each other.’ (10a) shows that when the embedded clause is overt, this fronted anaphor can be reconstructed under and bound by the embedded subject. However, the unacceptability in (10b) demonstrates that once the embedded clause is removed, such reconstruction is no longer possible. This suggests that, in the example where the embedded clause is elided, the fronted phrase has never been in the position below the embedded subject in the course of the derivation, contrary to the structure in (9). Otagai cannot be associated with the matrix subject in (10), because, being a reciprocal anaphor, it requires a plural binder and the matrix subject in this example does not. But once this requirement is met, the reading where the fronted anaphor is bound by the matrix subject is always available, even when the embedded clause is elided. To see this, consider the example below. (11) Bill-to Taroo-gai [Ziroo-ga a. Otagaii -no biru-kara-waj each other-GEN building-from-FOC Bill-and Taro-NOM Ziro-NOM ej detekuru-tokoro]-o mikaketa-si, saw-and exit-C-ACC ‘From each other’si buildingj , Bill and Taroi saw [Ziro come out ej ] and,’ Mary-to Hanako-gai [Ziroo-ga b. otagaii -no ie-kara-waj each other-GEN house-from-FOC Mary-and Hanako-NOM Ziro-NOM ej detekuru-tokoro]-o mikaketa. exit-C-ACC saw ‘From each other’si housej , Mary and Hanakoi saw [Ziro come out ej ].’ Just as in (11a), the fronted anaphor in (11b) can be associated with the matrix subjects, Mary and Hanako, even though the embedded clause is elided. The contrast in (10) and (11) suggests that the fronted phrase was originally in a position that is above the embedded subject but below the matrix subject. The same pattern is observed with the reflexive anaphor zibun. When the fronted phrase contains zibun, this anaphor can be bound by either the embedded subject or the matrix subject in (12a). However, when the embedded clause is elided as shown in (12b), the anaphor can be bound by the matrix subject but not by the embedded subject. That is, the fronted phrase can be reconstructed below the matrix subject but not below the embedded subject.5 5 One might wonder if the example with zibun in (12b) truly indicates the availability of reconstruction below the matrix subject, given that zibun is known to allow backward binding (Akatsuka 1976; Oshima 2009; Nishigauchi 2014), as shown in (i). Is it not possible to consider that the anaphor zibun is never reconstructed but gets backward binding by the matrix subject in a similar 188 (12) IKAWA, YAMADA AND MIYAMOTO Taroo-gai [Hana-gaj ek a. Zibuni/j -no biru-kara-wak building-from-FOC Taro-NOM Hana-NOM self-GEN detekuru-tokoro]-o mikaketa-si, exit-C-ACC saw-and ‘From self’si/j buildingk , Taroi saw [Hanaj come out ek ] and ...’ Ziroo-gai [Hana-gaj ek b. zibuni/∗ j -no ie-kara-wak house-from-FOC Ziro-NOM Hana-NOM self-GEN detekuru-tokoro]-o mikaketa. exit-C-ACC saw ‘From self’si/∗j housek , Ziroi saw [Hanaj come out ek ].’ We have so far shown that, only when the embedded clause is elided, the fronted phrase fails to be reconstructed to the position below the embedded subject. Again, this pattern is unexpected, if we assume that the fronted phrase can move out of the elided site, as shown in (9): Why does reconstruction become unavailable only when the embedded clause is omitted? On the other hand, this observation naturally follows if we assume that overt extraction out of the elided clausal argument is in fact not allowed, in line with Shinohara (2006), Saito (2007), and Sakamoto (2018), but against O&T or Takahashi (2020). When the embedded clause is not elided, there are two ways to derive the fronting of a phrase: (i) movement of the fronted phrase out of the embedded clause, more specifically, from the position below the embedded subject, or (ii) some other way. Only the former option would allow the reconstruction of the fronted phrase below the embedded clause. Consider the version in which the embedded clause is marked with an ellipsis. If extraction from the elided position is banned, as has traditionally been pointed out, the former option is no longer available. As only the former option makes reconstruction possible, this accounts for the pattern observed in this section. To sum up, we have shown that the availability of extraction out of the elided clause is a disguised one. The next question is how disguised extraction from an ellipsis site is derived. We will give an answer to this question in the next section. 3 Analysis The previous section has shown that the apparent extraction from the elided clause is a disguised one. Then how are the sentences with such an apparent extraction, way to (i)? (i) [Kyoozyu-ga zibuni -o inyoo-si-ta-koto]-ga Takasii -o utyooten-ni si-ta. professor-NOM self-ACC cite-do-PST-C-NOM Takasi-ACC crazy-COP make-PST “The fact that the professor cited selfi made Takasii crazy.” (Nishigauchi, 2014: 188 (79)) It is worth noting that the backward binding of the type in (i) is known to be licensed only in the clausal subject of some limited predicates, the majority of which are psych predicates (but see Oshima (2009) for another type of backward binding that is not directly relevant for the current data). Because the example in (12) does not fit this description, we believe it involves c-command of zibun by the matrix subject rather than backward binding. Japanese clausal argument ellipsis 189 as in (5b) above, derived? We propose that such sentences can be derived via base generation of the fronted phrase outside of the elided domain. In this section, we develop this idea and show that it can naturally account for the limitation on the extraction depending on the involved C head. Then we discuss how the current analysis further supports the view that ellipsis licensing can be long-distance. 3.1 Embedded clause periphery We start by clarifying our assumption about the periphery of the clause. Developing the split CP analysis originally proposed by Rizzi (1997), Saito (2021) proposes the following structure for Japanese clause periphery, which is parallel to what has been proposed for Italian (Rizzi & Bocci 2017). (13) [ [ [ [ [ [ [ [TP ... ] Fin] Top*] Focus] Top*] Int] Top*] Force/Report] One of the motivations for this structure comes from the observation that Japanese allows several complementizers to be stacked to yield a complex embedded clause periphery, as shown in (14) (Rizzi & Bocci 2017; Yamada 2019); no is analyzed as a realization of Fin, ka realizes Int, and to realizes Force/Report. (14) Taroo-wa Hanako-ni [CP [CP [CP [TP kare-no imooto-ga soko-ni Taro-TOP Hanako-DAT he-GEN sister-NOM there-at i-ta] no] ka] to] tazune-ta. be-PST no ka to asked ‘Taro asked Hanako if his sister was there.’ (Saito 2021:3 (13)) Extending this line of argument, we consider that tokoro is also another lexical item that realizes Fin, the same position as no. The example in (15) supports this view: first, tokoro cannot occcur with no. Second, it is followed by ka and to in exactly the same order as no, as demonstrated below. (15) [[TP isogasi i ] (*no) tokoro (*no) ka to] omoi-masu-ga, (onegaisimasu). busy PRS (*no) tokoro (*no) ka to think-POL-but, I.beg.you ‘I guess you are busy (but I beg you)’ In the rest of this paper, we assume the structure of the clausal periphery in (16), where no and tokoro are the Fin head, ka the Int head, and to the Force/Report head. (16) ForceP/ReportP TopP* Force/Report | Top to IntP Int | Top ka TopP* FocP* TopP* FinP TP Top Fin | no/tokoro FocP 190 IKAWA, YAMADA AND MIYAMOTO 3.2 Base-generation analysis Now we return to the ellipsis data. If the fronted phrase is not base-generated within the elided domain, it should be generated somewhere else. We claim that there is a pro inside the ellipsis site that is coindexed with the fronted phrase outside the ellipsis site, generating the interpretation where the referent of the fronted phrase serves as an argument to the embedded predicate. (17) [ Fronted Phrasei . . . [XP [TP Subj proi V T ] ] ] But if the fronted phrase is base-generated outside the ellipsis site, where exactly does it originate? The examples with anaphors again help us answer this question. The example involving the apparent extraction of a phrase involving an anaphor zibun out of the elided tokoro clause is repeated in (18). (18) Zibuni/∗ j -no ie-kara-wak Ziroo-gai [Hanako-gaj ek self-GEN house-from-FOC Ziro-NOM Hanako-NOM detekuru-tokoro]-o mikaketa. exit-C-ACC saw ‘From self’s i/∗j buildingk , Ziroi saw [Hanakoj come out ek ].’ As previously mentioned, when the embedded clause undergoes clausal argument ellipsis, the anaphor zibun cannot be bound by the embedded subject, but can be bound by the main subject. We already mentioned that the unavailability of binding by the embedded subject naturally follows if extraction out of an ellipsis site (and hence reconstruction into the ellipsis site) is not allowed. What is notable at this point is the availability of the reading where zibun is bound by the matrix subject. This reading suggests that the base-generation position of the fronted phrase is lower than the main subject. Combining the aforementioned split-CP analysis with this observation, we propose that the base-generation position of the fronted phrase is the left periphery of the embedded clause: the fronted phrase is generated as a dangling topic phrase in the embedded clause periphery (Shyu 1995), as illustrated in (19). (19) VP TopP V saw Top’ DP ⌥ ⌅ [self’s building]i ⌃ FinP ⇧ Top Hanako exited proi Fin tokoro As suggested in (17), the interpretation of this dangling topic phrase as the source argument of exit (the embedded verb) is derived from the presence of a null pronoun inside FinP, which is coindexed with the dangling topic. In (18), the FinP headed by tokoro (the boxed phrase in (19)) undergoes ellipsis. Note that the dangling topic Japanese clausal argument ellipsis 191 is not included in this elided position and is thus accessible for movement. More specifically, this phrase gets moved to the Spec position of the matrix left periphery, yielding the surface word order. 6 This analysis naturally deduces the effect of the C head choice under the previously stated clausal periphery assumption. Remember that, unlike an elided tokoroclause and other elided clauses, an elided to-clause does not allow extraction, even with the contrastive prosody. The example is repeated in (20). (20) a. Kono-biru-kara-wai Taroo-ga [Hanako-ga ei detekita-to] this-building-from-TOP Taro-NOM Hanako-NOM exited-C omotta-si, thought-and ‘From this building, Taro thought [Hanako came out’ b. ∗ ano-biru-kara-wai Ziroo-ga [Hanako-ga ei detekita-to] that-building-from-TOP Ziro-NOM Hanako-NOM exited-C omotta. thought ‘From that building, Ziro thought [Hanako came out ei ].’ (O&T:6 (28)) 6 Our analysis that the extraction out of an ellipsis site is a disguised one and extraction out of an ellipsis site is generally banned contrasts with Takahashi (2020), who first brought up the data in (5) and concluded that movement out of the ellipsis site is possible. In relation to this, it is worth noting that Takahashi (2020) additionally (in fact, mainly) deals with the observation that clefting out of an elided clausal argument in Japanese is possible as shown below. (i) a. Harry-ga [Ginny-ga t detekuru no]-o mokugekisita no]-wa kono biru kara] Harry-NOM [Ginny-NOM exit C ]- ACC witnessed C - TOP this building from da. be ‘It was from this building that Harry witnessed Ginny came out.’ b. Ron-ga [∆] mokugekisita no-wa ano-biru-kara da. Ron-NOM testified C - TOP that-building-from be ‘lit. It was from that building that Ron witnessed [∆].’ (Takahashi 2020:56) Given our conclusion that overt extraction is not available in general, we expect (i) to not involve overt movement out of an extraction site. In fact, theoretically speaking, the current analysis can capture the cleft data in (i) as well. The clefted phrase can be base-generated in the left periphery of the elided clause, and the ellipsis site includes a pro coreferrring to it, rather than its trace. (ii) [Harry-ga [ti [Ginny-ga proi detekuru no]]-o mokugekisita no]-wa [kono biru Ginny-NOM exit C - ACC witness C - TOP this building Harry-NOM kara]i ] da. from be ‘It is from this building [that Harry witnessed Ginny came out ti ].’ However, since the syntax of the cleft construction is not well understood at this point, we leave this issue for future research, merely noting that the examples in (i) do not pose any problem for the current proposal. 192 IKAWA, YAMADA AND MIYAMOTO Following (16), the structure in (21) is assumed for the ellipsis of the to-clause. Now recall that, for our account of the apparent extraction in the tokoro-sentence in (18), it was crucial that the Spec of Top remains outside the ellipsis site. But, given that to is the head of ReportP and ReportP is the highest head in the clause periphery, if the to-clause (i.e., ReportP, the boxed phrase) is entirely elided in (21), the Spec of TopP is also included in the ellipsis site. This means that the dangling topic is no longer accessible for extraction, and hence the fronting in (20) is disallowed. (21) VP ⌥ ⌅ ⌃ ⇧ ReportP TopP* Top IntP TopP* Int Top FocP TopP* Report | to V omow ‘think’ Foc DP/PP fronted phrase TP FinP Top Fin | tokoro Note that this analysis also captures the observation that other complementizers such as ka and no behave like tokoro, unlike to (see (7)–(8)): They are in Int and Fin, respectively, and hence should allow ellipsis leaving TopP. We have seen so far that there has to be an option to base-generate the fronted phrase below the matrix subject. But one might wonder whether there is an additional option to base-generate the fronted phrase somewhere higher. Is it, for example, possible to base-generate it in the matrix’s left periphery from the first, as shown in (22)? (22) [ Fronted Phrasei [TP Subj [emb-clause [TP Subj proi V T ] ] V] Notice that the ban on an apparent extraction out of a to-clause suggests that the base-generation of the fronted phrase in an embedded clause periphery instead of the base-generation in a matrix clause periphery is not only possible, but is required: If a phrase coindexed with a pro inside the elided embedded clause can be basegenerated in the matrix clause periphery as shown in (22), the apparent extraction should be possible no matter the size of the embedded clause. We will come back to this issue in Section 3.4 after we discuss how the ellipsis is licensed. 3.3 Ellipsis licensing The previous section has proposed that when there is a disguised extraction out of a tokoro/no-clause or a ka clause, FinP or IntP are omitted, leaving the TopP above Japanese clausal argument ellipsis 193 them. A natural question that arises is how the ellipsis of FinP or IntP is licensed in such examples. More specifically, what is the head that licenses the ellipsis of these projections? Recall that, under our analysis, the clausal argument ellipsis sometimes targets the complement of the Top head (FinP/IntP), while it sometimes targets RepP, which is the complement of the matrix verb. If we accept Merchant’s (2001) view that ellipsis is only licensed in a sister-node relationship, we must assume that the licensorhood is distributed arbitrarily: When a tokoro-clause or ka-clause is elided with apparent extraction, the Top is the licensor. On the other hand, when a toclause is elided (without apparent extraction), the licensor is V. While this type of analysis can capture the observation, it is just a mere restatement of descriptive generalization, adding an unnecessary complexity to the theory without making any significant contribution to our understanding of natural language. If, on the other hand, we take the long-distance licensing approach as advocated by Aelbrecht (2010), we can provide a more principled explanation. Regardless of the difference among FinP, IntP, and ReportP, what licenses the ellipsis is always the matrix verb that selects for the embedded clause. It elides the closest phrase whose head is overtly realized. For example, in the case of the to-clause, the ReportP is elided, because it is the highest and closest to the embedding verb among the phrases whose heads are pronounced. But in the other cases (e.g., the ka-clause and the tokoro-clause with apparent extraction), it skips the ReportP and TopP because, in a ka-clause and a tokoro-clause, the heads of ReportP and TopP are not pronounced and elides the IntP and FinP, respectively. (23) VP . ( .. ) TopP V (=licensor) DP Top FinP [this building]i (=ellipsis site) Hanako exited proi Fin tokoro The support for the idea that the matrix verb is always the licensor of the ellipsis also comes from the empirical domain. First, if Merchant (2001) is right and anything (for example, a head of TopP) can freely license ellipsis, it is predicted that the FinP in the matrix clause can be elided, as grammatically as the FinP in the embedded clause. However, consider the example in (24b): (24) Context: Taroo is looking for his textbook. You say to Taroo: a. Saikin kenkyuusitu-no mono-no ooku-wa [Hanako-ni recently office-GEN stuff-GEN most-TOP Hanako-by sute-rare-tei-te], dispose-PASS-ASP-te 194 IKAWA, YAMADA AND MIYAMOTO ‘Recently, most of the stuff in the office [have been disposed of by Hanako and’ b. ∗ kimi-no kyookasyo-wa [Hanako-ni sute-rare-tei-ru]. 2sg-GEN textbook-TOP Hanako-by dispose-PASS-ASP-PRS ‘your text book has been disposed by Hanako.’ Given the topic marker -wa and the context, it is clear that kimino kyookasyo ‘your textbook’ is a topic, so this phrase is considered in the Spec of TopP. Nonetheless, the sentence is illicit; Top itself is not capable of licensing an ellipsis of its complement. The second piece of evidence comes from an adjunct use of the tokoro-clause. Just as with the English when-clause, the tokoro-clause can be placed in an adjunct position as well as in a complement position, as exemplified in (25b). (25) a. Taroo-wa nakidasita. Taro-TOP cried out ‘Taro cried out.’ b. Taroo-wa [Hana-ga kono-biru-kara detekita-tokoro] nakidasita. Taro-TOP Hana-NOM this-building-from exited-C cried out ‘Taro cried out [when Hana came out of this building].’ Being an adjunct clause, the FinP in this clause is not c-commanded by the matrix verb cry out, in stark contrast to the examples with apparent extraction out of an elided tokoro-clause we examined above. If Top itself is the ellipsis-licensor, ellipsis of a tokoro clause with a dangling topic is predicted to be possible, irrespective of the complement/adjunct distinction; in contrast, if the embedding V is the ellipsis licensor, it is expected that an extraction is only possible from a complement tokoro-clause. Thus, we have a good testing ground to see whether the ellipsis is licensed by the Top head or by the matrix predicate. With this in mind, consider the following examples: (26) a. Kono biru-kara-wai Taroo-ga [Hana-ga ei detekita-tokoro] this building-from-FOC Taro-NOM Hana-NOM exited-C nakidasita-si, cried out ‘From this building, when Hana came out proi , Taro cried out.’ Ziroo-gai [Hana-ga ei detekita-tokoro] b. * ano biru-kara-wai that building-from-FOC Ziro-NOM Hana-NOM existed-C nakidasita. cried out ‘From that building, when Hana came out proi , Ziro cried out.’ This example in (26a) shows that fronting of kono-biru-kara ‘from this building’ out of the tokoro-clause is available when it is not elided. However, as indicated by (26b), when the tokoro-clause is elided, an extraction is illicit. This suggests that the ellipsis is sensitive to the adjunct/complement distinction, thus supporting Japanese clausal argument ellipsis 195 the view that the licensor of ellipsis is the embedding V that selects the embedded clause.7 The ellipsis licensing has to be long-distance, supporting the view that ellipsis licensing can occur long-distance via Agree. 3.4 The existence of a structure in narrow syntax Now that we have seen our analysis thoroughly, let us come back to our main claim: Overt extraction out of an elided clausal argument is in fact not possible, in spite of the presence of apparent extraction. This point is worth noting given that the possibility of extraction from an ellipsis site has been used as an argument for the existence of the structure (Depiante 2000; Merchant 2013; Sakamoto 2018): If an overt extraction is possible, there should be a structure in the narrow syntax. If not, there should not be one. Especially in the context of Japanese argument ellipsis, Sakamoto (2018) observes that Japanese bans overt extraction but allows covert extraction from an ellipsis site. He claims that the ellipsis site does not have a structure in the narrow sytnax but has one at LF, supporting the LF-copy analysis of ellipsis (Oku 1998). Taking these views at face value, the current observation seems to suggest that the elided clausal argument does not have a syntactic structure in the narrow syntax. Aelbrecht (2010), however, notes that it is too simplistic to say that the impossibility of extraction indicates the correctness of LF copy analysis. According to her analysis, the ellipsis site is frozen once the licensor of the ellipsis is introduced into the structure. Thus, even if the ellipsis site has a structure in narrow syntax, extraction out of it is possible only if there is a phase head, hence an escape hatch for the extracted phrase in between the licensor and the elided phrase, as in (27). (27) ellipsis YP licensor escape hatch Y phase head .. . ellipsis site Now the question is whether this idea from Aelbrecht can capture the current case or not. The answer depends on which head in the C domain serves as a phase 7 Regarding the idea that the ellipsis of the tokoro-clause is licensed by the matrix predicate that selects the tokoro-clause, one might wonder about the following example: (i) a. [Hana-ga deteki-ta-tokoro], Taroo-ga nai-ta. Hana-NOM exit-PST-C Taro-NOM cry-PST ‘When Hanako came out, Taro cried.’ b. Ziroo-mo nai-ta. Ziro-too cry-PST ‘Ziro cried, too.’ While this example allows the interpretation that Ziroo also cried when Hana came out, we consider this interpretation to come from the pragmatic effect instead of the ellipsis of the tokoro-clause. 196 IKAWA, YAMADA AND MIYAMOTO head and we do not attempt to give a full answer in this paper. Our data, on the other hand, appear to be at least compatible with the possibility that an elided clausal argument has a structure in the narrow syntax to be deleted at PF: If the phase head is located relatively low in the C domain, such that it is included in the ellipsis site for any size clausal argument, there would no phase edge as an escape hatch available in between the ellipsis site and the matrix predicate. In this case, we can capture the ban on extraction even under the view that there is a full-fledged syntactic structure for the ellipsis site in the narrow syntax. In fact, the view that the ellipsis site is syntactically active, but only until the ellipsis licensing head merges, can answer the question we left unanswered above: At the end of Section 3.2, based on the ban on extraction out of a to-clause, we argued that the fronted phrase in the apparent extraction examples has to be basegenerated in the periphery of the embedded clause and cannot be base-generated in the periphery of the matrix clause directly. Recall that, under our analysis, there is a pro in the ellipsis site that is coindexed with the fronted phrase. Suppose here that pro in the ellipsis site needs to establish a syntactic relationship with its antecedent. Now, if the freezing occurs when the ellipsis licensor enters the structure, the antecedent has to enter the structure before the merger of the ellipsis licensing head (i.e., the matrix predicate). This is what happens when the fronted phrase is base-generated in the specifier of the embedded TopP. However, establishment of a binding relationship is no longer possible if the fronted phrase is base-generated in the matrix’s left-periphery, which is far above the matrix predicate. Again, giving a full-fledged answer with regard to the existence of syntactic structure is out of the scope of this paper, but this observation seems to lend some support to the PF deletion approach to the clausal argument ellipsis. 4 Conclusions The disguised examples of overt fronting out of an elided clause served as the study’s starting point. Based on such examples, Takahashi (2020) and Otani & Tatsumi (2021) observed that fronting some phrases from an elided clause is permitted. On the other hand, we have argued that this is not a genuine example of an extraction out of an elided clause by showing that this is base-generated in the embedded clause periphery, giving support to the generalization that an overt extraction out of an elided clause is not possible (Shinohara 2006; Saito 2007; Sakamoto 2018). More specifically, we claimed that the ellipsis in question leaves the Spec of TopP in the embedded clause, where the moved phrase is externally merged. We further showed that this gives support to the availability of long distance licensing of ellipsis. Acknowledgements This work is supported by the following grants: The JSPS Core-to-Core Program, A. Advanced Research Networks “International Research Network for the Human Language Faculty" (#JPJSCCAJ221702004) and JSPS KAKENHI Grant Numbers JP 22K20030 and JP 22K00507. The usual disclaimers apply. Japanese clausal argument ellipsis 197 References Aelbrecht, L. 2010. The syntactic licensing of ellipsis, volume 149. John Benjamins Publishing. Akatsuka, N. 1976. Reflexivization: A transformational approach. In Japanese generative grammar, 51–116. Brill. Depiante, M. A. 2000. The syntax of deep and surface anaphora: A study of null complement anaphora and stripping/bare argument ellipsis. University of Connecticut. Hoji, H. 2006. Otagai. Theoretical and Empirical Studies of Reference and Anaphora—Toward the establishment of generative grammar as an empirical science, a report of the Grant-in-Aid for Scientific Research (B), Project 126–138. Lobeck, A. C., et al.. 1995. Ellipsis: Functional heads, licensing, and identification. Oxford University Press on Demand. Merchant, J. 2001. The syntax of silence: Sluicing, islands, and the theory of ellipsis. Oxford University Press. Merchant, J. 2004. Fragments and ellipsis. Linguistics and philosophy 27. 661–738. Merchant, J. 2013. Diagnosing ellipsis. In Diagnosing syntax, 537–542. Nishigauchi, T. 2014. Reflexive binding: Awareness and empathy from a syntactic point of view. Journal of East Asian Linguistics 23. 157–206. Oku, S. 1998. A theory of selection and reconstruction in the minimalist perspective. University of Connecticut. Oshima, D. 2009. A note on backward anaphora in japanese. Ibaraki Daigaku Ryuugaku-sentaa Kiyou (Ibaraki University International Student Center bulletin) 7. 79–89. Otani, S., & Y. Tatsumi. 2021. The pf-deletion approach to missing nominal clauses. Presentation at Ellipsis workshop. Rizzi, L. 1997. The fine structure of the left periphery. In Elements of grammar, 281–337. Springer. Rizzi, L., & G. Bocci. 2017. Left periphery of the clause: primarily illustrated for Italian. In The Wiley Blackwell Companion to Syntax [second edition], ed. by M. Everaert & H. van Riemsdijk, 1–30. Hoboken, NJ: John Wiley & Sons. Saito, M. 1992. Long distance scrambling in japanese. Journal of East Asian Linguistics 1. 69–118. Saito, M. 2007. Notes on East Asian argument ellipsis. Language Research 43. 203–227. Saito, M. 2021. wh-phrases without quantificational particles. Saito, M., T.-H. J. Lin, & K. Murasugi. 2008. N’-ellipsis and the structure of noun phrases in chinese and japanese. Journal of East Asian Linguistics 17. 247–271. Saito, M., & K. Murasugi. 1990. N’-deletion in japanese: A preliminary study. Japanese/Korean Linguistics 1. 285–301. Sakamoto, Y. 2018. Overtly empty but covertly complex. Linguistic Inquiry 50. 105–136. Shinohara, M. 2006. On some differences between the major deletion phenomena and japanese argument ellipsis. MS. Nanzan University. Shyu, S. 1995. The syntax of focus and topic in Mandarin Chinese. University of Southern California dissertation. Takahashi, D. 2020. Derivational argument ellipsis. The Linguistic Review 37. 47–74. Yamada, A. 2019. The syntax, semantics and pragmatics of Japanese addressee-honorific markers. Georgetown University dissertation. Proceedings of CLS 58 (2022), 199-211 © Chicago Linguistic Society 2023. All rights reserved. 199 Covert raising and finite ECM in Japanese, Turkish and Uyghur Robin Jenkins University of Connecticut 1 Introduction This paper investigates a point of cross-linguistic variation regarding the structural position of embedded subjects in finite exceptional case-marking (FECM) constructions in Japanese, Turkish, and Uyghur. As in a number of languages, Japanese, Turkish and Uyghur instantiate a configuration where the subject of an embedded finite clause may receive Case from a higher clause head. While embedded subjects may also receive Case from within the embedded clause in all three languages, I observe a split regarding the structural position of embedded subjects in FECM configurations. In Japanese FECM, embedded subjects must always raise out of the embedded clause to a matrix position. If they do not do so overtly, then they do it covertly. Meanwhile, in Turkish and Uyghur FECM no obligatory raising of embedded subjects is observed. While embedded subjects may optionally raise overtly, i.e. scramble, to the higher clause, no covert raising is observed. I show that this split with respect to obligatory raising of embedded subjects correlates with another difference between Japanese, and Turkish and Uyghur FECM: the base-generation position of embedded subjects. In Japanese FECM, the embedded subject is base-generated at the edge of the embedded clause, SpecCP; in Turkish and Uyghur the embedded subject is base-generated lower in the embedded clause and raises to SpecCP. I propose that it is due to this difference regarding base-generation positions that raising is forced in Japanese FECM constructions. Specifically, I propose that due to a general constraint on DP-chains regarding the structural position of DP-copies, raising is forced in Japanese FECM contexts, but not in Turkish and Uyghur FECM contexts. 2 Finite ECM in Japanese, Turkish and Uyghur As shown below, in Japanese (1), Turkish (2), and Uyghur (3) the case morphology of subjects in embedded finite CPs alternates between NOM and ACC. (1) John-ga [CP Bill-ga/o tensai da to] omotteiru John-NOM Bill-NOM / ACC genius COP C think.PROG ‘John thinks that Bill is a genius.’ (Tanaka 2002) diye] duymuş (2) Pelin [CP Can-∅/ı öldü hear.EV. PST.3 SG Pelin Can-NOM / ACC die.PST.3 SG C ‘Pelin heard that Can died.’ 200 ROBIN JENKINS (3) Mahinur [CP Adil-∅/ni polu eti dep] oylaydu Mahinur Adil-NOM / ACC polu make.PST.3 SG C think.NPST.3 SG ‘Mahinur thinks that Adil made polu.’ While the cases where the embedded subject surfaces with NOM illustrate a typical embedded subject configuration (where the subject DP is licensed by the embedded T and located in SpecTP), the instances where the subject is marked with ACC are generally thought to be instances of ECM into a finite CP, i.e. FECM.1 Here, the embedded subject DP is licensed by the matrix v. In Japanese (4a), Turkish, (5a), and Uyghur (6a), FECM is not possible with a passivized matrix predicate, as indicated by the impossibility of ACC on the embedded subject (while NOM is still available (4b,5b,6b)). (4) (5) (6) a. *[CP Bill-o tensai da to] omowareta Bill-ACC genius COP C think.PASS . PST ‘It was thought that Bill is a genius.’ b. [CP Bill-ga tensai da to] omowareta a. *[CP Pelin-ı öldu diye] biliniyormuş Pelin-ACC die.PST.3 SG C know.PASS . PROG . EV. PST.3 SG ‘It was thought that Pelin died.’ b. [CP Pelin-∅ öldu diye] biliniyormuş a. *[CP Mehmet-ni kitap-ni oqudi dep] oylaildi Mehmet-ACC book-ACC read.PST.3 SG C think.PASS . PST.3 SG ‘It was thought that Mehmet read the book.’ b. [CP Mehmet-∅ kitap-ni oqudi dep] oylaildi Assuming v does not assign [ACC] under passivization, the impossibility of ACC on the embedded subjects in (4a,5a,6a) indicates that the matrix v is the source of [ACC]-assignment. Moreover, embedded subjects receive [ACC] while located in the embedded CP. In (7,8,9), the ACC embedded subjects may linearly follow temporal adverbs that modify the embedded clause. (7) John-ga [CP mada Mary-wo kodomo da to] omotta COP C think. PST still Mary-ACC child John-NOM ‘John thought that Mary was still a child.’ (Hiraiwa 2005) diye] girdi Mert-i sınava (8) Pelin [CP dün yesterday Mert-ACC exam.DAT enter.PST.3 SG C Pelin biliyor know.PRS .3 SG ‘Pelin thinks that yesterday Mert took an exam.’ (Şener 2008) 1 For Japanese see: Kuno (1976), i.a.. For Turkish see: George & Kornfilt (1981), i.a.. For Uyghur see: Major (2021), i.a.. Covert raising and finite ECM 201 (9) Mahinur [CP ete Ayghül-ni ketidu dep] oylaydu Mahinur tomorrow Ayghül-ACC leave.NPST.3 SG C think.NPST.3 SG ‘Mahinur thinks that tomorrow Ayghül will leave. Assuming that in (7,8,9) the temporal adverbs adjoin to a phrase within the embedded CP, the fact that ACC-marked embedded subjects may follow such adverbs indicates that they can receive [ACC] while located in the embedded CP.2 Evidence from pronoun binding indicates that the position where the embedded subject DP receives [ACC] is SpecCP of the embedded CP. For Japanese (10), binding between a matrix subject DP and embedded ACC-subject pronoun is impossible (binding is possible with NOM-subject pronoun in non-ECM constructions). (10) Tarooi -ga [CP karei -ga/*o tensai da to] omotteiru Taro-NOM 3 SG - NOM / ACC genius COP C think.PROG ‘Taroi thinks that hei is a genius.’ (Taguchi 2015) The absence of binding between the matrix DP and the ACC-subject DP in (10) indicates that the ACC-marked subject is in the same binding domain as the matrix DP (and higher than NOM-subjects which may be bound). Assuming that CPs delineate binding domains, where only their edges, i.e. SpecCP, may be part of a higher domain, the impossibility of pronoun binding in (10) indicates that ACC-marked DP is located in SpecCP, where it receives [ACC]. Similar tests with reciprocal binding indicate that in Turkish and Uyghur ACCmarked subject DPs receive [ACC] in SpecCP as well. As shown for Turkish (11) and Uyghur (12), when the embedded subject DP is a reciprocal anaphor, it must be ACC-marked. Reciprocal binding is impossible with NOM-subjects here.3 (11) (12) viski-yi içti diye] Bizi [CP birbirimizi -*∅/i 1 PL each.other-NOM / ACC whiskey-ACC drink. PST.3 SG C biliyoruz know.PRS .1 PL ‘Wei think of each otheri to have drunk the whiskey.’ (Şener 2008) ketidu dep] [Mehmet bilen Ayghül]i [CP birbirii -*∅/ni each.other- NOM / ACC leave.NPST.3 SG C Mehmet with Ayghül oylaydu think.NPST.3 SG ‘[Mehmet and Ayghül]i think each otheri will leave.’ Given that the reciprocal must be in the same domain as its binder, (11,12) show that the ACC-subject DPs are located in SpecCP here as well. While in all three languages, the ACC subject may remain inside the embedded CP, they may also scramble to a matrix position. In Japanese (13), Turkish (14), and 2 Note that the above facts in (7,8,9) also show that ACC-subjects are not base-generated outside the embedded CP, i.e. proleptic arguments (Salzmann 2017). 3 Due to interfering factors Japanese reciprocal otagai ‘each other’ cannot be used as a diagnostic here (for discussion, see Nishigauchi (1992)). 202 ROBIN JENKINS Uyghur (15), the ACC-subject may precede the manner-adverb, which presumably is adjoined to the matrix vP/VP, indicating that the ACC-subject can scramble to a matrix position. (13) John-ga Billi -o orokanimo [CP ti tensai da to] omotteiru John-NOM Bill-ACC stupidly genius COP C think.PROG ‘John stupidly thinks Bill is a genius.’ (Tanaka 2002) diye] duymuş (14) Pelin Cani -ı sık sık [CP ti öldü hear.EV. PST.3 SG die.PST.3 SG C Pelin Can-ACC often ‘Pelin often heard that Can died.’ dep] oylaydu (15) Mahinur Adili -ni da’im [CP ti polu eti polu make.PST.3 SG C think.NPST.3 SG Mahinur Adil-ACC often ‘Mahinur often thinks that Adil made polu.’ In light of the above facts, I conclude that Japanese, Turkish and Uyghur FECM constructions instantiate the configuration in (16) ((Taguchi 2009; Şener 2008), cf. (Hiraiwa 2005)). As shown below, in FECM constructions the embedded subject DP is in SpecCP where it receives [ACC] from the matrix v and may optionally scramble out (see Taguchi (2015) for arguments that this movement is scrambling). (16) [vP DP[uK: ACC] [v0 [VP [CP DP[uK: ACC] [C’ . . . C] ] V ] v [ACC] ] ] 3 Covert raising The previous section showed that Japanese, Turkish, and Uyghur instantiate FECM configurations, where the embedded subject DP is licensed with [ACC] by the matrix v while located in SpecCP. In this section, I observe a split between Japanese FECM, and Turkish and Uyghur FECM regarding the position of the ECMed subject. I argue that in Japanese FECM, the ECMed subject actually always raises out of the embedded CP, in particular, if it does not overtly, then it must covertly.4 Conversely, in Turkish and Uyghur FECM there is no obligatory raising, or covert raising, of the ECMed DP. In Japanese, embedded CPs, with a NOM-subject, may be extraposed as in (17a). In FECM constructions, CP-extraposition is blocked in cases where the ECMed subject has overtly moved out of the CP, as in (17b). Interestingly, CPextraposition is also not possible in cases where the ECMed DP has not scrambled out the embedded CP, as in (17c). (17) 4 a. John-ga ti itta [CP Bill-ga sono hon-o katta to]i John-NOM said Bill-NOM that book-ACC buy C ‘John said that Bill bought the book.’ See Taguchi (2009, 2015) for further arguments to this effect for Japanese. Covert raising and finite ECM 203 b. *John-ga Billj -o ti itta [CP tj sono hon-o katta to]i John-NOM Bill-ACC said that book-ACC buy C ‘John said that Bill bought the book.’ (Tanaka 2002) c. *John-ga ti itta [CP Bill-o sono hon-o katta to]i Given that extraposed CPs are islands (Ross 1967; Johnson 1985), the impossibility of scrambling the ECMed DP in (17b) is unsurprising in this respect. The fact that (17c) patterns with (17b) (but not (17a)), I argue, indicates that the ECMed DP must move out of the embedded CP in (17c) as well. More generally, the above facts suggest that even in cases where the ECMed DP appears overtly in the embedded CP, i.e. (17c), covert movement must occur (cf. (17a)). Thus, the general impossibility of CP-extraposition for Japanese FECM constructions.5 Contrasting with Japanese, Turkish and Uyghur FECM constructions show no evidence that the ECMed DPs covertly raise. As in Japanese, CP-extraposition is possible with NOM-subjects, as in (18a) for Turkish and (19a) for Uyghur. Further, as with Japanese, extraposed CPs disallow overt movement of the ECMed DP out of the embedded CP in Turkish (18b) and Uyghur (19b). Interestingly, unlike Japanese, CP-extraposition is possible in cases where the ECMed DP overtly remains in the embedded CP in both Turkish (18c) and Uyghur (19c). (18) (19) 5 a. Berk ti duymuş [CP Mete sınfta kaldı diye]i Berk hear.EV. PST.3 SG Mete class.LOC fail.PST.3 SG C ‘Berk apparently heard, that Mete flunked the class.’ b. *Berk Metej -yi ti duymuş [CP tj sınfta kaldı Berk Mete-ACC hear.EV. PST.3 SG class.LOC fail.PST.3 SG diye]i C ‘Berk apparently heard, that Mete flunked the class.’ c. ?Berk ti duymuş [CP Mete-yi sınf-ta kaldı diye]i (Şener 2008) a. Mahinur ti oylaydu [CP Adil polu eti dep]i Mahinur think.NPST.3 SG Adil polu make.PST.3 SG C ‘Mahinur thinks, that Adil made polu.’ b. *Mahinur Adilj -ni ti oylaydu [CP tj polu eti dep]i Mahinur Adil-ACC think.NPST.3 SG polu make.PST.3 SG C ‘Mahinur thinks, that Adil made polu’. Note that extraposed CPs are also islands for wh-in-situ in Japanese, as in (ia,b). i. a. *John-wa ti itta ka [Mary-ga nani-o katta to]i ? John-TOP said Q Mary-NOM what-ACC bought C ‘What did John say that Mary bought?’ b. *Kimi-wa ti omoimasu ka [Hanako-ga naze hashitteiru to]i ? 2 SG - TOP think Q Hanako-NOM why running C ‘For what reason/why do you think that Hanako is running?’ 204 ROBIN JENKINS c. Mahinur ti oylaydu [CP Adil-ni polu eti dep]i Unlike Japanese FECM, the possibility of CP-extraposition in (18c,19c) where the ECMed DP overtly remains in the CP indicates that no covert movement occurs here (otherwise (18c,19c) should be ungrammatical, on a par with (17c)). Additionally, CP-fronting is possible in Japanese (20a). But it is not possible to move a phrase out of the CP before CP-fronting (20b,20c) (Saito 1992). (20) a. [CP Hanako-ga sono hon-o yonda to]i Taro-ga ti sitta Hanako-NOM that book-ACC read C Taro-NOM said ‘That Hanako read that book, Taro said.’ b. *[CP Hanako-ga tj yonda to]i Taro-ga sono hon-oj ti sitta yonda to]i Taro-ga Hanakoj -o ti sitta c. *[CP tj sono hon-oj that book-ACC read C Taro-NOM Hanako-ACC said ‘That Hanako read that book, Taro said.’ Crucially, FECM is unacceptable even if the ECMed DP remains within the fronted CP (21), which also follows if the ECMed DP has to move of the FECM clause (covertly, if not overtly). (21) *[CP Hanako-o sono hon-o yonda to]i Taro-ga ti sitta Hanako-ACC that book-ACC read C Taro-NOM said ‘That Hanako read that book, Taro said.’ Crucially, the Turkish (22) and Uyghur (23) counterparts of (21) are acceptable. diye]i Pelin ti biliyor içti ?[CP Ayşe-yi viski-yi Pelin know.PRS .3 SG Ayşe-ACC whiskey-ACC drink.PST.3 SG C ‘That Ayşe drank the whiskey, Pelin thought.’ dep]i Mehmet ti anglidi (23) [CP Ayghül-ni Adil-ni söydu Mehmet hear.PST.3 SG Ayghül-ACC Adil-ACC love.NPST.3 SG C ‘That Ayghül loves Adil, Mehmet heard.’ (22) Evidence from scope possibilities further indicates the in Turkish and Uyghur the ECMed DP does not covertly raise to a matrix position. As shown below, in Turkish (24) and Uyghur (25) the ECMed DP only receives a surface scope interpretation relative to the matrix DP; an inverse scope interpretation is impossible. (24) Biri [CP herkes-i öldü diye] biliyormuş someone everyone-ACC die.PST.3 SG C know.PROG . EV. PST.3 SG ‘Someone thought everyone died’ (∃ > ∀; *∀ > ∃) (25) Birsi [CP hemme adem-ni oldi dep] oylaydu someone every person-ACC die.PST.3 SG C think.NPST.3 SG ‘Someone thinks that every person died.’ (∃ > ∀; *∀ > ∃) Covert raising and finite ECM 205 In contrast to Turkish and Uyghur, Japanese ECMed DPs do show scopal ambiguity with respect to the matrix DP. As (26a) show, both surface and inverse scope interpretations of the ECMed DP relative to the matrix DP are possible (cf. 26b). (26) a. Dareka-ga [CP minna-o tensai da to] omoteiru someone-NOM everyone-ACC genius is C think.PROG ‘Someone believes everyone to be a genius’ (∃ > ∀; ∀ > ∃) b. Dareka-ga [CP minna-ga tensai da to] omoteiru someone-NOM everyone-NOM genius is C think.PROG ‘Someone believes everyone to be a genius’ (∃ > ∀; *∀ > ∃) (Taguchi 2015) The contrast can be accounted for if the ECMed DP moves into the matrix clause in Japanese, but not in Turkish and Uyghur (Tanaka (2002), in fact, interprets the above ambiguity regarding ECMed DPs in Japanese as evidence of raising to the matrix clause). Further, note that such ambiguity must be due to covert movement of the ECMed DP and not the result of optional overt movement, i.e. scrambling. If the ambiguity in (26a) were due to scrambling into the matrix clause, then the Turkish and Uyghur constructions, (24), (25) respectively, should be ambiguous as well (given that scrambling to the matrix clause is possible (14,15)). The reason why the scrambling derivation is blocked here is because string-vacuous scrambling is quite generally not possible (Hoji 1985), hence (26a) cannot be a case of scrambling. To summarize, the above facts indicate that Japanese FECM diverges from Turkish and Uyghur FECM regarding the position of ECMed DPs. In Japanese, the ECMed DP always raises to the matrix clause, overtly or covertly. In Turkish and Uyghur FECM, there is no obligatory raising of the ECMed DP, and there is no covert movement of the ECMed DP. 4 Base-generation positions The previous section showed that Japanese FECM diverges from Turkish and Uyghur FECM regarding obligatory raising. This section shows that this split regarding obligatory raising correlates with a key difference between Japanese ECMed DPs and Turkish and Uyghur ECMed DPs: base-generation positions. As shown for Japanese below in (27), while NOM subject DPs can participate in idiom chunks, ECMed DPs cannot. Here, only a literal interpretation of the embedded CP is possible. (27) John-wa [CP te-ga/*-o soko-made mawar-anai to] omotteiru John-TOP hand-NOM / ACC there-to get.around-not C think.PROG Idiom:‘John thinks that he can’t take good care of it.” (Takano 2003) In the case with the NOM-subject DP, which has raised to SpecTP from a lower, TP-internal, base-generated position (e.g. SpecvP), participation in idiom chunks is expected given that the NOM-DP can reconstruct to the lower position for an idiomatic interpretation (Sportiche 2006). In the case of the ECMed DP, which is in SpecCP, no idiomatic interpretation is possible. Given the absence of an idiomatic interpretation under Japanese FECM, this indicates that the ECMed DP has 206 ROBIN JENKINS not undergone movement from a TP-internal position (leaving a copy) to SpecCP, otherwise reconstruction for an idiomatic interpretation should be possible. Thus, following Taguchi (2009), Saito (2018), and Bošković (to appear), I adopt the proposal that for Japanese the ECMed DP is base-generated in SpecCP and coindexed with a pro in SpecTP, as in (28).6 (28) [VP [CP DPi [C’ [TP proi [T’ . . . T] ] C ] ] V] Contrasting with Japanese, Uyghur permits idiomatic construals of ECMed DPs with embedded predicates, as in (29). (29) Ayghül [CP burut-ung-(ni) xet tartiptu dep] Ayghül mustache-2 POSS-ACC letter pull.INDIR . PST.3 SG C angilidi hear.PST.3 SG Idiom:‘Ayghül heard that you’ve become a man’ (adapted from (Major 2021)) Given that ECMed DPs can participate in idiom chunks in Uyghur, this indicates that ECMed DPs must be able to reconstruct to a lower, TP-internal, position to participate in an idiomatic interpretation. Thus, indicating the presence of a lower DP copy. Thus, I propose that Uyghur FECM constructions involve base-generation of the ECMed DP below SpecCP, e.g. SpecvP, which then undergoes movement to the edge of the embedded clause, SpecCP, as in (30) (cf. also (16)). (30) [VP [CP DP [C’ [TP [vP t [v’ . . . v] ] T] C ] ] V] In Turkish, ECMed DPs resist reconstruction for idiomatic construals, as in (31). However, close examination shows that Turkish ECMed DPs are base-generated lower than SpecCP (as in Uyghur). (31) Öǧretmen [CP Pelin-in etekleri-(*ni) tutuştu diye] teacher Pelin-GEN skirt. PL .3 POSS - ACC catch.fire.PST.3 SG C duymuş hear.EV. PST.3 SG Idiom:‘The teacher heard that Pelin was very anxious.’ (Şener 2008) Evidence that Turkish ECMed DPs are base-generated lower than SpecCP comes from the availability of default agreement on the embedded verb in FECM constructions, which is also observed in Uyghur FECM. In Uyghur, the ECMed DP cannot agree with the embedded verb, which must surface with default 3 SG agreement, as in (32). Turkish displays a similar pattern, as in (33). Here, the embedded verb can optionally agree with the ECMed DP or surface with default 3 SG agreement too. 6 Taguchi (2009) and Saito (2018) provide additional diagnostics which indicate Japanese ECMed DPs are base-generated in the embedded SpecCP. Covert raising and finite ECM 207 (32) Adil [CP seni Tursun-ni söy-y-du/*siz dep] anglidi Adil 2 SG . ACC Tursun-ACC love-NPST-3 SG /2 SG C hear.PST.3 SG ‘Adil heard you love Tursun.’ diye] (33) Pelin [CP seni Timbuktuya gitti-∅/n Pelin 2 SG . ACC Timbuktu.DAT go.PST-3 SG /2 SG C biliyormuş know.PROG . EV. PST.3 SG ‘Pelin thought that you went to Timbuktu.’ (Şener 2008) If it were the case that the absence of reconstruction effects in Turkish FECM were due to the ECMed DP being base-generated in SpecCP coindexed with pro in SpecTP, then since Turkish has an agreeing subject pro, pro should always be agreeing with the verb in (33), i.e. no default agreement. Thus, I conclude that Turkish FECM instantiates the configuration in (30) where the subject DP moves to SpecCP from a lower, TP-internal, position. 7 Regarding the absence of reconstruction effects, following Şener (2008), I suggest that in Turkish the embedded subject DP undergoes topicalization movement to SpecCP, which resists reconstruction in Turkish (Kornfilt 2005).8 Evidence for topicalization movement comes from the impossibility ECMed NPIs (34), which resist topicalization movement (Lasnik & Uriagereka 1988), (cf. Japanese (35) and Uyghur (36), where NPIs can be ECMed). 7 Regarding the apparent optionality of agreement morphology, there has been some dispute as to whether it reflects a point of interspeaker variation (Kornfilt 1977) or an instance of true optionality (Kural 1993; Aygen 2002; Zidani-Eroǧlu 1997; Şener 2008). I suggest that agreement morphology here is a truly optional PF effect, given that for the speakers I have consulted the presence/absence of agreement does not appear to have any syntactic effects, i.e. the presence/absence of agreement does not correlate with a deeper structural/syntactic difference. Thus, as in (i), CP-extraposition is possible with and without agreement. If the presence of agreement correlated with the ECMed DP being base-generated in SpecCP (with a pro in SpecTP), then (i) should be ungrammatical on a par with Japanese FECM (17c) which instantiates this configuration (28). Given that (i) is grammatical in both instances, I conclude, with the aforementioned works, that agreement with the ECMed DP is a case of low level morphological optionality. i Pelin ti biliyormus [CP seni Timbuktu-ya gitti-∅/n diye]i Pelin know.PROG . EV. PST.3 SG 2 SG . ACC Timbuktu-DAT go.PST-3 SG /2 SG C ‘Pelin thought that you went to Timbuktu.’ 8 As shown below in (ia), the anaphor, which has undergone topicalization to SpecCP, cannot reconstruct below the subject DP to its base-position, (cf. (ib)). i a. ??/*kendinei,1 Ahmet1 her akşam ti bir içki hazırlar Ahmet every evening a drink prepare.AOR SELF . DAT ‘Ahmet1 prepares every evening a drink for himself1 .’ b. Ahmet1 her akşam kendine1 bir içki hazırlar (Kornfilt 2005) 208 ROBIN JENKINS (34) *Pelin [CP kimse-yi bu kitab-i okumadi diye] Pelin nobody-ACC this book-ACC read.NEG . PST.3 SG C biliyor know.PRS .3 SG ‘Pelin thinks that nobody read this book.’ (Şener 2008) baka da to-mo] omotteinai (35) Taroo-wa orokanimo [CP dare-o who-ACC stupid COP C-either think.NEG . PROG Taro-TOP stupidly ‘Taro stupidly doesn’t believe that anyone is stupid.’ (Hiraiwa 2005) dep] umid.qilimen (36) Men [CP hechkim-ni ketmidi 1 SG nobody-ACC leave.NEG . PST.3 SG C hope.NPST.1 SG ‘I hope that nobody left.’ (Major 2021) To summarize, this section showed that obligatory raising of the ECMed DP correlates with the DP’s base-generation position. In Japanese FECM, the embedded subject DP is base-generated in SpecCP and must raise to the matrix clause. In Turkish and Uyghur, the embedded subject DP is base-generated lower in the embedded clause, e.g. SpecvP, where it moves to SpecCP (in Turkish this movement is topicalization), from SpecCP the DP may optionally scramble to the matrix clause or remain in SpecCP. 5 Deriving the split As the previous sections showed, the base-generation position of the ECMed DP correlates with whether raising to the matrix clause must occur in FECM constructions. In these two respects the ECMed DP in Japanese differs from the ECMed DP in Turkish and Uyghur. Considering this correlation, I propose that what underlies obligatory raising in Japanese FECM is the following licensing condition on DP-chains: (37) A DP which receives structural Case must have at least one member of its chain, i.e. a copy, in an A-position. The key point to (37) is that whether raising is forced depends on whether the ECMed was base-generated in SpecCP or lower in the embedded clause, e.g. SpecvP, Ā and A-positions respectively. In Japanese, the ECMed DP always raises, either overtly or covertly, to the matrix clause. Here, the embedded subject is base-generated in SpecCP, which is an Ā-position, where it receives [ACC] from the matrix v. From this position, the ECMed DP may undergo scrambling to the matrix SpecvP. In such cases, (37) is satisfied with overt movement. 9 However, if the DP does not overtly scramble to 9 Further evidence from binding indicates that movement to SpecvP in Japanese FECM is an instance of movement to an A-position. As shown below in (ib), the ECMed DP can scramble over the matrix DP and bind the anaphor, thereby alleviating the Condition A violation in (ia). i a. *Otagai-no sensei-ga karerai -o [CP ti bakada da to] omotteiru each.other-GEN teach-NOM them-ACC fool COP C think.PROG Covert raising and finite ECM 209 SpecvP, then it must raise covertly to SpecvP in order to satisfy (37), since otherwise no member of the DP-chain will be in an A-position.10 In Turkish and Uyghur FECM, no covert raising to SpecvP is triggered since (37) is always satisfied due to embedded subject DPs being base-generated in a lower A-position. Here, the DP is base-generated lower, e.g. SpecvP, and moves to SpecCP (in the case of Turkish this movement is topicalization). From SpecCP, the DP receives [ACC] from the matrix v and may optionally scramble to the matrix SpecvP. However, if the ECMed DP does not scramble out, then no covert raising is forced given that the DP-chain has one member in an A-position, e.g. SpecvP. 6 Conclusion In this paper, I investigated a point of cross-linguistic variation regarding the structural positions of ECMed DPs in FECM constructions. I showed that that while Japanese, Turkish and Uyghur FECM constructions pattern alike in several respects, they diverge regarding the position of ECMed DPs. I showed that in Japanese ECMed DPs always raise, either covertly or overtly, to matrix SpecvP. In Turkish and Uyghur ECMed DPs can either optionally raise overtly to the matrix clause or remain in the embedded SpecCP. Crucially, there is no covert movement. Moreover, this split with respect to obligatory raising correlates with the base-generation position of ECMed DPs. Namely, in Japanese, the ECMed DP is base-generated in SpecCP. In Turkish and Uyghur, the ECMed DP is base-generated lower and raises to SpecCP. Finally, I proposed that this split regarding the position of ECMed subjects between Japanese, and Turkish and Uyghur is due to the requirement on DPlicensing which requires at least one member in a DP-chain to be in an A-position, which forces raising in the Japanese FECM context. ‘Each other ’s teachers thinks of them as fools.’ sensei-ga ti [CP ti bakada da to] omotteiru b. karerai -o otaga-no them-ACC each.other-GEN teach-NOM fool COP C think.PROG ‘Each other ’s teachers thinks of them as fools.’ (Tanaka 2002) Assuming that only A-movement creates new binding relations, and that Ā-movement cannot feed A-movement (Chomsky 1973, 1981; i.a.), the fact that the scrambled ECMed DP in (ib) can bind the anaphor under A-scrambling over the matrix DP indicates that the previous movement step, i.e. SpecCP to SpecvP, must be A-movement as well (Saito 1992). 10 Interestingly, (37) (and its application to Japanese FECM) may shed light on the Improper Movement Condition (Chomsky 1973, 1981; i.a.). Traditional cases of Improper Movement violations involve movement chains where a DP undergoes A-to-Ā-to-A movement. However, it has been proposed that the Improper Movement Condition may reduce to a generalized ban on Adependencies with a DP in an Ā position (Keine 2020; Poole 2022). Interestingly, Japanese FECM involves an A-dependency from an Ā-position, i.e. SpecCP-to-SpecvP movement. Given that Āto-A movement is licit in Japanese FECM (in fact, it is required per (37)), the correct statement of the Improper Movement Condition should be a ban on A-to-Ā movement feeding A-movement, i.e. A-to-Ā-to-A, rather than a ban on A-dependencies from an Ā-position generally (i.e. Ā-to-A is allowed). 210 ROBIN JENKINS References Aygen, G. 2002. Finiteness, Case and Clausal Architecture. Cambridge: Harvard University dissertation. Bošković, Ž. to appear. The comp-trace effect and contextuality of the EPP. In Proceedings of WCCFL 39. Chomsky, N. 1973. Conditions on transformations. In A Festschrift for Morris Halle, ed. by S. Anderson & P. Kiparsky, 232—286. New York: Holt, Reinhard and Winston. Chomsky, N. 1981. Lectures on Government and Binding. Dordrecht: Foris. George, L., & J. Kornfilt. 1981. Finiteness and boundedness in Turkish. In Binding and filtering, ed. by F. Heny, 105–127. Cambridge: MIT Press. Hiraiwa, K. 2005. Dimensions of symmetry in syntax: Agreement and clausal architecture. Cambridge: Massachusetts Institute of Technology dissertation. Hoji, H. 1985. Logical form constraints and configurational structures in Japanese. Seattle: University of Washington dissertation. Johnson, K. 1985. A case for movement. Cambridge: Massachusetts Institute of Technology dissertation. Keine, S. 2020. Probes and their horizons. Cambridge: MIT Press. Kornfilt, J. 1977. A note on subject raising in Turkish. Linguistic Inquiry 8. 736–742. Kornfilt, J. 2005. Asymmetries between pre-verbal and post-verbal scrambling in Turkish. In The free word order phenomenon: Its syntactic sources and diversity, ed. by J. Sabel & M. Saito, 163–180. Berlin: Mouton de Gruyter. Kuno, S. 1976. Subject raising. In Japanese generative grammar, ed. by M. Shibatani, 17–49. Leiden: Brill. Kural, M. 1993. V-to(-I-to)-C in Turkish. In UCLA Occasional Papers in Linguistics, ed. by F. Beghelli & M. Kural, 17–54. Los Angeles: UCLA Department of Linguistics. Lasnik, H., & J. Uriagereka. 1988. A course in GB syntax. Cambridge: MIT Press. Major, T. 2021. On the nature of" say" complementation. Los Angeles: University of CaliforniaLos Angeles dissertation. Nishigauchi, T. 1992. Syntax of reciprocals in Japanese. Journal of East Asian Linguistics 1. 157–196. Poole, E. 2022. Improper case. Natural Language & Linguistic Theory 1–51. Ross, J. R. 1967. Constraints on variables in syntax. Cambridge: Massachusetts Institute of Technology dissertation. Saito, H. 2018. The monster tells where you are. In Proceedings of WCCFL 35, ed. by W. G. Bennett, L. Hracs, & D. R. Storoshenko, 341–348, Sommerville. Cascadilla Proceedings Project. Saito, M. 1992. Long distance scrambling in Japanese. Journal of East Asian Linguistics 1. 69–118. Salzmann, M. 2017. The Wiley Blackwell Companion to Syntax, volume 5, chapter Prolepsis, 1–42. Hoboken: Wiley-Blackwell, 2nd edition. Şener, S., 2008. Non-canonical case licensing is canonical: Accusative subjects of CPs in Turkish. ms., University of Connecticut. Sportiche, D. 2006. The Blackwell companion to Syntax, volume 4, chapter Reconstruction, binding, and scope, 35–93. Malden: Blackwell Publishing. Taguchi, S. 2009. Japanese ECM as embedded bare topicalization. In Proceedings of NELS 38, ed. by A. Schardl, M. Walkow, & M. Abdurrahman, 415–426, Amherst. GLSA, University of Massachusetts. Taguchi, S. 2015. Syntactic operations on heads and their theoretical implications. Storrs: University of Connecticut dissertation. Takano, Y. 2003. Nominative objects in Japanese complex predicate constructions: A prolepsis analysis. Natural Language & Linguistic Theory 21. 779–834. Tanaka, H. 2002. Raising to object out of CP. Linguistic Inquiry 33. 637–652. Zidani-Eroǧlu, L. 1997. Exceptionally case-marked nps as matrix objects. Linguistic Inquiry 28. 219–230. Proceedings of CLS 58 (2022), 211-230 © Chicago Linguistic Society 2023. All rights reserved. 211 Comparative tonal text-setting in Mandarin and Cantonese popular song James P. Kirby Institute of Phonetics and Speech Processing, LMU Munich 1 Introduction A common question asked by speakers of European languages with a naïve (and even not-so-naïve) understanding of tone languages is, “how is it possible to sing in a tone language?” This question highlights the fact that in a tone language linguistic and musical uses of pitch appear to conflict: if a Thai singer sings the word ‘far’ (klāj) on a descending musical melody, won’t listeners understand ‘near’ (klâj) instead? But clearly, people sing in all languages, including tone languages, so there must be a way that artists and composers meet this challenge. There are three important components to the answer to this question. The first is what might be called top-down cues, which is to say, context. Language is redundant, and song texts are often formulaic, predictable, and/or repetitive. Loss of tone information may not affect comprehension. In many cases, the other words in the lyric, or knowledge of the likely subject material, or other kinds of structural linguistic knowledge, will enable the listener to decode the message. Another important component of the answer is residual phonetic cues, that is, pitch modulations laid on top of the melody contour, which may assist the listener in recovering the underlying lexical tone (see Schellenberg & Gick 2020 for a recent review). The focus of the present study, however, is on a third component: the textsetting constraints which govern the productive ability to set texts to songs. There is a long tradition of work on stress-beat text-setting in non-tonal languages, but a body of research over the past few decades suggests that a wide variety of musical traditions in many—possibly all—tone languages enforce some sort of congruence between musical melody and tonal sequences. In other words, given a vocal melody, different sequences of words are judged to “fit” the melody better than others, depending (in part) on the tones of the words in that sequence. One research challenge then becomes to determine which principles determine the goodness of this fit within and across languages. 1.1 Text-setting in non-tonal languages In languages like French and English, it has long been observed that native speakers familiar with a particular tradition of sung or chanted verse can readily set novel lines to existing rhythms with a significant degree of agreement (Jakobson 1960; Kiparsky 1977; Halle & Lerdahl 1993; Dell & Halle 2009; Hayes 2009 inter alia). To take just one example, in English (as well as in other languages) it is important for major stressed syllables to occur on strong musical beats, and perhaps even more important for unstressed syllables not to occur on strong beats. To see this preference in action, consider the third line of the “Happy Birthday” song, the 212 JAMES P. KIRBY line in which a name needs to be inserted. If the name is a trochee like Susan (Figure 1a), this is unproblematic, but with an iamb like Suzanne (Figure 1b), the setting which results from the same strategy—1-to-1, left-to-right association of notes and syllables—is generally judged suboptimal by most native speakers of English. The repair, which is to insert an upbeat preceding the final measure and realize the stressed syllable melismatically (Figure 1c), is perhaps still somewhat forced, but nevertheless clearly preferable to setting the unstressed syllable of Suzanne on the strong downbeat of the third measure. This intuition reflects what we might call the basic English text-setting constraint: unstressed syllables should not be set to metrically prominent positions in the verse. (a)    Hap (b)    Hap (c)    Hap  py  py  py      birth day, dear Su san      birth day, dear Su zanne      birth day, dear Su zanne     Figure 1: Three examples of setting English texts to the “Happy Birthday” meter. (a): a well-set text. (b): a suboptimal setting, with an unstressed syllable set to a strong beat. (c): a melismatic repair, which ensures that only stressed syllables are associated with strong metrical positions. 1.2 Text-setting in tonal languages There is now a considerable body of work which establishes that a good match between tone and melody is not simply or even primarily a matter of aesthetics, but that structural-linguistic principles are also involved. While text-setting has been studied in an increasingly broad range of tonal languages, here I will focus on just two, Mandarin Chinese [cmn] and Cantonese [yue]. The tonal inventory of Mandarin consists of four tones: a level tone (T1), a rising tone (T2), a low-falling tone (T3) and a high-falling tone (T4). In addition, some function words, like the adjective suffix 的 /de/ or the perfective verb suffix 了 /le/, are usually realized with a so-called neutral tone, whose surface realization is largely predictable on the basis of the preceding tone (Cao 1992). Cantonese, on the other hand, contrasts 6 tones in sonorant-final syllables (Bauer & Benedict 1997), although some of these are presently undergoing various types of mergers (Mok et al. 2013). In this paper, Mandarin is transliterated using hànyǔ pı̄nyı̄n, with tones represented using diacritics, and Cantonese using the jyutping system, with tones represented as Chao Comparative tonal text-setting 213 numbers. Part of the reason for restricting our focus to these two languages is that one of the most intensely studied instances of tonal text-setting is that of ‘Cantopop’ (粵語流行音樂), an umbrella term for Cantonese-language popular music produced from the late 1970s onward. Text-setting in Cantopop was first studied by Marjorie Chan in a 1987 BLS paper (Chan 1987a) and accompanying UCLA working paper (Chan 1987b), and has subsequently been the focus of several further studies (Wong & Diehl 2002; Ho 2006; Ho 2010; Schellenberg 2011; Lo 2013). Given that this line of work has informed a great deal of subsequent research, it is worth considering the findings in some detail. Chan made two important observations, both of which have been corroborated many times since. First, when there is divergence between the tones of the syllables in different verses set to the same positions in the melodic line, the difference seems to only be in the tonal onset. This can be seen clearly in the first and last syllables of the texts shown in Figure 2: the first syllable in the first verse has a high tone /55/, while the first syllable in the second verse has a rising tone /25/. Similarly, the syllables set to the last note in the phrase have the opposite tones. What appears to be critical is not that the tones across verses are the same, but rather that they end ⿊⾊午夜 high.                擁 jung25 ⾞ ce55 內 noi22 橫巷 waang22hong22 內 noi22 藏 cong21 著 zoek2 愛侶抱 oi33 leoi23 pou23 ⽕ fo25 焰 jim22 無忌 mou21 gei22 地 dei22 搖 jiu21 盪 dong22 你我眼中 nei23 ngo23ngaan23 zung55 Figure 2: Excerpt from 黑色午夜 by Leslie Cheung. Adapted from Lo (2013). Second, Chan observed a striking degree of parallelism between the direction that the musical note sequence takes in relation to the direction of the associated tone sequences. It seems that in Cantopop, rather than a particular lexical tone being associated with a particular note (a strategy employed in some classical and religious genres: see e.g. Yung 1983; Tanese-Ito 1988), for any given sequence of two syllables, the direction of linguistic pitch change from the first syllable to the second tends to be mirrored in the corresponding musical note sequence. For example, a low tone-high tone sequence (such as 眼中 /ngaan23-zung55/ in Figure 2) will tend to be set to a rising melodic sequence (G# rising to A), while a high tone-low tone sequence (such as 車內 /ce55-noi22/) will tend to be set to a falling sequence (A falling to D). In what follows, I will refer to sequential pairs of lexical tones as tonal bigrams, or just bigrams for short. A subsequent study by Wong & Diehl (2002) presented a small quantitative survey of the way tonal bigrams are actually treated melodically in text-setting. For the purposes of defining pitch direction from one tone to the next, they grouped the six tones of Cantonese into three sets of high, mid, and low based on their relative  214 JAMES P. KIRBY pitch offsets. Defining pitch direction on this basis, they observed that the direction of the tone sequence matched that of the note sequence in over 90% of cases. To describe the various possible correspondences between the directions of musical melody and linguistic tone, Ladd & Kirby (2020) repurposed some standard terms from classical Western music theory, exemplified in Figure 3. They use the term similar setting to refer to the tendency described by Chan and Wong and Diehl for tone and melody to move together in the same direction, oblique setting for when either the tonal or melodic contour stays level while the other rises or falls, and contrary setting to refer to instances where one of the sequences rises while the other falls. This space of possibilities defines the three-by-three matrix shown in Table 1. a. b. contrary setting similar setting c. d. oblique setting 1 oblique setting 2 Figure 3: Different possibilities for setting bigrams to musical melodies. Contrary setting (a): a rising tonal sequence set to a falling musical sequence (or vice versa). Similar setting (b): both tonal and musical sequences are rising (or falling). Oblique setting 1 (c): consecutive syllables have different tones but are set to the same note. Oblique setting 2 (d): consecutive syllables have the same tone but are set to different notes. Adapted from Ladd and Kirby (2020). Tone seq. up down level Note seq. up down level Similar Contrary Oblique Contrary Similar Oblique Oblique Oblique Similar Table 1: Similar, contrary, and oblique settings, as defined by the relation between pitch direction in a sequence of two tones and the two corresponding musical notes. Subsequent work by Ho (2006, 2010) and Lo (2013) further refined Wong and Diehl’s classifications, but all studies have observed the same basic constraint at work, which we might describe as “avoid contrary settings”. Note that, much as the basic English text-setting constraint formulated in Section 1.1 above referred to avoidance (“unstressed syllables should not be set to metrically strong positions”) rather than enforcement (“stressed syllables are preferred on metrically strong positions”), avoidance of contrary settings is subtly but importantly different from a preference for similar settings. This is because there is some indication that oblique settings are not universally avoided. For example, Lo (2013) found that in one specific context—sequences of high-final (/55/ or /35/) tones set to a descending melody line—oblique settings were not uncommon, and Kirby & Ladd Comparative tonal text-setting 215 (2016) found a significant number of level tone sequences set to rising or falling melodic lines in Vietnamese. For whatever reason, rather less attention seems to have been directed at tonemelody correspondence in Mandarin popular music. The general consensus appears to be that tone-tune correspondence is largely ignored in Mandarin popular song, an observation going back at least to Chao (1956). For example, Ho (2006) writes that “many Mandarin songs, if not all, exhibit disagreement between tone and tune. Neither individual lexical tone contour nor tonal target transition is reflected in the melody systematically”, while Chan claims that “word tones are often suspended in favour of the overall melody” (Chan 1987b:163). On the other hand, Wee (2007) suggests that tone-tune correspondence in Mandarin is brought about by preserving contrasts in the tonal and melodic structures at positions of metrical prominence. Specifically, if a note sequence is falling, the first syllable must end high (i.e. be tone 1 or 2); if a note sequence is rising, the first syllable must end low (i.e. be tone 3 or 4). Wee argues that these constraints are observed more strictly in structurally prominent positions. However, it should be noted that Wee’s study considered Mandarin folk songs, rather than acculturated song, and difference in genre can have a considerable impact on tone-tune correspondence (see e.g. List 1961 on Thai). 1.3 The present study If it is true that modern Mandarin music shows little evidence of tone-tune correspondence, why might this be the case? One possibility is that this simply represents a point of cross-linguistic divergence: there is no a priori reason that avoidance of contrary settings is necessarily a universal constraint active to the same extent in all tonal languages. However, there are a number of facts about the realization of tone in Mandarin that may also play a role. The goal of the present study was to take into account three aspects of Mandarin in order to assess their effect (or lack thereof) on tone-melody correspondence: 1. Influence of (word and phrase) boundaries. Virtually all off the instances of contrary settings observed in Ho (2010) and Lo (2013) for Cantonese were found to occur across syntactic and/or musical phrase boundaries. More critical still for Mandarin might be word boundaries, given that a large proportion of the lexicon is minimally disyllabic—at least 50% according to Wu et al. (2020) and perhaps as much as 67% according to Li & Thompson (1981). To the best of my knowledge, no previous work on tonal text-setting in either Mandarin or Cantonese has explicitly coded for word boundaries or compared correspondence counts when word-internal transitions were included versus excluded. 2. Neutral tones. Unlike Cantonese, in Mandarin there are certain grammatical morphemes, including some suffixes, particles and kinship reference terms, which lack tone at both underlying and surface levels (Cao 1992). For example, the nominalizer 子 /zi/ surfaces with a different tone depending on what it is affixed to: compare 钉子 /dı̄ng zi/ [ding55 zi2 ] ‘nail’ vs. 脑子 /nǎo zi/ [nao21 zi4 ] ‘brain’. Whether or not lyricists take the surface tone of neutraltoned syllables into account is an open question. Chan (1987b) explicitly disregarded neutral toned syllables, but did not compare her results with an 216 JAMES P. KIRBY analysis that included them, so it is not yet clear to what extent they do or do not impact tone-melody correspondence. 3. Tone sandhi. In Mandarin, there is a well-known phonological alternation in which (low-falling) T3 changes to (high-rising) T2 when followed by another T3. Impressionistically, this phenomenon is not uncommon, although at the time of writing I have been unable to find any source which provides frequency counts of tonal bigrams in running text. An obvious example is the greeting 你好 /nı́ hǎo/ [ni35 hao214 ], pronounced with a rising tone on the first syllable: in isolation, both syllables (你 nı̌ and 好 hǎo) would be produced with the low or low-rising tone. As with the neutral tone, it is unclear whether bigrams consisting of two underlying T3s are treated with respect to tone-melody correspondence (i.e. as level or falling tonal transitions). In addition to exploring the role of these features, the present study also attempts to control for differences in melodic transitions by holding the melodic component constant. One problem when attempting to compare rates of tone-melody correspondence between any two languages is that not only do the tone systems differ, but the songs do as well. This means that it is unclear whether the differences observed have to do with the languages themselves, or with (incidental or systematic) properties of the melodies of the song corpora. To attempt to mitigate the potentially confounding effects of changing both melody and text, the present study focuses on a corpus of 8 song pairs for which versions exist with both Mandarin and Cantonese lyrics. The practice of releasing bilingual versions of popular songs is one that enjoys a long tradition in the Chinese-speaking popular music world, due perhaps in part to the fact that song lyric writing is generally considered as a separate skill from musical song composition (Mitchell 2006). This tradition has produced a rich body of song texts in both languages which are set to the same musical melodies, providing a natural laboratory in which to isolate the influence of linguistic, as opposed to musical, factors on rates of tone-melody correspondence. Alongside the three factors mentioned above, the present work considers the role of interval step size, as some authors (e.g. Ho 2010; McPherson & Ryan 2017) have suggested that contrary mappings are avoided more stringently for larger musical steps than smaller ones. Holding the melodic corpus constant makes it possible to see if the same differences in step size have different effects when setting texts in different languages, i.e. if the step size constraint is more highly ranked by Cantonese or Mandarin lyricists. 2 Materials and methods 2.1 Corpus The corpus was chosen opportunistically for a master’s thesis (Lin 2018). As seen in the Appendix, the Cantonese lyricists were primarily from Hong Kong; the Mandarin lyricists were primarily Mandarin native speakers, with a few Hong Kong-based bilinguals. Many of the Cantonese songs feature lyrics written by Lam Chik (Albert Leung), although there is greater diversity among the composers. The songs are typically written with particular performers in mind, so it is potentially also relevant that Eason Chan appears three times and Faye Wong twice. The Can- Comparative tonal text-setting 217 tonese versions of the songs in this corpus were typically released first, although in many cases the same artist recorded both versions. The exception is song 8《痴心绝对》‘Absolutely Infatuated’, which was a Mandarin-language hit by 李圣杰 Sam Lee in 2002 and only later appeared as a Cantonese-language cover version《残酷游戏》‘Cruel Game’ by 卫兰 Wai Lan (Janice Vidal) in 2009. Figure 4 shows the first two phrases of the song ‘Red Rose’ (in Mandarin, examples 1-2) or ‘White Rose’ (in Cantonese, examples 3-4), popularized by Eason Chan, who sings both of the versions analyzed here. Glosses and translations for the first phrase of each language are given in (1-4). (1) 红是朱砂痣烙印心口 hóng shì zhū shā zhì lào yìn xı̄n kǒu red COP scarlet birthmark brand.print heart.mouth ‘Red, the scarlet beauty mark over your heart1 ⋯’ (2) 红是蚊子血般平庸 hóng shì wén zi xuè bān píng yōng red COP mosquito blood kind common ‘Red is as common as mosquito’s blood ⋯’ (3) 白如白忙莫名被摧毁 2 21 2 21 2 21 baak jyu baak mong mok ming bei22 ceoi55 wai35 white as IDIOM inexplicably PASS destroy ‘White like working in vain2 as I’m inexplicably cut down ⋯’ (4) 得到的竟已非那位 5 33 5 35 23 dak dou dik ging ji fei55 naa23 wai25 get arrive POSS actually already not DEM PASS ‘Who I catch is not actually “the one” ...’ 2.2 Transcription and annotation For each text/melody pair, song melodies were transcribed as fixed do solfège values (e.g. A, B, C...) along with accidentals and octave range, which were then automatically converted to MIDI note numbers. The corpus was then enriched with word and phrase boundary annotations, to enable comparison of the counts of similar, contrary, and oblique settings depending on whether or not transitions which crossed a boundary were ignored. Annotation of word boundaries was fairly conservative; anything that occurred as a dictionary headword was marked as a compound, so this encompassed examples such as 领悟 lı̌ngwù ‘understand’ (understand+comprehend), 伤心 shāngxı̄n ‘sad, grieved’ (hurt+heart), 蝴蝶 húdié ‘butterfly’ (non-decomposible), or 号码牌 hàomǎpái ‘number plate’ (number+code+plate). 烙印 can also be ‘indelible mark, lasting mark, stigma’; 心口, literally ‘heart and mouth’ also has a figurative reading ‘words and thoughts’. 2 While awkward in translation, this idiom works because the other verses are ‘white as white teeth’, ‘white sugar’ (the Cantonese pronunciation of ‘red’ is /hung21 /) and because “white” has a connotation like “wiped clean” (Alan Yu, p.c.). 1 218 红玫瑰/⽩玫瑰 JAMES P. KIRBY 李焯雄    Mandarin:          痣烙印⼼红是砂朱 hong35 shi53 zhu55 sha55 zhi53 lao53 yin53 xin55 红玫瑰/⽩玫瑰莫名被摧如⽩忙 jyu21baak2 mong21mok2 ming21bei22 ceoi55 Cantonese: ⽩ baak2 ⼜ kou21                   所  有  刺 激剩   下疲 乏   红  是 suo21蚊 pi35 fa35  yuo21 庸般  xia53 平 ⼦ci53  ⾎ji55 sheng53 李焯雄  ⼼ n53 xin55 被摧 i22 ceoi55    的 de3  印尘⼼俗⼜ yin53 xin55zuk2 kou21 21 can21 被摧 1bei22 ceoi55     的 de3  ping35yong55 zue21 ban55 hong35 shi53 wen35 zi31 Mandarin:⽩红如是⽩朱蛾砂潜痣回烙红印尘⼼俗⼜ zhi53 lao53 yin53 xin55 kou21 hong35 shi53 zhu55 sha55 baak2 jyu21baak2 ngo21 cim21 wui21hung21can21 zuk2 那到位已⾮得的竟 dak5 dou33 ⽩dik5 如 ging35⽩ ji23 fei55 莫名 naa23 被 wai25 摧毁 Cantonese: 忙  ming21bei22 ceoi55 wai35  baak2 jyu21baak2   mong21  mok2                    ⽆动于再痛衷     zai53 dong53 wu35 yu35 tong53 zhong55   所 有 刺激 剩  下 疲  乏 的毁 wai35      ji55 sheng53 xia53 pi35 fa35 ⾎过般庸位平灵 ping35yong55 zue21 ban55 回 wai25 蛾gwo33潜ling21 红尘 ngo21 cim21 wui21hung21can21 位那已⾮ naa23 wai25 ji23 fei55     疲乏 pi35 fa35 的 de3 痛 tong53 再 zai53 ⽆ wu35 动于 dong53 yu35 衷 zhong55 红尘 ung21can21 俗 zuk2 世 sai33 俯 fu35 瞰 ham33 灵过 gwo33 ling21 位 wai25      3 suo21 yuo21 ci53 蚊⼦瞰世红俯是 hong35fu35 shi53 ham33 sai33 ⽩如 zi31 ⽩ wen35 baak2 jyu21baak2 到的得竟 dak5 dou33 dik5 ging35  红蚊⼦是 hong35 shi53 wen35 z  毁红玫瑰/⽩玫瑰 wai35  乏 5 fa35  ⼜ kou21   毁到得的红玫瑰/⽩玫瑰梁翘柏 wai35 dak5 dou33 dik5 3   痛 tong53   再 zai53 到得的 dak5 dou33 dik5 梁翘柏   痛 tong53 再 zai53 俗 zuk2 世 sai33 俯 fu35 Figure 4: First two phrases of 白玫瑰/红玫瑰 performed by 陳奕迅 Eason Chan. A musical phrase boundary occurs at the ends of the 2nd and 4th lines.  蚊世红俯是 hong35 shi53 wen35 sai33 fu35 de3  竟 gin  Comparative tonal text-setting 219 Next, musical phrase boundaries were added by musical proficient annotators. Again, the aim was to be maximally conservative, so in cases of annotator uncertainty or disagreement, a boundary was annotated. This made it possible to compute transitions over the entire corpus, or just within musical phrases, since previous work makes clear that constraints on tonal text-setting are often ignored or relaxed at phrase edges. Finally, each corresponding syllable of the lyric was first annotated with its dictionary tone value. In Cantonese, 变音 pinyām tone modifications (Bauer & Benedict 1997; Alderete et al. 2022), which are not easily predicted by general rule, were hand-annotated by native speakers; for Mandarin, tone changes were applied by rule. T3 was treated as low-falling /21/ unless it was phrase-final (which occurred very rarely, less than 10% of all occurrences of T3 and only 2% of all total transitions) or if it occurred before another T3, in which case it was treated as high-rising /35/ (81 instances representing about 20% of all lexical T3s, but less than 5% of the total corpus). This modification was applied both within and across word boundaries within utterances (Shih 1997). Neutral tones (of which there 96, or again around 5% of the total corpus) were assigned surface values based on those of the preceding syllable (Cao 1992), and these were hand-checked by native speaker annotators. All corpus materials and data analysis scripts are available at this paper’s accompanying OSF archive (https://osf.io/k6uma). 2.3 Analysis One methodological issue that arises in the analysis of tone-melody correspondence when contour tones are involved is that it is not immediately obvious whether a given tonal bigram should be encoded as rising, level, or falling (Ladd & Kirby 2020:679-680). For example, the tonal sequence /35-53/ (as in 红是 hóng shì, the first two lyrics of the Mandarin text given in Figure 4) might be classified as level (the transition from offset 5 to onset 5), rising (the transition from onset 3 to onset 5) or falling (the transition from offset 5 to offset 3). In the present study, transitions were computed in two ways: first, in the manner of Chan (1987a) and Wong & Diehl (2002), i.e. by considering the offsets of any two syllables in a musical bigram; and second, by considering the offset of the first syllable and the onset of the second. Using this latter method, a tonal sequence like /35-53/ would count as a level tonal transition, meaning the setting would be marked as oblique. As offset-offset mappings always resulted in higher rates of similar (or at leat non-opposing) settings in both languages, the remainder of the paper considers the first method only. 3 Results 3.1 Tone-melody correspondence In Cantonese, we find the expected high occurrence of similar settings: out of around 1700 transitions, nearly 80% were similar (Table 2); contrary settings comprised just 5% of the total. The remaining transitions are oblique, although as seen in other studies of Cantopop, sequences of identical tones far outnumber sequences 220 JAMES P. KIRBY of identical notes: that is, there are many more oblique level tone-sequences than oblique level note-sequences. In Mandarin, there are slightly fewer total transitions, but similar and oblique settings each make up around 38% of the total, with the remaining 25% contrary settings (Table 3). Note seq. Tone seq. up down level up 530 39 30 down 45 486 26 level 85 165 314 Table 2: Frequencies of similar (bold), contrary (italic), and oblique (plain) settings in the Cantonese song corpus. Similar settings: 77% (1330); contrary settings: 5% (84); oblique settings: 18% (306) out of 1720 total transitions. Tone seq. up up 244 down 204 level 173 Note seq. down level 218 126 272 121 179 107 Table 3: Frequencies of similar (bold), contrary (italic), and oblique (plain) settings in the Mandarin song corpus. Similar settings: 38% (623); contrary settings: 25% (416); oblique settings: 37% (605) out of 1644 total transitions. Similar Contrary Oblique 1 Oblique 2 Cantonese 77% 5% 4% 15% no internal trans. 77% 5% 4% 15% Mandarin 38% 25% 15% 22% no internal trans. 36% 28% 16% 20% no neutral tones 38% 24% 15% 23% Table 4: Comparison of similar, contrary, and oblique settings in the two song corpora. Oblique 1 is level melody; Oblique 2 is level tone (see Figure 3). Table 4 compares the rates of similar, contrary, and oblique settings in the two corpora, as well as how the percentages change when disregarding word-internal transitions and, in Mandarin, neutral tones. In effect, nothing changes; indeed, in Mandarin the number of similar settings decreases slightly when word-internal transitions are disregarded. So from this we may conclude that contrary and oblique settings truly seem to be more common in Mandarin than in Cantonese, and that this cannot be trivially accounted for by the existence of a primarily disyllabic lexicon or neutral tones. Comparative tonal text-setting 221 3.2 Interval size effects One suggestion that has been made by a number of other researchers (e.g. Ho 2006; McPherson & Ryan 2017) is that oblique or even contrary settings are more likely when the difference between melody steps is small. While step size was not systematically manipulated in the present study, we can nonetheless make some preliminary observations. Figure 5 shows a simple histogram of the distribution of melodic interval step sizes (which of course are the same for the two corpora). Generally speaking, the empirical preference appears to be to go up or down by single whole steps. Figures 6 and 7 give the distribution of interval sizes by direction of tonal transitions. In Cantonese, we can see that there is some indication of a slightly stronger preference to avoid contrary settings with larger melodic steps—as the interval step size increases, the number of contrary settings decrease (i.e., the distributions for the falling and rising tonal transitions are clearly skewed). In Mandarin, however, this tendency does not appear to be as pronounced. The tendency for rising tone bigrams to be set to rising note sequences and falling tone bigrams to be set to falling note sequences is still discernible, but the differences between 1-2 up/down and larger step sizes is closer the general distribution of interval sizes seen in Figure 5. So while melodic interval step size may well be an active constraint in Cantonese, it does not appear that it will help to explain the large discrepancy between the setting percentages in Cantonese and those in Mandarin, either. count 750 500 250 0 5+ down 3-4 down 1-2 down same 1-2 up 3-4 up 5+ up Interval size (binned) Figure 5: Distribution of musical interval step sizes between melodic bigrams. 4 Discussion 4.1 Frequency effects and the problem of contours The foregoing investigation indicates that, despite the fact that the Mandarin lexicon contains a rather larger percentage of polysyllabic words than does the Cantonese lexicon, rates of tone-melody correspondence do not change appreciably regardless of whether or not word-internal transitions and/or neutral tones are disregarded. In this section, I will suggest a different possibility for the divergence between languages, namely that spoken Mandarin contains a rather larger percentage of contour tones and that these are inherently more difficult to set. 222 JAMES P. KIRBY falling level rising Count 400 200 5+ d 3 - ow 4 n d 1 - ow 2 n do w n sa m 1- e 2 u 3- p 4 up 5+ up 5+ d 3 - ow 4 n d 1 - ow 2 n do w n sa m 1- e 2 u 3- p 4 up 5+ up 5+ d 3 - ow 4 n d 1 - ow 2 n do w n sa m 1- e 2 u 3- p 4 up 5+ up 0 Interval size Figure 6: Distribution of musical interval step sizes between melodic bigrams by tonal bigram type, Cantonese texts. falling level rising Count 300 200 100 5+ d 3- ow 4 n d 1- ow 2 n do w n sa m 1- e 2 u 3- p 4 up 5+ up 5+ d 3- ow 4 n d 1- ow 2 n do w n sa m 1- e 2 u 3- p 4 up 5+ up 5+ d 3- ow 4 n d 1- ow 2 n do w n sa m 1- e 2 u 3- p 4 up 5+ up 0 Interval size Figure 7: Distribution of musical interval step sizes between melodic bigrams by tonal bigram type, Mandarin texts. The intuition which underlies this suggestion begins with the observation (originally due to Bob Ladd: see Ladd in press) that there appear to be far fewer Cantonese syllables with high-rising tones in Cantopop songs than one would expect, given their frequency in running text. To explore this a bit more quantitatively, Table 5 gives the (unigram) tone frequencies for the six Cantonese tones computed from the Hong Kong Cantonese Adult Language Corpus (Leung & Law 2001; Leung et al. 2004), compared to the frequencies of the same tones in the current song corpus. While most of the tones have similar frequencies in both the song and text corpora, the high-rising tone (tone 2) does indeed occur rather less often Comparative tonal text-setting Tone 1 2 3 4 5 6 Description Value Example High level 55 /si1/ 詩 High rising 35 /si2/ 史 Mid level 33 /si3/ 嗜 Mid-low falling 21 /si4/ 時 Mid-low rising 23 /si5/ 市 Mid-low level 22 /si6/ 視 ‘poem’ ‘history’ ‘fond of’ ‘time’ ‘market, city’ ‘to view’ 223 Corpus Song 21% 22% 17% 10% 19% 17% 11% 15% 11% 13% 21% 23% Table 5: Frequency distributions in text vs. song: Cantonese. Corpus frequencies from Leung and Law (2004). Tone 1 2 3 4 Description Value Example High level 55 /fū/ 孵 High rising 35 /fú/ 福 Low 21(4) /fǔ/ 腐 High fall 53(51) /fù/ 父 ‘to hatch’ ‘good fortune’ ‘to ferment’ ‘father’ Corpus Song 25% 16% 24% 28% 16% 20% 35% 37% Table 6: Frequency distributions in text vs. song: Mandarin. Corpus frequencies from Wu et al. (2020). in the song corpus. This suggests the possibility that text-setters may be trying to avoid this tone when possible, perhaps because it is the most dynamic tone in the Cantonese repertoire.3 What is interesting if we compare with Mandarin (Table 6) is not so much the extent to which the tones occur in the song corpus relative to running text (frequencies from the corpus study of Wu et al. 2020), but instead the frequency distribution of the tones in the speech corpus. As shown by Wu et al. (2020), highfalling tone 4 is the most common tone in Mandarin, followed by tones 1 and 2. But together, tones 2 and 4—the high-rising and high-falling contour tones, whose dynamic changes in F0 are the critical cues for their perception (Jongman et al. 2012)—make up nearly 60% of the lexicon. Avoiding contour tones in Mandarin would make much of the lexicon off-limits, and presumably make for rather awkward and unnatural sounding lyrics. In Cantonese, however, the nature of the tone system makes it much easier to avoid potentially problematic contours, since there is really just one potentially troublesome tone. This is not to imply that Mandarin artistic traditions ignore tone tout court; this is clearly not the case, as shown by e.g. Chao (1956) and Stock (1999). Instead, it seems that the fact that the Cantonese tone system consists of (a) more and (b) primarily less dynamic tones gives it an advantage at this particular task. 4.2 Observed vs. expected ratios of Mandarin bigram sequences If we consider the observed/expected (O/E) ratios for the tonal bigrams set to rising, falling, and level melodic transitions (Tables 7a-7d), we see something interesting. Recall that the O/E ratio is a simple way of calibrating our expectations 3 Anecdotally, a number of Vietnamese songwriters have told me in informal discussions that they consciously try to avoid the high-rising “broken” tone (ngã) because they feel it is especially difficult to set. 224 JAMES P. KIRBY (a) T1-T1 T1-T2 T1-T3 T1-T4 Rise 0.58 0.80 0.49 0.80 T3-T1 T3-T2 T3-T3 T3-T4 Rise 1.41 1.39 1.21 1.43 (b) Fall 1.07 1.55 1.41 1.21 Level 1.62 0.30 1,12 0.95 Fall 0.66 0.76 0.86 0.61 Level 0.93 0.78 0.90 0.99 T2-T1 T2-T2 T2-T3 T2-T4 Rise 1.40 0.88 0.82 0.89 T4-T1 T4-T2 T4-T3 T4-T4 Rise 0.95 0.84 0.84 1.02 Fall 0.51 0.88 1.35 1.04 Level 1.13 1.43 0.63 1.13 Fall 1.00 1.13 1.11 1.05 Level 1.08 1.03 1.07 0.87 (d) (c) Table 7: O/E ratios for Mandarin melodic bigrams where corresponding texts are headed by (a) T1 /55/, (b) T2 /35/, (c) T3 /21/, (d) T4 /53/. about co-occurrences—-here, how often we should expect to see a particular tone sequence would be set to a particular melodic transition. When O/E is about 1, then a particular tonal bigram is set to a particular melodic transition about as often as one would expect if bigrams are independent of melodic transitions; values greater than or less than 1 thus indicate over- or underrepresentation respectively, under this assumption. O/E ratios for the Mandarin data were computed by tallying the number of tone bigrams (T1-T1, T1-T2...) that occurred with each type melodic transition (level, rising, or falling) and using these counts to generate the number of times that bigram would be expected to occur with a given transition. For example, there are 83 instances of T1-T4 sequences in the Mandarin corpus out of a total of 1447 total bigrams (ignoring transitions involving neutral tones or spanning phrase boundaries), of which 547 are rising note transitions; thus T1-T4 is expected to occur with a rising note transition (83 × 547)/1447 ≈ 31 times. Since this bigram is only set to rising transitions 25 times in the corpus, the O/E ratio is slightly less than 1 (0.8), meaning that this bigram occurs somewhat less often than we might expect with rising transitions, were the distributions of tonal bigrams and melodic sequences independent. First, consider the melodic bigrams headed by the high-level tone T1 (Table 7a). What we observe is that bigrams headed by T1 occur rather less than would be expected with rising note sequences, and rather more than would be expected with falling note sequences, some more than others. In other words, if the melodic transition is rising, tonal bigrams headed by T1 are disprefered, presumably because these would always involve some kind of fall on the tonal tier. Conversely, when the melodic transition is falling, sequences headed by T1 are overrepresented, the exception being oblique T1-T1 sequences. These same sequences are overrepresented on level musical transitions (O/E = 1.62). If we then consider the melodic bigrams headed by T3 (Table 7c)—that’s the low or low-falling tone, which only rises phrase-finally—we see the opposite pattern as we observed for bigrams headed by T1: namely, rises are overrepresented, while falls are underrepresented. (Note that T3-T3 sequences here are based on the underlying/lexical tone values of the Comparative tonal text-setting 225 syllables, not their surface sandhi values.) For comparison, we can also look at the bigrams headed by the dynamic contour tones, rising T2 and falling T4 (Tables 7b and 7d). Here, the pattern, if there is any, is far less obvious. It appears that T2 may be treated as a low tone when followed by T1 (overrepresentation of T2-T1 sequences on rising melodies) but high when followed by T3 (overrepresentation of T2-T3 sequences on falling melodies). However, bigrams headed by falling T4 are set to falling melodic transitions no more than would be expected if bigrams and melodic transitions were independent, nor do they appear to be especially disprefered with rising transitions. These differences between T1/T3 on the one hand and T2/T4 on the other suggest that, when Mandarin lyricists can avoid contrary motion and prefer similar motion, they do; it’s just that the facts about the tone distribution in this language make it difficult for them to do so very often, leading to correspondingly lower rates of tone-tune correspondence. 4.3 General constraints on tone-tune correspondence Recall that in Section 1.2, the fundamental tonal text-setting constraint was formulated as something like “avoid contrary settings” (*N ONPARALLEL4 in the parlance of McPherson & Ryan 2017). At least in Cantonese, this may also interact with a dispreference for large musical step size transitions, what McPherson & Ryan (2017) characterize as *S TEP. To this, we might add a third constraint penalizing the use of dynamic contour tones in texts, something like “avoid contour tones” (*C ONTOUR). As we are still at what we might call the constraint-discovery stage in the study of tone-melody correspondences, a formal treatment of how these constraints interact would be premature. Nevertheless, the Mandarin data indicate that a general dispreference for setting dynamic contour tones may well be important for the broader understanding of tonal text-setting and should be investigated in more detail in other languages, such as Thai, where the general rates of similar motion have been found to be low (List 1961; Saurman 1999; Ketkaew & Pittayaporn 2014; Kirby 2021). Moreover, the observed/expected analysis of Mandarin suggests that in at least some languages, even an optimal ranking of the relevant constraints may result in a sizeable number of contrary and oblique settings. 5 Conclusion and perspectives This paper has considered the extent to which three characteristic properties of Mandarin Chinese—a primarily polysyllabic lexicon, the existence of neutral tones, and the frequent application of tone sandhi—might account for its seemingly lax adherence to tone-melody correspondence constraints. Even when discounting transitions between syllables with neutral tones, as well as those which spanned word and musical phrase boundaries, the relative rates of tone-melody correspondence remained constant. This was also true of texts set in Cantonese to the exact 4 Implicitly or explicitly, terms such as “nonparallel” or “non-opposing” suggest that the notion of “oblique” may be reasonably treated as a kind of neutral, which as mentioned briefly in Section 1.2, seems unlikely to be empirically valid. For more discussion on this point, see Ladd (in press). 226 JAMES P. KIRBY same melodies, underscoring that the fact that the movements of tones and tunes are more closely aligned in Cantonese than they are in Mandarin is not due to differences in melody. Instead, I have suggested that the explanation may actually be primarily an accident of the languages’ tone systems, the relevance of which was also remarked on by Chan (1987b). However, while Chan drew a parallel between the effects of intonation and the “dominance” of melody over word tone in Mandarin, the current study has highlighted how the canonical phonetic realizations of the tones themselves may also play an important role. Cantonese contains (a) more tones which (b) mostly occupy single registers, with just one tone, the mid-high rising tone, involving a considerable pitch excursion, and this tone occurs less in the song corpus than one would expect based on its frequency in the lexicon (or in running text). Mandarin, on the other hand, has a smaller tone inventory overall, but well over half of the most commonly occurring tones are contour tones. To the extent that texts involving contour tones will more frequently result in non-similar settings, Cantonese lyricists are at a distinct advantage, given that syllables bearing contour tones make up a rather smaller proportion of the lexicon. In Mandarin, on the other hand, it is much more difficult to compose a text of any sort without making extensive use of contour tones. As emphasized by Ho (2010) (and more recently from an ethnomusicological perspective by Li 2021), low rates of tone-melody correspondence do not necessarily mean that non-corresponding sequences are judged as “ill-formed” or otherwise unacceptable to listeners. Other, potentially more subtle considerations are often at play. Nevertheless, text-setting constraints are clearly taken into account by lyricists and are salient to listeners, so it is worth studying them in detail. It is hoped that this paper has made a small contribution to exploring the space of possibilities, and may serve as an inspiration for more extensive future investigations. 6 Acknowledgments This study revises and extends an analysis based on data originally gathered by Ruoqi Lin for her 2018 University of Edinburgh MSc thesis “A comparison of tonal text-setting in Mandarin and Cantonese popular songs”. Thanks to Bob Ladd, as well as audiences at CLS58, the Chinese University of Hong Kong, and the 41st DGfS workgroup “Prosody from a cross-domain perspective: How language speaks to music (and vice versa)” for discussion, comments, and suggestions. Special thanks to Matthew Sung and Grace Wenling Cao for assistance with transliteration, data annotation and proofreading. The author remains solely responsible for any errors of fact or interpretation. This project was funded in part by AHRC grant AH/M005240/1 “Singing in tone: text-setting constraints in Southeast Asia”. Comparative tonal text-setting 227 References Alderete, J., Q. Chan, & S.-i. Tanaka. 2022. The morphology of Cantonese “changed tone”: Extensions and limitations. 言語研究（Gengo Kenkyu) 161. 139–169. Bauer, R. S., & P. K. Benedict. 1997. Modern Cantonese Phonology. De Gruyter Mouton. Cao, J. 1992. On neutral-tone syllables in Mandarin Chinese. Canadian Acoustics 20. 49–50. Chan, M. K. M. 1987a. Tone and melody in Cantonese. In Proceedings of the Thirteenth Annual Meeting of the Berkeley Linguistics Society (1987), p. 26–37. Chan, M. K. M. 1987b. Tone and melody interaction in Cantonese and Mandarin songs. UCLA Working Papers in Phonetics 68. 132–169. Chao, Y. R. 1956. Tone, intonation, singsong, chanting, recitative, tonal composition, and atonal composition in Chinese. In For Roman Jakobson: Essays on the occasion of his sixtieth birthday, ed. by M. Halle, H. Lunt, H. McLean, & C. H. van Schooneveld, p. 52–59. The Hauge: Mouton. Dell, F., & J. Halle. 2009. Comparing musical textsetting in French and in English songs. In Towards a typology of poetic forms: from language to metrics and beyond, ed. by J.-L. Aroui & A. Arleo, Language Faculty and Beyond (LFAB): Internal and External Variation in Linguistics, p. 63–78. Amsterdam: John Benjamins. Halle, J., & F. Lerdahl. 1993. A generative textsetting model. Current Musicology 55. 3–23. Hayes, B. 2009. Textsetting as constraint conflict. In Towards a typology of poetic forms: from language to metrics and beyond, ed. by J.-L. Aroui & A. Arleo, p. 43–62. Amsterdam: John Benjamins. Ho, W.-S. V. . 2006. The tone-melody interface of popular songs written in tone languages. In Proceedings of the 9th International Conference on Music Perception and Cognition (ICMPC 2006), ed. by M. Baroni, A. R. Addessi, R. Caterina, & M. Costa, p. 1414–1422. Ho, W.-S. V. . 2010. A phonological study of the tone-melody correspondence in Cantonese pop music. Hong Kong: University of Hong Kong dissertation. Jakobson, R. 1960. Closing statement: Linguistics and poetics. In Style in language, ed. by T. A. Sebok, p. 350–377. New York; London: The Technological Press; John Wiley & Sons. Jongman, A., Y. Wang, & C. B. Moore. 2012. Perception and production of Mandarin Chinese tones. In Handbook of East Asian psycholinguistics, Volume 1: Chinese, ed. by J. A. Sereno, P. Li, L. H. Tan, E. Bates, & O. J. L. Tzeng, p. 209–217. Cambridge: Cambridge University Press. Ketkaew, C., & P. Pittayaporn. 2014. Mapping between lexical tones and musical notes in Thai pop songs. In Proceedings of the 28th Pacific Asia Conference on Language, Information and Computation (PACLIC 28), p. 160–169. Kiparsky, P. 1977. The rhythmic structure of English verse. Linguistic Inquiry 8. 189–247. Kirby, J. 2021. Towards a comparative history of tonal text-setting practices in Southeast Asia. In Transcultural music history, ed. by R. Strohm, 291–312. Berlin: Berliner WissenschaftsVerlag. Kirby, J., & D. R. Ladd. 2016. Tone-melody correspondence in Vietnamese popular song. In Proceedings of the 5th International Symposium on Tonal Aspects of Languages (TAL-2016), p. 48–51, Buffalo. Ladd, D. R. In press. Two problems in theories of tone-melody matching. Studies in Prosodic Grammar (韵律语法研究) 8. Ladd, D. R., & J. Kirby. 2020. Tone-melody matching in tone language singing. In The Oxford Handbook of Language Prosody, ed. by C. Gussenhoven & A. Chen, p. 676–687. Oxford: Oxford University Press. Leung, M.-T., & S.-P. Law. 2001. HKCAC: the Hong Kong Cantonese adult language corpus. International Journal of Corpus Linguistics 6. 305–326. Leung, M.-T., S.-P. Law, & S.-Y. Fung. 2004. Type and token frequencies of phonological units in Hong Kong Cantonese. Behavior Research Methods, Instruments, & Computers 36. 500–505. Li, C., & S. Thompson. 1981. Mandarin Chinese: A Functional Reference Grammar. Berkeley, CA: University of California Press. 228 JAMES P. KIRBY Li, E. K. C. 2021. Cantopop and speech-melody complex. Music Theory Online 27. Lin, R. 2018. A comparison of tonal text-setting in Mandarin and Cantonese popular songs. University of Edinburgh MSc thesis. List, G. 1961. Speech melody and song melody in central Thailand. Ethnomusicology 5. 16–32. Lo, T. C. 2013. Correspondences between lexical tone and music transitions in Cantonese pop songs: a quantitative and analytic approach. Hons dissertation, The University of Edinburgh. McPherson, L., & K. M. Ryan. 2017. Tone-tune association in Tommo So (Dogon) folk songs. Language . Mitchell, T. 2006. Tian ci - Faye Wong and English songs in the Cantopop and Mandopop repertoire. In Access All Eras: Tribute Bands and Global Pop Culture, ed. by S. Homan, p. 215–228. Open University Press. Mok, P. P. K., D. Zuo, & P. W. Y. Wong. 2013. Production and perception of a sound change in progress: Tone merging in Hong Kong Cantonese. Language Variation and Change 25. 341– 370. Saurman, M. E. 1999. The agreement of Thai speech tones and melodic pitches. Notes on Anthropology 3. 15–24. Schellenberg, M. 2011. Tone contour realization in sung Cantonese. In Proceedings of the 17th International Congress of the Phonetic Sciences, p. 1754–1757. Schellenberg, M., & B. Gick. 2020. Microtonal variation in sung Cantonese. Phonetica 77. 83–106. Shih, C. 1997. Mandarin third tone sandhi and prosodic structure. In Studies in Chinese Phonology, ed. by J. Wang & N. Smith, 81–123. Berlin, Boston: De Gruyter. Stock, J. P. J. 1999. A reassessment of the relationship between text, speech tone, melody, and aria structure in Beijing Opera. Journal of Musicological Research 18. 183–206. Tanese-Ito, Y. 1988. The relationship between speech-tones and vocal melody in Thai court song. Musica Asiatica 5. 109–39. Wee, L. H. 2007. Unraveling the relation between Mandarin tones and musical melody. Journal of Chinese Linguistics 35. 128–144. Wong, P. C. M., & R. L. Diehl. 2002. How can the lyrics of a song in a tone language be understood? Psychology of Music 30. 202–209. Wu, Y., M. Adda-Decker, & L. Lamel. 2020. Mandarin lexical tones: A corpus-based study of word length, syllable position and prosodic position on duration. In Interspeech 2020, p. 1908–1912. Yung, B. 1983. Creative process in Cantonese opera I: the role of linguistic tones. Ethnomusicology 27. 29–47. Comparative tonal text-setting 229 Appendix: Song data Annotated song data and R scripts for analysis are available at https://osf.io/k6uma. Cantonese Pair Title 1 白玫瑰 2 回旋木马的终端 4 给自己的情书 5 下一站天后 6 明年今日 7 K 歌之王 8 残酷游戏 9 暗涌 Year 2006 2003 2000 2003 2002 2000 2009 1997 Performance 陈奕迅梁咏琪王菲 Twins 陈奕迅陈奕迅卫兰王菲 Lyrics 李焯雄林夕林夕黄伟文林夕林夕林夕林夕 Composition 梁翘柏林一峰江志仁伍乐城陈小霞陈辉阳蔡伯南陈辉阳 Arrangement 梁翘柏张人杰江志仁伍乐城陈辉阳陈辉阳雷颂德陈辉阳 Mandarin Pair Title 1 红玫瑰 2 遇见 4 笑忘书 5 莫斯科没有眼泪 6 十年 7 K 歌之王 8 痴心绝对 9 暗涌 Year 2007 2003 2001 2005 2003 2001 2002 2013 Performance 陈奕迅孙燕姿王菲 Twins 陈奕迅陈奕迅李圣杰杨丞琳 Lyrics 李焯雄易家扬林夕许常德林夕林夕蔡伯南林夕 Composition 梁翘柏林一峰江志仁伍乐城陈小霞陈辉阳蔡伯南陈辉阳 Arrangement 梁翘柏 Terrence Teo 江志仁 Mac Chew 陈辉阳陈辉阳陈飞午陈辉阳 Proceedings of CLS 58 (2022), 231-243 © Chicago Linguistic Society 2023. All rights reserved. 231 On the Adjectivalizer -Si in the Reduplicated and Deverbal Adjectives in Japanese Kazuya Kudo and Koji Shimamura Ryukoku U. and Kanazawa Gakuin U./Kobe City U. of Foreign Studies 1 Introduction The morphological process of reduplication is widely attested in many languages, and Japanese is no exception, where it exhibits a high degree of productivity. One example that will concern us in this paper is a case of reduplication that renders adjectives such as (1b). (1) a. Taroo-wa waka-i. Taro-TOP young-COP. PRES ‘Taro is young.’ b. Taroo-wa waka.waka-si-i. Taro-TOP young.young-ADJ - COP. PRES ‘Taro looks young.’ In (1a), the adjectival base waka- ‘young’ can be used as an adjective without reduplication, but when reduplicated as in (1b), it must be accompanied by -si, and the derived adjective assumes an additional meaning ‘look young’. Thus, Taro must be young regarding his actual age in (1a), but in (1b) he can be old to the extent that he looks young.1 One issue that we’d like to point out at this juncture is that -si is also used for deverbal adjectives. Consider (2).2 (2) a. Hanako-wa Taroo-no seikoo-o urayam-dei-ru. Hanako-TOP Taro-GEN success-ACC envy-ASP - PRES ‘Hanako envies Taro’s success.’ b. Hanako(-ni)-wa Taroo-no seikoo-ga urayam-a-si-i. Hanako-DAT- TOP Taro-GEN success-NOM envy-A - ADJ - COP. PRES ‘Hanako is envious of Taro’s success.’ In (2b), -si is added to the verbal stem urayam- ‘envy’ in order for the deverbal adjective urayam-a-si-i ‘envious’ to be derived. Then, a natural question that will 1 In some out-of-blue context, we cannot say (1a) if Taro is in his 60s or 70s. Alternatively, old as he may be, we can utter it if there is any objective evidence for Taro’s youthfulness, e.g., he runs a marathon regularly. 2 In (2b) the particle -a is inserted between urayam- and -si. This particle is found in various forms in other derivations of Japanese adjectives, and for the sake of brevity, we assume it to be inserted for phonological reasons. 232 KUDO AND SHIMAMURA arise for (1b) is why -si is possible with reduplication. Relevantly, we cannot have it with a base adjective, so *waka-si-i is impossible. At face value, -si in (1b) looks like a nonderivational (non-category-changing) suffix that simply adds the “look” meaning to the base adjective. However, this does not go through in (2b) since it changes the category of the base word (i.e. verb) into an adjective. Furthermore, the deverbal adjective in (2b) does not have the “look” meaning as in the reduplicated adjective in (1b). In this paper, we will present a uniform analysis of -si in (1b) and (2b) in the framework of Distributed Morphology (Halle & Marantz 1993). Based on the structure of Japanese adjectives put forth by Nishiyama (1999), we contend that -si in the above examples is a morphological realization of an adjectival head a (cf. Arad 2003; Marantz 1997; Marantz 2000). Evidence of our analysis comes from the phonological process called Sequential Voicing (SV), or Rendaku (Otsu 1980; Tatsumi 2022), as well as the interpretations of the reduplicated and deverbal adjectives. This paper is organized as follows. In Section 2, we will go over the analysis of Japanese adjectives proposed by Nishiyama (1999). Then, in Section 3, adopting Nishiyama’s copular structure, we will look into the pertinent two cases of adjectival formation: the reduplicated and deverbal adjectives, with a special attention to -si. Section 4 provides two pieces of evidence that support our analysis of the reduplicated adjective: the obligatory SV (Section 4.1) and its idiomaticity (Section 4.2). Then, in Section 5, we will consider what category can be a root in rendering a reduplicated adjective, and show that nominal items can also enter into the process. In Section 6, we will discuss why *waka-si-i is impossible even if our analysis apparently allows it to be derived, arguing with Bobaljik (2017) that constructing the structure of such a nonexistent adjective per se is possible but Spelling-Out it is impossible. Finally, Section 7 will conclude. 2 The Structure of Japanese Adjectives: Nishiyama (1999) Nishiyama (1999) argues that Japanese has two kinds of adjectives: Canonical Adjectives (CA) and Nominal Adjectives (NA). (3) Canonical Adjectives (CA) a. Yama-ga {taka-i/takak-at-ta}. mountatin-NOM high-COP. PRES/high-COP - PAST ‘The mountain {is/was} high.’ b. Miti-ga {hiro-i/hirok-at-ta}. road-NOM wide-COP. PRES/wide-COP - PAST ‘The road {is/was} wide.’ (4) Nominal Adjectives (NA) a. Yoru-ga {sizuka-da/sizuka-dat-ta}. night-NOM quiet-COP. PRES/quiet-COP - PAST ‘The night {is/was} quiet.’ The adjectivalizer -si 233 b. Hon-ga {kirei-da/kirei-dat-ta}. book-NOM pretty-COP. PRES/pretty-COP - PAST ‘The book {is/was} pretty.’ (based on Nishiyama 1999: 183) Nishiyama proposes that CA and NA both involve Predicate Phrase (PredP) in the sense of Bowers (1993), which we assume is the source of the theme argument of these adjectives (cf. Baker 2003). Thus, the structures for the adjectives in (3a) and (4a) (as used in the past forms) are as follows: (5) b. The Structure of NA a. The Structure of CA TP TP VP V PredP DP T Pred0 T VP -ta -ar V PredP Pred0 DP AP Pred AP Pred taka -k(u) sizuka -d(e) -ar In (5), -ar, which Nishiyama argues is a dummy copula to support the tense, undergoes a phonological change of gemination, becoming -at, and the Pred heads -ku and -de become -k and -d, respectively, when they are used in the past form. Then, the CA taka-ku-ar-ta and the NA sizuka-de-ar-ta are pronounced in reality as taka-k-at-ta and sizuka-d-at-ta, respectively. In the present form, this dummy copula is not used for CA, but once we have a focus particle like -mo ‘also’ between Pred and V, it surfaces: (6) a. taka-ku-mo-ar-u high-PRED-also-DUM . COP - PRES ‘be also high’ b. sizuka-de-mo-ar-u qietness-PRED-also-DUM . COP - PRES ‘be also quiet’ Then, Nishiyama argues that CA in the present form does not require a dummy copula, and that Pred must be morphologically null. Therefore, the structure of CA in the present tense can be described as in the following: -ta 234 KUDO AND SHIMAMURA (7) The Structure of CA (Present Tense) TP PredP T Pred0 DP AP Pred taka -k ⇒ φ -i According to Nishiyama, CA like taka-i is underlyingly taka-k-i due to the fact that Old Japanese used to have -k overtly. In what follows, we assume that Nishiyama is on the right track regarding the copula structure in Japanese adjectives, and will be exclusively concerned with the stem formation of CA. 3 -Si in Japanese Adjectives With Nishiyama’s (1999) analysis at hand, let us consider the pertinent adjectival formation given in (1), repeated here as (8). (8) a. Taroo-wa waka-i. Taro-TOP young-COP. PRES ‘Taro is young.’ b. Taroo-wa waka.waka-si-i. Taro-TOP young.young-ADJ - COP. PRES ‘Taro looks young.’ With respect to the morphosyntactic properties of the reduplicated adjectives, it is obvious that we now have an additional suffix, -si. The conjugational patterns of this (bound) morpheme follow those typical to the CA waka-i in that -k(u) appears in the past and negative forms as well as where -mo ‘also’ is inserted: (9) a. waka.waka-si-k-at-ta young.young-ADJ - PRED - DUM . COP - PAST ‘looked young’ b. waka.waka-si-ku-nai young.young-ADJ - PRED - NEG . PRES ‘not look young’ c. waka.waka-si-ku-mo-ar-u young.young-ADJ - PRED-also-DUM . COP - PRES ‘also look young’ Then, considering the adjectival structure proposed by Nishiyama (1999), the morpheme order in (9) tells us that the reduplicated adjective plus -si is in the place of The adjectivalizer -si 235 AP. More specifically, we argue with the spirit of Distributed Morphology that the √ stem of the reduplicated adjective is composed of a root ( ) and an adjectivalizer a. Thus, the structure of the reduplicated adjective in Japanese is: (10) The Structure of Reduplicated Adjectives TP PredP T -i Pred0 DP Pred aP √ √ WAKA . WAKA WAKA √ WAKA a -k ⇒ φ -si Here, we propose that -si is a category-determining head which takes a root and categorize it as adjectival. For reduplicated adjectives, therefore, what is reduplicated is a root, not an AP. This root complex is then merged with -si, an adjectivalizer, forming an aP. Finally, an aP is merged with a Pred, which is morphologically null but appears as -k(u) in the past tense with the dummy copula ar-, according to Nishiyama (1999). As already alluded to above, we assume that the theme argument is merged to Spec-PredP so that aP does not require an argument (cf. Baker 2003). The same analysis is carried over to deverbal adjectives in (2), repeated here as (11). (11) a. Hanako-wa Taroo-no seikoo-o urayam-dei-ru. Hanako-TOP Taro-GEN success-ACC envy-ASP - PRES ‘Hanako envies Taro’s success.’ urayam-a-si-i. b. Hanako(-ni)-wa Taroo-no seikoo-ga Hanako-DAT- TOP Taro-GEN success-NOM envy-A - ADJ - COP. PRES ‘Hanako is envious of Taro’s success.’ For deverbal adjectives, what is combined with -si is a verb, not a root. Then, the structure of the deverbal adjective urayam-a-si-i ‘envious’ is given in (12).3 3 In (12) we assume that -m is an verbalizer in Japanese, since Japanese has many examples of deverbal adjectives whose verbal stem ends with -m: isam- ‘cheer up’, itam- ‘hurt’, urayam- ‘envy’, utom- ‘distance oneself’, konom- ‘prefer’, nayam- ‘bother’, yam- ‘get sick’, etc. Etymologically, however, urayam- ‘envy’ can be decomposed into ura ‘mind’ and yam- ‘get sick’. 236 (12) KUDO AND SHIMAMURA The Structure of Deverbal Adjectives TP PredP T -i Pred0 DP Pred aP a vP √ URAYA v -k ⇒ φ -si -m(a) Here we propose that -si is also a category-changing head which takes a vP unit and turns it into an adjective. Also, we assume that the theme argument for the deverbal adjective is still selected by a Pred, not by a v, just as in the case of the reduplicated adjective in (10). A piece of evidence for this structure comes from the fact that the theme argument (e.g. Taroo-no seikoo in (11b)) cannot bear an accusative case, irrespective of the case of the subject. (13) * Hanako(-ni)-wa Taroo-no seikoo-o urayam-a-si-i. Hanako-DAT- TOP Taro-GEN success-ACC envy-A - ADJ - COP. PRES Intended ‘Hanako is envious of Taro’s success.’ According to Ura (2000), the experiencer argument can be a dative subject in the potential construction as in (14). (14) a. b. c. * Taroo-ni-wa doitugo-o hanas-e-ru. Taro-DAT- TOP German-ACC speak-can-PRES Intended: ‘Taro can speak German.’ Taroo(-ni)-wa doitugo-ga hanas-e-ru. Taro-DAT- TOP German-NOM speak-can-PRES ‘Taro can speak German.’ Taroo-wa doitugo-o hanas-e-ru. Taro-TOP German-ACC speak-can-PRES ‘Taro can speak German.’ What is of importance here is that when the dative subject appears, the theme argument must bear a nominative case, whereas the accusative theme can be used if the subject does not bear a dative case. Then, the ungrammaticality of the accusative theme in (13) is different from that in (14a) by nature, since it is ungrammatical The adjectivalizer -si 237 anyway whether the experiencer bears a dative case or not. We suggest that the ungrammaticality of the accusative case for the theme argument in (13) comes from the fact that the theme argument of the deverbal adjectives is independent of any verbal projection, as the structure in (12) shows. 4 Evidence for Root Compounding 4.1 Sequential Voicing As we have just proposed, reduplicated adjectives such as (8b) are structured as in (10), where -si is a category-determining head, and we argue that it constitutes a phase (Marantz 1997). A piece of evidence for this claim comes from a phonological process called Sequential Voicing (SV), or Rendaku, where the first consonant of the latter of the two words combined becomes voiced (Otsu 1980). For instance, combining the nominal kara ‘dryness’ and the verbal huk- ‘wipe’ creates a compound noun, kara-buki ‘dry-wiping’, where the bilabial voiceless fricative [h] ([F], more precisely) turns into the bilabial voiced stop [b]. Tatsumi (2022) recently proposes that in Japanese root-level compounds must undergo SV when it is possible.4 Details aside, since our analysis in (10) employs the root-level compound structure, it is predicted that SV must be applied for reduplicated adjectives. This prediction is indeed borne out: (15) a. karu-i ⇒ karu.garu-si-i (*karu.karu-si-i) light-COP. PRES light.light-ADJ - COP. PRES ‘light → {careless/thoughtless}’ b. samu-i ⇒ samu.zamu-si-i (*samu.samu-si-i) cold-COP. PRES cold.cold-ADJ - COP. PRES ‘cold → {looks cold/desolate}’ We argue that the reduplicated adjective in (15a) should be analyzed as in (16). (16) TP PredP T -i Pred0 DP Pred aP √ √ 4 KARU . GARU KARU √ KARU a -k ⇒ φ -si Nishiyama (2017) also discusses the correlation between root-level compounds and SV. 238 KUDO AND SHIMAMURA If karu- is reduplicated after it becomes an adjective via merging the phasal head a to the root, a should block SV as Tatsumi (2022) proposes. Thus, to the extent that Tatsumi’s analysis of SV is correct, the obligatory SV in (15) supports our analysis. 4.2 Idiomaticity Turning to the meaning of reduplicated adjectives, we see that they can create special meanings that are unavailable to their base forms. For example, karu.garu-si-i in (15a) and samu-zamu-si-i in (15b) yield the idiomatic interpretations of ‘careless’ and ‘desolate’, respectively. Adopting the phase-based Distributed Morphology approach toward these sorts of interpretational facts (Marantz 1997; Marantz 2000), we take this rather unpredictable nature of the meanings of reduplicated adjectives as indicating that their special/idiomatic meanings are determined when the first category-determining node a is merged: (17) The Domain of the Idiomatic Interpretation √ [P reP [aP . . . a ] Pred ] Our structure in (16) fits the idea put forth by Marantz (1997). The meaning of the relevant√adjective is determined when a is merged to the root: reduplication is applied to KARU, not after but before it becomes an adjective. 5 What Can Be the Root? What we have seen concerning the reduplicated adjectives has the adjectival base form. However, nominal bases can also enter into the pertinent reduplication. Consider (18). (18) a. baka.baka-si-i fool.fool-ADJ - COP. PRES ‘look/sound foolish’ b. me.me-si-i woman.woman-ADJ - COP. PRES ‘sissy/coward’ The base baka ‘fool’ in (18a) can be used as a noun, and me ‘woman’ in (18b) is a bound morpheme which cannot be used by itself. Also, witness the examples in (19), which exhibit the reduplication of nominal stems: (19) a. sora ⇒ sora.zora-si-i sky sky.sky-ADJ - PRES ‘sky → {untrue/nonrealistic}’ b. kai ⇒ kai.gai-si-i worthiness worthiness.worthiness-ADJ - PRES ‘worthiness → {efficient/brisk/devoted}’ The adjectivalizer -si 239 c. hana ⇒ hana.bana-si-i flower flower.flower-ADJ - PRES ‘flower → {glorious/splendid/spectacular}’ Note that all the examples in (19) involve SV, so that we can make sure that they are root-level compounds. Also notable is that the resulting adjectives in (19) do not retain the original meanings of their base nouns. They rather obtain special/idiomatic readings (albeit not completely unrelated to the meanings of the bases). These facts can be explained by assuming that what is reduplicated here is not a word but a root yet to be merged with a category-determining head. 6 On the Non-existence of *Waka-si-i: -Si or φ So far, we have been concerned with the morphological makeup of -si adjectives. However, merging -si to the root or vP does not always come for free, and there are distributional asymmetries: that is, almost all the reduplicated adjectives do not have their non-reduplicated/base adjectival counterparts, as shown in (20). (20) a. waka.waka-si-i/*waka-si-i young.young-ADJ - PRES/young-ADJ - PRES ‘look young/(Intended) young’ b. ui.ui-si-i/*ui-si-i first.firrst-ADJ - PRES/first-ADJ - PRES ‘innocent/(Intended) first’ c. im-a.im-a.si-i/*im-a-si-i hate-A.hate-A - ADJ - PRES/hate-A - ADJ - PRES ‘irritating/(Intended) hateful’ We assume in line with the framework of Distributed Morphology that making such adjectives itself is possible as far as syntax is concerned, but there are no corresponding words to realize them morphologically (Bobaljik 2017). In principle, the DM framework does not block the formation of the underlying structure for making the non-reduplicated adjectives in (20), so that the sole structure-building engine, i.e. syntax, does compose the non-reduplicated waka- ‘young’ with -si. However, it is simply not a word in Japanese, just like we do not have a causative version of the verb arrive or a verbal use of the noun cat in English. In short, Distributed Morphology predicts that the non-reduplicated adjectives in (20) should be possible, but they are not, at least for many Japanese.5 5 Interestingly, a quick Google search takes us to a few examples of waka-si-i. For instance: (i) mineraru-yutaka-na waka-si-i kyube mineral-abundant-COP. ADN young-SI - PRES cuvée ‘cuvée (wine) that is rich in minerals and not aged ’ We are not sure whether (i) is just made due to a mistake, but if it is created as it is intended, it gives a case where waka-si-i is possible (at least for a small set of Japanese speakers). 240 KUDO AND SHIMAMURA In this connection, we argue that waka-i ‘young’ is also derived via the a head, which is however morphologically null, so the structure is as follows: (21) TP T PredP -i Pred0 DP Pred aP √ WAKA a -k ⇒ φ φ Thus, we suggest that all Japanese adjectival heads appear in either -si or φ. It is difficult to completely predict which adjectives will appear in which form, but the tendency seems to be such that they often appear in φ when they describe states based on objective judgments, such as aka-i ‘red’, taka-i ‘high’ and waka-i ‘young’, and in -si when they describe states based on subjective judgments, such as ure-si-i ‘happy’, kana-si-i ‘sad’ and tano-si-i ‘fun’.6 Our analysis is supported by the fact that if there is a phase boundary between the two adjectives, SV does not apply: (22) a. ure-si-i + tano-si-i ⇒ happy-ADJ - PRES delightful-ADJ - PRES ure-si.tano-si-i (*ure-si.dano-si-i) happy-ADJ.delightful-ADJ - PRES ‘happy and delightful’ b. ito-si-i + koi-si-i ⇒ beloved-ADJ - PRES longed.for-ADJ - PRES ito-si.koi-si-i (*ito-si.goi-si-i) beloved-ADJ.longed.for-ADJ - PRES ‘beloved and longed for’ Albeit this type of compound formation is not so productive, the resulting amalgams do nothing semantically but just a coordination of two adjectives. If the stems of these adjectives such as uresi- ‘happy’ and tanosi- ‘delight’ are a root, then we would predict that (22a) and (22b) undergo SV, contrary to fact. We thus suggest that this is a case of aP-aP compounding as in (23). 6 The only exception of the complementary distribution of -si and φ in Japanese adjectives is the pair of chika-i ‘close’ and chika-si-i ‘close’. Historically, chika-si-i was derived from chika-i, and chika-i is now used to describe the physical closeness of distance or time while chika-si-i expresses the psychological closeness of relationship. The adjectivalizer -si (23) 241 aP3 aP1 √ URE aP2 a √ TANO -si a -si This is also true to the coordination of adjectives which does not have -si. Witness: (24) a. ama-i + kara-i ⇒ ama.kara-i (*ama.gara-i) sweet-PRES spicy-PRES sweet.spicy-PRES ‘sweet and spicy’ b. kimo-i + kawai-i ⇒ kimo.kawai-i (*kimo.gawai-i) weird-PRES cute-PRES weird.cute-PRES ‘weird but cute’ Again, SV cannot be applied in (24) and the compound adjectives are semantically a simple concatenation of two adjectives in the sense that they don’t have any peculiar/unpredicted meanings from the combined items. Thus, we can carry over the analysis in (23) to (24), but in this case, the a head is null as in (25). (25) aP3 aP1 √ AMA aP2 a φ √ KARA a φ Since the phasal head a demarcates the two combined adjectives for their meanings and SV domains, the semantic and phonological properties of (22) and (24) are explained in the same fashion. 7 Conclusion In Japanese, -si appears in both reduplicated and deverbal adjectives; the former is seemingly a case of non-category-changing affixation, and the latter, that of category changing (i.e. derivational) affixation. We have proposed a uniform analysis of -si in these adjectives, contending that -si is an adjectival head a as postulated in Distributed Morphology, and given a couple of consequences of our analysis in terms of SV (Rendaku) and idiomaticity of derived adjectives. To be specific, we have argued that -si is a category-determining head a that can select a root or a vP. Insofar as our analysis of -si is on the right track, the pertinent morpheme categorizes root items as, or category-changes verbs into, adjectives. We have also discussed how adjectives without -si is formed, and propose that such adjectives 242 KUDO AND SHIMAMURA also have a category-determining head a but it is morphologically null. Supporting evidence, as we have seen, concerns a certain case of adjectival coordination, which does not undergo SV. Then, it follows from our analysis that Japanese has two instances of the adjective-making head a, -si or φ, the choice of which depends on the semantic subjectivity/objectivity of the resulting adjectives. References Arad, M. 2003. Locality constraints on the interpretation of roots: The case of Hebrew denominal verbs. Natural Language and Linguistic Theory 21. 737–778. Baker, M. C. 2003. Lexical Categories: Verbs, nouns and adjectives. Cambridge: Cambridge University Press. Bobaljik, J. D. 2017. Distributed morphology. In Oxford Research Encyclopedia of Linguistics. Oxford University Press. Bowers, J. 1993. The syntax of predication. Linguistic Inquiry 24. 591–656. Halle, M., & A. Marantz. 1993. Distributed morphology and the pieces of inflection. In The View from Building 20, ed. by K. Hale & S. J. Keyser, 111–176. Cambridge, MA: MIT Press. Marantz, A. 1997. No escape from syntax: Don’t try morphological analysis in the privacy of your own lexicon. In Proceedings of the 21st Annual Penn Linguistics Colloquium, ed. by A. Alexiadou, L. Siegel, C. Surek-Clark, & A. Williams, volume 4, 201–225. Marantz, A. 2000. Roots: The universality of root and pattern morphology. paper presented at the conference on Afro-Asiatic languages, University of Paris VII. Nishiyama, K. 1999. Adjectives and the copulas in Japanese. Journal of East Asian Linguistics 8. 183–222. Nishiyama, K. 2017. Phrasal compounds in Japanese. In Further investigations into the nature of phrasal compounding, ed. by C. Trips & J. Kornfilt, 149–183. Berlin: Language Science Press. Otsu, Y. 1980. Some aspects of rendaku in Japanese and related problems. In MIT Working Papers in Linguistics: Theoretical Issues in Japanese Linguistics, ed. by Y. Otsu & A. Farmer, 207– 228. Cambridge, MA: MIT, Department of Linguistics and Philosophy, MITWPL. Tatsumi, Y. 2022. Structural restrictions on sequential voicing in Japanese N-V compounds. In Proceedings of Japanese/Korean Linguistics 29, ed. by K. Horie, K. Akita, Y. Kubota, D. Y. Oshima, & A. Utsugi., 99–112. Stanford, CA: CSLI Publications. Ura, H. 2000. Checking theory and grammatical functions in universal grammar. Oxford: Oxford University Press. Proceedings of CLS 58 (2022), 243-254 © Chicago Linguistic Society 2023. All rights reserved. 243 Murrinhpatha number mismatch as partial Agree Elango Kumaran University of Southern California 1 Introduction Number marking on verbs in Murrinhpatha (non-Pama-Nyungan, Australia) has two striking properties. First, apparent mismatches occur: singular marking is used for some dual subjects and dual marking is used for some paucal subjects. Second, the singular-dual mismatch is conditioned by the position of an apparently unrelated element, the non-sibling marker ngintha/nintha. In (1), ngintha is prefixal and the word-initial prefix does not match the subject’s number, whereas in (2), ngintha is suffixal and there is no mismatch. Nordlinger and Mansfield (2021) argue that this pattern diverges from known morphotactic principles reported in the literature. (1) (2) bangintha- ngkardu -nu 1SG.13.FUT- NSIB.DU . F- see -FUT (Nordlinger and Mansfield 2021:6) ‘We (two non-siblings) will see it.’ ngubanhingkardu -nu -ngintha 1DU.13.FUT- 2SG.OBJ- see -FUT -NSIB . DU . F ‘We (two non-siblings) will see you.’ (Nordlinger and Mansfield 2021:6) In this paper, I propose that the facts emerge solely from the action of Agree (Chomsky 2000) operating over a featurally complex representation of number (Harbour 2014). Mismatches result from partially-defective intervention: Agree takes place not with the subject itself, but with the non-sibling marker – an intervening element which bears a subset of the subject’s features. The analysis hinges on some interesting theoretical assumptions, listed in (3). (3) a. b. c. d. Number is encoded using Harbour’s (2014) feature system. All else equal, learners will posit the smallest set of probe features necessary to encode the contrasts present in the agreement morphology (building on Preminger 2019). Morphemes expone meanings denoted by feature sets (rather than exponing features directly). Mobile affixation is syntactic (at least in this instance). The paper is structured as follows. Section 2 presents the relevant generalizations, following Nordlinger and Mansfield 2021. Section 3 introduces assumptions (3a) and (3b). Section 4 proposes that mismatched number agreement is derived I am especially grateful to Sam Zukoff and Rachel Nordlinger for helpful feedback and discussion. Thanks also to Roumyana Pancheva, USC’s S Side Story, my LING 547 classmates at USC, the attendees of PLC 46 and CLS 58, and reviewers for PLC 46, WCCFL 40, and CLS 58. 244 ELANGO KUMARAN from partially-defective intervention. I show that this analysis is able to account for the full pattern, provided we adopt assumptions (3c) and (3d). Section 5 discusses a puzzle that the mobile affixation pattern raises for linearization and modularity. Section 6 briefly concludes. 2 The pattern Murrinhpatha is a polysynthetic language with radical pro-drop. It is a member of the Southern Daly language family (northern Australia; see Green and Nordlinger n.d. for information on the languages of the Daly region). There is a four-way number system in the pronoun morphology: singular, dual, paucal, plural. Verbs take a prefix called the classifier (bolded in (4)) – a portmanteau encoding conjugation class1 (glossed as a numeral), tense, aspect, mood, subject person, and subject number. This paper focuses on the number-marking function of the classifier. (4) a. b. bangkardu -nu 1SG.13.FUT- see -FUT ‘I will see it’ ngubangkardu -nu 1DU.13.FUT- see -FUT ‘We (two siblings) will see it.’ (Nordlinger and Mansfield 2021:4) (Nordlinger and Mansfield 2021:4) Strikingly, kinship (specifically, non-siblinghood2 ) of dual and paucal subjects is obligatorily marked in the verbal complex.3 (4b), which lacks a non-sibling marker, is obligatorily interpreted as having a sibling subject. A dual non-sibling subject requires a non-sibling marker, as in (5).4 (5) bangintha- ngkardu -nu 1SG.13.FUT- NSIB.DU . F- see -FUT (Nordlinger and Mansfield 2021:6) ‘We (two non-siblings) will see it.’ Note that (5) exhibits a number mismatch: the classifier is singular, even though the subject is dual. I will propose that this type of mismatch is due to intervention of the non-sibling marker. (Compare to (5) to (4b), which lacks the non-sibling marker and accordingly lacks a number mismatch.) For expository purposes, I will use constructed examples to illustrate the full pattern. All generalizations are drawn from Nordlinger and Mansfield 2021, and all constructed examples are based on Mansfield 2019. 1 The classifier may be a sort of light verb, with each ‘conjugation class’ associated with a different light-verb semantics (see Nordlinger and Mansfield 2021 and references therein). 2 This paper follows a terminological convention well established in the literature on Murrinhpatha (and other Australian languages) whereby siblinghood is to be understood as a relation which holds not only between children of the same parents, but also between children of brothers and between children of sisters (see Davidson 2018:54 for discussion). 3 For an analysis of the diachronic evolution of non-sibling marking, see Blythe 2010. 4 Non-sibling marking can also be controlled by the object (Nordlinger 2015:506-507). I will not be discussing this here, but it should be looked at in future work. Murrinhpatha number mismatch as partial Agree 245 The classifiers exhibit a three-way number contrast, conflating paucal and plural: singular, dual, paucal/plural. With a singular subject, the singular classifier is used: (6) ngurdungkarl -nu 1SG.29.FUT- bring.back -FUT ‘I will bring it back.’ (SG subject) (constructed example) With a dual sibling subject, the dual classifier is used: (7) ngurdrda- ngkarl -nu 1DU.29.FUT- bring.back -FUT ‘We will bring it back.’ (DU sibling subject) (constructed example) And with a pacual sibling subject or a plural subject, the paucal/plural classifier is used. There is no non-sibling marking for plural subjects (which makes functional sense since there is no ambiguity – plurals are always non-sibling, practically speaking): (8) ngurdrdungkarl -nu 1PC / PL.29.FUT- bring.back -FUT ‘We will bring it back.’ (PC sibling or PL subject) (constructed example) It is with dual and paucal non-sibling subjects that number mismatches arise. With a dual non-sibling subject, the dual non-sibling marker appears. The form of the dual non-sibling marker is sensitive to gender: ngintha (F) vs. nintha (M). The masculine is only used for uniformly masculine groups, with the feminine used otherwise. By default, the dual non-sibling marker surfaces as a prefix, as in (9). The prefixal dual non-sibling marker triggers a number mismatch: the classifier is singular in (9), even though the subject is dual. (9) ngurdungintha- ngkarl -nu 1SG.29.FUT- NSIB . F. DU- bring.back -FUT ‘We will bring it back.’ (DU non-sibling subject) (constructed example) The paucal non-sibling markers are -ngime (F) and -neme (M). The paucal nonsibling marker always surfaces as a suffix, and always triggers a number mismatch: the classifier is dual in (10), even though the subject is paucal. (10) ngurdrda- ngkarl -nu -ngime 1DU.29.FUT- bring.back -FUT -NSIB . F. PC ‘We will bring it back.’ (PC non-sibling subject) (constructed example) It is important to note that even when the classifier shows a mismatch in number with the subject, is not the case that the number of the classifier is independent of the number of the subject. The number of the subject still affects the number of the mismatched classifier: (9) has a singular mismatched classifier because the subject is dual, whereas (10) has a dual mismatched classifier because the subject is paucal. Popp (2022a,b) posits that the singular classifier is actually a default 246 ELANGO KUMARAN form uninflected for number, and that the apparent singular-dual mismatch in cases like (9) reflects a complete lack of number on the classifier. This analysis cannot extend to the dual-paucal mismatch in cases like (10), which Popp does not address. Accordingly, I will pursue a different line of analysis, treating the singular-dual and dual-paucal mismatches as instances of partial agreement, i.e. agreement with a subset of the subject’s features. What we have seen so far is summarized in (11). (11) SUBJECT SIBLINGS ? singular dual dual paucal paucal plural no yes no yes no no NON - SIBLING MARKER ngintha (F) / nintha (M) ngime (F) / neme (M) CLASSIFIER EX . singular (6) dual (7) singular (9) paucal/plural (8) dual (10) paucal/plural (8) There is an exception to (11): in some cases with a dual non-sibling subject, a non-mismatching dual classifier appears. This happens when the prefixal templatic slot that the dual non-sibling marker would otherwise occupy is unavailable due to object agreement. There is no object agreement for third person singular direct objects, but other objects do trigger agreement. The object agreement marker surfaces in the prefixal slot. When object agreement is present, the dual non-sibling marker cannot appear in prefixal position; instead, it is suffixal, as in (12). Somehow, this displacement of the dual non-sibling marker eliminates the number mismatch: the classifier in (12) is dual. (Compare to (9), which has a singular classifier.) Nordlinger and Mansfield (2021) identify this phenomenon of ‘positional dependency’ – i.e. a change in the form of one element (the classifier) conditioned by the linear position of another (the dual non-sibling marker) – as divergent from known morphotactic principles reported in the literature.5 (12) ngurdrda- nhingkarl -nu -ngintha 1DU.29.FUT- 2SG . OBJ- bring.back -FUT -NSIB . F. DU ‘We will bring you back.’ (DU non-sib. subject) (constructed example) In sections 3 and 4, I develop an Agree-based account of (11). The account also predicts (12), provided we stipulate that affix mobility is syntactic. This raises some issues for linearization and modularity, however, which I dicuss in section 5. 3 The feature specification of goals and probes 3.1 The featural encoding of number I adopt Harbour’s (2014) feature encoding of number, which is designed to capture the attested typology of number systems more accurately than other feature systems. In Harbour’s model, the four-way number system of Murrinhpatha can be encoded using three binary features. The features and their meanings are listed in (13).6 5 6 See also Popp 2022a,b for a Stratal Optimality Theory analysis. For precise denotations of these features, I refer the reader to Harbour 2014. Murrinhpatha number mismatch as partial Agree (13) 247 [+atomic] = singular [+additive] = plural [+minimal] = as small as possible, taking into account any other features combined with [–atomic], [–additive], [–minimal] = not [+atomic], not [+additive], not [+minimal] respectively [±atomic] separates singular (atoms) from non-singular. [±additive] separates plural (the only number closed under addition – every union of plurals is plural) from non-plural. [±minimal] is more interesting, because of the way it combines with other features. If [±minimal] exists in isolation, without any other feature present, [+minimal] will correspond to singular and [-minimal] will correspond to non-singular. If, on the other hand, [±minimal] combines with [–atomic], for instance, then [+minimal] will correspond to dual and [-minimal] will correspend to paucal/plural. This behavior of the [±minimal] feature will play a key role in deriving the number mismatch pattern. The number feature specification of subjects is as follows:7 (14) singular subject: [+atomic][+minimal][-additive] dual subject: [-atomic][+minimal][-additive] paucal subject: [-atomic][-minimal][-additive] plural subject: [-atomic][-minimal][+additive] 3.2 A lexical economy condition on probes Preminger (2019) argues (based on the distribution of Person Case Constraint effects) that in languages or clause types which lack morphological agreement altogether, there is no abstract syntactic agreement. It seems that the learner’s default assumption is that all heads lack probes. The learner only posits a probe when presented with clear, direct evidence for a probe. I propose that this principle applies not only to the presence or absence of agreement altogether, but also to individual features on a head – learners prefer not to posit unnecessary probe features. The result is that each head bears the smallest set of probe features necessary to encode the constrasts present in its agreement morphology. This means that the classifier head, which conflates paucal and plural in the morphology, does not have a [ additive] probe feature. The [±additive] feature is solely responsible for differentiating between plural and non-plural, which is not a necessary contrast for the classifier morphology. The [±atomic] and [±minimal] features, on the other hand, are both necessary to encode the three-way number constrast present in the classifier morphology, so the classifier does have [ atomic] and [ minimal] probe features. More importantly, the non-sibling marker only makes a two-way number distinction, so it only needs one number probe feature. When the subject is dual, ngintha/nintha is used, and when the subject is paucal, ngime/neme is used. The fea7 Note that the feature combinations listed are the only valid combinations of these three features. All other logically possible combinations are semantically incoherent. See Harbour 2014. 248 ELANGO KUMARAN ture which tracks this contrast is [± minimal], so I assume the non-sibling marker bears a [ minimal] probe feature.8 4 Mismatched number agreement as partially-defective intervention The fact that the non-sibling marker has fewer features than the classifier is crucial in deriving mismatched agreement. I propose that mismatched agreement occurs when the classifier receives the subject’s features from the non-sibling marker instead of from the subject itself. Since the non-sibling marker is missing the [±atomic] feature (following section 3.2.), incomplete information is passed on to the classifier about the subject. Non-mismatched agreement occurs when the classifier Agrees directly with the subject, i.e. when the subject is the highest element bearing number features in the classifier’s c-command domain. Non-mismatch scenarios with dual and paucal subjects are shown in (15). The classifier ends up with the features [– atomic][+minimal] if the subject is dual and [–atomic][–minimal] if the subject is paucal. (Disclaimer: the trees in (15)-(16) are highly schematic syntactic representations. No attempt is made to show morpheme order / head movement / affixation.) (15) a. Dual subject with no mismatch: Classifier ngurdrda [ – atomic] [ + minimal] b. ... DP proSUBJ [–atomic] [+minimal] [–additive] ngkarl proOBJ Paucal subject with no mismatch: Classifier ngurdrdu [ – atomic] [ – minimal] 8 ... ... DP proSUBJ [–atomic] [–minimal] [–additive] ... ngkarl proOBJ One might wonder by what synchronic mechanism the non-sibling marker comes to be inserted in the first place in clauses with a dual or paucal non-sibling subject. This is an interesting question, but one that I unfortunately have no special insight into. The fact is that the non-sibling markers exist and are sensitive to number, and all I am concerned with here is how this number sensitivity might be featurally encoded given the lexical economy principle that I have proposed. Murrinhpatha number mismatch as partial Agree 249 Mismatched agreement occurs when the classifier Agrees with the non-sibling marker, i.e. when the non-sibling marker is syntactically positioned between the classifier and the subject. Mismatch scenarios with dual and paucal subjects are shown in (16). The classifier ends up with the feature [+minimal] if the subject is dual and [–minimal] if the subject is paucal. The classifier’s [ atomic] probe remains unvalued. (16) a. Dual subject with mismatch: Classifier ngurdu [ atomic] [ + minimal] b. ... NonSibling ngintha [ + minimal] ... ... DP proSUBJ [–atomic] [+minimal] [–additive] ngkarl proOBJ Paucal subject with mismatch: Classifier ngurdrda [ atomic] [ – minimal] ... NonSibling ngime [ – minimal] ... DP proSUBJ [–atomic] [–minimal] [–additive] ... ngkarl proOBJ In (16), it is crucial that even though the non-sibling marker lacks a [±atomic] feature, it still blocks the classifier’s [ atomic] probe from Agreeing with the subject. That is, the non-sibling marker acts as a defective intervener for the [ atomic] probe (to use Chomsky’s (2000) terminology). I make no particular commitment as to how exactly this should be modeled. There are many conceivable technical implementations of this defective intervention effect, and I have no evidence in favor of any one in particular. For instance, one could appeal to disjunctive satisfaction (Roversi 2020: Agree halts as soon as the probing head finds a goal which meets any of its needs) or the Condition on Agree Domains (Puškar-Gallien 2019: a probe cannot search past a goal that another probe on the same head has Agreed with). (17) lists the set of number features which end up on the classifier according to 250 ELANGO KUMARAN my proposal in each possible context. The classifier morph which appears in each context is also listed. CONTEXT SUBJECT MISMATCH ? (17) singular dual dual paucal paucal plural no no yes no yes no CLASSIFIER FEATURES MORPH [+atomic][+minimal] singular [–atomic][+minimal] dual [+minimal] singular [–atomic][–minimal] paucal/plural [–minimal] dual [–atomic][–minimal] paucal/plural TREE (15a) (16a) (15b) (16b) Just using standard feature exponence, it is impossible to write down lexical entries consistent with (17) for the different classifier morphs. To see why, first consider the dual classifier. It corresponds to the feature sets [–atomic][+minimal] and [–minimal]. Those sets have no features in common, so if we were to write a feature exponence rule for the dual morph, it would have to expone no number features at all. This in itself is not a problem, but now consider the singular classifier. It corresponds to the feature sets [+atomic][+minimal] and [+minimal]. The only feature that those sets have in common is [+minimal], so it cannot be the case that the singular classifier expones any number features other than [+minimal]. Furthermore, it cannot be the case that the singular classifier expones no number features, because we have already said that the dual classifier expones no number features. So it must be that the singular classifier expones [+minimal] and no other number features. But this incorrectly predicts that the singular classifier should be preferred for insertion over the dual classifier in the context of the feature set [–atomic][+minimal] (due to the Subset Principle). I propose that instead of exponing feature sets, the classifiers expone meanings denoted by feature sets. The singular classifier corresponds to the feature sets [+atomic][+minimal] and [+minimal], both of which denote singular. The paucal/plural classifier corresponds to the feature set [–atomic][–minimal], which denotes paucal/plural. And the dual classifier corresponds to both the feature set [– atomic][+minimal], which denotes dual, and the feature set [–minimal], which denotes underspecified nonsingular. This means that (17) can correctly be accounted for if we adopt the semantic insertion rules in (18). (18) Classifier insertion rules: – Insert the singular classifier if the feature set denotes singular – Insert the paucal/plural classifier if the feature set denotes paucal/plural – Insert the dual classifier otherwise. Although this semantic exponence mechanism is nonstandard, it does not seem unreasonable to me. Crosslinguistically, there is a range of number agreement phenomena which seem to challenge simple feature exponence and suggest a level of semantic computation: for instance, plural agreement with collective singular nouns (e.g. Corbett 1979), cumulative number agreement with coordinated arguments and in right-node-raising constructions (e.g. Lyskawa 2021), and cumulative subject+object number agreement (e.g. Gluckman 2018). Murrinhpatha number mismatch as partial Agree 251 5 Morphological position classes and syntactic affix mobility Recall that the dual non-sibling marker only triggers mismatched agreement when it is prefixal, as in (19). When it is suffixal, as in (20), there is no mismatch. (19) (20) bangintha- ngkardu -nu 1SG.13.FUT- NSIB.DU . F- see -FUT ‘We (two non-siblings) will see it.’ (Nordlinger and Mansfield 2021:6) ngubanhingkardu -nu -ngintha 1DU.13.FUT- 2SG.OBJ- see -FUT -NSIB . DU . F ‘We (two non-siblings) will see you.’ (Nordlinger and Mansfield 2021:6) Nordlinger and Mansfield (2021) note that this is a non-canonical morphotactic pattern – it is remarkable that the linear position of the non-sibling marker affects the form of another element, the classifier. In a way, one might say that this pattern is evidence in favor of my analysis. My Agree-based analysis predicts that mismatched agreement occurs only when the nonsibling marker syntactically intervenes between the classifier and the subject. If the non-sibling marker were to be syntactically positioned higher than the classifier, no mismatch would occur. The existence of (20), then, seems to validate my analysis: we can just assume that the syntactic position of the non-sibling marker is different in (20) than in (19) – an assumption which seems validated by the difference in their linear position – and the agreement non-mismatch is correctly predicted. This raises two issues, however. The first issue is that we have to assume a lack of uniformity between the paucal and dual non-sibling markers, both in terms of syntactic position and in terms of linearization. In terms of syntactic position, we have to assume that the paucal non-sibling marker is always positioned low, between the classifier and the subject, since it always triggers an agreement mismatch; whereas the singular non-sibling varies between that low position and a higher position. And terms of linearization, the paucal non-sibling marker’s low syntactic position is mapped to the suffixal position (see (10)); whereas for the dual nonsibling marker, the low syntactic position is mapped to the prefixal position, while the higher position is mapped to the suffixal position. This is summarized in (21). (21) NON - SIBLING MARKER SYNTACTIC POSITION LINEAR POSITION dual dual paucal low high low prefixal suffixal suffixal The second, more significant issue is a modularity paradox. The displacement of the dual non-sibling marker in (20) appears to be motivated by a morphological position class constraint (Nordlinger and Mansfield 2021) which gives the object agreement marker priority for the relevant templatic slot.9 The assumption that this purely morphological constraint is able to determine the syntactic position of the 9 Popp (2022a,b) proposes that position classes are not in fact involved – instead, there is a constraint which favors linearizing person morphology to the left. Regardless, the constraint is postsyntactic, which is all that is relevant here. 252 ELANGO KUMARAN non-sibling marker seems countercyclic, and hence undesirable. One possibility worth considering is that the displacement of the non-sibling marker is not in fact conditioned by a postsyntactic constraint, but is instead conditioned in the narrow syntax. Nordlinger and Mansfield (2021:8) argue that this cannot be the case because what matters is the overt presence of an object agreement marker. They state that in (19), the object’s features are present in the syntax, but just happen not to be overtly realized. Note, though, that this does not strictly speaking logically rule out a purely syntactic explanation for the difference in position of the dual non-sibling marker in (19) vs. (20). It is possible that the presence of third person singular direct object features in the syntax in (19), as opposed to second person singular direct object features in (20), is somehow responsible for the variable position of the dual non-sibling marker. Another possibility is that, contrary to what Nordlinger and Mansfield assume, there is in fact no syntactic object agreement in (19), for instance because the relevant probe’s relativization is such that it ignores third person singular direct objects. Maybe the position of the dual non-sibling marker could somehow be tied to the presence or absence of narrowsyntactic object agreement. I am not sure how realistic these possibilities are, but I thought I should at least mention them. Suppose, though, that the difference in position of the dual non-sibling marker between (19) and (20) is indeed due to a postsyntactic constraint on linear order. I want to point out that to whatever extent this seems to raise issues for modularity, I do not think these issues are unique to my analysis. My analysis treats the variable linear position of the dual non-sibling marker as reflecting variablility in its syntactic position. This may seem like an undesirable assumption given that the variability seems to be conditioned by postsyntactic factors. However, abandoning this assumption would not eliminate the modularity paradox. The fact is that the number mismatch is present in (19) but goes away in (20). If we were to assume that the variablility in linear order of the dual non-sibling marker is not syntactic, we would have to assume that the linear position of the dual non-sibling marker (i.e. a PF property) is able to affect the number (i.e. morphosyntactic features) of the classifier – which still seems countercyclic (and this is the puzzle that Nordlinger and Mansfield highlight in their paper, though they do not discuss it in exactly these terms). The variable-syntactic-position hypothesis should therefore not be dismissed. On the contrary, I think it is favored. The analysis that I have proposed in this paper allows a principled explanation of the (non-)mismatches, becased solely on the operation Agree (Chomsky 2000) acting upon a typologically well-motivated feature set (Harbour 2014) alongside a lexical economy condition on probes (Preminger 2019). If we were to abandon the hypothesis that the variable linear position of the dual non-sibling marker is syntactically determined, we would have to abandon my analysis, leaving us without an explanation for how and why the singular-dual and dual-paucal mismatches occur. 6 Conclusion I have presented an Agree-based analysis of number mismatches between the subject and verb in Murrinhpatha. My proposal is that the singular-dual and dual- Murrinhpatha number mismatch as partial Agree 253 paucal number mismatches conditioned by the non-sibling marker are not arbitrary, but are the consequence of the feature decompostion of number (Harbour 2014) and a learner bias against positing probe features (Preminger 2019). The analysis also relies on the mechanisms of semantic exponence (18) and partially-defective intervention (16). A prediction of this analysis is that mismatched agreement requires the nonsibling marker to occupy a particular syntactic position. On the surface, this prediction seems to be validated by the fact that rightward displacement of the dual non-sibling marker eliminates the agreement mismatch (Nordlinger and Mansfield 2021), though the pattern raises problems for linearization and modularity which I have not solved here. References Blythe, J. 2010. From ethical datives to number markers in Murriny Patha. Grammatical Change: Theory and Description, edited by R. Hendery and J. Hendriks, 157-184. Canberra: Pacific Linguistics. Chomsky, N. 2000. Minimalist inquiries: the framework. Step by Step: Essays on Minimalist Syntax in Honor of Howard Lasnik, edited by R. Martin, D. Michaels, and J. Uriagereka, 89-155. Cambridge MA: MIT Press. Corbett, G. G. 1979. The agreement hierarchy. Journal of Linguistics 15.2:203-224. Davidson, L. K. 2018. Allies and Adversaries: Categories in Murrinhpatha speaking children’s talk. Doctoral dissertation, University of Melbourne. Gluckman, J. 2018. Decomposing Number Features. Manuscript, University of Kansas. September 7, 2018 version. Green, I. and R. Nordlinger. n.d. The Daly Languages (Australia). dalylanguages.org Harbour, D. 2014. Paucity, abundance, and the theory of number. Language 90.1:185-229. Lyskawa, P. 2021. Coordination without grammar-internal feature resolution. Doctoral dissertation, University of Maryland. Mansfield, J. 2019. Murrinhpatha Morphology and Phonology. Boston/Berlin: De Gruyter. Nordlinger, R. 2015. Inflection in Murrinh-Patha. The Oxford Handbook of Inflection, edited by M. Baerman, 491-519. Oxford: Oxford University Press. Nordlinger, R. and J. Mansfield. 2021. Positional dependency in Murrinhpatha: expanding the typology of non-canonical morphotactics. Linguistics Vanguard 7.1. Popp, M.-L. 2022a. Does anything go in Murrinhpatha morphotactics? Not with phonology. GLOW 45 poster. Popp, M.-L. 2022b. Morphotactics in Affix Ordering: Typology and Theory. Doctoral dissertation, Universität Leipzig. Preminger, O. 2019. What the PCC tells us about “abstract” agreement, head movement, and locality. Glossa 4.1:13. Puškar-Gallien, Z. 2019. Resolving polite conflicts in predicate agreement. Glossa 4.1:33. Roversi, G. 2020. How to satisfy probes: person/number hierarchy effects in Äiwoo. NELS 50 preprint. Proceedings of CLS 58 (2022), 255-267 © Chicago Linguistic Society 2023. All rights reserved. 255 On Syntactic Tenselessness in Singlish: Evidence from Eventivity Si Kai Lee University of Connecticut 1 Introduction Singlish is a contact variety spoken in Singapore, which has largely been considered a creole with an English superstrate, and a myriad of different substrate varieties, including Bahasa Melayu (or, Malay), which is indigenous to the region, as well as a number of Sinitic varieties which have featured prominently in the (relatively recent) development of modern Singapore, and thus, by extension, of Singlish. Of the Sinitic varieties, the most salient are two of the Southern Min varieties, Teochew and Hokkien, Cantonese, and, in more recent times, Mandarin. One characteristic of Singlish, which sets it apart from Standard English(es) is that the subject-agreement verbal morphology appears to be optional, as exemplified in (1) vs. (2). (1) John knows Mary. (2) (%) John know Mary. [=(1)] Note, however, that for some speakers, the absence of verbal agreement in this class of constructions, which I labelled AGREEMENT- DROP in previous work (Lee 2022), leads to some pragmatic infelicity, which is indicated in (2) through the use of the % notation, which is adopted from Tsai (2008), who uses it to mark a parallel effect in Mandarin Chinese. The intuition for such speakers is that the sentence requires some degree of additional contextualisation for it to be fully acceptable, and in the absence of such contextualisation, i.e. in a vacuum or out of the blue, these sentences generate a feeling of ‘incompleteness’. This is one indication that the two constructions are not truly in free alternation, at least for this subset of speakers, contrary to what has been reported in prior literature (e.g. Wee & Ansaldo 2004). In Lee (2022), I further demonstrated that the alternation also has both syntactic and semantic reflexes, and argued that the PF-based treatment of the phenomenon which has been suggested in the literature (see, e.g. Sato 2014,2016) could not possibly be maintained for this population of speakers, and a syntactic account was instead in order. On the basis that the reflexes observed were closely paralleled in the classical reflexes of the syntactic operation of topicalisation, I argued that this implicated the projection of a topicalisation structure in the left periphery of agreement-drop constructions. In this paper, I demonstrate that my use of the label “agreement-drop” for this class of constructions is not entirely appropriate, in that it only describes a subset of the constructions that share the properties, and thus, the syntactic configuration, 256 SI KAI LEE reported in Lee (2022). Namely, it does not adequately describe the constructions where the omitted morpheme is not the agreement-morpheme per se, but rather the past tense morpheme, which, owing to the impoverished English inflectional system which Singlish inherited, does not overtly inflect for phi-feature agreement (3a vs. 3b). (3) a. John knew Mary. b. (%) John know Mary. [interpretable either as past (=3a) or present (cf. 2)] Owing to the existence of these cases, the more appropriate label for the class of constructions as a whole would therefore be “tense/agreement-drop constructions”, though for ease of exposition, I will continue to refer to the entire class as “agreement-drop constructions” through the remainder of the present paper; these agreement-drop constructions are juxtaposed against the “fully-agreeing” constructions, which, for our present intents and purposes, are entirely equivalent to basic Standard English clauses.1 Notably, the linear string in (3b), which conveys meaning equivalent to the overtly past tense (3a), is identical to the string presented in (2), which is translated as bearing present tense; there is therefore homomorphism across tenses in agreement-drop constructions, which is not found in their fully-agreeing counterparts. In other words, agreement-drop constructions are ambiguous between a past and present interpretation. Given the empirical data, there are two potential analyses for the non-realisation of overt tense in agreement-drop constructions, which I will refer to as the underspecification analysis, where Tense as a syntactic node is projected within the derivation, but underspecified with respect to its semantic content, and the tenseless analysis, where Tense is syntactically absent, respectively. I present novel data which demonstrates the existence of a hitherto undocumented asymmetry between eventive and stative predicates with respect to their compatibility with agreement-drop, and argue that this asymmetry, alongside the noted incompleteness effects that are baseline, can best accounted under the tenseless analysis, rather than the underspecification analysis. I further argue that the insights drawn from the Singlish data presented can be used to settle a longstanding debate in the Sinitic literature regarding the very same issue. 2 The Data As noted in section 1, agreement-drop constructions are semantically indeterminate, with the exact same string being compatible with both past and present tense interpretations. In the baseline example (3a), repeated below in (4a), this means that the sentence can be conveying information either about John’s present knowledge of Mary, or John’s past knowledge of Mary. 1 There is some discussion to be had as to whether the fully-agreeing constructions that surface in Singlish is not Singlish per se, but rather instances of intersentential code-switching between bona fide Singlish (i.e. agreement-drop constructions) and Standard English. For reasons of scope, I set this issue aside in the present paper. Syntactic tenselessness in Singlish (4) 257 a. (%) John know Mary. [interpretable either as past or present] In the absence of any additional contextual information (setting aside for the time being the accompanying incompleteness effect), the most natural interpretation of the sentence involves the present tense interpretation, and not the past tense interpretation. Indeed, the past tense interpretation is extremely difficult to access in a vacuum. However, it becomes substantially more accessible when the appropriate context is provided. An example of such a context, i.e. one which is able to coerce the past tense interpretation of the agreement-drop construction in (4a), is provided in (5). (5) Context: John has recently passed away and the present discussion involves identifying an executor for his will. Under such a context, the Standard English (=fully agreeing) present tense clause (6) is patently infelicitous, and this follows from standard assumptions about the (in)ability of the deceased to know, well, anything. (6) John knows Mary. (# under context 5) Since the present-tense interpretation is ruled out, the only felicitous reading of our agreement-drop construction in (4a) necessarily involves past temporal reference, ergo agreement-drop constructions are genuinely temporally ambiguous, even as pragmatic considerations oftentimes rule out the past tense interpretation absent any further manipulation of the context of evaluation. (7) a. I knew Mary. b. (%) I know Mary. [=(7a)] Singlish Fully-agreeing Singlish Agreement-drop In my previous examination of Singlish agreement-drop constructions in Lee (2022), since the scope of the investigation was limited to cases where agreement morphology is dropped, the impoverished system of (overt) subject agreement which Singlish inherited from English limited the range of observable alternations to exactly those involving third person singular subjects in the present tense. Given that the presence/absence of past tense morphology is not subject to this same deficiency, being invariant across all the possible phi-feature valuations of the subject, the generalisations made in Lee (2022) can be extended to cover the entire class of constructions, and evaluated accordingly. In this way, even sentences like (7b), which at first appear string-identical to the fully-agreeing present-tense construction, given that agreement with the 1st-person singular subject results in a zero-allomorph in the present tense, can be evaluated for (tense/)agreement-drop, since we can check for the availability of the past tense interpretation of the overtly present-tensed clause. Granted, as before, the past interpretation remains less accessible than the present tense interpretation, and indeed, this problem is further compounded in this particular example because of the fact that the sentence is most naturally parsed as a fullyagreeing present tense statement, in part because the fully-agreeing constructions do not induce any of the incompleteness effects that are found with their agreementdrop counterparts; consequently, the fully-agreeing (i.e. present tense) reading of 258 SI KAI LEE the same string is more accessible, for processing (and accommodation) related reasons. To overcome this natural bias and force the past interpretation to surface, the appropriate contextual setup is required. Consider first the context in (8): (8) Context: Mary has recently passed away and I am explaining my decision to attend her wake. Given such a context, the Standard English (9) is infelicitous; in the same way that a deceased individual cannot be asserted to presently know anything, one cannot claim to presently know a deceased individual; only the past-tensed counterpart (7a) is deemed acceptable here. (9) I know Mary. (# under context 8) Standard English The felicity of the exact same string in Singlish (7b) under this context is therefore indicative of agreement-drop having covertly applied, thereby opening the door to a past-tensed interpretation of an overtly present-tensed fully-agreeing construction. Further evidence that these cases, where agreement-drop has covertly applied, behave on par with canonical agreement-drop constructions (i.e. present tense constructions with third-person singular subjects) comes from the fact that the same incompleteness effect that marks other simplex agreement-drop constructions arises in sentences such as (7b) when they are interpreted as past, but not when they are interpreted as present (since here, there is nothing distinguishing them from their fully-agreeing counterparts). We can therefore say at this point that agreement-drop constructions in Singlish have the twin properties of (i) being temporally ambiguous, and (ii) requiring additional contextual support for full acceptability. Given that temporal information is not overtly encoded in Singlish agreementdrop constructions, as demonstrated by the temporal ambiguity of those exact constructions, Singlish is therefore a member of the class of languages that have been argued to be tenseless, such as Yucatec Maya (Bohnemeyer 2002, 2009), Kalaallisut (Bittner 2005, 2008), various Slavic languages (see, e.g., Todorovic 2016), Turkish (Zanon 2014), Korean (Kang 2014), and most pertinently, given their intimate relationship to Singlish as major substrates, the Sinitic languages exemplified by Mandarin (e.g. Sybesma 2007; Lin 2005, 2010). However, the absence of morphological tense (and the corresponding temporal indeterminacy that arises) is insufficient for determining whether tense should be projected within the structure or not, with the two alternative lines of reasoning being either (i) the absence of morphological tense is a direct consequence of the absence of syntactic tense, and (ii) that the absence of morphological tense corresponds to a possibly impoverished tense projection within the structure. This has been hotly debated in the literature surrounding the other languages that have been argued to be tenseless, though perhaps the fiercest debate has centred around the syntactic status of tense in the Sinitic varieties in particular (e.g.Sybesma 2007, Lin 2005, 2010, Huang 2015). The determination of whether Singlish agreement-drop constructions, which are similarly “tenseless” in that they lack overt tense marking and are temporally ambiguous, are syntactically tenseless, could settle the debate in Sinitic, under the Syntactic tenselessness in Singlish 259 assumption that the absence of tense in Singlish was inherited from its Sinitic substrates. Independently, establishing the absence of syntactic tense in Singlish would also add to the body of cross-linguistic evidence against the universality of Tense within the syntax. 2.1 Eventive vs. Stative Predicates The fact of the matter is that agreement-drop is not blindly applicable to all predicates in Singlish in the same way. In the examples presented above, I have demonstrated that we minimally get an incompleteness effect. However, all of the examples presented thus far have been constructions involving the stative predicate know, and in fact, when agreement-drop is applied to other predicates, we can get a difference in acceptability, with the sharpest contrasts arising with eventive predicates in particular, which are straight up incompatible with agreement-drop (absent any further modifications). Consider, for example, the minimal pair (10) and (11), which differ only in the choice of predicate that has undergone agreement-drop; (11), which has an eventive predicate eat, is ungrammatical, in contrast to (10), which instead contains the stative predicate like; while still incomplete, (10) is markedly more acceptable than (11). (10) (11) (%) John like meat. [‘John likes meat.’ or ‘John liked meat.’] * John eat meat. [Intended: ‘John eats meat.’ or ‘John ate meat.’] What is the source of this contrast? That is to say, why should eventive predicates be more adversely affected by the application of agreement-drop as compared to stative predicates? To answer this question, we must first step aside from the present point of contrast between eventive and stative predicates in the context of agreement-drop and consider some of the ways in which they pattern together instead, such possible rescue operations which allow for agreement-drop constructions of both the eventive and stative varieties to be fully acceptable. 2.2 The particle ‘one’ One such repair involves the merger of the sentence-final particle ‘one’, which addresses both the weaker incompleteness with stative bare agreement-drop constructions, and the stronger ungrammaticality induced by eventive bare agreement-drop constructions, as in (12). (12) a. John know Mary one. [≈‘John knows Mary.’] b. John eat meat one. [≈‘John eats meat.’] Note, however, that the merger of ‘one’ is not an innocent rescue operation, in that it has semantic/pragmatic effects; i.e. the particle itself has interpretational contributions, and is not therefore more generally available as a repair, unlike dosupport in standard English, for example. 260 SI KAI LEE In particular, ‘one’ has the semantics of a present-tense stativiser. The presenttense meaning borne by ‘one’ is evidenced by the bleeding of the past-tense interpretation of its complement, as indicated by the absence of the past-tense interpretation in the translation of (12). This is not to say that ‘one’ constructions are wholly incompatible with complements which carry some past-tense meaning, however, as co-occurrence of ‘one’ with a temporal adverbial which bears past temporal information, such as ‘last time’, can be perfectly felicitous, regardless of whether the predicate that has undergone agreement-drop is eventive or stative, as demonstrated by the full acceptability of both (13) and (14). (13) John last time know Mary one. [≈‘John knew Mary.’] (14) John last time eat meat one. [≈‘John used to eat meat.’] This state of affairs appears to suggest that ‘one’ is not inherently present-tense bearing, but this is an illusion, that follows in part from the interaction between past events and present states: consider that the existence of a past event results in a present state wherein the past event happened. That is to say: if I ate meat at some point in the past, for example, then I have the present state of having eaten meat in the past. This ties into the stativising effect of ‘one’, which can also be seen when we consider the meaning of the ‘one’ sentences as well as their distribution. In particular, constructions such as (12b), where the eventive predicate has undergone agreement-drop alongside the merger of the sentence-final particle, are only compatible with a habitual or repetitive reading, and not a truly eventive reading, a distributional contrast which is not entirely reflected in the provided translations. This is because the English simple present, which is used in the translations are ambiguous between the two types of readings, albeit with the truly eventive interpretation being difficult to induce even in standard English. The most surefire way of inducing this reading is through the establishment of a context of Free Indirect Discourse, i.e. where the sentences surface within some larger narration of an ongoing sequence of events; the ‘one’ sentences are not available under such contexts, so perhaps a better translation of (12b) would be as follows : (15) John eat meat one. [‘John has the (present) property of eating meat.’] One other interesting property of ‘one’ is that it is in complementary distribution with overt tense/agreement marking, both when the predicate bears past tense (16), which is in direct conflict with the proposed present tense semantics of ‘one’, and, more surprisingly, even when it also bears present tense (17), i.e. there is harmony in temporal content between the predicate and the particle. (16) (17) a. b. a. b. * John knew Mary one. [Int: = (13)] * John ate meat one. [Int: = (14)] * John knows Mary one. [Int: = (12a)] * John eats meat one. [Int: = (12b)] Even more surprisingly, ‘one’ is incompatible even with present-tense constructions where the subject is in the 1st or 2nd person, which could be parsed as either being fully-agreeing or having undergone vacuous agreement-drop. Furthermore, Syntactic tenselessness in Singlish 261 the exact same string can also be parsed as past tense agreement-drop constructions, as demonstrated previously, the constraints on the attachment of the particle ‘one’ suggests that transparent (canonical) agreement-drop is a prerequisite for the attachment of the particle ‘one’, i.e. where the fully-agreeing option is a possible parse, the particle is ruled out. This is ultimately orthogonal to the present discussion but has implications for the overall pattern of agreement-drop and the architecture of Singlish more generally, and I must therefore leave further consideration of the particle to future research. (18) (19) a. b. a. b. * I know Mary one. [Int: ‘I know Mary.’] * I eat meat one. [Int: ‘I eat meat.’] * You know Mary one. [Int: ‘You know Mary.’] * You eat meat one. [Int: ‘You eat meat.’ For the purposes of the present paper, suffice to say that the incompatibility of ‘one’ with overtly agreeing constructions, especially those which carry presenttense semantics, is not straightforwardly predicted if the absence of tense morphology were not reflected by a difference in the underlying structure. That is not to say that this alone is a knockdown argument in favour of syntactic tenselessness, however: it is possible that the distinction in structure that governs the attachment of ‘one’ is instead found at the extreme left periphery instead, higher than TP, i.e. ‘one’ is sensitive to the presence/absence of tense/agreement morphology only to the extent that it corresponds to differences in the structure higher than TP. 2.3 Temporal and Quantificational Adverbs An alternative strategy for ameliorating both eventive and stative agreement-drop constructions involves the merger of a clearly temporal adverb, such as ‘last night’ and ‘last year’ in (20) and (21) respectively. (20) (21) Last night he eat the cake. [‘Last night, he ate the cake.’] Last year she love dancing. [‘Last year, she loved dancing.’] Notably, with eventive predicates, the temporal adverb completely addresses the degradation, as well as any possible residual incompleteness. However, it should be noted that this amelioration only takes effect when the attachment site for the adverb is at the left periphery: (22) (23) * He <last night> eat the cake <last night>. [Int:=(20)] * She <last year> love dancing <last year>. [Int:=(21)] Relatedly, yet another way of ameliorating the degradation associated with bare eventive agreement-drop constructions involves the merger of a quantificational temporal adverb, such as always as in (24). Here, however, the resultant construction is only ameliorated to the point where it is on par with the bare stative agreement-drop construction, i.e. it suffers only from the incompleteness effect which is baseline to agreement-drop constructions, and can be addressed via any of the mechanisms which repair incompleteness, including contextual enrichnment. (24) (%) John always eat meat. ‘John always eats meat.’ 262 SI KAI LEE 2.4 Contextual Amelioration Note that I previously characterised incompleteness as involving a deficiency in terms of contextualisation; this implicates that contextual enrichment can also work towards the alleviation of the incompleteness effect. This is borne out empirically: for example, the (stative) agreement-drop construction is perfectly acceptable if either (i) it itself constitutes an interrogative, or (ii) it constitutes an answer to an interrogative: (25) A: Who know Mary? B: John know Mary. This crucially is not available for eventive agreement-drop constructions, with both options being ruled out, as indicated by the ungrammaticality of both question and answer in the following dialogue. (26) A: *Who eat meat? B: *John eat meat. Note that the ungrammaticality of the answer is not derived from the ungrammaticality of the question; switching out the agreement-dropped question for its fully-agreeing counterpart in the stative (25) is unproblematic, whilst the ungrammaticality of the agreement-drop answer in the eventive example is maintained, as demonstrated by the persistent contrast between (27) and (28). (27) A: Who knows Mary? B: John know Mary. (28) A: Who eats meat? B: *John eat meat. Other forms of contextual manipulation, such as contextual enrichment, have parallel effects. For example, if we return to our earlier context in which John has died, i.e. (5), the stative agreement-drop construction in (29) is fully acceptable, and does not induce any sense of incompleteness, whilst the eventive agreement-drop construction (30) is ungrammatical. (29) John love soft toys. (30) * John buy soft toys. 2.5 Interim Summary Taking stock of the situation, then: we have seen that agreement-drop results in temporal ambiguity, and a sense of incompleteness that must be addressed through either contextual or lexical means. Additionally, there is clear evidence that agreementdrop is inherently incompatible with eventive predicates, and the examination of two of the key repair mechanisms that are able to reconcile agreement-drop and eventive predicates, namely the attachment of the particle ‘one’, and the merger of quantificational adverbs such as ’always’, reveals that repairing this degradation essentially involves the neutralisation of eventivity to derive stative predication. The one exception to this is the merger of an overt temporal adverb, such as ‘last night’, though the fixed attachment site of the adverb suggests that there is something else at play in those cases. In the following section, I argue that these properties of agreement-drop constructions can best be addressed if tense were not syntactically projected in agreementdrop constructions in Singlish. Syntactic tenselessness in Singlish 263 3 Analysis The properties of agreement-drop constructions can be classified into two distinct categories; the eventive/stative split appears to be a deeper issue that underlies the incompleteness effects, seeing as the means of ameliorating the degradation found in eventive predicates are unable to address the incompleteness that is still present afterwards. Additionally, the fact that the incompleteness that is observed with stative agreement-drop constructions (as well as eventive agreement-drop constructions that have undergone some form of repair) can be eliminated through contextual manipulation, tells us that this effect is semantic/pragmatic in nature. I will return to this issue later, but for now, I want to focus on the eventive/stative asymmetry that is observed, and I will argue that the asymmetry is predicted if tense is not present within the structure of agreement-drop constructions. In the literature on asymmetries between eventive and stative predicates, the claim that eventive predicates, unlike statives, contain an open variable, has been made by Enç (1991). This open variable is in turn subject to a licensing condition, which can only be satisfied by something else in the derivation which bears temporal content. One such element would be Tense itself, and in Singlish, Tense appears to be able to license this open variable across clausal boundaries, as shown by the grammaticality of (31), where the matrix clause has a fully-agreeing predicate which includes overt present-tense morphology; consequently, the embedded eventive predicate is able to undergo agreement-drop without too much issue.2 (31) ? Mary knows that Peter eat meat. ‘Mary knows that Peter eats meat.’ This idea has since been formalised as a syntactic effect by other researchers, i.e. that eventive predicates are distinguished from stative predicates by the presence of additional internal structure (see, e.g., Travis 2010, Vander Klok & Déchaine 2014). In particular, Vander Klok & Déchaine (2014) proposes that the inner syntax of eventive predicates includes an InnerAsp projection, which is situated between little v and V. This InnerAsp projection is argued to be the locus for the regulation of telicity, and is itself sensitive to the temporal specification of higher syntactic material. For ease of exposition, I will be treating InnerAsp as being abstractly equivalent to the open variable postulated in Enç (1991). If any case, if the presence/absence of InnerAsp is indeed the source of the asymmetry between eventive and stative predicates, then any account of the Singlish facts must establish how the licensing of InnerAsp in eventive predicates comes to fail. If, in the most basic case, Tense is in principle available to serve as a licensor for InnerAsp, then this implicates that Singlish agreement-drop constructions either do not have Tense, or that any Tense that is present within the syntax of agreementdrop constructions is deficient to the extent that it is unable to perform the requisite licensing. Consider the possible dimensions of deficiency that can be postulated of Tense, if it were syntactically projected. To account for the semantic ambiguity that arises in the absence of overt tense/agreement morphology, one dimension of deficiency that must apply is that the hypothetical Tense head must be underspecified with 2 Note that these are not quite the same as the cases with which Enç (1991) was concerned. 264 SI KAI LEE respect to its temporal content. The fact that the interpretation is dependent on contextual factors indicates that the underspecified Tense head must essentially inherit a value from the context, in order for the sentence to be interpreted with the relevant past or present meaning. However, this is insufficient, as it would predict that the contextually valued Tense head would no longer be underspecified, and would therefore behave on a par with a regular T head, e.g. of the standard English variety. This wrongly predicts that we should no longer see any of the effects of agreementdrop if the context were sufficiently enriched, contrary to empirical evidence. This is compounded by the fact that contextual amelioration, as a rescue operation, is available only to stative and not eventive predicates. For any advocate of the underspecification analysis, accounting for this failure of the underspecified T head to erase the eventive/stative distinction implicates that the deficiencies that afflict T extend beyond simple semantic underspecification. In other words, the licensing of InnerAsp cannot be purely semantic, or we would not expect the eventive/stative distinction to survive when contextual enrichment has applied. We must therefore say that InnerAsp is licensed syntactically, and T is both semantically and syntactically impoverished. The overall situation for the deficient T position is therefore as follows: T is semantically deficient in that it does not bear any temporal content, and this semantic deficiency is necessarily overcome through pragmatic means when interpreting agreement-drop constructions. T is also morphologically impoverished, as it does not have an overt morphological exponent. Finally, T is syntactically deficient, at least at the point in the derivational cycle where the licensing of InnerAsp in eventive predicates is established. Conversely, consider instead an analysis wherein T is not deficient, but rather simply absent from the structure: the effects that are described above would immediately and naturally fall out under this approach; any temporal information that would usually be borne by T is not introduced into the derivation, because T itself is not introduced into the derivation. The morphological non-realisation of tense morphology follows from the fact that there is no syntactic projection that would be morphologically realised. The inability of T to license InnerAsp follows too: T is not available as a potential licensor, as it is itself absent from the structure. That temporal adverbs are able to perform the licensing function of T can be attributed to the semantic content borne by the temporal adverbs themselves; they are able to license InnerAsp because they carry the relevant interpretable features that are able to bind the open variable introduced by InnerAsp in lieu of T. Given the right background assumptions and auxiliary stipulations, the two systems can be made to have the same empirical coverage. However, one analysis appears to be much more elegant in the following sense: if T does not fulfill any functions that T usually fulfills, bearing no semantic information, having no morphological reflex, and performing none of its usual licensing duties (possibly as byproduct of the aforementioned semantic deficiency), then why should we say that T is present at all? The empirical evidence for the existence of T in the structure appears to be lacking. Though absence cannot be proven per se, especially given the option of extreme deficiency, it should be noted that the empirical evidence in the case of Singlish agreement-drop constructions follows naturally if we take T to simply be absent. Syntactic tenselessness in Singlish 265 4 Conclusion and implications To conclude, I have presented a variety of novel data points pertaining to examples drawn from the contact language Singlish, and established that in the absence of overt tense/agreement morphology, the so-called agreement-drop constructions are temporally ambiguous, and display a hitherto undocumented contrast in terms of distribution when considering eventive versus stative predicates. In particular, eventive predicates are unable to undergo agreement-drop unless accompanied by some form of repair, namely either through the overt merger of a temporal adverb or the sentence final particle ‘one’, or the merger of a quantificational adverb. In the latter case, the stative/eventive contrast evaporates, but a more general incompleteness effect is residual, and must be independently addressed, e.g. through contextual manipulation. I have considered how these facts can be accounted for, under the assumption that the eventive/stative distinction has a syntactic source and that eventive predicates, unlike stative predicates, are subject to a temporal licensing requirement, evaluating two analyses, the first being one in which the syntactic head T is impoverished, and the other being one in which T is absent. While both analyses are able to achieve empirical coverage, the tenseless analysis requires much fewer background assumptions and auxiliary stipulations to achieve this, and it is therefore on theoretical grounds that I consequently concluded that the tenseless analysis of Singlish agreement-drop constructions is superior. As the alternation between fully-agreeing and agreement-drop constructions in Singlish presented in the paper here is ostensibly a consequence of Singlish’s nature as a contact variety, and the individual configurations can directly be traced to specific input languages within the contact ecology of Singlish, namely English and the Sinitic varieties for fully-agreeing and agreement-drop constructions respectively, the observations and conclusions drawn in the present paper bear not only on the inner workings of Singlish itself, but also on the relevant source langauges. That is to say, the arguments for syntactic tenselessness in Singlish agreement-drop constructions contained here can therefore be taken as an argument for the treatment of Sinitic languages as being syntactically tenseless, in support of the position espoused by Lin (2010), a.o. This anticipated correspondence between Singlish and the Sinitic varieties aside, the contrasts between eventive and stative predicates observed in Singlish can also be evaluated in the Sinitic varieties, especially given that Tsai (2008) has already established the existence in Mandarin of the incompleteness effects that are central to many of the observations presented. If the correspondence I suggest is real, then we predict that these observations will be paralleled, and this may ultimately prove conclusive in settling the tenselessness debate in the Sinitic literature. References Bittner, M. 2005. Future Discourse in a Tenseless Language. Journal of Semantics 22. 339–387. Bittner, M. 2008. Aspectual universals of temporal anaphora. In Theoretical and Crosslinguistic Approaches to the Semantics of Aspect, ed. by S. Rothstein, 11–349. John Benjamins. Bohnemeyer, J. 2002. The Grammar of Time Reference in Yukatek Maya. LINCOM Europa. 266 SI KAI LEE Bohnemeyer, J. 2009. Temporal anaphora in a tenseless language. In The Expression of Time, 83–128. Mouton de Gruyter. Enç, M., 1991. On the Absence of the Present Tense Morpheme in English. Ms., University of Wisconsin, Madison. Huang, Z. N. 2015. On syntactic tense in Mandarin Chinese. In Proceedings of the 27th North American Conference on Chinese Linguistics, ed. by H. Tao, Y.-H. Lee, D. Su, K. Tsurumi, W. Wang, & Y. Yang, 406–423, Los Angeles. UCLA. Kang, J. 2014. On the Absence of TP and its Consequences: Evidence from Korean. Storrs, CT: University of Connecticut dissertation. Lee, S. K. 2022. On agreement-drop in Singlish: topics never agree. Glossa: a journal of general linguistics 45. Lin, J.-W. 2005. Time in a Language Without Tense: The Case of Chinese. Journal of Semantics 23. 1–53. Lin, J.-W. 2010. A Tenseless Analysis of Mandarin Chinese Revisited: A Response to Sybesma 2007. Linguistic Inquiry 41. 305–329. Sato, Y. 2014. Argument ellipsis in Colloquial Singapore English and the anti-agreement hypothesis. Journal of Linguistics 50. 365–401. Sato, Y. 2016. Remarks on the parameters of argument ellipsis: A new perspective from Colloquial Singapore English. Syntax 19. 392–411. Sybesma, R. 2007. Whether we tense-agree overtly or not. Linguistic Inquiry 38. 580–587. Todorovic, N. 2016. On the Presence/Absence of TP: Syntactic Properties and Temporal Interpretation. Storrs, CT: University of Connecticut dissertation. Travis, L. 2010. Inner Aspect: The Articulation of VP. Studies in Natural Language and Linguistic Theory. Springer Netherlands. Tsai, W.-T. D. 2008. Tense anchoring in Chinese. Lingua 118. 675–686. Vander Klok, J., & R.-M. Déchaine. 2014. Stative versus Eventive Predicates and vP-internal Structure. In Proceedings of the Annual Meeting of the Berkeley Linguistics Society, volume 40. Wee, L., & U. Ansaldo. 2004. Nouns and noun phrases. In Singapore English: A grammatical description, ed. by L. Lim, Varieties of English around the world G33, 57–74. Amsterdam/Philadelphia: John Benjamins. Zanon, K. 2014. On the Status of TP in Turkish. Studies in Polish Linguistics 9. Proceedings of CLS 58 (2022), 267-277 © Chicago Linguistic Society 2023. All rights reserved. 267 Linearly adjacent allomorphs and syntactic copies Soo-Hwan Lee New York University 1 Introduction This work raises the question of whether more than one allomorph can surface when multiple syntactic copies are spelled out via movement. I argue that nominal expressions in Korean are the ideal testing ground for investigating this question. It sheds light on the locality domain necessary for the realization of these allomorphs. It also provides a syntactic analysis which does not appeal to post-syntactic operations such as Lowering and Impoverishment (contra Choi 2019). The current work provides a parsimonious take on how allomorphs can be realized linearly adjacent to one another in syntax. An allomorph, call it α, is often realized in the absence of its elsewhere form, call it β. Quite interestingly, however, there are cases where α and β are realized together linearly adjacent to one another. (1), for instance, shows that the Yiddish predicate ‘to be’ is spelled out as zayn and bin via predicate clefting. (1) zayn bin ix a kirzhner. ZAY . INF am I a hatter ‘I am a hatter.’ (Yiddish; Travis 2003: 244) In this paper, I demonstrate that multiple allomorphs, α and β, are realized linearly adjacent to one another in certain Korean noun phrases as shown in (2). (2) a. sayng-il-nal birth-day-day ‘birthday’ b. hay-pyen-ka sea-side-side ‘seaside’ (Korean) I argue that the syntactic configuration of the allomorphs in (2) is subject to syntactic locality conditioning discussed in Distributed Morphology (DM; Halle & Marantz 1993, Bobaljik 2012, Harley 2014). I propose that spelling out multiple syntactic copies via movement best captures the realization of these linearly adjacent allomorphs in Korean. This work is organized as follows: Section 2 provides the empirical picture showcasing linearly adjacent allomorphs in Korean. Section 3 draws parallels among languages (Hiaki, Georgian, Creek, and Korean) in terms of how root suppletion is conditioned. Section 4 introduces a movement-based analysis to handling the primary phenomenon at issue. Section 5 provides cross-linguistic evidence from Ewe, 268 SOO-HWAN LEE Yiddish, and Polish in suggesting that the movement-based analysis is applicable elsewhere in the grammar. Section 6 demonstrates that post-syntactic operations such as Lowering and Impoverishment are not necessary in deriving multiple allomorphs. Section 7 concludes. 2 Linearly adjacent allomorphs Korean has phonologically unrelated morphemes which are semantically identical to one another (Choi 2019). These allomorphs, which I refer to as α and β, can co-occur inside a nominal domain without changing the base semantics of the noun as shown in (3). The inner allomorph α is obligatory whereas the outer allomorph β is not. There is a restriction on the ordering of α and β as well: the linear order α-β is possible whereas *β-α is not. (3) a. hyu-il(-nal) / *hyu-nal(-il) rest-day(-day) ‘holiday’ b. hay-pyen(-ka) / *hay-ka(-pyen) sea-side(-side) ‘seaside’ c. hyen-mi(-ssal) / *hyen-ssal(-mi) rough-rice(-rice) ‘rough uncooked rice’1 d. may-hwa(-kkoch) / *may-kkoch(-hwa) ume-flower(-flower) ‘an ume flower’ e. tan-pal(-meli) / *tan-meli(-pal) short-hair(-hair) ‘short hair’ f. che-ka(-cip) / *che-cip(-ka) wife-house(-house) ‘wife’s parents’ house’ g. ka-sa(-il) / *ka-il(-sa) house-work(-work) ‘housework’ (Korean; (3a), (3f), and (3g) are from Choi 2019) Between the two linearly adjacent allomorphs, α is Sino-Korean (SK) and β is Native-Korean (NK). SK morphemes are borrowed from Chinese whereas NK morphemes are genuinely Korean. The former are often categorized as bound mor1 As for hyen-mi-ssal ‘uncooked brown rice’ and hyen-mi-pap ‘cooked brown rice’, I assume the former involves spelling multiple allomorphs whereas the latter involves compounding two different morphemes. This seems to be on the right track given that hyen-mi on its own means ‘uncooked brown rice’ rather than ‘cooked brown rice’. Only ssal in hyen-mi-ssal displays vacuous semantics. Linearly adjacent allomorphs and syntactic copies 269 phemes which require the presence of another SK morpheme, call it χ.2 The latter are categorized as free morphemes that can stand alone. The allomorphs shown in (3) are classified as either α (SK) or β (NK) in (4): (4) Example (3a) (3b) (3c) (3d) (3e) (3f) (3g) α (SK) il ‘day’ pyen ‘side’ mi ‘rice’ hwa ‘flower’ pal ‘hair’ ka ‘house’ sa ‘work’ β (NK) nal ‘day’ ka ‘side’ ssal ‘rice’ kkoch ‘flower’ meli ‘hair’ cip ‘house’ il ‘work’ At first glance, the co-occurence of α and β seems to run counter to the notion of complementary distribution. However, we will see in Section 3 and Section 4 that the co-occurence of the root allomorphs in (3) is derived via movement of the same syntactic copy undergoing multiple spell out. 3 What triggers root suppletion? Cross-linguistic evidence from Hiaki (Harley 2014), Georgian (Lomashvili 2019), and Creek (Johnson 2016) suggests that the sister of a root triggers root suppletion. Harley (2014) demonstrates that a set of Hiaki verbs undergoes allomorphy based on the number feature (singular or plural) of its internal argument. (5), for instance, shows that the form of the verb mea∼sua ‘to kill’ is determined by its sister koowi(m) ‘pig(s)’: (5) a. Santos Hose-ta koowi-ta mea-ria-k. Santos Jose-ACC pig-ACC kill.SG-APPL-PRF ‘Santos killed a pig for Jose.’ b. Santos Hose-ta koowi-m sua-ria-k. Santos Jose-ACC pig-PL kill.PL-APPL-PRF ‘Santos killed pigs for Jose.’ (Hiaki; Harley 2014) Lomashvili (2019) adds weight to the claim that the sister of a root can trigger allomorphy. Evidence comes from Georgian. (6) shows that the verb gdo∼q’ar ‘to throw’ is suppletive and is conditioned by the number feature of the internal argument satamasho(-eb) ‘toy(s)’. Note that allomorphy is not triggered by an applied argument such as bavshv(-eb)-s ‘for the kid(s)’ as shown in (6c). This is not surprising given Lomashvili’s assumption that applied arguments are not in a sister relation with the root. (6) 2 a. Laura-m bavshv-s satamasho gada-u-gdo. Laura-ERG kid-DAT toy-ø PVB - PV -throw. SG ‘Laura threw away the toy for the kid.’ Choi (2019) mentions that there are a limited number of SK morphemes which can stand alone such as san ‘mountain’ and kang ‘river’. For now, I leave this issue aside. 270 SOO-HWAN LEE b. Laura-m bavshv-s satamasho-eb-i gada-u-q’ar-a. Laura-ERG kid-DAT toy-PL-NOM PVB-PV-throw.PL-3 S . SG ‘Laura threw away the toys for the kid.’ c. Laura-m bavshv-eb-s satamasho gada-u-gdo. Laura-ERG kid-PL-DAT toy-ø PVB - PV -throw. SG ‘Laura threw away the toy for the kids.’ (Georgian; Lomashvili 2019) Johnson (2016) also argues that the sister of a root can condition root suppletion in Creek. (7), for instance, shows that the verb léy∼ká:∼apó: ‘to sit/set’ is a suppletive triplet conditioned by the number feature (i.e., singular, dual, plural) of the object pú:si ‘cat(s)’. (7) a. pú:si -n tak- léyhc -ey -s (SG) cat -ACC LOC- set.SG.PFV -1SA -IND ‘I put the cat down.’ b. pú:si -n tak- ká:hy -ey -s (DU) cat -ACC LOC- set.DU.PFV -1SA -IND ‘I put (two) cats down.’ c. pú:si -n tak- apó:hy -ey -s (PL) cat -ACC LOC- set.PL.PFV -1SA -IND ‘I put the cats down.’ (Creek; Haas 1948, glossed by Johnson 2016) I argue alongside Harley (2014), Lomashvili (2019), and Johnson (2016) in terms of demonstrating that the sister of a root can condition root suppletion in Korean. Based on our discussion in Section 2, α (SK) has to first merge with another SK morpheme, χ, instead of β (NK). In other words, an SK morpheme needs to merge with another SK morpheme before an NK morpheme enters the derivation. Based on this assumption, the following prediction can be established: χ (SK) and β (NK) cannot be realized together in the absence of an SK morpheme such as α. This prediction is borne out: (8) / *hyu-nal / *il-nal a. hyu-il birthSK -daySK / daySK -dayNK / birthSK -dayNK / *hay-ka b. hay-pyen / *pyen-ka seaSK -sideSK / sideSK -sideNK / seaSK -sideNK / *hyen-ssal / *mi-ssal c. hyen-mi roughSK -riceSK / riceSK -riceNK / roughSK -riceNK / *may-kkoch / *hwa-kkoch d. may-hwa umeSK -flowerSK / flowerSK -flowerNK / umeSK -flowerNK / *tan-meli / *pal-meli e. tan-pal shortSK -hairSK / hairSK -hairNK / shortSK -hairNK / *che-cip / *ka-cip f. che-ka wifeSK -houseSK / houseSK -houseNK / wifeSK -houseNK / *ka-il / *sa-il g. ka-sa houseSK -workSK / workSK -workNK / houseSK -workNK Linearly adjacent allomorphs and syntactic copies 271 (8) suggests that α and β are not in a sister relation. Instead, χ and α form a constituent in the absence of β. The way that allomorphy is conditioned in (8) is consistent with how allomorphy is conditioned in Hiaki (5), Georgian (6), and Creek (7). In all of these cases, the root and its sister form the necessary domain for root suppletion. To put it in another way, χ conditions the realization of α in a sister relation. This suggests that the elsewhere form, namely β, resides outside of this domain. A tree structure for (3) is provided in (9): (9) Preliminary tree structure for (3) 2nd merge 1st merge β (NK) χ (SK) α (SK) With this in mind, we are now in a position to address where exactly α and β reside in syntax. In what follows, I put forward a movement-based analysis to capture the distribution of the two allomorphs. 4 Syntactic movement & copies As mentioned in Section 2, α is semantically identical to β. Recall that χ-α and χ-α-β are identical in semantics as exemplified in (3). This suggests that α and β cannot both contribute to the semantic make-up of χ-α-β. How might this be possible? I argue that a syntactic analysis provides an adequate explanation. I posit that α and β are the same syntactic copies subject to movement. Adopting DM, I establish the following Vocabulary Insertion (VI) rules for α and β. (10) Vocabulary Insertion (VI) rules at PF √ a. ↔ α / [√P χ —– ] √ b. ↔ β / elsewhere In (10), the realization of α and β are contextually determined. The co-occurence of α and β is simply a result of spelling out the lower and the higher copy in different parts of the structure. This correctly predicts that χ-α is semantically identical to χ-α-β. Simply put, α and β are the same element in syntax. √ I further posit that deriving α and β is possible via root ( )-to-n movement. First, this captures the correct morpheme order when the plural/associative marker -tul is introduced in the derivation. (11) shows that -tul follows α and β. Note that -tul cannot intervene between α and β. (11) hyu-il-nal-tul / *hyu-il-tul-nal rest-day-day-TUL / rest-day-TUL-day ‘holidays’ Kim & Melchin (2018) assume that -tul is an nP modifier. Based on Kim & Melchin (2018), the morphemes preceding -tul is in principle a property internal to nP. 272 SOO-HWAN LEE Hence, it is plausible to assume that the linearly adjacent allomorphs are a result √ of -to-n movement as shown in (12): (12) Tree structure adopting Kim & Melchin (2018) nP -tul nP √ X P √ χ α n 1 √ n 1 β Park (2022), on the other hand, posits that -tul is realized in Cl(assifier), which is a head structurally higher than n. Under her analysis, -tul individualizes mass nouns, which are not countable on their own. It is worth mentioning that the nominal expression hyen-mi-ssal ‘rough rice’, hosting α and β, is an inherently mass noun. It cannot be realized with -tul as shown in (13). This is the case even when β is absent in the derivation: (13) hyen-mi-ssal(*-tul) / hyen-mi(*-tul) rough-rice-rice-TUL ‘rough uncooked rice (mass)’ Based on (13), it becomes evident that the realization of -tul does not rely on the realization of β. This suggests that the movement involved in deriving β does not √ target Cl if Park is on the right track. Putting these pieces together, -to-n move√ ment rather -to-Cl movement seems to provide a plausible account for deriving α and β: (14) Tree structure adopting Park (2022)3 ClP Cl nP √ X P √ χ α -tul n 1 √ 1 n β For present purposes, it is not crucial to adjudicate between the two analyses on -tul. Neither of them poses a challenge to the current take on the linearly adjacent √ allomorphs in Korean. What is crucial, however, is that -to-n movement (rather than some other type of movement) accounts for the ordering of α, β, and -tul (see 3 ClP is absent for mass nouns such as (13). Linearly adjacent allomorphs and syntactic copies 273 Choi 2019: 300–301 for additional evidence that α and β cannot be intervened by other elements and also see Travis 2003 for an analysis on how head movement in syntax interacts with reduplication). The assumption made so far suggests that the √ √ locality domain for suppletion is P: α is realized inside P whereas β is realized outside of it. A question remains as to why the lower copy α is obligatory whereas the higher copy β is not. I argue that α (the marked form) must be realized in order to satisfy the SK featural requirement:4 a morpheme with the feature [SK] needs to be spelled out by having [SK] checked by another morpheme bearing [SK]. The higher copy β (the unmarked form) is not subject to this featural requirement, since the requirement has already been satisfied by its lower copy α. Optionally spelling out β via head movement is possible as it has no direct consequences to the semantics.5 In the next section, I provide cross-linguistic evidence from Ewe, Yiddish, and Polish suggesting that the realization of multiple syntactic copies subject to allomorphy is possible via head movement. 5 Cross-linguistic evidence for spelling out multiple copies The first evidence that head movement can induce the spell-out of multiple allomorphs comes from Ewe bipartite negation. As shown in (15), the preverbal negation marker mé- and the post-verbal negation marker o have to be realized together in order to induce negation. (15) Kofí mé-kp´ ame áéké o Kofi NEG1 -see person any NEG2 ‘Kofi didn’t see anybody.’ (Ewe; Collins et al. 2018: 341) Collins et al. (2018) argue that the post-verbal marker o is structurally higher than the pre-verbal marker mé-. Evidence comes from fragment answers. In (16b), which is a fragment answer to the question (16a), the negative polarity item (NPI) is obligatorily realized with o instead of mé-. (16) a. ame ka-é ne-kp´ person which-FOC 2 SG-see ‘Who did you see?’ b. ame áéké *(o) person any NEG2 ‘Nobody’ (Ewe; Collins et al. 2018: 350) According to Collins et al. (2018), (16b) involves TP-ellipsis. The post-verbal marker o survives the derivation because it is realized in C, unlike mé- which is realized in T. More crucially, Collins et al. (2018) posit that the realization of mé4 Choi (2019) argues that there is a reserved root class for SK morphemes in Korean. Note that in many cases head movement does not bear semantic consequences. The presence of V-to-T movement in French, and the lack of it in English (Pollock 1989), for instance, do not involve any obvious semantic differences. 5 274 SOO-HWAN LEE and o is a result of spelling out of the lower and the higher copy participating in head movement. Moreover, the realization of multiple allomorphs in Yiddish predicate clefting can be derived via head movement. (17), repeated from (1), shows that the predicate ‘to be’ is spelled out multiple times in different forms, namely zayn and bin. (17) zayn bin ix a kirzhner. ZAY . INF am I a hatter ‘I am a hatter.’ (Yiddish; Travis 2003: 244) Likewise, the empirical facts collected from Polish predicate clefting show allomorphy. In (18), for instance, the predicate ‘to be’ is spelled out as by and jeste. (18) {?by-ć / *jeste-ć} w domu, jeste-śmy. {?be-INF / *be.PRES-INF} in house be.PRES-1PL ‘As for being at home, we are.’ (Polish; Arregi & Pietraszko 2021: 280) Overall, the realization of multiple allomorphs observed in bipartite negation and predicate clefting provide independent evidence that multiple allomorphs can be realized via head movement. Here, parallels can be drawn between the clausal and the nominal domain: the co-occurence of allomorphs via movement is possible in both domains. If the current analysis is on the right track, a broader implication may be that all instances of multiple allomorphs are derived via movement. In this regard, the co-occurence of multiple allomorphs that Inkelas & Zoll (2005) report across languages (e.g., Paamese, Sye, Kawaiisu, and Central Vanuatu) may be framed under a syntax-driven approach. 6 A previous analysis relying on post-syntactic operations An alternative analysis to the one proposed in this paper involves applying multiple post-syntactic operations. Choi (2019) posits that a vacuous reduplication (VR) morpheme is syntactically merged in n. After spell-out, the VR morpheme undergoes Lowering. This forms a sister relation between the VR morpheme and the root α. The sister relation is subject to the Impoverishment rule given in (19). (19) Impoverishment rule (Choi 2019: 307) Delete the marked Class feature [SK] on a terminal root node in the context of VR morpheme on n. As a result of Impoverishment, the VR morpheme is deprived of its [SK] feature and the unmarked/default form β is inserted as its phonological form during VI. (20) fleshes out the derivation based on Choi’s analysis. Linearly adjacent allomorphs and syntactic copies (20) 275 Tree structure based on Choi (2019) nP √ P √ X χ nVR[SK] √ nVR[SK] α β In short, Choi (2019) postulates the following: (i) a VR morpheme, (ii) Lowering, and (iii) Impoverishment. The analysis put forward in this paper does away with these post-syntactic assumptions. Instead, syntactic copies and movement, which are robustly assumed elsewhere in the grammar, are employed to account for the same phenomenon. In this regard, the analysis provided in this work is parsimonious. Recall that Collins et al. (2018), Travis (2003), and Arregi & Pietraszko (2021) argue for a similar kind of derivation for bipartite negation in Ewe and predicate clefting in Yiddish and Polish. (see Section 5). Setting aside the issue of parsimony, Choi’s analysis suffers from a couple of challenges. Choi (2019) argues that the VR morpheme copies the semantic content of the root in PF. Here, a question arises: if a semantic feature is copied and its phonological form (β) is irrelevant to the phonological form of the root (α), why should this issue be taken up in PF? An additional question that remains to be answered is the timing of reduplication for Choi. According to Deal (2016), Haugen (2011), and Schreiner (2021), phonological reduplication takes place very late during the derivation in PF. In fact, they suggest that reduplication is an instance of a phonological rule which comes after VI. Choi’s reduplication, however, occurs before VI. This seems to be an assumption that is inconsistent with the general notion of reduplication and what has been discussed in the previous literature. Overall, relying on multiple post-syntactic apparatuses does not seem to be desirable: (i) it makes extra assumptions about the derivation and (ii) it is inconsistent with the general notion of reduplication put forward in the literature. 7 Conclusion I have argued that the linearly adjacent allomorphs in Korean noun phrases are generated by spelling out multiple copies via movement. This in many ways patterns together with the analyses suggested for deriving allomorphs in the clausal domain. The previous analysis adopting post-syntactic operations is not entirely satisfactory in that it requires stipulations that are neither well-motivated elsewhere in the grammar nor consistent with the notion of reduplication previously established in the literature. The implication here is that the realization of multiple allomorphs can be captured under a syntax-driven DM analysis, which provides a parsimonious take on the issue. 276 SOO-HWAN LEE References Arregi, K., & A. Pietraszko. 2021. The ups and downs of head displacement. Linguistic Inquiry 52. 241–290. Bobaljik, J. D. 2012. Universals in comparative morphology: Suppletion, superlatives, and the structure of words. Choi, J. 2019. Semantically Vacuous Semantic Reduplication in Korean: A Distributed Morphology Analysis. The Korean Journal of Linguistics 44. 293–314. Collins, C., P. M. Postal, & E. Yevudey. 2018. Negative polarity items in Ewe. Journal of Linguistics 54. 331–365. Deal, A. R. 2016. Plural exponence in the Nez Perce DP: A DM analysis. Morphology 26. 313–339. Haas, M. R. 1948. Classificatory verbs in Muskogee. International Journal of American Linguistics 14. 244–246. Halle, M., & A. Marantz. 1993. Distributed morphology and the pieces of inflection. In The view from building 20, 111–176. The MIT Press. Harley, H. 2014. On the identity of roots. Theoretical linguistics 40. 225–276. Haugen, J. D. 2011. Reduplication in distributed morphology. Inkelas, S., & C. Zoll. 2005. Reduplication: Doubling in morphology, volume 106. Cambridge University Press. Johnson, K. 2016. Suppletion in a three-way number system: Evidence from creek. Ms. University of Texas at Arlington . Kim, K., & P. B. Melchin. 2018. Modifying plurals, classifiers, and co-occurrence: The case of Korean. Glossa: a journal of general linguistics 3. Lomashvili, L. 2019. Root suppletion and the theory of allomorphic locality. World Journal of English Language 9. 34. Park, S.-Y. 2022. Two types of plurals and numeral classifiers in classifier languages: the case of Korean. Journal of East Asian Linguistics 1–39. Pollock, J.-Y. 1989. Verb movement, universal grammar, and the structure of IP. Linguistic inquiry 20. 365–424. Schreiner, S. L. 2021. Span–conditioned allomorphy and late linearization: Evidence from the Classical Greek perfect. Glossa: a journal of general linguistics 6. Travis, L. 2003. Reduplication feeding syntactic movement. In Proceedings of the 2003 annual conference of the Canadian Linguistic Association, 236–247. Proceedings of CLS 58 (2022), 277-289 © Chicago Linguistic Society 2023. All rights reserved. 277 Modal concord in American and British English: A register-based experimental study Mingya Liu and Stephanie Rotter Humboldt-Universität zu Berlin 1 Introduction Atomic sentences in natural language can take an unmodalized (UM) form such as in (1), contain a single modal (SM) element of various categories, such as in (2), or multiple modality (MM). Here, we propose MM as a general descriptive, pretheoretical notion covering at least the following three different sets of phenomena: First, the phenomenon of “modal concord” (MC), where the modal auxiliary and the co-occurring modal element (an adverb, typically) in a sentence are of the same modal flavour (epistemic modality) and modal force (possibility or necessity), and the resulting sentence has an interpretation of one modality (Geurts & Huitink 2006; Zeijstra 2007), as illustrated in (3). Secondly, it covers the phenomenon known as “multiple modals” in the literature on varieties of English (Halliday 1970; Huang 2011), as illustrated in (4). Thirdly, we also use it as a neutral term to include constructions with mismatching modals (MiM), i.e., where co-occurring modal elements of the same modal flavor have distinct modal forces, as illustrated in (5). (1) (2) (3) (4) (5) It is raining. It must be raining. / It is certainly raining. It must certainly be raining. I thought you might could help me. (Mishoe and Montgomery 1994) It may certainly be raining. In this paper, we will focus on the MC vs. SM alternation in English. While MC has been much discussed in the theoretical literature, we take a unique approach to examine its use and interpretation impacted by register, which is defined as “those aspects of socially recurring intra-individual variation that are influenced by situational and functional settings [...], adopting a broad understanding of situation and function” (Lüdeling et al. 2022, see also Pescuma et al. 2022). Our research is guided by one central question concerning a language user’s register knowledge and its interaction with their grammatical knowledge. In this regard, MC provides a useful case of study for several reasons: First, MC showcases similar variation as other redundancy phenomena in natural language, such as negative concord (e.g., I did not buy nothing/anything., see Eckert 2019) and double comparatives (e.g., I’ll speak (more) louder., see Mondorf 2003). In comparison to SM, MC has a more restricted use while they are ‘functional equivalents’, that is, SM and MC share the same semantics and can function as alternative choices in different situational contexts, often via enriched meanings. This raises the question of in what contexts a speaker would opt for MC over SM given no semantic 278 LIU AND ROTTER differences, and why. We assume that the related intra-individual variation might be partially register-driven, that is, a result from the interaction of language-internal and language-external factors. For instance, speakers might prefer MC or SM given the place of the conversation (e.g., office vs. home), the interlocutor relation (e.g., while taking to a stranger vs. family member), the topic or purpose of the conversation (e.g., temperature rise at a conference on climate change vs. warm weather for planning the family weekend). To characterize the situational-functional properties of the context that determine the speaker’s choice will, we believe, further our understanding of register and grammar. Secondly, despite of involved redundancy, MC is distinct from negative concord and double comparatives in nontrivial aspects. Across languages, negative concord is part of the standard variety for some languages (e.g., Romance and Slavic languages) while being part of the non-standard varieties for others (e.g., English); it is thus not necessarily a linguistic universal. In contrast, MC is attested in both standard and nonstandard varieties of natural language. Additionally, in the case of English, negative concord and double comparatives tend to occur more in the contexts of lower formality (i.e., lower register), while the native speakers we consulted with so far all relate MC to higher registers. Based on their intuition, sentences such as (3) are more appropriate when e.g., being uttered to a colleague in a conference room than to a family member at a kitchen table. Assuming that the MC/SM variation is a register phenomenon and MC tends to be used in higher registers in comparison to SM, we aim to investigate to what extent this is empirically valid. In the following, we report on two rating studies testing the acceptability and interpretation of MC in American vs. British English (AE/BE) and in relation to register. For the latter, we will focus on the formality of the utterance situation by manipulating only one situational variable (as the very first step), namely, interlocutor relation (Rotter & Liu 2022). Our hypotheses were: [H1a] MC would be rated as less appropriate than SM; [H1b] MC would be rated as more appropriate in formal than in informal contexts. [H2a] Based on the MC approach, MC would be interpreted similarly to SM, i.e., both would receive similar speaker commitment ratings (see Liu et al. 2021). [H2b] Furthermore, we would explore whether the interpretation of MC is sensitive to the formality of the context of use. The paper is structured as follows: In Section 2 we report the details and the results of the two experiments. Section 3 briefly discusses the results and concludes the paper. 2 Experiments 2.1 Experiment 1: American English 2.1.1 Participants 120 native speakers of AE participated in the study on the crowd-sourcing platform Prolific (https://www.prolific.co/) and received monetary compensation. The experiment took roughly 35 minutes. We received informed consent from all participants. Based on the data quality (see 2.1.4), we removed the entire data of 8 participants. The remaining data set consisted of 112 participants (mean Modal concord in American and British English 279 age=36.8, SD=12.2; female N=58, male N=54, non-binary N=0). The participants of the study originate from 35 of the 50 American states. In total, 9.8% of the participants reported to be dialect speakers with three different dialect backgrounds: African American Vernacular English (N=1), Midwestern American English (N=3), Southern American English (N=6), dialect information missing (N=1). The majority have completed college (46.3%); 33.0% have completed high school; 18.8% hold a graduate degree, and below 1% did not finish high school. More than half of the participants grew up in a suburban environment (57.1%), while roughly the same amount grew up in a rural (20.5%) or urban (22.3%) environment. 2.1.2 Material and Design The experiment was implemented using PCIbex (Zehr & Schwarz 2018) and hosted on the PCIbex farm platform (https://farm.pcibex.net/). It used a 2x2x2 factorial design with the factors1 FORMALITY (formal/FL vs. informal/IFL), MODAL FORCE (possibility/3 vs. necessity/2), and MODAL NUMBER (MC vs. SM), henceforth, simply FORCE and NUMBER. Furthermore, we added an UM sentence as a control condition and a MiM sentence for explorative reasons2 . We used 32 critical and 70 filler items in the form of short stories, see (6). (S1) and (S2) set a context. (S3) contains the formality manipulation via the interlocutor relations (formal/informal: public/private). (S4) contains the FORCE and NUMBER manipulation. (Q1) asks about the appropriateness of (S4) in the given context, and (Q2) asks about speakers’ belief in the given proposition in (S4). (6) (S1) (S2) (S3) (S4) Cathy Evans walks through a store. The passer-by seems familiar. She says to the cashierFL / her sisterIFL : .. SM: “I might3 / must2 have seen my co-worker.” MC: “I might maybe3 / must certainly2 have seen my co-worker.” UM: “I have seen my co-worker.” MiM “I might certainly have seen my co-worker.” (Q1) Is the sentence appropriate within the given context? (Q2) Does Cathy Evans believe that she has seen her co-worker? In addition, we included two forced choice tasks at the end of the study, one with 3 and the other with 2 choices, see (7) and (8). (7) Imagine a situation in which you have attended a meeting in person, and you want to tell someone that you feel unwell. Which of the following would you be more likely to say? a. I might have caught a flu. b. I might maybe have caught a flu. c. Either (a) or (b). 1 Factors and their levels are all written in capital letters. Both conditions are not included in the main analysis, but we provide the descriptive statistics for the interested reader, see also the Appendix for a secondary analysis involving UM. 2 280 LIU AND ROTTER (8) Imagine a situation in which you have seen a movie trailer, and you want to tell someone that the movie seems familiar. Which of the following would you be more likely to say? a. I must have seen the movie before. b. I must certainly have seen the movie before. c. Either (a) or (b) 2.1.3 Procedure Participants gave their informed consent before first read the instructions of the experiment and then started with four practice stories to familiarize themselves with the setup. They indicated that they had read the sentence by pressing on the space key before the next one appeared in the middle of the screen. (S1) to (S3) were shown separately after one another. (S4) remained on the screen while (Q1) and (Q2) appeared separately under the sentence together with the scale. The answer to the question was given on a 7-point Likert scale in which all points were labelled, i.e., “Certainly/Probably/Possibly not” (1/2/3), “unclear” (4), “Possibly/Probably/Certainly yes” (5/6/7). Participants used the according keys to indicate their answer. After each story, a fixation cross appeared for 400 ms in order to enforce a proper distinction between the stories. After the experiment, the participants took part in a short demographic survey including their language background, and lastly conducted the forced choice post-task. 2.1.4 Data analysis The data was processed and analyzed with the open source software ‘R’ (R Core Team 2022) in the RStudio environment (RStudio Team 2022). We removed the data from participants who did not match the inclusion criteria of being English native speakers or aged between 18 to 65 years. The critical measurements were ratings for (Q1/2). We also collected data of reading times for (S1) to (S4) and of reaction times for (Q1/2). Furthermore, we used the reading times for (S1) to (S4) and the reaction time for (Q1/2) to exclude entire trails if they were below 400ms — meaning that participants did not pay attention to the sentence. If more than 25% of a participant’s trail set had to be removed, we excluded the entire data of the participant (low attendance rate). The remaining data was analyzed using the package ‘ordinal’ (Christensen 2019) in the cumulative link function model framework (Howcroft & Rieser 2021; Liddell & Kruschke 2018). We sum coded the main effect FORMALITY with its levels FL (0.5) and IFL (-0.5), FORCE with its levels 3 (0.5) and 2 (-0.5), and NUMBER with its levels MC (0.5) and SM (-0.5). The models included all possible 2-way interactions as well as the 3-way interaction. If a 2-way interaction turned out significant, we conducted a sub-analysis by splitting the data along the significant main factor of the interaction. The sub-analysis included the same coding of the main effects as well as the possible 2-way interaction. The link functions were identified by comparing the loglikelihood values of each of the five possible link functions (i.e., probit, logit, cauchit, loglog, and cloglog) with each other. The function with the highest value was chosen for the Modal concord in American and British English 281 model (Christensen 2019). This method has the advantage that the chosen link function follows the shape of the data and accounts for skewness to maximize the model’s fit. The used link functions for each model of the analyses are indicated in the result section. Random effect structures were obtained with the help of the most parsimonious model approach; the used models are indicated in the result section; p-values were obtained with the help of loglikelihood ratio test comparisons of nested models (Bates et al. 2015). All statistical values of means, estimates and the like are rounded to the second decimals except for p-values smaller than 0.01. 2.1.5 Results In the post-task, see (7) and (8), the majority of the participants indicated that they prefer SM (N3 =108, N2 =102). Two participants (non-dialect speakers) indicated that they prefer the 3-MC construction, and two indicated no preference between 3-SM and -MC (Southern American dialect speaker (N=1), non-dialect speaker (N=1)). Two participants indicated that they prefer the 2-MC construction (nondialect speaker (N=2)), and 8 indicated no preference between 2-SM and -MC (Southern American dialect speaker (N=1), non-dialect speaker (N=7)). The descriptive statistics of the ratings in (Q1) and (Q2) are shown in Table 1 and depicted in Figure 1. The output for the models with (Q1) and (Q2) as dependent variables is shown in Table 2. Figure 1: Means and error bars of appropriateness (Q1) and interpretation (Q2) ratings on a 7-point Likert scale (1=Certainly not, 7=Certainly yes) for Experiment 1 (AE, top) and 2 (BE, bottom) 282 LIU AND ROTTER (Q1). The logit-link function model with random subject intercepts and slopes for NUMBER and the 2-way interaction of FORCE*NUMBER, as well as random item intercepts with slopes for FORCE fit the data best. The results showed an effect of NUMBER with lower ratings in MC than SM conditions (β̂=-1.30, χ (1)=40.33, p<0.0001) — we will label this as Finding AE-1: [Q1/NUMBER: MC<SM]. The 2-way interaction between FORCE*NUMBER turned out significant (β̂=-0.63, χ (1)=24.58, p<0.0001); thus, we conducted a sub-analysis by splitting the data into the levels SM and MC of the factor NUMBER. First, the SM-model included the loglog link function, random subject intercepts and item intercepts with slopes for FORCE. The results showed no significant effect. Second, the MC-model included the logit link function, random subject intercepts with slopes for FORCE and item intercepts. The results showed a significant effect of FORCE with lower ratings in 3 (might maybe) than 2 conditions (must certainly) (β̂=-0.70, χ (1)=14.74, p=0.0001). This suggests that the significant 2-way interaction FORCE*NUMBER of the main analysis stems from lower 3 ratings in the MC condition and absence of a difference in the SM condition — Finding AE-2: [Q1/FORCE*NUMBER: MC → 3<2; SM → 3=2]. condition FL 3 SM FL 3 MC FL 2 SM FL 2 MC FL UM FL MiM IFL 3 SM IFL 3 MC IFL 2 SM IFL 2 MC IFL UM IFL MiM n 253 245 244 241 247 248 240 248 239 252 246 241 mean 6.23 5.04 5.92 5.46 6.00 5.17 6.23 5.03 5.94 5.50 6.06 5.29 Q1 sd 1.23 2.11 1.55 1.81 1.45 1.97 1.25 2.08 1.50 1.84 1.30 1.92 se 0.08 0.13 0.10 0.12 0.09 0.13 0.08 0.13 0.10 0.12 0.08 0.12 mean 5.77 5.58 6.14 6.25 6.52 5.88 5.68 5.57 6.17 6.19 6.48 5.84 Q2 sd 1.13 0.99 1.03 1.05 1.07 1.02 1.08 1.08 1.01 1.18 1.11 1.03 se 0.07 0.06 0.07 0.07 0.07 0.06 0.07 0.07 0.07 0.07 0.07 0.07 Table 1: Descriptive statistics for Q1 and Q2 in Experiment 1 (AE) (Q2). The logit-link function model with random subject intercepts and slopes for FORCE, NUMBER, and the 3-way interaction, as well as random item intercepts fit the data best. The results showed an effect of FORCE with lower ratings in 3 than 2 conditions (β̂=-1.68, χ (1)=86.70, p<0.0001); NUMBER and FORMALITY did not show an effect — which we label respectively as Finding AE-3: [Q2/FORCE: 3<2]; Finding AE-4: [Q2/NUMBER: MC=SM]; Finding AE-5: [Q2/FORMALITY: FL=IFL]. The 2-way interaction between FORCE*NUMBER (β̂=-0.48, χ (1)=25.70, p<0.0001) turned out significant; thus, we conducted a sub-analysis by splitting the data into the levels 3 and 2 of FORCE. First, the model with 3 conditions included the loglog link function, random subject intercepts with slopes for NUMBER, and item intercepts. The result showed a significant effect of NUMBER with lower ratings in MC than SM conditions (β̂=- Modal concord in American and British English 283 0.25, χ (1)=10.16, p=0.001). Second, the model with 2 conditions included the logit link function, random subject intercepts with slopes for NUMBER, and item intercepts. The result showed a significant effect of NUMBER with higher ratings in MC than SM conditions (β̂=0.53, χ (1)=8.71, p=0.003). This suggests that the significant interaction FORCE*NUMBER of the main analysis stems from crossover effect of NUMBER: While MC received lower ratings than SM in the 3 conditions, MC received higher ratings than SM in the 2 conditions — Finding AE-6: [Q2/FORCE*NUMBER: 3 → MC<SM; 2 → MC>SM] . Q1 Q2 Fixed effects Estimates β̂ se Formality (F) -0.01 0.09 Force (MF) -0.09 0.13 Number (MN) -1.30 0.19 3-int 0.03 0.09 2-int:F*MF -0.02 0.09 2-int:F*MN 0.05 0.09 2-int:MF*MN -0.63 0.12 Formality (F) 0.04 0.09 Force (MF) -1.68 0.16 Number (MN) -0.01 0.12 3-int -0.11 0.10 2-int:F*MF 0.05 0.09 2-int:F*MN -0.03 0.09 2-int:MF*MN -0.48 0.10 z-value -0.13 -0.65 -6.91 0.28 -0.19 0.49 -5.19 0.46 -10.73 -0.11 -1.05 0.56 -0.33 -5.04 Model comparison χ (1) p-value 0.02 0.90 0.42 0.52 40.33 <0.0001 0.08 0.78 0.04 0.85 0.24 0.62 24.58 <0.0001 0.21 0.65 86.70 <0.0001 0.01 0.92 1.12 0.29 0.31 0.58 0.11 0.74 25.70 <0.0001 Table 2: Model output of the main analysis for Q1 and Q2 in Experiment 1 (AE) 2.2 Experiment 2: British English 2.2.1 Participants 120 native speakers of BE were recruited and all participants gave their informed consent. Based on the data quality (see 2.1.4), we removed the entire data of 5 participants. The remaining data set consisted of 115 participants (mean age=37.2, SD=11.4; female N = 58, male = 57, non-binary = 0). In total, 14.8% of the participants reported to be dialect speakers with different dialect backgrounds: Bristolian (N=1), Brummie (N=1), Cardiff (N=1), Derbyshire (N=1), Doric (N=1), Estury (N=1), Geordie (N=1), Leicester (N=1), Midlands (N=1), Norfolk (N=1), Scottish (N=3), Sussex (N=1), Welsh (N=2), and Yorkshire (N=1) English dialect. The majority have a graduate degree (47.0%); 29.6% hold a college degree; 22.6% have completed high school; and below 1% did not finish high school. More than half of the participants grew up in a suburban environment (53%), while 32.2% grew up in an urban and 14.8% in a rural environment. 2.2.2 Results In the post-task, the majority of the participants indicated that they prefer the SM construction (N3 =111, N2 =104). Two participants indicated that they prefer the 3MC construction (non-dialect speaker (N=2)), and two indicated no preference be- 284 LIU AND ROTTER tween 3-SM vs. -MC (Yorkshire (N=1) and Doric (N=1) English dialect speaker). Three participants indicated that they prefer the 2-MC construction (non-dialect speaker (N=2), Scottish dialect speaker (N=1)), and 8 indicated no preference between 2-SM or -MC (Yorkshire English dialect (N=1), no dialect indicated (N=1), non-dialect speaker (N=6)). The descriptive statistics of the ratings in (Q1) and (Q2) are shown in Table 3 and depicted in Figure 1. The output for the models with (Q1) and (Q2) as dependent variables is shown in Table 4. condition FL 3 SM FL 3 MC FL 2 SM FL 2 MC FL UM FL MiM IFL 3 SM IFL 3 MC IFL 2 SM IFL 2 MC IFL UM IFL MiM n 245 257 253 260 256 256 258 255 250 262 252 256 mean 5.91 5.00 5.50 5.27 5.69 4.84 6.09 4.84 5.55 5.30 5.89 5.12 Q1 sd 1.33 1.80 1.49 1.73 1.54 1.98 1.21 1.93 1.49 1.68 1.34 1.79 se 0.08 0.11 0.09 0.11 0.10 0.12 0.08 0.12 0.09 0.10 0.08 0.11 mean 5.47 5.34 5.94 6.00 6.29 5.60 5.54 5.37 5.88 5.91 6.37 5.59 Q2 sd 0.99 0.99 0.95 1.08 1.07 1.04 0.92 0.92 1.06 1.08 1.03 0.99 se 0.06 0.06 0.06 0.07 0.07 0.06 0.06 0.06 0.07 0.07 0.06 0.06 Table 3: Descriptive statistics for Q1 and Q2 in Experiment 2 (BE) Q1 Q2 Fixed effects Estimates β̂ se Formality (F) -0.04 0.05 Force (MF) 0.12 0.07 Number (MN) -0.60 0.09 3-int 0.07 0.05 2-int:F*MF -0.02 0.05 2-int:F*MN 0.17 0.06 2-int:MF*MN -0.41 0.07 Formality (F) 0.04 0.08 Force (MF) -1.21 0.09 Number (MN) -0.10 0.08 3-int -0.03 0.08 2-int:F*MF -0.08 0.08 2-int:F*MN 0.08 0.08 2-int:MF*MN -0.23 0.08 z-value -0.77 1.75 -6.72 1.31 -0.36 2.81 -6.19 0.49 -13.87 -1.24 -0.37 -0.93 0.90 -2.69 Model comparison χ (1) p-value 0.59 0.44 3.05 0.08 40.42 <0.0001 1.71 0.19 0.13 0.72 7.55 <0.01 35.06 <0.0001 0.24 0.62 201.59 <0.0001 1.53 0.22 0.14 0.71 0.86 0.35 0.81 0.37 7.27 <0.01 Table 4: Model output of the main analysis for Q1 and Q2 in Experiment 2 (BE) (Q1). The loglog-link function model with random subject intercepts and slopes for FORCE, NUMBER, and the 2-way interaction of FORMALITY*NUMBER as well as random item intercepts fit the data best. The results showed an effect of Modal concord in American and British English 285 NUMBER with lower ratings in MC than SM conditions (β̂=-0.60, χ (1)=40.42, p<0.0001) — Finding BE-1: [Q1/NUMBER: MC<SM]. The 2-way interactions FORMALITY*NUMBER (β̂=0.17, χ (1)=7.55, p=0.006) and FORCE*NUMBER (β̂=-0.41, χ (1)=35.06, p<0.0001) turned out significant; thus, we conducted a subanalysis by splitting the data into the levels SM and MC of the factor NUMBER. First, the SM-model included the loglog link function, random subject intercepts and item intercepts with slopes for FORMALITY. The results showed a significant effect of FORCE with higher ratings for 3 than 2 (β̂=0.43, χ (1)=33.59, p<0.0001). Furthermore, there was a marginally significant effect of FORMALITY with lower ratings in FL than IFL (β̂=-0.18, χ (1)=2.91, p=0.09). Second, the MCmodel included the logit link function, random subject as well as item intercepts with slopes for FORMALITY. The results showed a significant effect of FORCE with lower ratings in 3 than 2 conditions (β̂=-0.54, χ (1)=10.28, p=0.001). Furthermore, there was a marginally significant effect of FORMALITY with higher ratings in FL than IFL (β̂=0.22, χ (1)=3.29, p=0.07). This suggests that the significant 2-way interaction FORCE* NUMBER stems from the cross-over effect in the NUMBER condition: While 3 received higher ratings than 2 in SM conditions, the reverse applies to MC conditions — Finding BE-2: [Q1/FORCE*NUMBER: MC → 3<2; SM → 3>2]. Furthermore, the significant 2-way interaction FORCE*NUMBER stems from the cross-over effect of FORMALITY in that FL conditions are rated marginal significantly lower than IFL in SM conditions, the reverse applies to MC conditions — which we will label as an tentative Finding BE-3: [Q1/FORMALITY*NUMBER: MC → FL>IFL; SM → FL<IFL] . (Q2). The logit-link function model with random subject intercepts fit the data best. The results showed an effect of FORCE with lower ratings in 3 than 2 conditions (β̂=-1.21, χ (1)=201.59, p<0.0001); NUMBER and FORMALITY did not show an effect — which we label respectively as Finding BE-4: [Q2/FORCE: 3<2]; Finding BE-5: [Q2/NUMBER: MC=SM]; Finding BE-6: [Q2/ FORMALITY: FL=IFL]. As the 2-way interaction FORCE*NUMBER (β̂=-0.23, χ (1) =7.27, p=0.007) was significant, we conducted a sub-analysis by splitting the data into the levels 3 and 2 of the factor FORCE. First, the 3-model included the logit link function, random subject intercepts and item intercepts with slopes for FORMALITY. The results showed a significant effect of NUMBER with lower ratings in MC than SM conditions (β̂=-0.38, χ (1)=9.63, p=0.002). Second, the 2-model included the logit link function, random subject and item intercepts as well as the 2-way interaction of FORMALITY*NUMBER. The results showed no significance. This suggests that the significant 2-way interaction FORCE*NUMBER stems from the significant effect of NUMBER in the 3 conditions, which is absent in the 2 conditions — Finding BE-7: [Q2/FORCE*NUMBER: 3 → MC<SM; 2 → MC=SM]. 3 Discussion and conclusion To sum up, the current study has the following findings: 286 LIU AND ROTTER [H1a] was confirmed: In both AE and BE, MC was rated less appropriate than SM (cf. Finding AE-1 / BE-1). Furthermore, across both data sets, possibility MC was rated less appropriate than necessity MC (cf. Finding AE-2 / BE-2), which could be due to the specific items we used. For this, we conducted a preliminary search in the iWeb corpus (https://www.english-corpora.org/ iweb/). The results show among others that might maybe has a much lower frequency than must certainly. Thus, we are planning to conduct additional studies using the more frequent combinations such as may possbily3 . Beyond these similarities, AE and BE show one difference in that whereas might and must were rated as equally appropriate in AE, might was rated as more appropriate in BE (cf. Finding AE-2 vs. BE-2). [H1b] was not confirmed: We did not find register effects in the AE data. However, the marginally significant effect in BE shows that MC tends to be more appropriate in formal than in informal contexts (cf. tentative Finding BE-3), the direction of the tendency being as predicted. It is to note that in our study we only used one situational variable ‘interlocutor relation’ to manipulate the level of formality; future work may consider different variables or the combinations thereof for targeting the register effects simulating a more realistic setting of natural language communication. A further methodological question we ask is whether the appropriate rating measure is sensitive enough for the targeted effects. [H2a] was confirmed: In both AE and BE, MC and SM did not differ in interpretation, in line with the concord analysis (cf. Finding AE-4 / BE-5). However, sub-analyses show that in both AE and BE, possibility-MC received lower speaker commitment ratings than its SM-variant. This means that MC has a more weakening effect than SM. In the case of necessity modals, on the other hand, MC received higher ratings than SM in AE (with no difference in BE), meaning that it has an additive (strengthening) effect. The latter two findings pose questions for the MC analysis of the tested constructions (cf. Finding AE-6 / BE-7). Thus, further studies need to consider the possibility that MC and SM after all are not functionally equivalent but have subtle differences in terms of semantics (as well as enriched meanings). Concerning the explorative question in [H2b], in both data sets, there was no effect of register on the interpretation of MC (cf. Finding AE-5 / BE-6). While the experimental study of MC is preliminary and the results about the register effects are unclear, the register-based approach brings us to a few interesting and puzzling findings about possibility and necessity MC in two varieties of English, calling for further investigation. Data availabiltiy statement The data sets presented in this study can be found in online repositories: https://osf.io/v8ktb/. Acknowledgement This research is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – SFB 1412, 416591334. 3 In fact, the MiM construction might certainly has a much higher frequency than might maybe, in line with the experimental results, see Figure 1. For scope reasons, we will not further discuss the contrast between MC and MiM in the paper. Modal concord in American and British English 287 References Bates, D., R. Kliegl, S. Vasishth, & H. Baayen. 2015. Parsimonious Mixed Models. arXiv 1506. Christensen, R. H. B. 2019. ordinal—regression models for ordinal data. R package version 2019.12-10. https://CRAN.R-project.org/package=ordinal. Accessed on 2022-01-28. Eckert, P. 2019. The limits of meaning: Social indexicality, variation, and the cline of interiority. Language 95(4). 751–776. Geurts, B., & J. Huitink. 2006. Modal concord. In Proceedings of the ESSLLI 2006 Workshop Concord Phenomena at the Syntax-Semantics Interface, ed. by P. Dekker & H. Zeijlstra, 15– 20. University of Malaga. Giannakidou, A., & A. Mari. 2017. Epistemic future and epistemic must: Nonveridicality, evidence, and partial knowledge. In Mood, Aspect, Modality Revisited: New Answers to Old Questions, ed. by J. Blaszczak, A. Giannakidou, D. Klimek-Jankowska, & K. Migdalski, 75– 118. University of Chicago Press. Halliday, M. A. K. 1970. Functional diversity in language as seen from a consideration of mood and modality in english. Foundations of Language 6. 322–361. Howcroft, D. M., & V. Rieser. 2021. What happens if you treat ordinal ratings as interval data? Human evaluations in NLP are even more under-powered than you think. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing . Huang, N. 2011. Multiple modals. Yale Grammatical Diversity Project: English in North America 322–361. http://ygdp.yale.edu/phenomena/multiple-modals. Accessed on 2021-12-07. Liddell, T. M., & J. K. Kruschke. 2018. Analyzing ordinal data with metric models: What could possibly go wrong? Journal of Experimental Social Psychology 79. 328—-348. Liu, M., S. Rotter, & A. Giannakidou. 2021. Bias and modality in conditionals: Experimental evidence and theoretical implications. Journal of Psycholinguistic Research 50. 1369–1399. Lüdeling, A., A. Alexiadou, A. Adli, K. Donhauser, M. Dreyer, M. Egg, A. H. Feulner, N. Gagarina, W. Hock, S. Jannedy, F. Kammerzell, P. Knoeferle, T. Krause, M. Krifka, S. Kutscher, B. Lütke, T. McFadden, R. Meyer, C. Mooshammer, S. Müller, K. Maquate, M. Norde, U. Sauerland, S. Solt, L. Szucsich, E. Verhoeven, R. Waltereit, A. Wolfsgruber, & L. E. Zeige. 2022. Register: Language Users’ Knowledge of Situational- Functional Variation– Frame Text of the First Phase Proposal for the CRC 1412. Register Aspects of Language in Situation (REALIS) 1(1). 1–57. Mondorf, B. 2003. Support for more-support. In Determinants of Grammatical Variation in English, ed. by G. Rohdenburg & B. Mondorf, 251–304. Berlin, New York: De Gruyter Mouton. Pescuma, V. N., D. Serova, J. Julia, A. Sauermann, R. Schäfer, A. Adli, F. Bildhauer, M. Egg, K. Hülk, A. Ito, S. Jannedy, V. Kordoni, M. Kuehnast, S. Kutscher, R. Lange, N. Lehmann, M. Liu, B. Lütke, K. Maquate, C. Mooshammer, V. Mortezapour, S. Müller, M. Norde, E. Pankratz, A. G. Patarroyo, A.-M. Plesca, C. Rodríguez-Ronderos, S. Rotter, U. Sauerland, B. Schulte, G. Schüppenhauer, B. M. Sell, S. Solt, M. Terada, D. Tsiapou, E. Verhoeven, M. Weirich, H. Wiese, K. Zaruba, L. E. Zeige, A. Lüdeling, & P. Knoeferle. 2022. Situating language register across the ages, languages, modalities, and cultural aspects: Evidence from complementary methods. Frontiers in Psychology 13. R Core Team. 2022. R: A Language and Environment for Statistical Computing. https://www.R-project.org/. Accessed on 2022-11-01. Rotter, S., & M. Liu. 2022. Interlocutor relation predicts the formality of the conversation: an experiment in American and British English. Submitted manuscript/under review, HumboldtUniversität zu Berlin. RStudio Team. 2022. RStudio: Integrated Development Environment for R. http://www.rstudio.com/. Accessed on 2022-08-14. Zehr, J., & F. Schwarz. 2018. PennController for Internet Based Experiments (IBEX). Zeijstra, H. 2007. Modal concord. Proceedings of Semantics and Linguistic Theory 317–332. 288 LIU AND ROTTER Appendix We provide an additional analysis on SM and UM here for the interested reader. In both data sets, UM received significantly higher speaker commitment ratings than SM, and 2-SM received higher speaker commitment ratings than 3-SM. This finding (more−committed <UM, MUST, MIGHT>less−committed ) is in line with the speaker commitment scale , i.e., more−committed <UM, MUST-P, POSSIBLY-P>less−committed as proposed in Giannakidou & Mari (2017). Data analysis: For both data sets, we conducted 1-factorial analyses of UM and SM of 3 and 2 with two models. First, we used UM as reference level plus a slope for 3 and 2. Second, we used 3 as reference level plus a slope for UM and 2, see Section 2.1.4 for modelling details. The results are shown in Table 5. Experiment 1 Model 1 Model 2 Experiment 2 Model 1 Model 2 Intercept UM 3 UM 3 Fixed effects Estimates β̂ 3 -1.75 2 -1.06 UM 1.75 2 0.69 3 -2.84 2 -1.61 UM 2.84 2 1.23 se 0.11 0.11 0.11 0.09 0.27 0.19 0.27 0.17 z-value -16.08 -9.65 16.08 7.92 -10.63 -8.33 10.63 7.36 Model comparison χ (1) p-value 305.49 p<0.0001 100.23 p<0.0001 305.49 p<0.0001 64.45 p<0.0001 92.77 p<0.0001 70.80 p<0.0001 92.77 p<0.0001 48.28 p<0.0001 Table 5: Model output of the second analysis for Q2 Experiment 1 (AE) and 2 (BE) Experiment 1: Model 1 fit the data best with loglog link function and random subject intercepts. The results showed a significant effect of 3 with lower ratings in 3 than UM conditions (β̂=-1.75, χ (1)=305.49, p<0.0001). 2 showed a significant effect with lower ratings in 2 than UM conditions (β̂=-1.06, χ (1)=100.23, p<0.0001). Model 2 fit the data best with the loglog link function and random subject intercepts. The results showed a significant effect of UM with higher ratings in UM than 3 conditions (β̂=1.75, χ (1)=305.49, p<0.0001). 2 showed a significant effect with higher ratings in 2 than 3 conditions (β̂=0.69, χ (1)=64.45, p<0.0001). Experiment 2: Model 1 fit the data best with logit link function and random subject intercepts with 3 and 2, as well as random item intercepts. The results showed a significant effect of 3 with lower ratings in 3 than UM conditions (β̂=2.84, χ (1)=92.77, p<0.0001). 2 showed a significant effect with lower ratings in 2 than UM conditions (β̂=-1.61, χ (1)=70.80, p<0.0001). Model 2 fit the data best with logit link function and random subject and intercepts as well as 3 and 2 random slopes for subjects. The results showed a significant effect of UM with higher ratings in UM than 3 conditions (β̂=2.84, χ (1)=92.77, p<0.0001). 2 showed a significant effect with higher ratings in 2 than 3 conditions (β̂=1.23, χ (1)=48.28, p<0.0001). Proceedings of CLS 58 (2022), 289-301 © Chicago Linguistic Society 2023. All rights reserved. 289 Genre formation as enregisterment Aidan Malanoski CUNY Graduate Center 1 Introduction: Enregisterment and musical genre In this paper, I argue that the formation of musical genres is a process of enregisterment. Enregisterment refers to the processes by which a semiotic repertoire—a set of potentially meaningful signs—becomes recognizable and acquires social meaning (Agha 2003, 2005). The product of enregisterment is a register, a socially recognized linkage between a semiotic repertoire and particular social meanings. Often, the meaning of a register coalesces into a persona (or characterological figure), a “holistic, ideological social type[] that [is] recognizably linked with ways of being and speaking” (D’Onofrio 2020). An example of a register is Yat English, which links a “Yat” persona—a White, working-class New Orleanian associated with tackiness and lack of education, but also with authentic local identity—to a repertoire that includes features such as th-stopping, r-lessness, and COIL - CURL reversal (Carmichael & Dajko 2016). A register is transmitted in a speech chain, a chronological series of speech events in which the receiver of one message is the sender of the next message (Agha 2003). Participants in the speech chain link particular qualities, values, or personae to particular signs. Different speakers may establish different linkages, so a register’s repertoire and social meaning may evolve as it is transmitted in the speech chain (Agha 2005). Enregisterment is not a language-specific process. Agha (2003, 232) noted early on that the processes involved in enregisterment “apply to. . . any other cultural form,” and in general, “potentially meaningful acts of any kind” can be enregistered (Johnstone, 2016, 634). Thus, the application of enregisterment to the formation of musical genres is entirely natural. This is an emic approach to genre, intended to model musical genres as they are understood by everyday social actors. Furthermore, approaching genre formation as enregisterment requires a combination of a presentist approach (which “retroactive[ly] group[s] . . . texts into a genre based on a presumed stylistic consistency and critical consensus”; Brackett 2016, 5) and a historicist approach (which “stud[ies] . . . the conflictual meanings of categories via a reconstruction of a historical horizon of meaning”; ibid.). This is because both historical and contemporary viewpoints participate in the speech chain that transmits a genre. In this respect, the theory of genre formation as enregisterment contrasts with musicological approaches to genre, which tend to be presentist (Brackett 2016). If genre formation is enregisterment, then the creation of a genre depends on discourse in a speech chain. It is not enough for performers to create similar music; only through discourses about music will certain musical texts be grouped together and recognized as distinct from other groupings of musical texts. Lena (2012, 6) 290 AIDAN MALANOSKI draws the same conclusion: “a genre exists where there is some consensus that a distinctive style of music is being performed.” The dependence of genre on a speech chain is why “generic conventions . . . are constantly being modified by each new text that participates in the genre” (Brackett 2016, 13): musical texts are themselves links in a speech chain, implicitly (and sometimes explicitly) linking certain signs (the stylistic characteristics of the text) with certain social meaning (the social categories inhabited by the performers, for example). However, they are not the only links in a speech chain; speech events involving fans, critics, musicians, marketers and other actors all contribute to the development of the genre, as Brackett himself notes. Of course, if genres form through enregisterment, then they are registers. Crucially, a register is not a repertoire; it is the linkage between a repertoire and particular social meanings. Consequently, both the repertoire and meaning associated with a genre may change over time without being seen as a break in the genre. Conversely, the genre label can change without any corresponding change in repertoire or meaning (Agha 2005). Music that was formerly considered rock can become pop, for example (Fornäs 1995, Brackett 2016). Furthermore, different individuals may hold “competing models” of a register (Agha 2005, 56). Different individuals may associate a genre with different musical characteristics—that is, with different repertoires—giving rise to genre debates (Rockwell 2012). They could also associate a genre with the same repertoire, but disagree about which features are sufficient or necessary to invoke a genre (see Blommaert & Varis 2011). And just as a genre’s repertoire can vary across individuals, so too can its meaning. As seen in the discussion of carimbó chamegado below, not all individuals associate the genre with caboclo ethnic identity. Another consequence of this proposal is that the features that form a genre’s repertoire need not be musical; the repertoire may contain meaningful signs of any type. For example, the country music repertoire appears to include rhoticity: whereas American singers avoid rhoticity in most genres of popular music (Trudgill 1983), the American country singers investigated by Duncan (2017) all produce coda [r] at least 60% of the time. Likewise, the difference between heavy metal and white metal lies not in musical characteristic but in subject matter, indicating that the latter genre’s evangelical themes form part of its repertoire (Moore 2001). If genres are registers, then they are linked to sociocultural meaning, a point that was mentioned but not defended above. Some genres seem to be transparently linked to social personae; for genres such as outlaw country, punk rock, and gangsta rap, the very name of the genre points to a specific social type. More broadly, psychological research has demonstrated that listeners associate different genres of music with social and psychological traits (e.g., Rentfrow & Gosling 2007, Shevy 2008). Through a genre’s social meaning, we can understand the relationship between genre and identity. Brackett (2016, 25), following Born (2000), distinguishes between homologous genre-audience relations, which are “most deeply felt and experienced as ‘natural”’, and imaginary relations, which “parallel those aspects of ourselves that we are conscious of learning how to ‘perform’.” From the perspective of enregisterment, homologous relations are those where the audience sees themself (or is seen) as “matching” the persona linked to a genre: the audience is recognized Genre formation as enregisterment 291 to have the characteristics associated with the persona. Although homologous relations are felt to be natural, they are not (Brackett 2016): the link between genre and identity is the ideological output of a speech chain, however inevitable that link appears in retrospect. Imaginary relations involve a mismatch between audience and persona: some of the characteristics of the persona are at odds with those of the audience. Importantly, both homologous and imaginary relations are subjective: individuals may disagree as to whether a given audience matches the persona linked to a genre. Consequently, all genre-audience relations are imaginary to some extent, existing not in the real world but in the minds of individuals. What distinguishes so-called imaginary relations from homologous ones is whether they are felt to be real. Of course, an individual may produce or consume various genres of music. In some cases, they will have a more homologous relation with the persona associated with the genre; in other cases, they will have a more imaginary relation. By engaging in a homologous relation, an individual can reinforce aspects of their own identity (or the identity projected onto them). By engaging in an imaginary relation, an individual can take on aspects of an identity seen as different, which may “prefigure, crystallize or potentialize emergent, real forms of sociocultural identity” (Born & Hesmondhalgh 2000, 35; Brackett 2016 refers to this as an emergent relation). Conversely, an individual engaging in an imaginary relation may dissociate themself from the genre persona, as in parody, in order to reinforce their own identity or avoid undesired characteristics of another identity. Such an imaginary relation serves to “inscribe and reinscribe existing boundaries of self and other” (Born & Hesmondhalgh 2000, 35; Brackett 2016 refers to this as an exoticist relation). Through acts of identity, the meaning of a genre can change. If a group of individuals repeatedly engages in an imaginary relation with a genre, then eventually the genre may come to be associated with (the characteristics of) that group, changing its meaning. In this way, what was once a performative relation may become natural over time—an imaginary relation may become homologous. Conversely, homologous identification may serve to preserve the genre in its current form. In either case, “fantasies of identity continue to sustain the genre” (Gunn 1999, 39). Even as the meaning of a genre changes, so too may its repertoire. If the meaning and repertoire change in tandem, this may be interpreted as the birth of a new genre, rather than a continuation of the old genre. Brackett discusses the case of blues rock, a genre associated with White people. Blues rock, of course, is based on blues, a genre associated with Blackness. What originated as an exoticist relationship between White audiences and blues developed into a homologous relationship between White audiences and blues rock. In short, rather than blues itself becoming associated with Whiteness, it gave way to new genre associated with Whiteness. Practically speaking, not all identity relations are equally available to the individual. Certain relations may be censured (e.g., those subject to accusations of cultural appropriation), and institutions may aim to solidify one potential identification at the expense of another. Since both individual acts of identification and institutional constraints form part of the speech chain, we must attend to the role of both top-down and bottom-up forces in the development of a genre. I have proposed that musical genres form through enregisterment, and are therefore registers themselves. This proposal accounts for (i) the ability of genres (or 292 AIDAN MALANOSKI their labels) to change; (ii) the fact that different individuals may disagree about the properties of a genre; (iii) the role of non-musical features in a genre; and (iv) the sociocultural associations of genres. Moreover, the analysis of genres as registers provides an account of the relation between genre and identity. 2 Case study: carimbó chamegado To illustrate the proposal that genre formation is enregisterment, I present a case study of carimbó chamegado, also called chamego, an emergent genre spearheaded by Dona Onete.1 Ionete da Silveira Gama, known by her stage name Dona Onete, is a singer-songwriter from Pará, a state in the North Region of Brazil largely covered by the Amazon rainforest. According to Lamen (2013), Pará “has historically been viewed as isolated and provincial” (133) and occupies a “liminal position on the periphery of the Brazilian nation” (150). Dona Onete identifies as cabocla (G.Lab 2021). The term caboclo (feminine: cabocla) historically referred to indigenous peoples, and then to those of partial indigenous heritage (Lisansky 2008). However, caboclo is now a pejorative term associated with rurality and low socioeconomic status, akin to the English word hick (ibid.). Despite the history of the term, the categories caboclo and indigenous are today seen as mutually exclusive and even opposed, with indigenous peoples “understood as standing outside the [Brazilian] political economy, the [caboclo] standing precariously on its margins (Lamen 2013, 147). The caboclos’ marginalization goes beyond the socioeconomic. According to Adams et al. (2009, 4), “nationally, the caboclo represents an unfinished project of the creation of a Brazilian culture that broke with its European, African and Indigenous antecedents”. Given that the “founding myth of Brazilian nationality” presents the Brazilian nation as “the product of ‘racial admixture’ brought into being through the meeting of Europeans, Africans, and Amerindians” (Santos et al. 2009, 798), Adams et al.’s claim effectively identifies caboclos as unfinished, imperfect Brazilian citizens. Thus, the caboclos are symbolically marginal as a representation of an imperfect past incarnation of the Brazilian nation. As noted above, Dona Onete calls her music carimbó chamegado, identifiying it as a subgenre of carimbó, a genre of music practiced in Pará. Along with siriá, carimbó is one of genres in Pará “most explicitly associated with blackness” (Lamen 2013, 148). Carimbó has also been identified as a tradition “available for assembly as caboclo culture” (Nugent 1997, 45). As we will see below, Dona Onete’s carimbó chamegado embraces the ethnic associations of carimbó while rejecting the marginalization of the caboclo. 1 The word chamego has a number of meanings: ‘very intimate friendship’, but also ‘excitement for a lustful act’ and ‘violent passion or irresistible physical attraction’ (these definitions are from the Michaelis dictionary). Chamegado is the perfect participle of chamegar, a verb derived from chamego. Genre formation as enregisterment 293 2.1 Carimbó chamegado according to Dona Onete Dona Onete has released three albums: Feitiço Caboclo in 2012, Banzeiro in 2017, and Rebujo in 2019.2 She has also released two standalone singles, “Boto Namorador” (’flirtatious dolphin’) and “Homenagem de Dona Onete para Rainha Soberana da Amazônia” (’Dona Onete’s homage for [the] sovereign queen of Amazonia’). As discussed above, musical texts are links in a speech chain, since they implicitly musical characteristics with social meaning. However, many of Dona Onete’s lyrics explicitly describe her own music—that is, they explicitly link musical characteristics to social meaning. Consequently, these songs are expected to play a key role in the construction of carimbó chamegado, and will be a central focus of the analysis. First, let us consider how Dona Onete defines the repertoire of carimbó chamegado. On the song “Carimbó Chamegado” from Feitiço Caboclo, Onete describes the genre as being performed on makeshift instruments built from materials available in the rural Amazon. These include a banjo made from a cooking pot, a wooden flute and viola, and percussion instruments made from animal skins and plants found in Amazonia. Onete also describes carimbó chamegado as a mixture of musical and cultural traditions of Pará, including lundu, bangué, carimbó, siriá, tambor de nagô, and boi bumba. On “Carimbó Arrepiado” (‘carimbó with goosebumps’, from Rebujo), Onete identifies the saxophone as another instrument of her carimbó while reaffirming the role of the banjo and of various percussion instruments mentioned on “Carimbó Chamegado.” Moreover, she specifies how carimbó chamegado is played: the drum has a strong thump, the banjo does an exciting dance, and the saxophone runs wild and improvises notes. On “Vem Chamegar” (‘come chamegar’, from Rebujo), Onete defines carimbó in opposition to other genres, identifying suingue (‘swing’, understood in the context to refer to chamego, given the title of the song and the fact that suingue “swing” replaces chamego in the second verse) as something distinct from the Latin American genres of merengue, cumbia, zouk (a genre originating in the French Antilles), and chá-chá-chá. Thus, Onete excludes foreign influences from the repertoire of carimbó chamegado. By and large, there is no explicit metacommentary on Dona Onete’s lyrics within those lyrics. However, the recurrence of certain themes across Dona Onete’s lyrics implicitly link those themes to the repertoire of carimbó chamegado. These implicit links are supplemented by Onete’s explicit commentary on her lyrics and in interviews and other fora. One major theme of carimbó chamegado is love and sexuality, a theme present in the very name of the genre (see footnote 1). Indeed, songs about desire recur throughout her oeuvre, from Feitiço Caboclo’s “Poder da Sedução” (‘power of seduction’) to Banzeiro’s “No Sabor do Beijo” (‘in the taste of the kiss’) and Rebujo’s “Galante Sedutor” (‘gallant seducer’). Indeed, Onete asserts that the difference between carimbó chamegado and other types of carimbó is that the former “talks 2 Feitiço caboclo means ‘caboclo spell’. Banzeiro and rebujo are both regional terms: a banzeiro is a wave caused by passing ships (Ferreira 2016), and rebujo refers to debris from the bottom of a river that has been brought to the surface by the movement of fish (Natura 2019). 294 AIDAN MALANOSKI about love, in a slower and very sensual rhythm” (Brasil Música e Artes 2019). A second major topic of Dona Onete’s lyrics is life in Pará, with songs about cuisine (e.g., “Jamburana”, about an herb used in the region’s cuisine), music popular in Pará (e.g., “Faceira”, an homage to other siriá and carimbó musicians), and various Paraense locales (e.g., “Lua Jaci”, where Onete recounts a night spent singing carimbó along the banks of the Guamá river). In her own words, she “talks about the waters, the river, the trees, the fruit, the magic . . . [she] takes a lot of inspiration from the people who live on the rivers” (G.Lab 2021). As for the meaning of carimbó chamegado, Onete links it to caboclo identity. On “Vem Chamegar,” she calls her suingue a “nice caboclo rhythm,” and on “Carimbó Chamegado,” she sings, “[The] caboclo is a dancer / The caboclo is a musician . . . [The] caboclo is a composer,” identifying chamego as a music produced by caboclos. Furthermore, the caboclo is the object of chamego’s sexual themes: on “Fogo na Aldeia” (‘fire in the village’), Onete sings of “kissing a caboclo mouth.” Likewise, the song “Mistura Pai D’Égua” (‘cool mix’) closes with the lines “What caboclo beauty / He’s handsome, that caboclo.” Furthermore, Dona Onete identifies both carimbó chamegado and caboclos themselves as the products of racial mixture. On “Carimbó Chamegado,” when singing the supposed origin story of her music, she says that “the Black man, White man, and Indian [índio] left everything mixed.” On the same song, however, she identifies the caboclo as the creator of carimbó chamegado, as discussed above. If chamego was created by caboclos, but is also a mixture created by “the Black man, White man, and Indian,” this implies that caboclos themselves are a mixture of Black, White, and indigenous influences. Similarly, on “Mistura Pai D’Égua,” Onete describes her chamego as the outcome of a“cool mix / [that] happened in Pará / a mix of races / a mix of colors,” and specifically identifies it as negro (roughly ‘Black’), loiro (‘blond’), moreno (‘dark-haired”or ‘dark-skinned’), crioulo (‘creole’ or ‘Black’), mulato (‘mulatto’), mestiço (‘mestizo’), and caboclo, in that order. By listing caboclo last in this list of racial descriptors, Dona Onete implicitly presents caboclos as a category encompassing the preceding racial terms. This is also implied by the dual categorization of carimbó chamegado: carimbó chamegado is caboclo, but it is also the product of racial mixture, insinuating that being caboclo means being the product of racial mixture. Furthermore, Pará is presented as the site of that mixture. Given that racial mixture is the founding myth of the Brazilian nation (Santos et al. 2009), by presenting caboclos and Pará as the product and the site of racial mixture, Dona Onete identifies them as quintessentially Brazilian. In doing so, she creates a counternarrative that challenges the peripheral position of caboclos within Pará and of Pará within Brazil: Pará epitomizes Brazil, and the caboclo epitomizes Pará. This section has presented carimbó chamegado as constructed by Dona Onete— that is, as presented in those links in the speech chain for which Dona Onete is responsible. The repertoire of carimbó chamegado includes instruments such as the banjo, the flute, the viola, the saxophone, and various percussion instruments. It also includes elements of other genres of Pará while eschewing foreign influences. The repertoire is also characterized by lyrics about love and sexuality on the one hand, and about life in Pará on the other. The genre is linked to a caboclo persona. Genre formation as enregisterment 295 The persona is associated with sexual desire—specifically, it is presented as the object of that desire—and is identified as racially mixed. This implicitly links the caboclo persona to Brazilianness, alongside a more transparent connection to Pará. 2.2 The transmission of carimbó chamegado Let us now turn to the transmission of carimbó chamegado beyond the words and music of Dona Onete. I make no attempt to provide a comprehensive overview of discourse about carimbó chamegado. The goal of the present section is to demonstrate how to analyze genre through the lens of enregisterment, specifically by examining how a genre is constructed and transmitted in a speech chain. Given these goals, it is sufficient to consider only a small section of that speech chain. Many aspects of carimbó chamegado as presented in Dona Onete’s music are recirculated in the descriptions of her albums provided on Bandcamp, though sometimes with modifications. These album descriptions are written in English, a language that Dona Onete does not speak, and refer to Dona Onete in the third person, so it is unlikely that Onete is directly responsible for their content. Rather, they were likely written by someone working at Dona Onete’s record label, Mais Um, which is based in the United Kingdom. The album descriptions identify sexual lyrics as part of the genre’s repertoire. The page for Feitiço Caboclo, for example, describes how her “songs talk about the delights of seducing men” (Mais Um Discos 2020), while the page for Banzeiro says, “Whether she’s. . . singing about the delights of indecent proposals or praising a former lover for his ‘crazy ways of making love’, Banzeiro is defined by Onete’s honest reflections on life, love and sex” (Mais Um Discos 2017). Likewise, the album descriptions recapitulate Onete’s depiction of carimbó chamegado as a mixture of various musical genres, for example referring to it as a “hybrid genre” (Mais Um Discos 2020). That said, the Bandcamp descriptions differ in what they present carimbó chamegado as a hybrid of. The descriptions for both Banzeiro and Rebujo read, “the music [Dona Onete] sings is a unique mix of rhythms from native Brazilians, African slaves and the Caribbean” (Mais Um Discos 2017, Mais Um Discos 2019). This contrasts with the description of the genre on “Vem Chamegar,” which opposed chamego to cha-cha-chá and merengue, genres of Caribbean origin. Thus, the Bandcamp descriptions assign a place to Caribbean influences, whereas Dona Onete excluded them. The preceding quotation also demonstrates the theme of racial mixing: carimbó chamegado is a “mix of rhythms from native Brazilians, [and] African slaves.” Although European influence is absent from this description, it is invoked on the description for Feitiço Caboclo: “Onete sings carimbó, an indigenous rhythm and dance from Pará. . . influenced by both African and European traditions” (Mais Um Discos 2020). Thus, the album descriptions present Dona Onete’s music as a fusion of indigenous, African and European influences, although they tend to focus on the indigenous and African aspects of the music. The Bandcamp descriptions also reproduce the importance of Amazonian culture. However, unlike in Dona Onete’s music, the emphasis is on the Amazon as a whole rather than on Pará specifically. The only reference to Pará itself is in the above quotation, where carimbó is referred to as “an indigenous rhythm and dance 296 AIDAN MALANOSKI from Pará.” Elsewhere, Dona Onete is described as the “Queen of the Amazon” (Mais Um Discos 2019) and “an ambassador of Amazonian culture” (Mais Um Discos 2017, Mais Um Discos 2019), and is said to sing about “her delight in the everyday pleasures of life in the Amazon” (Mais Um Discos 2017). This contrasts with Onete’s music, which—with one exception—never mentions the Amazon river or Amazonia and does not, to my knowledge, reference any place in the Amazon outside of Pará.3 Most egregiously, the Bandcamp descriptions make no mention of caboclos, with the word only occurring in the name of the album Feitiço Caboclo and its title track. Whereas Dona Onete’s music combatted the marginalization of caboclos, the album descriptions reproduce that marginalization, erasing caboclos from the presentation of Dona Onete’s music. The genre is reconfigured yet again on the album reviews left by Bandcamp users. The themes of sexuality and racial mixing are absent from the reviews, and although one review preserves the theme of musical hybridity, it describes Dona Onete’s music as a “mix of quite raw sounding Brazilian Carimbo [sic] with some full production samba,” which is not a characterization that I have seen Dona Onete use for her music. Additionally, as on the Bandcamp album descriptions, the reviews make no specific reference to Pará. Instead, they reference the Amazon, for example calling Onete the “Musa da Amazônia” (‘Muse of Amazonia’), or Brazil as a whole, as in a review of Banzeiro calling it “one of the most energetic and fun albums Brazil has given us in the past 10 years.” Likewise, there are no references to Caboclos in the album reviews. Brazilian journalism describes carimbó chamegado fairly similarly. Multiple articles discuss the hybridity of carimbó chamegado, with one describing Dona Onete as making “carimbó, guitarrada [another genre practiced in Pará], samba” (Araujo 2012), another describing her as making “carimbó, lundu, siriá and other roots genres from the Brazilian North” (Wanderley 2016). Some articles also reaffirm the role of lyrics about sexuality, with one describing her as singing “subjects that go from love to religion, passing through the rural world” (Araujo 2012). However, like the Bandcamp descriptions and user reviews, Brazilian press coverage sometimes diverges from Dona Onete’s own characterization of her music. One writer describes her as a singer of boleros (Mans 2019). Although Dona Onete says that her music includes boleros, that is certainly not the only style of music she produces (Wanderley 2016). Another writer highlights Caribbean influences on Dona Onete’s music, contrasting the characterization established by Dona Onete herself (Muniz 2021). And many writers link Dona Onete’s music to the Amazon more broadly, rather than Pará specifically (Wanderley 2016, Gabriel 2018, Mans 2019). This contrasts with Dona Onete’s own attitude: she only says that she is from the Amazon (rather than Pará specifically) as a necessary concession to foreigners, who “mix up everything” (Wanderley 2016). Additionally, most articles do not mention caboclos whatsoever. Thus, as carimbó chamegado circulates, both its repertoire and its meaning are 3 The exception is “Homenagem de Dona Onete para Rainha Soberana da Amazônia.” The “sovereign queen” of song’s title is the Virgin Mary. Genre formation as enregisterment 297 changing. For some individuals, the repertoire has expanded to include Caribbean influences. Additionally, the lyrical theme of life in Pará has been replaced with a one of life in the Amazon more broadly. And for most individuals, the genre has lost its link to a caboclo persona, instead being associated with Amazonian-ness, or even with Brazilian-ness. 3 Discussion I have argued that genre formation is enregisterment; this in turn entails that genres are registers. In other words, genres are sociocultural linkages between a semiotic repertoire and certain social meanings, and are constructed and transmitted in a “speech” chain (which may include non-linguistic texts such as songs). I have illustrated this proposal through a brief presentation of a genre known as carimbó chamegado. As presented by Dona Onete, its main practitioner, the genre’s repertoire is characterized by instruments such as the saxophone and banjo; the fusion of different genres of Pará; and lyrics about sexuality and life in Pará. This repertoire is linked to a caboclo persona, associated with sexual desire, racial mixture, and Brazilian identity. The genre has changed as it has circulated. For some, its repertoire has expanded to include Caribbean influences. Furthermore, for most individuals, the genre is no longer connected to a caboclo persona, and is not associated with Pará in particular. Instead, it has come to be associated with the Amazon more broadly, or even with Brazil as a whole. The extension of enregisterment to musical genre has consequences for our understanding of enregisterment itself. As implied by the name, carimbó chamegado is a subgenre of carimbó (Gabriel 2018). But if a genre is a register, what is a subgenre? The relationship between genre and subgenre does not reduce to subsethood: both the repertoire and social meaning of the subgenre may differ from that of the “parent” genre. For example, carimbó chamegado differs from other carimbó in its lyrical focus on sexuality. Instead, the relation between a genre and its subgenre appears to be ideological. During the enregisterment of the subgenre, it is identified as an instantiation—another “level” (Brackett 2016)—of the parent genre, despite differences in form and meaning. We can refer to this process as nested enregisterment, since it nests one register within another, and refer to its outcomes as the sub-register (a subgenre, in the context of music) and the super-register. Nested enregisterment is not limited to musical genre. For example, consider the following discussion of accent identification (Wells 1982; in Agha 2003, 233) . . . a Liverpool working-class accent will strike a Chicagoan primarily as being British, a Glaswegian as being English, an English southerner as being northern, an English northerner as being Liverpudlian, and a Liverpudlian as being working-class. As Agha notes, none of the descriptions of the accent are wrong, but rather reflect different levels of geographical specificity. Under the present perspective, the same speech can belong to multiple registers, because one register can exist “within” another: working-class Liverpudlian English is a subregister of Liverpudlian English, 298 AIDAN MALANOSKI which is a subregister of northern English, which is a subregister of the English of England, which is a subregister of British English. The present example also motivates an extension of the notion of a speech chain. As discussed above, a speech chain is a “historical series of speech events” such that “the receiver of the message in the (n)th speech event is the sender of the message in the (n+1)th speech event” (Agha 2003, 247). As Agha discusses, this notion is complicated by the fact that receiver and sender are not individuals, but roles. A speech event may be produced by multiple individuals, and may be heard by multiple individuals—in other words, multiple individuals may occupy the sender role or receiver role of a single speech event. Even with this complication in mind, the idea of the speech chain must be extended to account for the present data. As argued above, musical texts form links in a speech chain. Most albums of music consist of multiple musical texts (i.e., multiple songs). Consequently, texts about an album—for example, the descriptions of Dona Onete’s albums on Bandcamp, or reviews of her albums in the media— simultaneously respond to multiple speech events. In the case of an album, these multiple speech events are effectively simultaneous (although some musical texts may have a prior life as singles), but this need not be the case. For example, one may argue that user reviews on Dona Onete’s Bandcamp page respond both to Dona Onete’s music, and to the descriptions of her music on the Bandcamp page—which themselves respond to Dona Onete’s music. And although multiple individuals may share the receiver role of a speech event, they might not share the sender role of a subsequent event. That is, a single speech event may spawn multiple responses, each with different individuals occupying the role of sender. The proliferation of senders and receivers is likely part of the reason that different individuals come to have different models of a register. Given these considerations, a speech chain cannot be a linear sequence of senders and receivers, even with the caveat that multiple individuals may occupy the role of sender or receiver. Instead, a speech chain is better represented as a directed acyclic graph, as in Figure 1. The nodes in the graph are speech events, and the edges (the lines connecting the nodes) represent the change of role from receiver to sender. Thus, the arrow from S1 to R1 indicates that a receiver of speech event S1 is the sender of speech event R1 . Given the nature of the response relationship—if speech event A is a response to speech event B (or a response to a response to B, etc.), then speech event B cannot be a response to A—the speech chain is acyclic. Additionally, I argued in section 1 that genre repertoires need not consist only of musical signs. Lyrics, linguistic features, fashion choices, even orthographic choices can figure in a genre’s repertoire. This was a prediction of the definition of a register. A register links social meaning to a semiotic repertoire; the repertoire need not be purely musical. Nor, of course, need it be purely linguistic. Consequently, the linguist must consider the possibility that non-linguistic features figure in a register’s repertoire. Without this awareness, they may fall into the trap of categorizing a speaker’s language based not on its own characteristics, but on the characteristics of the speaker. Finally, although the present paper has focused on the enregisterment of musical genre, the framework can be extended to other cultural forms, as discussed in section 1. For example, it is likely that styles of dance can also be enregistered Genre formation as enregisterment 299 S1 R1 C1 S2 R2 C2 … … … Figure 1: A representation of a hypothetical speech chain. To make the figure clearer, different line styles (solid, dotted, dashed) are used for edges from different nodes. The choice of line style has no other significance. (Ainoa Martínez Cuervo, p.c.). Recognizing this commonality among various cultural forms—they are all registers, formed through enregisterment—facilitates the transmission of knowledge across academic domains: what is true of one type of register (say, a linguistic register) may be true of another type (say, a musical genre). Thus, the extension of enregisterment started here may help development a holistic, interdisciplinary view of culture, society, and individual identity. References Adams, C., R. Murrieta, W. Neves, & M. Harris (eds.) 2009. Amazon Peasant Societies in a Changing Environment: Political Ecology, Invisibility and Modernity in the Rainforest. Springer. Agha, A. 2003. The social life of cultural value. Language and Communication 23. 231–273. Agha, A. 2005. Voice, footing, enregisterment. Journal of Linguistic Anthropology 15. 38–59. Araujo, B. 2012. Crítica: Dona Onete faz disco regional, graças a Deus. O Globo . https://oglobo.globo.com/cultura/critica-dona-onete-faz-disco-regional-gracas-deus5811700. Last accessed 2023-01-11. Blommaert, J., & P. Varis. 2011. Enough is enough: The heuristics of authenticity in superdiversity. Tilburg Papers in Culture Studies 2. 1–13. Born, G., & D. Hesmondhalgh. 2000. Introduction: On difference, representation, and appropriation in music. In Western music and its others: Difference, representation, and appropriation in music, ed. by G. Born & D. Hesmondhalgh, 1–58. Berkeley/Los Angeles/London: University of California Press. Brackett, D. 2016. Categorizing sound: Genre and twentieth-century popular music. Oakland: University of California Press. Brasil Música e Artes. 2019. Dona Onete se apresenta pela primeria vez na Womex. https://bma.org.br/2019/10/08/dona-onete-se-apresenta-pela-primeira-vez-na-womex/. Last accessed 2021-04-13. Carmichael, K., & N. Dajko. 2016. Ain’t dere no more: New Orleans language and local nostalgia in Vic & Nat’ly comics. Journal of Linguistic Anthropology 26. 234–258. D’Onofrio, A. 2020. Personae in sociolinguistic variation. WIREs Cognitive Science 11. Duncan, D. 2017. Australian singer, American features: Performing authenticity in country music. Language & Communication 52. 31–44. 300 AIDAN MALANOSKI Ferreira, T. 2016. Dona Onete e o vulto permanente de seu carimbó. Na Mira do Groove . https://namiradogroove.com.br/blog/destaque/ dona-onete-banzeiro. Last accessed 2023-01-15. Fornäs, J. 1995. The future of rock: Discourses that struggle to define a genre. Popular Music 14. 111–125. Gabriel, J. 2018. Rainha do carimbó, Dona Onete abre o jogo sobre sua trajetória na música paraense. A Crítica . G.Lab. 2021. Dona Onete: “Me perguntam como posso ser compositora, cantora e poeta. Eu não sei, misturo tudo e deixo virar sopa”. O Globo . Gunn, J. 1999. Gothic music and the inevitability of genre. Popular Music and Society 23. 31–50. Johnstone, B. 2016. Enregisterment: How linguistic items become linked with ways of speaking. Language and Linguistics Compass 10. 632–643. Lamen, D. 2013. Claiming Caribbeanness in the Brazilian Amazon: Lambada, critical cosmopolitanism, and the creation of an alternative Amazon. Latin American Music Review 34. 131–161. Lena, J. C. 2012. Banding together: How communities create genres in popular music. Princeton/Woodstock: Princeton University Press. Lisansky, J. 2008. Caboclo. In Encyclopedia of Latin American History and Culture, ed. by J. Kinsbruner & E. D. Langer, volume 2, p. 3. Charles Scribner’s Sons, 2 edition. Mais Um Discos. 2017. Banzeiro | Dona Onete. https://donaonete.bandcamp.com/ album/banzeiro. Mais Um Discos. 2019. Rebujo | Dona Onete. Mais Um Discos. 2020. Feitiço Caboclo (2020 remaster) | Dona Onete. https://donaonete.bandcamp.com/album/feiti-o-caboclo-2020-remaster. Mans, M. 2019. Crítica: ‘Amazônia Groove’ é mergulho na música regional brasileira. Esquina da Cultura . https://www.esquinadacultura.com.br/post/critica-amazonia-groove-e-mergulhona-musica-regional-brasileira. Last accessed 2023-01-11. Moore, A. F. 2001. Categorical conventions in music discourse: Style and genre. Music & Letters 82. 432–442. Muniz, E. 2021. “Para nós, tudo é rio”. Continente . Natura. 2019. Rebujo, de Dona Onete, dá tom pop a ritmos tradicionais do Pará. Nugent, S. 1997. The coordinates of identity in Amazonia: At play in the fields of culture. Critique of Anthropology 17. 33–51. Onete, D. 2012. Feitiço Caboclo [Album]. Na Music [dist. Tratore]. Onete, D. 2017. Banzeiro [Album]. Mais Um Discos. Onete, D. 2019. Rebujo [Album]. Mais Um Discos. Rentfrow, P. J., & S. D. Gosling. 2007. The content and validity of music-genre stereotypes among college students. Psychology of Music 35. 306–326. Rockwell, J. 2012. What is bluegrass anyway? Category formation, debate, and the framing of musical genre. Popular Music 31. 363–381. Santos, R. V., P. H. Fry, S. Monteiro, M. C. Maio, J. C. Rodrigues, L. Bastos-Rodrigues, & S. D. J. Pena. 2009. Color, race, and genomic ancestry in Brazil: Dialogues between anthropology and genetics. Current Anthropology 50. 787–819. Shevy, M. 2008. Music genre as cognitive schema: extramusical associations with country and hip-hop music. Psychology of Music 36. 477–498. Trudgill, P. 1983. On Dialect: Social and Geographical Perspectives, chapter Acts of conflicting identity: The sociolinguistics of British pop-song pronunciation, 141–160. Oxford/New York: Basil Blackwell and New York University Press. Wanderley, F. 2016. Dona Onete: a história da a ex-professora que ‘ensinou’ carimbó aos gringos. A Crítica . Wells, J. C. 1982. Accents of English. Cambridge: Cambridge University Press. Proceedings of CLS 58 (2022), 301-315 © Chicago Linguistic Society 2023. All rights reserved. 301 Two strategies for PP-fronting from certain island environments in English Andrew McInnerney University of Michigan 1 Introduction The sentences in (1) have a surface structure that suggests topicalization of PP. (1) a. Of the employees, several were fired. b. In your parents, I wonder how much trust you have left. The labeled bracketing in (2) below illustrates relevant aspects of this analysis.1 In (2a), the initial PP of the employees may appear to have been fronted from a base position within the bracketed subject NP. Similarly, the fronted PP in (2b) in your parents may appear to have been fronted from within the bracketed wh-phrase of the embedded clause. (2) a. [PP Of the employees], [ [NP several __ ] were fired]. b. [PP In your parents], I wonder [ [NP how much trust __ ] you have left]. Yet, if the PP-fronting in sentences like (2) is a type of movement, this would be puzzling, for at least two reasons. First, this movement would be an acceptable case of pied-piping (with the analogous p-stranding derivations sharply degraded, as we will see), highly unexpected given the general pattern for English. Second, the apparent extraction sites in these sentences are contained within island domains. In (2a), the gap site is part of a subject, and in (2b), it is part of a fronted wh-phrase; both domains are typically opaque for extraction. There are two analytical options in response to these problems. Either (i) there is something exceptional about the constructions in (1) which reverses patterns obtaining in most other extraction paradigms, or (ii) the constructions in (1) involve some derivation(s) other than that depicted in (2). I pursue the second option here, according to the following outline. First, in section 2, I expand on the problems confronting (2), considering pied-piping and p-stranding in other A-bar movement constructions (beyond topicalization shown in (1), including violations of island constraints). Then, in section 3, I develop an analysis involving base-generation of PP, which accounts for a subset of the relevant data. Section 4 develops an alternative analysis based on extraposition of PP, which also accounts for a subset of the relevant data. Section 5 provides evidence that both analyses are needed. I consider consequences for the analysis of subject islands, focusing especially on the discourse-pragmatic approach of Abeillé et al. (2020) in section 6. 1 In my notation, underscores indicate the canonical position of the boldfaced (displaced) element. 302 ANDREW MCINNERNEY 2 P-stranding and pied-piping in English Both p-stranding and pied-piping (of PP)2 occur in written English. However, piedpiping is well-known to have a much more restricted distribution than p-stranding in English. In most circumstances, p-stranding is more acceptable than analogous pied-piping. To take a classic case, consider p-stranding vs. pied-piping in embedded wh-questions. Kayne (1994, p25) gives the following judgments. The analogous examples with p-stranding are all perfectly natural. (3) a. b. c. d. e. * We want to know [ [PP about what] you’re thinking __ ]. * Tell me [ [PP at who(m)] you were looking __ ]. * Nobody knows [ [PP to what school] he goes __ ]. * I can’t figure out [ [PP for which kid] you bought it __ ]. * Tell us [ [PP from where] you got that __ ]. In some cases, in particularly formal registers, pied-piping in matrix wh-questions (and certain relative clauses) can sometimes appear to be possible. Even in those cases where pied-piping appears possible, p-stranding is typically the more natural option. Consider (4), for instance. Here, ‘>’ indicates that (4a) is more acceptable than (4b). (4) a. What topic are they thinking [PP about __ ]]? b. > [PP About what topic] are they thinking __ ]. The preference for p-stranding is not restricted to wh-questions. Examples (5)(7) illustrate the preference in restrictive relatives, free relatives, and it-clefts respectively. (5) a. That’s the topic [ which they’re thinking [PP about __ ]]. b. > That’s the topic [ [PP about which] they’re thinking __ ]. (6) a. [Whichever topic they’re thinking [PP about __ ]], it doesn’t matter. b. > [ [PP About whichever topic ] they’re thinking __ ], it doesn’t matter. (7) a. It’s this topic (not that one) [which they’re thinking [PP about __ ]]. b. > It’s this topic (not that one) [ [PP about which] they’re thinking __ ]. See Cable & Harris (2011), Günther (2021), McInnerney & Sedarous (2023) for experimental corroboration of the general preference for p-stranding in English. Pied-piping is also thought to be restricted even more tightly in certain contexts in English. For example, PPs resist Across-the-Board extraction and fail to license parasitic gaps (Cinque 1990, Munn 1992, 1993). Crucially for this paper, one of the relevant restricted contexts involves islands. A widespread characterization of weak islands refers to categorial asymmetries in extraction. Hence, while (some) nominals can (sometimes - perhaps marginally) 2 Generally speaking, the term pied-piping covers more than just PP-extraction (extraction with whose is pied-piping, for instance). For the purposes of this paper, however, pied-piping abbreviates pied-piping of PP. Two strategies for PP-fronting 303 be extracted from weak islands, PPs cannot be. I illustrate here with relativization in several types of weak islands,3 including infinitival relative clauses, indefinite complex NPs, and gerundive ‘adjunct’ clauses. (8) (9) (10) a. ? That’s the grant [ which we need to find someone to apply for __ ]] b. * That’s the grant [ for which we need to find someone to apply __ ]] a. ? b. * That’s the grant [ for which I heard rumors that you’re applying __ ]] a. ? b. * That’s the grant [ for which they upset their advisor by applying __ ]] That’s the grant [ which I heard rumors that you’re applying for __ ]] That’s the grant [ which they upset their advisor by applying for __ ]] We have thus established two descriptive generalizations for p-stranding vs. pied-piping in English. First, p-stranding is preferred to pied-piping virtually across the board. Second, nominal-extraction from islands is sometimes possible, but PPextraction from islands remains degraded even when nominal-extraction is possible. The data of (1) clash with both generalizations, at least under the analysis sketched in (2). Compare (apparent) pied-piping with p-stranding in these contexts. (11) a. b. Of the employees, several were fired. * The employees, several of were fired. (12) a. b. In your parents, I wonder how much trust you have left. * Your parents, I wonder how much trust in you have left. If (11a) and (12a) involve pied-piping (as in (2)), then these examples would represent constructions where (i) pied-piping is preferable to p-stranding, and (ii) PP-extraction from an island is possible while nominal-extraction is not. I take these conflicts to call for an alternative analysis of (11)/(12). I propose that (11a) and (12a) each involve a distinct, non-island-violating derivation consistent with the descriptive generalizations adduced in this section. I develop these two distinct analyses in turn in sections 3 and 4. 3 A base-generation analysis I will now outline a non-island-violating analysis of sentences like (11a). The central idea is that fronted of -phrases can be base-generated in initial position. The PP-fronting in (11a) and similar sentences can thus be accomplished via basegeneration, with no need for potentially island-violating pied-piping. Evidence for an analysis along these lines comes from Akmajian & Lehrer (1976), who identify examples with fronted of -phrases where there is simply no possible extraction site lower in the clause. They offer the sentences (13a) and (13b). Further similar examples can easily be constructed; I add (13c) and (13d). 3 The pattern is quite general. See Cinque (1990), Szabolcsi (2008), among others, for further examples and discussion. 304 ANDREW MCINNERNEY (13) a. b. c. d. Of the stories about Watergate (that have so far appeared), only yesterday’s was truly shocking. Of the current committee members, only the chair wants to vote against the proposal. Of all the competitors, I can’t believe it was me who won. Of cars on display, only the newest of the corvettes still works. In their words (p401), ‘the point to notice about these examples is that there is no position within the main clause from which the of -phrase could have been extracted.’ By analogy with (11a), one might consider whether the of -phrases could have been extracted from within the grammatical subjects (e.g. yesterday’s). But this would lead to ungrammaticality, as (14) shows. (14) a. b. c. d. * Only yesterday’s of the stories about Watergate (that have so far appeared) was truly shocking. * Only the chair of the current committee members wants to vote against the proposal. * I can’t believe it was me of all the competitors who won. * Only the newest of the corvettes of all the cars on display still works. The fronted of -phrases in (13) occupy a high topic position. The ‘topical’ nature of the position can be seen from the incompatibility of base-generation with focus. A base-generated of -phrase cannot be targeted for A-bar movement in whquestions. I assume this is because the focus requirements that would be imposed on the questioned of -phrase would be incompatible with its topicality. (15) a. b. * Of which stories about Watergate was only yesterday’s truly shocking? * Of which committee members does only the chair want to vote against the proposal? At the same time, a base-generated of -phrase can be targeted for A-bar movement in relativization. This is because, unlike wh-phrases, the extracted constituent in a relative clause does not necessarily have a focus interpretation. (16) a. b. The stories about Watergate, [ of which only yesterday’s was truly shocking], were somewhat uninformative. The committee members, [ of whom only the chair wants to vote against the proposal], all have ulterior motives. These facts suggest an analysis of (11a) along the lines of (17).4 4 Many fine-grained aspects of the structure are irrelevant for present concerns and are therefore omitted from the tree diagram here. Two strategies for PP-fronting (17) 305 TopP TP PP of the employees T0 DP several T VP were fired What’s crucial is that the TopP projection is to be formed by External Merge, not Internal Merge. 4 An extraposition analysis The base-generation analysis developed in the previous section is not sufficient on its own to account for the full range of data represented by (1). (17) shows how the of -phrase in (1a) could be base-generated, but what about (1b), repeated here? (18) In your parents, I wonder how much trust you have left. This example has two properties which make a base-generation analysis inappropriate. Before discussing those properties, however, I should clarify that the base-generation analysis could apply to a sentence like (19). (19) Of the students, I wonder [ [how many __ ] will pass the test]. The fronted PPs in the examples discussed in section 3, like (17), involve apparent (and only apparent) extraction from subject. However, the base-generation analysis is also appropriate for examples like (19), involving apparent extraction from a fronted wh-phrase. Example (18) also involves apparent extraction from a fronted wh-phrase. That is not where the problems lie. Instead, one problem is that the initial PP of (18) can occur in focus positions. Specifically, it can be targeted for wh-movement. (20) In whose parents do you wonder how much trust you should have? If the initial PP were base generated high here, then we would not expect it to be able to be targeted for wh-movement, as discussed in section 3. Second, the fronted PP is in a relationship of L-selection with the noun trust.5 The L-selection relation is syntactically very local, typically holding between a lexical head (e.g. a noun or a verb) and the head of its complement (in this case, the P head of a complement PP). If the initial PP in your parents were base-generated high, then the local configuration between trust and in would not be able to be established. 5 I.e., ‘Lexical’ selection (Pesetsky 1995). This is another way of saying that the noun-preposition combination trust in is idiomatic. 306 ANDREW MCINNERNEY These same concerns also arise in certain sentences with apparent pied-piping from subject, as in (21) from Zyman (2021). Note that the head of the fronted PPs are L-selected by the subject nouns (obedience in (21a), and extraction in (21b)). (21) a. To whom is obedience demanded by the law? b. From which islands is extraction most readily tolerated? An alternative analysis (other than the base-generation analysis discussed in section 3) is therefore needed for these data. Because of the L-selection relation that can hold in these sentences, I assume that the fronted PPs in sentences like (21) are externally merged with the selecting subject nouns (or with a selecting whfronted noun in sentences like (20)). However, instead of being fronted by Internal Merge directly from this base position, the PPs are first extraposed. The derivation of, e.g., (21) would not necessarily need to proceed along the lines of (22), but could proceed as in (23) instead. (22) CP PP1 to whom C TP is DP obedience T PP1 VP V demanded (23) DP obedience PP1 CP PP1 to whom C TP is DP obedience PP1 T VP PP1 V demanded DP obedience PP1 Two strategies for PP-fronting 307 A digression is required at this point to explain a potential contradiction in reasoning. One of the major points of section 2 was that p-stranding is generally preferred to pied-piping in English. This is central to the problematization of analyses like (22), because corresponding p-stranding derivations lead to ungrammaticality. Yet, the analysis (23) would also be ungrammatical with p-stranding. (24) * Who is obedience demanded by the law to? Shouldn’t (23) then be problematic for the same reasons as (22), i.e. as another exception to the general preference for p-stranding? Here’s the critical difference between (23) and (22) concerning the p-stranding preference. While the internal constituents of the PP are inaccessible in both cases, the PP itself is accessible in (23) but not in (22). In (22), the internal constituents of the PP are inaccessible because the whole PP itself is inaccessible, being contained within an island. But in (23), the internal constituents of the PP are inaccessible because the PP is frozen; the PP itself remains accessible. I assume that the p-stranding preference holds only when p-stranding and pied-piping are equally viable options. When p-stranding is independently blocked, pied-piping can be more acceptable if it remains unaffected. For example, p-stranding is sometimes lexically blocked (i.e. certain prepositions don’t allow extraction of their complements). In such cases, pied-piping is possible (see Stanton 2016 for discussion of potential mechanisms for such effects). (25) a. I don’t know [ via which path to proceed __ ]. b. I don’t know [ with respect to which problem to make a decision __ ]. Pied-piping is required, appearing to subvert the p-stranding preference, in cases like (23) because p-stranding is made independently unavailable by the well-known freezing effect. 5 Ambiguity and disambiguation There are thus two ways of generating the sentences in (26) which are consistent with both (i) the overall preference for p-stranding over pied-piping in English and (ii) the categorial asymmetries of extraction from islands (PP-extraction being generally blocked). For (26a), the fronted PP is base-generated in topic position. For (26b), the fronted PP is first extraposed, then extracted from extraposed position. (26) a. Of the employees, several were fired. b. In your parents, I wonder how much trust you have left. These examples were chosen precisely to disambiguate between the two different analyses. The base-generation analysis applies to (26a), but it could not apply to (26b) because the fronted PP is L-selected by the noun trust. The extraposition 308 ANDREW MCINNERNEY analysis applies to (26b), but it could not apply to (26a) because the of -phrase in that example cannot be extraposed (consider *Several were fired of the employees). Both analyses are therefore needed: the base-generation analysis to account for (26a), and the extraposition analysis to account for (26b). However, there are many cases where both options would in principle be available. In this section, I provide evidence supporting a dual analysis (assuming both the base-generation analysis and the extraposition analaysis) by illustrating various ways of disambiguating between the two options. For example, we expect that, if extraposition is unavailable, then L-selection and wh-movement of the fronted PP should be unavailable (because these are not compatible with the base-generation analysis). Similarly, we expect that, if basegeneration is unavailable, then only those PPs that can be extraposed will be able to be fronted. These predictions appear to be correct. To illustrate, first consider the case of self-embedded nominals. Example (27) shows two nominals, each contained within a larger nominal. In (27a), the embedded nominal contains an of phrase of the children. In (27b), the embedded nominal contains an L-selected PP to the children. (27) a. [Several of the best behaved of the children] are quite well-liked. b. [Several of the longest letters to the children] have been stolen. The underlined PPs cannot be extraposed, as (28) shows. Therefore, the extraposition analysis is not available for fronting of these PPs. (28) a. b. * [Several of the best behaved] are quite well-liked of the children. * [Several of the longest letters] have been stolen to the children. Even though the extraposition analysis is unavailable, the base-generation analysis is still available for fronting of the of -phrase in (27a). Thus, (29a) is acceptable. However, the base-generation analysis cannot apply for fronting of the L-selected PP in (27b). Thus, (29b) is unacceptable. (29) a. b. Of the children, [several of the best behaved] are quite well-liked. * To the children, [several of the longest letters] have been stolen. Next, consider the case of partitive vs. non-partitive structures. Example (30a) gives a partitive subject several of the employees, while example (30b) gives a nonpartitive subject several pictures of the employees. (30) a. [Several of the employees] were fired. b. [Several pictures of the employees] were put up. Extraposition of the of -phrase from a partitive is not possible in English (Selkirk 1977). However, extraposition of of -phrases from non-partitive nominals like the boldfaced one in (30b) is sometimes possible. (31) a. b. * [Several] were fired of the employees. [Several pictures] were put up of the employees. Two strategies for PP-fronting 309 Given this, the extraposition analysis is applicable for fronting of -phrases in non-partitive structures like (30b), but not for partitives as in (30a). Therefore, we expect that a fronted of -phrase with a non-partitive nominal should be able to be targeted for wh-movement (since the extraposition analysis is applicable), but a fronted of -phrase with a partitive should not be able to be wh-moved (since the extraposition analysis is not applicable). This is correct, as (32) illustrates. (32) a. b. * Of which employees were [several] fired? Of which employees were [several pictures] put up? It is also well-known that the availability of extraposition from subjects is (broadly) correlated with transitivity. PPs can be extraposed from the subjects of passives and unaccusatives, but typically cannot be extraposed from the subjects of unergatives and active transitives (see e.g. Johnson 1985). Examples (33) and (34) are illustrative of the general pattern. In (33), extraposition from the subject of passive be proposed is preferred over extraposition from the subject of active transitive save. In (34), extraposition from the subject of unaccusative arrive is preferred over extraposition from the subject of unergative resign. (33) a. b. [A solution] will soon be proposed to this problem. * [A solution] will save us all to this problem. (34) a. b. [The governor] will soon arrive of the state of California. * [The governor] will soon resign of the state of California. Similar contrasts have been noted to play a role in acceptability of wh-extraction from subjects, as well. Notably, passives and unaccusatives are claimed to allow extraction from their subjects, while unergatives and active transitives are argued not to (Chomsky 2008, Polinsky et al. 2013, Haegeman et al. 2014, Zyman 2021). (35) a. b. ? To which problem will [a solution] soon be proposed? * To which problem will [a solution] save us all? (36) a. b. ? Of which US state will [the governor] soon arrive? * Of which US state will [the governor] soon resign? These contrasts are expected under the analysis proposed in this paper. Whextraction is not compatible with the base-generation analysis, meaning sentences like (35) and (36) must be derived via the extraposition analysis. Wh-extraction is therefore only possible in contexts where the necessary extraposition is also possible (i.e. in passive (35a) and (36a)). In fact, the correlatiom between transitivity and extraposition is not exact. There are some circumstances in which extraposition from the subject of a transitive/ unergative predicate is possible. Johnson (1985, p110), drawing on Guéron (1980), for example, notes the case of ‘presentational’ predicates, in which ‘the subject is the unmarked focus.’ Unergative predicates tend to allow extraposition from their subjects if they are ‘presentational’ in this sense. Consider that the nuclear pitch accent (indicated with small caps) falls on the subject in (37a), but on the verb phrase in (37b). I take this to indicate that (37a) is presentational but (37b) is not. 310 (37) ANDREW MCINNERNEY a. [The governor of the state of C ALIFORNIA] [just walked by]. b. [The governor of the state of California] [just VOMITED]. Thus, extraposition from subject is possible for (37a), but not for (37b). This is illustrated in (38) (38) a. b. ? The governor just walked by of the state of California. * The governor just vomited of the state of California. The analysis developed in this paper predicts the patterns of extraposition to feed wh-extraction. PP-fronting under wh-movement should be possible for subjects of presentational predicates like (38a), but not for non-presentational predicates like (38b). Example (39) illustrates this prediction. (39) a. b. ? Of which US state did the governor just walk by? * Of which US state did the governor just vomit? As predicted (because the base-generation analysis is incompatible with whquestions), only the extraposable PP of (39a) can be targeted for wh-movement. The non-extraposable PP of (39b) cannot be fronted, because it cannot be extraposed. 6 Consequences of the dual analysis The previous section discussed a range of examples suggesting that both the basegeneration analysis of section 3 and the extraposition analysis of section 4 are needed to account for apparent PP-fronting from subjects in English. When extraposition is not available, the fronted PP must be an of -phrase, and cannot be targeted for wh-movement, indicating that base-generation remains available as an option. When base-generation is not available, the fronted PP must be able to be extraposed, suggesting that extraposition remains available as an option. I will now discuss some consequences of the conclusions drawn here. First, the arguments here conflict with claims of Abeillé et al. (2020). They conducted a series of acceptability-judgment surveys showing that English speakers consistently judge sentences such as (40a) to be more acceptable than sentences such as (40b), with this difference disappearing (even reversing) for pairs like (41), with relatively (41a) more acceptable than (41b). (40) a. Of which car did the color delight the baseball player because of its luminance? b. < Of which car did the baseball player love the color because of its luminance? (41) a. The dealer sold a sportscar, of which the color delighted the baseball player because of its luminance. b. > The dealer sold a sportscar, of which the baseball player loved the color because of its luminance. Two strategies for PP-fronting 311 Abeillé et al. interpret these results as indicative of a subject island effect (SIE) in wh-question formation, and a lack of an SIE in relativization. According to them, this motivates a reduction of the SIE to discourse-pragmatic principles. In particular, they advocate what they call the ‘Focus-Background-Conflict Constraint’ (FBC), as in (42). (42) The Focus-Background-Confict Constraint: A focused element should not be part of a backgrounded constituent. The idea is that (40a) is out because (i) subjects are backgrounded and (ii) whfronting in wh-questions places focus on the fronted constituent; (40a) thus requires focus within the (backgrounded) subject, violating the FBC. On the other hand, (40b) is possible because, although a constituent is fronted from within the backgrounded subject, this fronted constituent is not necessarily focused. As I mentioned, Abeillé et al. imply that this pragmatically-based analysis invalidates previous work placing more emphasis on the syntactic properties of the SIE. I quote from their p12: . . . the attempt to explain constraints on extraction via syntactic principles may be on the wrong track, and . . . more general discourse coherence principles may explain the differences between constructions. Once the discourse function of these constructions is taken into account, these constraints may fall under more general cognitive principles and become learnable. See McInnerney & Sugimoto (2022) and Kobzeba et al. (2022) for some critical discussion of these conclusions. Relevant for present purposes, all of Abeillé et al.’s critical English items used fronted of -phrases. I have argued in section 3 of this paper that fronting of of -phrases in relative clauses like (41a) can be achieved via base-generation, not requiring extraction from subject at all. I also argued that this base-generation strategy fails to generate fronting of of -phrases in wh-questions like (40b). Importantly, the extraposition analysis for PP-fronting (section 4) is not available for (40)/(41), because the verbs are active transitives (as were all the critical English items in Abeillé et al.’s stimuli). As discussed, active transitives resist extraposition from subject, making base-generation the only option to derive (40)/(41). As such, under the analysis of PP-fronting in this paper, the results represented by (40)/(41) are entirely consistent with a syntactic approach to the SIE which rejects the FBC. This would further explain Abeillé et al.’s finding that any suggestion of the FBC disappears under p-stranding. As illustrated by (43)/(44), p-stranding from subject is worse than p-stranding from object in both relative clauses and wh-questions. This could be because p-stranding as in (44) necessarily involves genuine extraction, while the apparent pied-piping in (41) can be derived with basegeneration of the of -phrase. (43) a. Which car did the color of delight the baseball player because of its luminance? 312 ANDREW MCINNERNEY b. < Which car did the baseball player love the color of because of its luminance? (44) a. The dealer sold a sportscar, which the color of delighted the baseball player because of its luminance. b. < The dealer sold a sportscar, which the baseball player loved the color of because of its luminance. The results of this paper thus cast doubt on Abeillé et al.’s discourse-pragmatic account of the SIE. To pivot to a different topic, recent years have seen efforts to reduce the SIE to an interaction of independent syntactic factors. Haegeman et al. (2014; see also Greco et al. 2017), for example, argue that eight distinct conditions factor into the SIE. One relevant factor is freezing, with extraction from moved constituents being less acceptable than extraction from constituents in-situ. Also, extraction from constituents at a phase edge is degraded, as is extraction from definite nominals. However, another of Haegeman et al.’s conditions is p-stranding. This condition is particularly important for their analysis, because the contributions of the independent principles they identify is much weaker when p-stranding is used than when pied-piping is used. Their characterization (p119-120) is that what they call the Preposition Stranding Condition (requiring, in English, that no preposition can be stranded inside a moved constituent) ‘is a violable constraint albeit relatively strong in English.’ Under the analysis advocated in this paper, critical examples with pied-piping may not involve genuine extraction from subject at all. Given that subjects seem to be (at least) weak islands in English, it would be highly surprising if a whole set of independent syntactic conditions could each cause significant swings in acceptability of pied-piping. Instead, it would be more likely for relevant conditions to each have a relatively small effect on such extractions, which would already be sharply degraded simply by virtue of involving pied-piping from an island. Therefore, in some cases, it may be necessary to reinterpret syntactic conditions as relevant for extraposition. For example, the edge condition, originally proposed by Chomsky (2008), holding that subextraction from constituents at phase edges is blocked, might be reinterpreted as holding that extraposition from phase edges is blocked.6 Alternatively, it may be necessary to assume a more minor role for proposed conditions, considering the effect relevant conditions have in the case of p-stranding. Haegeman et al. argue that the edge condition has a detectable but ultimately very subtle effect under p-stranding. It might be that this relatively minor effect under p-stranding (along with the other relatively minor effects of proposed conditions 6 Though, this condition would face a host of major that I cannot go into in detail here. It is widely believed, for example, that constituents can be extraposed from moved wh-phrases, as in (i), which would be unexpected given that Spec-CP is a phase edge. (i) [How many people] did you hire who you already knew? Two strategies for PP-fronting 313 under p-stranding) also characterizes the effect under pied-piping, with the true effect under pied-piping obscured by possible alternative derivations involving basegeneration or extraposition. 7 Conclusion I argued in this paper that the apparently island-violating pied-piping in sentences like (45) is illusory. (45) a. Of the employees, several were fired. b. In your parents, I wonder how much trust you have left. Instead of involving PP-extraction from within an island, in each of these sentences an alternative derivation is available which obeys conventional island constraints. In (45a), the PP can be base-generated as an initial topic. In (45b), the PP can be extraposed from subject, and then targeted from that island-external extraposed position. In some cases, both analyses may be applicable, as discussed in section 5. I discussed the consequences of this analysis for prior accounts of the Subject Island Effect (SIE), particularly those of Abeillé et al. (2020) and Haegeman et al. (2014). The consequences for other analyses of the SIE (e.g. Chomsky 2008, Zyman 2021, Bianchi & Chesi 2014, among others) should be considered as well, though I will not do so here. I will conclude with some general speculations on the SIE. The factors giving rise to the effect have been mysterious ever since the effect’s discovery; at no point has there been a satisfactory account of both (i) the general pattern of unnacceptability for extraction from subject and (ii) the vast range of exceptions to this general pattern. However, ever since Ross (1967), where acceptable (apparent) pied-piping from subject was discussed in the main text (p242), with unacceptable analogous pstranding unacceptable discussed in a footnote (p265), work on the SIE has tended to focus primarily on pied-piping over p-stranding. Given the results discussed here, I would argue instead that, where English is concerned, examples with p-stranding may in fact be more informative. References Abeillé, A., B. Hemforth, E. Winckela, & E. Gibson. 2020. Extraction from subjects: Differences in acceptability depend on the discourse function of the construction. Cognition p. 204. Akmajian, A., & A. Lehrer. 1976. NP-like quantifiers and the problem of determining the head of an NP. Linguistic Analysis 2. 395–413. Bianchi, V., & C. Chesi. 2014. Subject islands, reconstruction, and the flow of the computation. Linguistic Inquiry 45. 525–569. Cable, S., & J. A. Harris. 2011. University of massachusetts occasional papers in linguistics: Processing linguistic structure. In University of Massachusetts Occasional Papers in Linguistics: Processing Linguistic Structure, ed. by M. Grant & J. A. Harris, volume 38, 1–22. GLSA Publishing. 314 ANDREW MCINNERNEY Chomsky, N. 2008. On phases. In Foundational issues in linguistic theory: Essays in honor of Jean-Roger Vergnaud, ed. by R. Freidin, C. P. Otero, , & M. L. Zubizarreta, 133–166. Cambridge, MA: MIT Press. Cinque, G. 1990. Types of Ā-dependencies. MIT press Cambridge, MA. Greco, C., M. Marelli, & L. Haegeman. 2017. Deconstructing the Subject Condition in terms of cumulative constraint violation. The Linguistic Review 34. 479–531. Guéron, J. 1980. On the syntax and semantics of PP extraposition. Linguistic Inquiry 11. 637–678. Günther, C. 2021. Preposition Stranding vs. Pied-Piping—The Role of Cognitive Complexity in Grammatical Variation. Languages 6. 89. Haegeman, L., Ángel L. Jiménez-Fernández, & A. Radford. 2014. Deconstructing the Subject Condition in terms of cumulative constraint violation. . The Linguistic Review 31. 73–150. Johnson, K. 1985. A case for movement. Doctoral Dissertation, MIT. Kayne, R. S. 1994. The Antisymmetry of Syntax. Cambridge, MA: MIT Press. Kobzeba, A., C. Sant, P. T. Robbins, M. Vos, T. Lohndal, & D. Kush. 2022. Comparing island effects for different dependency types in Norwegian. Languages 7. 197. McInnerney, A., & Y. Sedarous. 2023. Effects of extractee category in adjunct-island strength: Nominals vs. Prepositional Phrases. Poster presented at the 97th annual meeting of the Linguistic Society of America. McInnerney, A., & Y. Sugimoto. 2022. On dissociating adjunct island and subject island effects. Proceedings of the Linguistic Society of America 7. 5207. Munn, A. 1992. A null operator analysis of ATB gaps. The Linguistic Review 9. 1–26. Munn, A. 1993. Topics in the syntax and semantics of coordinate structures. Doctoral Dissertation. University of Maryland, College Park. Pesetsky, D. 1995. Zero Syntax: Experiencers and Cascades. Cambridge, MA: MIT Press. Ross, J. R. 1967. Constraints on variables in syntax. Doctoral Dissertation, MIT. Selkirk, E. 1977. Some remarks on noun phrase structure. In Formal Syntax, ed. by P. W. Culicover, T. Wasow, & A. Akmajian, 285–316. New York: Academic Press. Stanton, J. 2016. Wholesale Late Merger in Ā-Movement: Evidence from Preposition Stranding. Linguistic Inquiry 47. 89–126. Szabolcsi, A. 2008. Strong vs. weak islands. In The Blackwell Companion the Syntax, Vol. 4, ed. by M. Everaert & H. van Riemsdijk, 479–531. Oxford: Blackwell. Zyman, E. 2021. Antilocality at the phase edge. Syntax 24. 510–556. Proceedings of CLS 58 (2022), 315-328 © Chicago Linguistic Society 2023. All rights reserved. 315 Predicate landing sites in verb-initial languages John Middleton University of Auckland 1 Introduction1 The derivation of verb-initial word order, either through V-movement or VPmovement, has raised the important question of where the landing site for the moved constituent is. For predicate-fronting languages, an early proposal was that the EPP feature on T˚ differed between SVO and V-initial languages (Massam and Smallwood 1997; Otsuka 2005). This was claimed to be a parametric contrast: in subject-initial languages, the feature is [+D], while in predicate-initial languages, the feature is [+pred]. The result is that in predicate-initial languages, the predicate raises to SpecTP to check the EPP[+pred] feature, leading to the verb-initial order. However, pre-verbal particles such as tense/aspect/modal (TAM), pre-verbal pronouns and negation have led other authors to reject SpecTP and propose other landing sites, such as SpecFP or SpecFinP. Doner (2021) claims TAM sits in T˚ in Tongan (thus below SpecTP), while Collins (2017) argues that pre-verbal pronouns raise to SpecTP, and NegP is dominated by TP, in Samoan. As TAM, pre-verbal pronouns and negation all precede the predicate, both analyses are forced to consider a new projection (FP) below TP (and NegP) for the landing site of the predicate. Meanwhile, Massam (2020) proposes that TAM and negation are generated in the high left periphery in Niuean, and the predicate raises to SpecFinP, in the low end of the left periphery. In these alternative models, the projection claimed to host the predicate lacks any overt head, and there is no explanation as to why the EPP feature resides on different heads in different languages. Tokelauan, a predicate-raising Polynesian language (Middleton and Syed 2022), also has pre-verbal TAM, subject pronouns and negation. However, unlike previous proposals, these may be accounted for without changing the EPP on T˚ analysis of predicate-movement. The specific claim is that negation and pre-verbal pronouns are clitics which attach to TAM, which thereafter raises to the left periphery. As such, this paper presents evidence that the predicate raises to SpecTP, without the need for ad hoc functional projections. This paper is ordered as follows. Section 2 introduces the concept of the EPP[+pred] feature, while the following subsections show how pre-verbal particles in several Polynesian languages derail the SpecTP landing site hypothesis. 1 This work could not have been undertaken without the generous help of Iutana Pue, whom I thank for sharing his language with me. I would also like to thank the audience of CLS 58 for their helpful comments with this research. Data is in Tokelauan, unless otherwise marked. Abbreviations used in the data follow the Leipzig Glossing Rules. Additional abbreviations include: ANP=anaphoric particle; CIA=agentive verbal suffix; DIR =directional particle; INT=intensifier; TAM=tense/aspect/modal particle. 316 JOHN MIDDLETON Section 3 examines Tokelauan and demonstrates that each of the pre-verbal particles can be accounted for without blocking SpecTP as the predicate-landing site. Section 4 concludes. 2 Predicate landing sites in other languages The EPP feature was originally suggested to trigger the obligatory raising of the subject DP to SpecTP in English (Chomsky 1982). The uninterpretable EPP[+D] feature resides on T˚, and probes for the closest constituent with a [+D] feature. This means the subject, generated in SpecvP (and therefore the closest DP to the EPP feature) is required to raise to the specifier position of TP, satisfying the EPP[+D] feature. In many analyses of verb-initial languages, an EPP feature is posited to cause the movement of verb or predicate. Massam and Smallwood (1997) suggest that the EPP feature of T˚ is parametrically different in verb-initial languages. They argue that instead of a [+D] feature, the EPP feature in verb-initial languages can be checked by a predicate. Expanding on this concept, Alexiadou and Anagnostopoulou (1998) claim the EPP feature may license the movement of a head or a phrasal constituent. For verb-initial languages, this meant the EPP feature could be adopted for both head-movement and predicate-movement models. A parametric difference between the EPP feature of T˚ alternating between [+D] and [+V] at first appeared completely suitable for verb-initial languages (for example, Lee 2000; Massam 2000). However, one of the major issues which became evident in later research was that the position of the raised verb did not always seem to match up with SpecTP. Most problematically, multiple particles which are argued to be generated lower than SpecTP surface preceding the raised verb. In verb-initial Polynesian languages, these include TAM, pre-verbal pronouns and negation. The presence of these particles has lead several authors to suggest that the EPP feature is actually found on a different functional projection, be that higher in the C-domain, or lower than TP. 2.1 Tongan In most Polynesian literature, the TAM particle is generated in T˚ (Massam 2000, 2001; Otsuka 2005; Collins 2017; Middleton 2021). Doner (2021) notes that the TAM particle in Tongan precedes the predicate (1). (1) Na‘e ‘alu ‘a Siale. go.SG ABS Siale ‘Siale went.’ (Tongan, Churchward 1953:56) TAM If the EPP feature which licences predicate-movement remains on T˚, we would expect the predicate to raise to SpecTP and precede the TAM particle in T˚.2 As 2 Note that while Doner (2021) argues for a predicate-movement account of Tongan, other authors have suggested that a head-movement model is more appropriate (Custis 2004; Otsuka 2005). Predicate landing sites in verb-initial languages 317 TAM precedes the predicate on the surface, Doner proposes that the EPP feature is instead found on a lower functional projection, FP, which is immediately dominated by TP. 2.2 Samoan Like Tongan, Samoan has a clause-initial TAM particle. Collins (2017) proposes that TAM raises to the complementiser position, via T-to-C movement.3 This removes the problematic ordering issue of the TAM particle and the predicate found in the analysis of Tongan above. The TAM particle raises above TP, meaning SpecTP is a viable landing site for the predicate. However, Samoan exhibits pre-verbal pronouns which surface between the TAM particle and the verb (2). ‘ou manatua ai nei fo‘i 1SG remember ANP now also ‘I now also remember the words.’ (Samoan, Mosel and Hovdhaugen 1992:333) (2) Ua PRF upu. word(PL) Collins (2017) argues these pronouns are DPs which raise to SpecTP due to an EPP[+D] feature. This feature only targets weak pronouns, which explains why not all subject pronouns are fronted. With the specifier of TP filled, Collins proposes that the EPP feature which fronts the predicate in Samoan is on FP, a functional projection below TP. As such, the predicate raises to SpecFP, rather than SpecTP. Negation is also found preceding the predicate (3), so Collins (2017) argues that FP must also be below NegP. ‘ou lē fia ‘ai. 1SG NEG want eat ‘I did not want to eat.’ (Samoan, Mosel and Hovdhaugen 1992:332) (3) Sā PST 2.2 Niuean Massam (2020) argues that TP does not exist in Niuean. Massam argues that the TAM particle in generated in the highest complementiser position in the left periphery (ForceP), due to various characteristics that TAM and complementisers share in Niuean. Furthermore, it is claimed there is no agreement in the language, which means the roles normally played by TP are not required in Niuean, and as such TP is made redundant. With TAM in ForceP, two other functional projections exist in the C-domain: PolP and FinP. Negation is argued to be generated in PolP (due to its complementary distribution with the question particle usually generated in this 3 Note that T-to-C movement has been proposed in some analyses of Tongan as well (Custis 2004; Otsuka 2005). 318 JOHN MIDDLETON position), just below ForceP. Negation is found following TAM, but preceding the predicate (4). nākai fetataiaki e PST NEG agree ABS ‘The experts didn’t agree . . . ’ (4) Ne tau tagata pulotu . . . PL person expert (Niuean, Massam 2020:28) With both TAM and negation preceding the predicate, the landing site cannot be in the specifier positions of ForceP or PolP. No TP exists, meaning the last remaining position for the predicate is SpecFinP. Therefore, Massam (2020) argues the EPP feature resides on Fin˚, forcing the moment of the predicate to SpecFinP. 3 Predicate landing sites in Tokelauan This paper presents data that demonstrates the pre-verbal particles in Tokelauan do not present a problem for a SpecTP predicate-landing site. In (5), three particles including TAM, negation and pre-verbal pronouns precede the predicate (in square brackets). Each of these will be discussed in the following subjections. kō [tuki-a] TAM NEG 1SG hit-CIA ‘I will not hit Rangi.’ (5) E hē ia ABS Rangi. Rangi 3.1 TAM movement Like the three Polynesian languages discussed above, Tokelauan has a clause-initial TAM particle, preceding the predicate. TAM is base-generated in T˚ (Middleton 2021; Otsuka 2005; Collins 2017). If TAM remains in-situ, then SpecTP is problematic for the landing site for the predicate, as SpecTP linearly precedes T˚. However, Middleton (2021) claims that TAM raises to the functional projection above TP, FinP (the lowest projection in the left periphery).4 A critical piece of evidence for TAM-raising is that TAM is in complementary distribution with the complementisers ke and oi, which are generated in Fin˚ (6). (6) Na taumafai ia John ke (*na) hao try ABS John COMP PST escape te vaka mai te afā. DEF boat from DEF hurricane ‘John tried to escape the ship from the hurricane.’ PST Middleton (2021) suggests that when a complementiser is generated in Fin˚, TAM is blocked from undergoing T-to-Fin movement, so it stays covert. This accounts 4 This movement is the equivalent to T-to-C movement proposed for other Polynesian languages (Otsuka 2005; Collins 2017). Predicate landing sites in verb-initial languages 319 for the distribution seen in (6). In summary, as TAM raises above SpecTP to Fin˚, SpecTP remains a potential landing site for the predicate. 3.2 Pre-verbal pronouns are clitics Pre-verbal pronouns precede the predicate and follow the TAM particle in Tokelauan (7a). Only ergative subject pronouns may appear in a pre-verbal position, and this is only optionally (7b). This makes Tokelauan pre-verbal pronouns the most restricted in the Polynesian family (Moyse-Faurie 1997). (7) a. Na ia velo-a te ika. 3SG spear-CIA DEF fish ‘He speared the fish.’ (Hooper 1993:62) PST b. Na velo e ia te ika. spear ERG 3SG DEF fish ‘He speared the fish.’ (Hooper 1993:62) PST In most Polynesian literature, pre-verbal pronouns are understood to be clitics (Moyse-Faurie 1997), in much the same way as the clitic pronouns in Romance languages (for example, Zwicky 1977).5 As they are clitics, the pre-verbal pronouns must attach to a clitic host. For Tongan, pre-verbal pronouns have been argued to attach to TAM (Otsuka 2000, 2005; Ball 2008) or the verb (Custis 2004). This paper adopts the stance that TAM is the clitic host, for various syntactic reasons. As clitics are D heads, they should display head-like properties, as opposed to a full DP which would display phrasal properties (Cardinaletti and Starke 1999). Only phrasal constituents are able to be coordinated (Sportiche 1996; Monachesi 1999), meaning that we predict that pre-verbal pronouns should not coordinate if they are clitics. As (8) shows, this is true. kē ma ia tunu nā ika. 2SG CONJ 3SG cook DEF.PL fish Intended: ‘You and him cooked the fish.’ (8) *Na TAM Furthermore, if pre-verbal pronouns are clitics, they would be weak pronouns, generated in D˚. As only strong pronouns bear case marking, we would expect clitic pronouns to be caseless (Doner 2021). This turns out to be correct: preverbal pronouns are ungrammatical with case-markers (9). (9) *Na ia velo-a te 3SG spear-CIA DEF Intended: ‘He speared the fish.’ PST 5 e ERG ika. fish Collins (2017) is the exception - he makes a case that in Samoan, pre-verbal pronouns are full DPs which undergo movement (see section 2.2). 320 JOHN MIDDLETON Additionally, pre-verbal pronouns cannot cooccur with post-verbal pronouns (10), as expected if they are clitics, rather than some kind of agreement marker (Otsuka 2005). (10) *Na te TAM ia velo-a (e) ia 3SG spear-CIA ERG 3SG Intended: ‘He speared the fish.’ DEF ika. fish As an adverb may intervene between the clitic and the verb (11), this paper suggests that pre-verbal pronouns are enclitics, attaching to TAM, rather than pro-clitics adjoined to the verb. (11) Na ia toe velo-a te 3SG again spear-CIA DEF ‘He speared the fish again.’ PST ika. fish The final piece of evidence is that clitics require an overt host. This makes a prediction that pre-verbal pronouns should require a host TAM particle if they are clitics. Therefore, in constructions without TAM, we expect pre-verbal pronouns will be ungrammatical. In Tokelauan, imperatives do not have TAM particles (12a), meaning there is no clitic host for the pronoun to attach to. As predicted, pre-verbal pronouns are not accepted in imperatives (12b). (12) a. Tipi te lakau! cut DEF wood ‘Cut the wood!’ b. (*Kē) tipi-a te 2SG cut-CIA DEF ‘You cut the wood!’ fafie! wood Analysing pre-verbal pronouns as clitics attaching to TAM in verbal sentences presents a problem when examining negation, which intervenes between TAM and pre-verbal pronouns (13). However, this issue becomes nullified if we analyse negation as another clitic. (13) Na hēki ia velo-a te PST NEG 3SG spear-CIA DEF ‘He didn’t spear the fish.’ 3.3 Negation is a clitic ika fish Predicate landing sites in verb-initial languages 321 In Tokelauan, negation sits between the TAM particle and the verb.6 Two main analyses exist for Polynesian negation. One is that negation is a particle (phrasal head), and the other is that negation is itself a predicate, which takes a verbal subordinating clause as its complement (Hovdhaugen and Mosel 1999). Although the negative predicate concept has been successfully argued for a variety of Polynesian languages (Hohepa 1969; Chung 1970, 1978, 2021; Waite 1987; Custis 2004; Ball 2008; Potsdam and Polinsky 2017; Clemens 2018), this paper argues that it is not appropriate for Tokelauan. We can rule negative predicates out by examining ko-topicalization, the process by which a topicalised nominal raises to a clause-initial position. In Tokelauan, kotopicalization is a mono-clausal process; arguments do not raise beyond the left periphery of the clause in which they originate (14). (14) Ko ko John John te TOP DEF TOP na lea say na PST mai DIR ika tunu fish PST cook ‘John said Rangi cooked the fish.’ e ERG Rangi. Rangi If negation was a bi-clausal structure (with a negative predicate and a subordinate verbal predicate), then a ko-topicalised argument from the verbal clause appear in the left periphery of the verbal clause, but not raise above the negation. As is demonstrated in (15), ko-topics actually raise beyond both the verb and negation, indicating a mono-clausal structure. (15) John e hēki kiki TOP John TAM NEG kick ‘John wasn’t kicked by Rangi.’ a. Ko b. *E e ERG Rangi. Rangi John na kiki John PST kick Intended: ‘John wasn’t kicked by Rangi.’ TAM hē ko e NEG TOP ERG Rangi. Rangi Therefore, this paper assumes Tokelauan negation is a particle, not a predicate. In addition, it is proposed that negation is a clitic attaching to TAM, just like the preverbal pronouns. Two different negative particles occur in Tokelauan: the imperfective hē negates states or situations and the perfective hēki negates events (Hooper 1993:55). The hēki negative can be also used to mean ‘not yet’. Although both may occur for certain contexts like (16), the two cannot be coordinated (17), indicating that negation has clitic-like properties (Sportiche 1996; Monachesi 1999). 6 However, unlike Samoan, Tokelauan negation precedes pre-verbal pronouns. An analysis of why this occurs, and the differing order for Samoan, is given in Middleton (to appear). 322 (16) JOHN MIDDLETON a. E TAM hē i NEG LOC ‘Rangi is not here.’ b. E hēki i NEG LOC kinei ia here ABS Rangi. Rangi kinei ia Rangi. here ABS Rangi ‘Rangi is not here yet.’ or ‘Rangi was not here.’ TAM (17) E hē ma hēki i ia TAM NEG CONJ NEG LOC ABS kinei here Intended: ‘Rangi was and is not here.’ Rangi. Rangi As clitics must have a host, we can use this as a diagnostic (just as we did for preverbal pronouns). If there are constructions where there is no potential clitic host, we expect negation to be unacceptable if it is indeed a clitic. Alternatively, if there is a clause-initial particle to act as the host, we expect negation to be possible. This latter scenario occurs for unmarked verbal clauses and nominalisations, which have a clause-initial TAM particle or determiner. As expected, negation may directly follow these clitic hosts: (18) a. E hēki velo-a e TAM NEG spear-CIA ERG ‘He didn’t spear the fish.’ b. te hē ia 3SG kai-ga o DEF NEG eating-NMLZ GEN ‘the not eating of the fish’ te DEF te DEF ika. fish ika fish However, imperatives do not have TAM particles in Tokelauan. Therefore, a clitic analysis of negation would suggest that negation cannot occur in imperatives. As observed in (19), this is indeed the case.7 (19) 7 *Hēki tipi-a te fafie. NEG cut-CIA DEF wood Intended: ‘Don’t cut the wood!’ Although several analyses have been proposed to deal with the ban on negative imperatives in head negation languages, none are viable for Tokelauan. In one, imperatives are formed when the verb head raises to TP/MoodP, and the negative head blocks head-movement of the verb (Zeijlstra 2004). As Tokelauan is a predicate-raising language, verbal head-movement does not occur, making this analysis unfeasible. Another analysis is that imperatives do not have a TP layer, and NegP sits above TP, so will not be found in imperatives (Zanuttini 1994). However, in Tokelauan, negation is dominated by the TP layer (as negation follows TAM), indicating this model is not suitable either. Predicate landing sites in verb-initial languages 323 Another structure which has no potential clitic host is an equative clause with a nominal predicate. These have a predicate marker ko, but no TAM particle (20a).8 Like imperatives, nominal predicates cannot be negated using the normal negative particle preceding the predicate (20b). (20) a. Ko he tautai te PRED INDF fisherman DEF ‘The man is a fisherman.’ b. *Hē ko tamaloa. man he tautai te tamaloa. NEG PRED INDF fisherman DEF man Intended: ‘The man is not a fisherman.’ For both imperatives and nominal predicates, this paper proposes the ban on negation is due to the lack of a suitable clitic host. Instead, some sort of last resort clitic host is needed in order to negate either construction. Negative imperatives are formed with nahe (21). I suggest the nahe particle is a lexicalisation of the TAM particle na and the stative negative form hē. The TAM particle provides the negative particle a host to attach to, which has become lexicalised into a single particle over time. (21) tipi-a te cut-CIA DEF ‘Don’t cut the wood!’ Nahe NEG.IMP fafie! wood Similarly, nominal predicates employ a last resort TAM particle when the clause is negated (22). This precedes the negative particle and acts as the clitic host. (22) E hē ko he TAM NEG PRED INDF ‘The man is not a fisherman.’ tautai te fisherman DEF tamaloa. man Assuming this is a last resort clitic host may account for the fact that TAM particles are unacceptable in positive nominal clauses, but must be found in negated nominal clauses. Furthermore, in cases when a suitable clitic host is available, we predict the last resort TAM will not be needed. As anticipated, in subordinate clauses with an overt complementiser, the last resort TAM particle is redundant, and disappears (23). 8 Note that the ko particle in nominal predicates is not the same as the topic ko particle found in kotopicalisation. This can be demonstrated by the fact that the argument of a nominal predicate may be ko-topicalised, resulting in two ko-marked nouns: the first being the topicalised argument and the second being the nominal predicate (Hooper 1993:208-209). 324 (23) JOHN MIDDLETON mafai ke hē ko Viliamu TAM possible COMP NEG PRED Viliamu ‘It is possible that William is not the teacher.’ E te faiaoga. teacher DEF One last piece of evidence links negation with pre-verbal pronominal clitics. The Cia affix is found in many Polynesian languages and appears for different reasons depending on language. For example, -Cia marks a passive verb in Eastern Polynesian languages (Sanders 1991). However, for Tokelauan, the function of Cia is unclear (Hovdhaugen 2000). In Tokelauan, the -Cia suffix appears on transitive verbs in four contexts. The suffix is observed when a transitive subject is ko-topicalised (24a), relativised (24b), or appears as a pre-verbal pronoun (24c). Note that in ko-topicalization and relativisation, pre-verbal pronouns are also obligatory. The other appearance of Cia is when a transitive verb is negated (24d). (24) a. Ko John na ia John PST 3SG ‘John built a canoe.’ TOP fau-a te build-CIA DEF b. te tamaloa na ia man PST 3SG ‘the man who kicked Rangi’ DEF c. Na ia kiki-a 3SG kick-CIA ‘He kicked himself.’ PST d. Na hēki ia ABS manatua-gia PST NEG remember-CIA ‘I didn’t remember the food.’ kiki-a kick-CIA vaka. boat ia ABS ia 3SG lava. e au 1SG ERG Rangi Rangi INT te DEF meakai. food The one thing that all of these construction have in common is that they either have a pre-verbal pronoun or are negated. I tentatively suggest that this suffix is therefore linked to cliticization, occurring when a clitic attaches to TAM. The following section discusses how TAM, pre-verbal pronouns and negation interact to form a single complex head. 3.4 SpecTP predicate landing side If we adopt a movement-approach for clitics (Kayne 1975, 1989; Uriagereka 1995; Anagnostopoulou 2003), we gain a relatively straightforward account of the preverbal particles in Tokelauan. Predicate landing sites in verb-initial languages 325 The clitic pronoun is generated in the D˚ head of the subject DP (Sportiche 1996). The subject is generated in SpecvP like normal, but the pronoun undergoes head-movement, raising to T˚ and attaching as an enclitic to TAM.9 The analysis for negation is very similar. The negative head undergoes headmovement to TAM, adjoining to the right. The ordering of TAM-NEG-pro falls out if we assume NegP is generated above SpecvP where the pronoun is formed (see Collins 2017 for evidence of this in Samoan). Therefore, the pronoun raises to Neg˚, right-adjoins, and subsequently the NEG-pro complex head raises and right-adjoins to TAM (see Middleton to appear for more detailed discussion). At this juncture, TAM-NEG-pro is a single complex head. With TAM undergoing T-to-Fin movement, all three particles are found in the left periphery, higher than TP. Nothing else intervenes between these three particles and the predicate, which means positing SpecTP as the predicate landing site is appropriate for this language. This allows us to maintain the EPP[+pred] or EPP[+D] parametric contrast first suggested by Massam and Smallwood (1997). For Tokelauan, this makes the concept of predicate-fronting much tidier than in closely related languages. Instead of positing new functional projections to deal with pre-verbal particles, this paper is able to account for the pre-verbal material while maintaining a standard SpecTP landing site for the predicate. 4 Conclusion This paper examines Tokelauan pre-verbal particles. I demonstrate that SpecTP is a suitable predicate-landing site, despite the pre-verbal material which has caused other authors to suggest new functional projections for the predicate to raise to. While TAM, pre-verbal pronouns and negation are all generated below TP, it is argued that all reside above TP post-movements. Pre-verbal pronouns are understood to be clitics attaching to TAM. Syntactic evidence and morphological parallels to pre-verbal pronouns suggests that negation is also a clitic. Both clitics form a complex head with TAM, and this head raises above TP to FinP. The result is that the predicate raises to SpecTP due to an EPP[+pred] feature, which follows TAM, negation and pre-verbal pronouns. This paper does not make any claims on the predicate landing site in other Polynesian languages like Tongan, Samoan and Niuean. However, it does suggest that the pre-verbal particles found in Polynesian languages do not necessarily rule out a SpecTP predicate landing site. Further research is needed to determine whether similar analyses are possible for related languages. References Alexiadou, Artemis, and Elena Anagnostopoulou. 1998. Parametrizing AGR: Word order, Vmovement and EPP-checking. Natural Language and Linguistic Theory 16(3): 491-539. Anagnostopoulou, Elena. 2003. The Syntax of Ditransitives: Evidence from Clitics. Berlin: Mouton de Gruyter. 9 Custis (2004) proposes a similar analysis for Tongan pre-verbal pronouns. 326 JOHN MIDDLETON Ball, Douglas. 2008. Clause structure and argument realization in Tongan. Doctoral dissertation, Stanford University, California. Cardinaletti, Anna, and Michal Starke. 2011. The typology of structural deficiency: A case study of the three classes of pronouns. Clitics in the languages of Europe, ed. Henk Van Riemsdijk, 145-234. De Gruyter Mouton. Chomsky, Noam. 1982. Some concepts and consequences of the theory of government and binding. Cambridge, Mass.: MIT Press. Chung, Sandra. 1970. Negative verbs in Polynesian. Doctoral dissertation, Harvard University, Cambridge, Massachusetts. Chung, Sandra. 1978. Case Marking and Grammatical Relations in Polynesian. Austin, TX: University of Texas Press. Chung, Sandra. 2021. Reaffirming Māori negatives as verbs. Polynesian Syntax and Its Interfaces, eds. Clemens, Lauren, and Diane Massam, 147-167. Oxford University Press. Churchward, C. Maxwell. 1953. Tongan Grammar. London: Oxford University Press. Clemens, Lauren. 2018. Niuean nākai as a Negative Verb: Implications for the Derivation of V1 Order, Festschrift in Honor of Lisa De Mena Travis: Working Papers in Linguistics, eds. Laura Kalin, Ileana Paul, and Jozina Vander Klok, 78–90. Montreal: McGill University. Collins, James. 2017. Samoan predicate initial word order and object positions. Natural Language & Linguistic Theory 35(1): 1-59. Custis, Tonya. 2004. Word order variation in Tongan: a syntactic analysis. Doctoral dissertation, University of Minnesota, Minnesota. Doner, Julianne. 2021. Predicate-EPP in Niuean, Tongan, and beyond. Polynesian Syntax and Its Interfaces, eds. Clemens, Lauren, and Diane Massam, 253-281. Oxford University Press. Hohepa, Patrick. 1969. The Accusative-to-ergative Drift in Polynesian Languages, Journal of the Polynesian Society, 78: 295–329. Hooper, Robin. 1993. Studies in Tokelauan syntax. Doctoral dissertation, University of Auckland, Auckland. Hovdhaugen, Even. 2000. Cia-words in Tokelauan. SICOL Proceedings of the Second International conference on Oceanic Linguistics: Vol. 2, Historical and descriptive studies, 363-368. Pacific Linguistics. Hovdhaugen, Even, and Ulrike Mosel, (eds.). 1999. Negation in Oceanic languages: typological studies. Vol. 2. Lincom Europa. Kayne, Richard. 1975. French syntax: The transformational cycle. Vol. 30. Cambridge, MA: MIT press. Kayne, Richard. 1989. Null Subjects and Clitic Climbing, The Null Subject Parameter, eds. O. Jaeggli and K. Safir, 239-261. Springer, Dordrecht, Lee, Felicia. 2000. VP remnant movement and VSO in Quiavini Zapotec. In The syntax of verb initial languages, ed. Andrew Carnie and Eithne Guilfoyle, 143–162. Oxford: Oxford University Press. Massam, Diane. 2000. VSO and VOS: Aspects of Niuean word order. In The syntax of verb initial languages, ed. Andrew Carnie and Eithne Guilfoyle, 97–116. Oxford: Oxford University Press. Massam, Diane. 2001. Pseudo noun incorporation in Niuean. Natural Language and Linguistic Theory, 19:153–197. Massam, Diane. 2020. Niuean Predicates and Arguments in an Isolating Language (Vol. 6). Oxford University Press, USA. Massam, Diane and Smallwood, Carolyn. 1997. Essential Features of Predication in English and Niuean, North East Linguistics Society, 27: 263-272 Middleton, John. 2021. Revisiting the clause periphery in Polynesian languages. Glossa: A Journal of General Linguistics, 6(1):87. Middleton, John. To appear. Two types of negation in Samoan and Tokelauan. Proceedings of Austronesian Formal Linguistics Association (AFLA) 29. University of Manchester. Predicate landing sites in verb-initial languages 327 Middleton, John, and Syed, Saurov. 2022. T→C movement in Polynesian: a case study of Tokelauan. Proceedings of Austronesian Formal Linguistics Association (AFLA) 28:83-94. Monachesi, Paola. 1999. A lexical approach to Italian cliticization. Stanford: CSLI Publications. Mosel, Ulrike, and Even Hovdhaugen. 1992. Samoan reference grammar. Oxford: Oxford University Press. Moyse-Faurie, Claire. 1997. Syntactic and pragmatic functions of pronominal arguments in some Western Polynesian languages. Oceanic Linguistics, 6-28. Otsuka, Yuko. 2000. Ergativity in Tongan. Doctoral dissertation, University of Oxford. Otsuka, Yuko. 2005. Two derivations of VSO: A comparative study of Niuean and Tongan. Verb first: On the syntax of verb-initial languages, eds. Andrew Carnie, Heidi Harley and Sheila Ann Dooley, 65–90. Amsterdam: Benjamins. Potsdam, Eric, and Maria Polinsky. 2017. A preliminary look at exceptives in Tahitian, Asking the right questions: Essays in honor of Sandra Chung, eds. Jason Ostrover, Ruth Kramer, and Joseph Sabbagh, 28–36. Santa Cruz: Linguistics Research Center, Sanders, Gerald. 1991. Levelling and reanalysis in the history of Polynesian passive formations. The Journal of the Polynesian Society 100(1): 71-90. Sportiche, Dominique. 1996. Clitic constructions. Phrase structure and the lexicon, eds. Rooryck, Johan, and Laurie Zaring, 213-276. Springer, Dordrecht. Uriagereka, Juan. 1995. Aspects of the Syntax of Clitic Placement in Western Romance. Linguistic Inquiry 26:79-124. Waite, Jeffrey. 1987. Negatives in Māori: A Lexical-Functional Approach, Te Reo 30:79–100. Zanuttini, Raffaella. 1994. Speculations on negative imperatives. Rivista di linguistica 6:67-89 Zeijlstra, Hedde. 2004. Sentential negation and negative concord. Netherlands Graduate School of Linguistics. Zwicky, Arnold. 1977. On clitics. Bloomington: Indiana University Linguistics Club. Proceedings of CLS 58 (2022), 329-344 © Chicago Linguistic Society 2023. All rights reserved. 329 Argument for Verb-Stranding Ellipsis by Blocking Verb-Raising in Japanese Yosho Miyata* University of Massachusetts, Amherst 1 Introduction Among previous analyses of ellipsis phenomena, the relation between head movement and ellipsis has received extensive discussion focusing on the nature of verb stranding ellipsis (Goldberg (2005), Gribanova (2017), among others). In the standard analysis, the verb is moved out of a verb phrase or TP and ellipsis applies to the VP. The main aim of this paper is to reveal the size of elidable elements by showing that an adverbial particle prevents verb-raising in Japanese. For this purpose, I examine data involving null adjunct and an adverbial particle, along the line of Funakoshi (2016), and propose that ellipsis targets a complement of a head that V raises to. Funakoshi (2016) argues for the verb stranding VP ellipsis (hereafter VVPE) analysis and proposes a generalization about the distribution of nulladjuncts: (1) In Japanese, an adjunct can be null only if the clause-mate object (or other VP-internal elements), if any, is also null. (Funakoshi 2016) He shows that VVPE can explain this generalization. Interestingly, however, I observe that an adjunct cannot be null when the clause-mate object is null if verb movement is blocked by an adverbial particle, as in (2). 0 This is a revised version of my first general paper. I would like to thank Faruk Akkuş, Ana Arregui, Christopher Davis, Brian Dillon, Yoshiki Fujiwara, Vera Gribanova, Shrayna Halder, Polina Kasyanova, Jason Merchant, Shota Momma (Negishi), Ellen Woolford and the audience at SSRG and Syntax-workshop at University of Massachusetts, Amherst (November, 13rd, 2020 and April 15th, 2022) and at CLS 58th for helpful comments. Special thanks to Rajesh Bhatt and Kyle Johnson for helpful, insightful comments and discussions of this idea. 330 YOSHO MIYATA (2) TP T XP Obj X Blocked by F * FP F vP V-v Adverb VP VP ... tObj ... tV ... Why is an adjunct disallowed to be null when an adverbial particle appears? This is the main puzzle that I address. 2 Previous Analyses Ellipsis phenomena are widespread but its mechanisms are still controversial (See Johnson (2001), Lobeck (2006), Merchant (2006) and among others). Japanese is not an exception. It allows arguments to be unpronounced, as in (3). (3) Taro-wa ringo-o tabe-ta kedo, Jiro-wa ∆ tabe-nakat-ta. Taro-TOP apples-ACC eat-PAST but Jiro-TOP eat-NEG-PAST ‘Taro ate apples but Jiro did not eat apples.’ (In this paper, an ellipsis site is indicated by ∆.) The object is not pronounced in the second conjunct and the unpronounced object refers to ringo ‘apples’. In order to derive sentences like these, the object is elided by VP-ellipsis under the VVPE analyses after the verb is moved (see Otani & Whitman (1991), Goldberg (2005), Funakoshi (2014) and among others). But in previous analyses of ellipsis, there are two other types of approaches to derive such sentences: pro analysis and Argument Ellipsis analysis. In the pro analysis, the null pronoun pro appears in the object position and it refers to the object (see Kuroda (1965), Hoji (1985), and among others). According to the Argument Ellipsis analysis, the object exists in syntax and it is deleted (see Oku (1998), Saito (2007), Sakamoto (2016), among others)1 . 1 Oku (1998) argues against the existence of VVPE and claims that, based on his intuition (indicated by (*)), the null adjunct reading is unavailable: (i) a. Bill-wa kuruma-o teineini arat-ta. Bill-TOP car-ACC carefully wash-PAST ‘Bill washed the car carefully.’ Verb-stranding ellipsis by blocking verb-raising 331 Funakohsi (2016) shows that an adjunct in Japanese can be null only when its clause-mate object (or other VP-internal elements) is also null. He gives four tips to derive null-adjunct reading: (i) negation in the antecedent sentence, (ii) the particle -mo ‘also’ in the ellipsis, (iii) disjunction kedo ‘but’ between the antecedent and the ellipsis sentence sentence, observed in Funakoshi (2014), and (iv) an appropriate context favoring null-adjunct reading. As we will see in (11), the null-adjunct reading is available in Japanese. An adjunct can be null only when the clause-mateobject is also null, as demonstrated below: (4) CONTEXT: Taro and Hanako decided to wash a car together. After Taro washed the body of the car, Hanako wiped the windows. Taro was thorough in his work, but Hanako was not. a. Taro-wa teineini shatai-o arat-ta. Taro-TOP carefully body.of.car-ACC wash-PAST ‘Taro washed the body of the car carefully.’ b. # demo hanako-wa ∆ madogarasu-o huka-nakat-ta. Hanako-ga but Hanako-TOP windows-ACC wip-NEG-PAST Hanako-NOM hui-ta ato-no madogarasu-wa kitanakat-ta. wipe-PAST after-GEN windows-TOP dirty-PAST ‘But Hanako did not wipe the windows. The windows that Hanako wiped were dirty.’ (Funakoshi 2016) Funakoshi (2014) proposes VVPE to capture generalization in (1) as in (5) (5) VVPE TP Sbj vP V-v-T Sbj VP Adverb tV -v VP Obj tV Ellipsis of either vP or VP derives the null-adjunct reading. Therefore, Funakoshi (2016) concludes that Japanese does not use two ellipses, ellipsis of adverb and ellipsis of object, to generates the null-adjunct reading, but it uses VVPE instead. b. (*) John-wa ∆ arawa-nakat-ta. John-TOP wash-NEG-PAST Int.:‘John did not wash the car carefully.’ (Oku 1988) According to my Japanese informants including me, however, the null-adjunct reading is available in (i). I leave open the issue of speaker variation. 332 YOSHO MIYATA 3 Proposal In this section, I propose Verb-Stranding XP-Ellipsis. Before seeing my proposal, I will examine a new observation which is contrary to Funakoshi’s (2016) generalization in , where XP contains vP/VP. 3.1 New Observation: Ban of VP-Ellipsis in Verb with Adverbial Particle Interestingly, the vP-internal adjunct, teineini ‘carefully’, cannot always be null even if its clause-mate object is also deleted as illustrated in (6)2 . (6) CONTEXT: Taro and Hanako decided to wash the body of their parents’ car and its windows together to get their allowance. At first, they washed the body. But they did not wash it carefully because the cars are really big. So it was also difficult for them to wipe the window carefully. So they did not finish their work. Taro-wa teineini madogarasu-o huki-mo shi-nakat-ta. Taro-TOP carefully windows-ACC wipe-and.then do-NEG-PAST ‘And then Taro did not wipe the windows carefully.’ B1: # Hanako-mo ∆ huki-mo shi-nakat-ta. Hanako-ga Hanako-also wipe-and.then do-NEG-PAST Hanako-NOM hui-ta madogarasu-wa mada kitanakat-ta. wipe-PAST windows-TOP still be.dirty-PAST ‘And then Hanako did not wipe the windows (at all). The windows that Hanako wiped are still dirty.’ shi-nakat-ta. B2: Hanako-mo teineini madogarasu-o huki-mo Hanako-also carefully windows-ACC wipe-and.then do-NEG-PAST Hanako-ga hui-ta madogarasu-wa mada kitanakat-ta. Hanako-NOM wipe-PAST windows-TOP still be.dirty-PAST ‘And then Hanako did not wipe the windows carefully. The windows that Hanako wiped are still dirty.’ A: The example in (6B1) is an ellipsis sentence which appears with a follow-up sentence favoring null-adjunct reading. An adverbial particle -mo ‘and then’ is attached onto the verb (Aoyagi (1996); Aoyagi (2006), and see also Kuroda (1965) for -wa). Importantly note that dummy verb su(-ru) ‘do’ is obligatorily inserted when the particle is attached onto verb. So it is safe to say that verb-raising does not happen when the adverbial particle appears with the verb. The sentence in (6B1) sounds contradictory when the adverb and an object are not pronounced. This indicates that an adjunct cannot be null when an adverbial particle appears with the verb, even if its clause-mate object is covert.3 2 Although (6) satisfies three tips out of four tips argued in Funakoshi (2016) to derive null adjunct reading, the null adjunct reading is unavailable. 3 One might think that (6B1) would be unsuitable because of two -mo on Hanako and verb huki ‘wipe’, not because of ellipsis. But a full sentence of (6B1) is grammatical and suitable as in (6B2). Verb-stranding ellipsis by blocking verb-raising 333 3.2 Proposal I propose a following principle and Verb-Stranding XP-Ellipsis: (7) Ellipsis targets a complement of a head that V raised to. (8) Verb-Stranding XP-Ellipsis * ZP 1 * 3 V-v-... X-Y-Z XP 2 YP tV -v-...X-Y Let us see how the proposal works. (9) Structure of (6) TP Sbj * XP 1 T Obj NegP FP * 2 vP NEG F -mo tSbj * 3 Adverb VP X do-insertion V-v tObj tV Note that I adopt an analysis proposed by Kobayashi (2014) that an adverbial particle projects its own projection, FP in (6) (cf. Aoyagi (1996); Aoyagi (2006)). Ellipsis of XP does not derive the null-adjunct reading because F prevents V-raising to a higher head. Another option, either ellipsis of either vP or VP, does not derive the null-adjunct reading. According to (1), only single ellipsis of adverb and its clause-mate object can derive the null-adjunct reading but neither vP-ellipsis nor VP-ellipsis target an object in Spec, XP. Therefore, any ellipsis option cannot derive the null-adjunct reading in (6)4 . Note that, an object is obligatorily moved into a higher position than NEG, adapting Shibata (2015a); Shibata (2015b). One might So the argument here can be maintained. 4 The proposed analysis incorrectly overgenerates the null-adjunct reading if VP is elided that contains only an adverb. I leave this issue open for a future research. 334 YOSHO MIYATA think that the null-adjunct reading is unavailable because the adverb appears in a higher position than an elliptical site. But stranding the adverb is not allowed in VP-fronting: (10) a. Taro-wa [F P teineini madogarasu-o huki-mo] shi-ta. Taro-TOP carefully windows-ACC wipe-and.then do-PAST ‘And then Taro wiped the windows carefully.’ b. [F P teineini madogarasu-o huki-mo]i Taro-wa ti shi-ta. carefully windows-ACC wipe-and.then Taro-TOP do-PAST ‘And then Taro wiped the windows carefully.’ Taro-wa teineini k ti c. * [F P tk madogarasu-o huki-mo]i windows-ACC wipe-and.then Taro-TOP carefully shi-ta. do-PAST ‘And then Taro wiped the windows carefully.’ The facts in (10) indicate that the adverb teineini ‘carefully’ is inside FP. As observed in Funakoshi (2016), the null-adjunct reading is available in Japanese when the adverbial particle does not appear. (11) CONTEXT: Taro and Hanako washed their parents’ cars to get their allowance. Taro was thorough in his work while Hanako was not. Taro-wa teineini kuruma-o arat-ta. TaroTOP carefully car-ACC wash-PAST ‘Taro washed the car carefully.’ B: Hanako-wa ∆ arawa-nakat-ta. Hanako-ga arat-ta TaroTOP wash-NEG-PAST Hanako-NOM wash-PAST ato-no kuruma-wa kitan-akatta. after-GEN car-TOP dirty-PAST ‘Hanako did not wash the car carefully. The car that Hanako washed was dirty.’ (Funakoshi 2016, 119) A: (12) Structure of (11B) Verb-stranding ellipsis by blocking verb-raising 335 TP Sbj OK XP 1 V-v-NEG-T Obj NEGP * vP 2 X tV -v-N EG tSbj * VP 3 Adverb tV -v tObj tV Verb-stranding XP ellipsis targets XP that V raises into. This analysis predicts nulladjunct reading to be available in (11B) because both an object and an adverb are elided by a single ellipsis of XP. This prediction is born out. Other ellipsis options should not derive the null-adjunct reading because the object in Spec, XP survives these ellipses. 3.3 Evidence: Interpretative (A)symmetry of (non) Ellipsis It has been observed in Goro (2007), Shibata (2015a); Shibata (2015b) and others that a disjunction obligatorily takes scope over negation in Japanese as in (13). (13) Taro-wa madogarasu-ka-hoiiru-o fuka-nakat-ta. Taro-TOP windows-or-wheels-ACC wipe-NEG-PAST a. ‘It is either windows or wheels that Taro did not wipe.’ b. * ‘Taro wiped neither windows nor wheels.’ D IS>N EG N EG>D IS Interestingly, Shibata (2015b) and Funakoshi (2017) observe that the scope relation changes once the disjunctive phrase becomes null as in (14). (14) A: Taro-wa madogarasu-ka-hoiiru-o fui-ta. Taro-TOP windows-or-wheels-ACC wipe-PAST ‘Taro wiped windows or wheels.’ B: John-wa ∆ fuka-nakat-ta. John-TOP wipe-NEG-PAST a. * ‘It is either windows or wheels that Taro did not wipe.’ D IS>N EG b. ‘Taro wiped neither windows nor wheels.’ N EG>D IS Furthermore, a conjunctive phrase also takes an obligatorily wide-scope over negation when it is overt, while the sentence containing them becomes ambiguous when it is covert, as observed in Funakoshi (2017): 336 YOSHO MIYATA (15) John-wa speingo-mo furansugo-mo hanasa-nai. John-TOP Spanish-also French-also speak-NEG.PRESENT Lit.: ‘John does not speak Spanish and French.’ a. ‘John does not speak Spanish AND he does not speak French.’ C ON>N EG b. * ‘John does not speak Spanish OR he does not speak French.’ N EG>C ON (Funakoshi 2017) (16) Mary-wa speingo-mo furansugo-mo hanasu-ga, Mary-TOP Spanish-also French-also speak-PRESENT-but John-wa ∆ hanasa-nai. John-TOP speak-NEG.PRESENT Lit.: ‘Mary speaks Spanish and French, but John does not speak Spanish and French.’ a. ‘John does not speak Spanish AND he does not speak French.’ C ON>N EG b. ‘John does not speak Spanish OR he does not speak French.’ N EG>C ON (Funakoshi 2017) Importantly note that only the wide-scope interpretation of the conjunction is available when it is overt as in (17). (17) Mary-wa speingo-mo furansugo-mo hanasu-ga, John-wa Mary-TOP Spanish-also French-also speak-PRESENT-but John-TOP speingo-mo furansugo-mo hanasa-nai. Spanish-also French-also speak-NEG.PRESENT Lit.: ‘Mary speaks Spanish and French, but John does not speak Spanish and French.’ a. ‘John does not speak Spanish AND he does not speak French.’ C ON>N EG b. * ‘John does not speak Spanish OR he does not speak French.’ N EG>C ON (Funakoshi 2017) overt D IS > N EG(=(13a)) *N EG > D IS(=(13b)) OK C ON > N EG(=(15a)) *N EG > C ON(=(15b)) OK Disjunction Conjunction null *D IS > N EG(=(14Ba)) OK N EG > D IS(=(14Bb)) OK C ON > N EG(=(16a)) OK N EG > C ON(=(16b)) Table 1: Distribution of available scope relations between disjunction/conjunction and negation in Japanese The observed facts in (13-17) are summarized in Table 1. These facts in Table 1 can be accounted for under verb-stranding XP ellipsis. Ellipsis applies to a complement that V raised to via NEG, establishing a structural relation that NEG asymmetrically c-commands a conjunctive/disjunctive phrase as in (18b). (18) Derivations of (13-17) under Verb-Stranding XP Ellipsis Verb-stranding ellipsis by blocking verb-raising a. Base-Structure/No-ellipsis TP 337 b. Ellipsis TP Sbj Sbj XP T XP OK Dis./Con. V-v-NEG-T Dis./Con. NEGP vP X NEGP NEG tSbj vP X tV -v-N EG tSbj VP tDis./Con. v V VP tDis./Con. tV -v tV I assume here that verb-raising does not happen in the non-ellipsis configuration in (18b). The proposed analysis, however, does not capture why the wide-scope interpretation of conjunction over negation in an elliptical sentence is available. How do we derive that reading? Funakoshi (2017) argues for pro analysis for it. According to Hoji (2003), pro can be plural-denoting as in (19), considering that only plural-denoting pronouns like karera ‘them’ or sorera ‘these’ can refer to splitantecedents (19) Toyota-ga1 Nissan-ni2 [ zeimusho-ga pro1+2 Toyota-NOM Nissan-DAT tax.office-NOM sirabeteiru to] tsuge-ta. investigate.PROGRESSIVE C inform-PAST ‘Toyota1 informed Nissan2 that the tax office was investigating them1+2 .’ (Hoji 2003) (20) Mary-wa speingo-mo furansugo-mo hanasu-ga, John-wa Mary-TOP Spanish-also French-also speak-PRESENT-but John-TOP sorera-o hanasa-nai. them-ACC speak-NEG.PRESENT ‘Mary speaks Spanish and French, but John does not speak these.’ (Funakoshi 2017) (20) is truth-conditionally equivalent to the wide-scope interpretation of the conjunctive phrase. Therefore, the pro analysis and the verb-stranding XP ellipsis analysis can capture all facts in Table 15 . 5 Kyle Johnson (p.c.) points out that the narrow-scope reading of the disjunction phrase becomes unavailable when the antecedent contains negation as in (i), which the proposed analysis overgenerates. I leave this issue open for future research. 338 YOSHO MIYATA 4 Argument Against Argument Ellipsis Argument Ellipsis analyses cannot capture the observed facts in Table 1, which the proposed analysis here can explain straightforwardly. The Argument Ellipsis analysis predicts that the wide-scope interpretation of the disjunction phrase over negation should be available but the wide-scope interpretation of negation over it should not regardless of whether ellipsis happens or not. This is so because Argument Ellipsis does not need verb-raising as (14), repeated as (21), and its structure under the Argument Ellipsis analysis is illustrated below: (21) A: Taro-wa madogarasu-ka-hoiiru-o fui-ta. Taro-TOP windows-or-wheels-ACC wipe-PAST ‘Taro wiped windows or wheels.’ B: John-wa ∆ fuka-nakat-ta. John-TOP wipe-NEG-PAST a. * ‘It is either windows or wheels that Taro did not wipe.’ D IS>N EG b. ‘Taro wiped neither windows nor wheels.’ N EG>D IS (22) Structure of (21B) under the Argument Ellipsis TP Sbj T XP Obj X NEGP vP NEG tSbj VP tObj v V The analysis that Argument Ellipsis is the sole ellipsis option available in Japanese cannot be maintained, which is also argued in Funakoshi (2016); Funakoshi (2017). Maeda (2019), however, argues against verb-stranding ellipsis, citing an observation in Sakamoto (2016): (i) A: John-wa speingo-mo furansugo-mo hanasa-nai. John-TOP Spanish-and French-ACC speak-NEG.PRESENT. ‘John does not speak Spanish and he does not speak French.’ B: Bill-wa ∆ hanasa-nai. Bill-TOP speak-NEG.PRESENT i. ‘John does not speak Spanish AND he does not speak French.’ ii. * ‘John does not speak Spanish OR he does not speak French.’ C ON>N EG N EG>C ON Verb-stranding ellipsis by blocking verb-raising (23) 339 A: John-wa speingo-ka furansugo-o hanasa-nai. John-TOP Spanish-or French-ACC speak-NEG.PRESENT. ‘John does not speak Spanish or French.’ B: Bill-mo ∆ hanasa-nai. speak-NEG.PRESENT Bill-also a. * ‘Bill speak neither Spanish nor French.’ N EG>D IS b. ‘it is either Spanish or French that Bill does not speak.’ D IS>N EG (Sakamoto 2016) In (23), the disjunction obligatory takes scope over negation in an elliptical sentence when its antecedent sentence contains negation. Maeda (2019) explains the obligatory wide-scope reading of the disjunction phrase over negation in (23) by Scope Economy and Parallelism in Fox (2000), adopting Takahashi (2008) arguing that Scope Economy and Parallelism can capture scope facts in Japanese. The analyses adopted Maeda (2019) are illustrated below: (24) Scope Economy Covert optional operations cannot be scopally vacuous (i.e., they must reverse the relative scope of two noncommutative quantificational expressions). (Fox 2000) (25) Parallelism In an ellipsis/phonological reduction construction the scopal relationship among the elements in β A must be identical to the scopal relationship among the parallel elements in β E . (Fox 2000) According to Maeda (2019), the fact in (21) and (23) can be explained by (24) and (25). In (23), when the antecedent contains negation and disjunction, it is unambiguously interpreted as the wide-scope reading of disjunction. The elliptical sentence also has negation and the disjunction and it shows the same scope relation as its antecedent, the wide scope interpretation of disjunction in this case, due to Scope Parallelism. On the contrary to this, (21) shows the obligatory wide-scope interpretation of negation when the antecedent does not contain negation. A covert movement of disjunction should create a new interpretative output by Scope Economy but such movement is not allowed because the antecedent is scopally uninformative, and then the elliptical sentence is disambiguated in favor of the surface scope. So Parallelism and Scope Economy force the disjunction to take scope below negation in the ellipsis sentence. Maeda (2019) predicts that an ellipsis sentence should show the same scope interpretation as its antecedent if the antecedent is scopally informative. This is, however, contrary to the fact in (26). (26) CONTEXT: According to car cleaning checklist, car owners should clean two parts out of three parts of the car, body, windows and wheels, in order to make it clean. Mary and Bill already washed the body of their own car carefully and they are tired. 340 YOSHO MIYATA A: Mary-wa [ madogarasu-ka hoiiru-o] arawa-naka-ta. demo Mary-TOP window-or wheels-ACC wash-NEG-PAST but Mary-no kuruma-wa kirei-da ne. clean-COP PRT Mary-GEN car-TOP Lit.: ‘Mary did not wash windows or wheels, but Mary’s car is clean.’ a. * ‘Mary washed neither windows nor wheels.’ N EG>D IS b. ‘it is either windows or wheels that Mary did not wash.’D IS>N EG B: demo Bill-wa ∆ arawa-naka-ta. dakara Bill-no kuruma-wa wash-NEG-PAST therefore Bill-GEN car-TOP but Bill-TOP kitanai yo. dirty PRT a. ‘Bill washed neither windows nor wheels. Therefore, Bill’s car is dirty.’ N EG>D IS b. * ‘it is either windows or wheels that Bill did not wash. Therefore, Bill’s car is dirty.’ D IS>N EG Importantly note that the subject in (26B) appears with a topic-marker wa, unlike Sakamoto’s (2016) example. Maeda (2019) predicts (26B) to be inappropriate in this context. This is so because the ellipsis sentence must have the same interpretation as (26A), the wide-scope reading of disjunction, due to Scope Parallelism, but that reading contradicts its follow-up sentence ‘Therefore, Bill’s car is dirty.’. This is, however, contrary to the fact6 . Note that pro cannot be used in (26) because (B) means that Bill wiped neither windows nor wheels that is NOT the same one as Hanako did. How do we explain the observation in (23)? As pointed out in Maede (2019), VVPE cannot explain why disjunction obligatory takes scope over negation. This is also problematic to verb-stranding XP ellipsis that this paper proposed. Instead of VVPE and Argument Ellipsis, let us see how the pro analysis works here. Christopher Davis (p.c.) points out that a null part can yield the disjunctive E-type reading (cf. Simons (1996)). A pronoun in Japanese can have it when it refers to a disjunctive phrase. (27) A: John-wa speingo-ka furansugo-o hanasu. John-TOP Spanish-or French-ACC speak.PRESENT ‘John speaks Spanish or French.’ B: Bill-mo sore-o hanasu. Bill-also it-ACC speak.PRESENT ‘John speaks the one that John speaks.’ The pronoun sore ‘it’ and an overt counterpart of the disjunctive E-type reading can appear in the ellipsis site in (23). (28) A: John-wa speingo-ka furansugo-o hanasa-nai. John-TOP Spanish-or French-ACC speak-NEG.PRESENT. 6 (26) is also a counter example to Takahashi (2008) arguing that Japanese obeys Scope Economy and Parallelism. I leave open the issue of why Scope Economy and Parallelism do not capture (26). Verb-stranding ellipsis by blocking verb-raising 341 ‘John does not speak Spanish or French.’ B: Bill-mo { pro / sore-o / John-ga hanasa-nai-no-o} Bill-also it-ACC John-NOM speak-NEG-PRESENT-LN-ACC hanasa-nai. speak-NEG.PRESENT Lit.: ‘Bill does also not speak {it/the one that John does not speak}.’ The available interpretation in (28B) is that Bill does not speak one of them that John does not speak, which is equivalent to the wide-scope reading of the disjunction phrase in that Bill does not speak one of them but he speaks another. In addition, the unavailability of the narrow-scope reading of the disjunction phrase reported in (23) can be explained by pro. As discussed in the previous section and Funakoshi (2017), a plural-denoting pronoun sorera ‘these’ can refer to the conjunction/disjunction phrases and a sentence with sorera ‘these’ is truth-conditionallyequivalent to the narrow-scope reading of the disjunction phrase as in (29). (29) Mary-wa supeingo-ka furansugo-o hanas-u ga, John-wa Mary-TOP Spanish-or French-ACC speak-PRESENT but John-TSC sorera-o hanasa-nai. these-ACC speak-NEG.PRESENT ‘Mary speakes Spanish or French, but John does not speak these.’ (Finakoshi 2017) Crucially, that reading becomes inappropriate when the subject appears with -mo as in (30). (30) A: John-wa speingo-ka furansugo-o hanasa-nai. John-TOP Spanish-or French-ACC speak-NEG.PRESENT. ‘John does not speak Spanish or French.’ B: # Bill-mo sorera-o hanasa-nai. Bill-also these-ACC speak-NEG.PRESENT ‘Bill does also not speak these.’ Therefore, pro can explain the interpretative asymmetry of (23), assuming that pro is a covert counterpart of an overt pronoun. If this analysis is on the right track, it is predicted that we should see a contrast between the overt disjunctive phrase and the null-object when a provided context does not favor a reading where the overt disjunctive phrase is interpreted. The prediction is born out as in (31). (31) CONTEXT: Godzilla and King Ghidorah rampage nearby two castles, Inuyama castle and Gifu castle. We can see one of them did collapse, but we do not know which. A: Godzilla-wa Inuyama-jyo-ka Gifu-jyo-o kowasa-nakat-ta. Godzilla-TOP Inuyama-castle-or Gifu-castle-ACC break-NEG-PAST ‘Godzilla either did not break Inuyama Castle or Godzilla did not break Gifu Castle.’ 342 YOSHO MIYATA B1: # King Ghidorah-mo Inuyama-jyo-ka Gifu-jyo-o King Ghidorah-also Inuyama-castle-or Gifu-castle-ACC kowasa-nakat-ta. break-NEG-PAST ‘King Ghidorah either did also not break Inuyama castle or King Ghidorah did also not break Gifu castle.’ B2: King Ghidorah-mo pro kowasa-nakat-ta. King Ghidorah-also break-NEG-PAST ‘King Ghidorah did also not break the one Godzilla did not break.’ (31B1) sounds contradict to (31A) and the context. (B1) with an overt disjunctive phrase means that King Ghidorah either did not break Inuyama castle or did not break Gifu castle, not that King Ghidorah broke neither of them. But by uttering the disjunctive phrase, King Ghidorah should break one of them and it should be the same castle that Godzilla broke because of -mo. So King Ghidorah should breaks the castle that Godzilla broke again. But it is infelicitous because an event of breaking the castle does not happen multiply in natural context. On the other hands, (31B2) is appropriate because it only means that King Ghidorah did not break the one that Godzilla did not break. The contrast in (31) runs counter to the Argument Ellipsis analysis because the elliptical site should be unambiguously interpreted as disjunction. Along the same line of Funakoshi (2017), at least verbstranding ellipsis and pro are necessary to explain the null-object phenomena in Japanese. 5 Conclusion This paper argues for an existence of verb-stranding Ellipsis in Japanese by examining data where verb-raising is blocked by adverbial particle. The (un)availability of the null-adjunct reading is accounted for straightforwardly under verb-stranding XP-ellipsis that a complement is elided that verb raised to. In addition to this, the proposed analysis successfully captures the interpretative asymmetry of disjunction and negation observed in Shibata (2015a); Shibata (2015b). This paper also shows that verb stranding ellipsis and pro can explain all the data examined in this paper. Along the same line of Funakoshi (2016); Funakoshi (2017), the analysis cannot be maintained in Japanese that the Argument Ellipsis is only one available option of ellipsis. As an implication of this paper, verb-raising should exist in Japanese, on the contrary to literatures arguing against it (Fukui & Takano (1998), Hoji (1998), Fukui & Sakai (2003), and among others). When adverbial particles prevent verbraising, the null-adjunct reading becomes unavailable. If the verb-raising does not exist in Japanese at all, the interpretative asymmetry of null-adjunct reading between an existence of adverbial-particle and its absence should not show up. Verb-stranding ellipsis by blocking verb-raising 343 References Aoyagi, H. 1996. On the Nature of Particles in Japanese and Its Theoretical Implications. University of Southern California dissertation. Aoyagi, H. 2006. Nihongo No Joshi to Kinoohanchuu (Particles and Functional Categories in Japanese). Tokyo: Hituzi Syobo. Fox, D. 2000. Economy and Semantic Interpretation.. Linguistic Inquiry Monographs: 35. Cambridge, MA: MIT Press. Fukui, N., & H. Sakai. 2003. The visibility guideline for functional categories: Verb raising in Japanese and related issues. Lingua 113. 321–375. Fukui, N., & Y. Takano. 1998. Symmetry in Syntax: Merge and Demerge. Journal of East Asian Linguistics 7. 27. Funakoshi, K. 2014. Syntactic Head Movement and Its Consequences. College Park: University of Maryland dissertation. Funakoshi, K. 2016. Verb-stranding verb phrase ellipsis in Japanese. Journal of East Asian Linguistics 25. 113. Funakoshi, K. 2017. Disjunction and Object Drop in Japanese. Florida Linguistics Papers 4. Goldberg, L. M. 2005. Verb-Stranding VP Ellipsis : A Cross-Linguistic Study. Montréal: McGill University dissertation. Goro, T. 2007. Language-Specific Constraints on Scope Interpretation in First Language Acquisition. College Park: University of Maryland dissertation. Gribanova, V. 2017. Head movement and ellipsis in the expression of Russian polarity focus. Natural Language & Linguistic Theory 35. 1079–1121. Hoji, H. 1985. Logical form constraints and configurational structures in Japanese. Doctoral dissertation, University of Washington . Hoji, H. 1998. Null Object and Sloppy Identity in Japanese. Linguistic Inquiry 29. 127–152. Hoji, H. 2003. Falsifiability and repeatability in generative grammar: A case study of anaphora and scope dependency in Japanese. Formal Japanese syntax and universal grammar: the past 20 years 113. 377–446. Johnson, K. 2001. What VP Ellipsis Can Do, and What it Can’t, But Not Why. The Handbook of Contemporary Syntactic Theory 439–479. Kobayashi, Y. 2014. The dummy verb insertion in Japanese: A selection- and feature-based approach. Master’s thesis, Meiji Gakuin University, Tokyo. Kuroda, S. 1965. Generative Grammatical Studies in the Japanese Language. Cambridge, MA: MIT dissertation. Lobeck, A. 2006. Ellipsis in DP. The Blackwell companion to syntax Vol. 2. 145–173. Maeda, M. 2019. Argument Ellipsis and Scope Economy in Japanese. Syntax 22. 419–437. Merchant, J. 2006. Sluicing, volume Vol.4, 271–291. New York: Blackwell Publishing. Oku, S. 1998. A Theory of Selection and Reconstruction in the Minimalist Perspective. Storrs: University of Connecticut dissertation. Otani, K., & J. Whitman. 1991. V-Raising and VP-Ellipsis. Linguistic Inquiry 22. 345–358. Saito, M. 2007. Notes on East Asian argument ellipsis. In Language Research 43, 203–227. Sakamoto, Y. 2016. Scope and Disjunction Feed an Even More Argument for Argument Ellipsis in Japanese. In Japanese/Korean Linguistics 23, ed. by Michael Kenstowicz, Theodore Levin, & Ryo Masuda, Stanford, CA. CSLI Publications. Shibata, Y. 2015a. Exploring Syntax from the Interface. Storrs: University of Connecticut dissertation. Shibata, Y. 2015b. Negative structure and object movement in Japanese. Journal of East Asian Linguistics 24. 217–269. Simons, M. 1996. Disjunction and Anaphora. Semantics and Linguistic Theory 6. 245–260. Takahashi, D. 2008. Quantificational Null Objects and Argument Ellipsis. Linguistic Inquiry 39. 307–326. Proceedings of CLS 58 (2022), 345-354 © Chicago Linguistic Society 2023. All rights reserved. 345 Optionality and variation in Camuno interrogatives: a syntactic analysis Anda Neagu York University (Toronto) 1 Introduction Camuno is a dialect of Eastern Lombard, an endangered language (Moseley, 2010) spoken in Northern Italy in the provinces of Bergamo and Brescia, and the northern areas of the provinces of Cremona and Mantua. As far as Eastern Lombard is concerned, it is spoken North of the La Spezia – Rimini isogloss, which divides the Gallo-Italic group, in the North, from the Italo-Romance group in the South (Biondelli, 1853; Tamburelli & Brasca, 2018). This makes Eastern Lombard a Gallo-Italic variety that presents substantial differences in its syntax when compared to standard Italian, an Italo-Romance variety. When it comes to A-bar phenomena, they are similar to those observed in other Northern Italian vernaculars, such as Veneto. Specifically, its syntax is characterized by a high degree of variation and optionality, two features that also characterize interrogative structures within this variety. This also applies to Camuno, an Eastern Lombard dialect spoken in Valle Camonica, an alpine valley in the province of Brescia. When it comes to interrogative constructions, there are three main strategies available, namely wh-doubling, wh-fronting, and a clause-internal whP. These strategies display some variation between the North and the South of the valley, with Monnese, a variety from the North, allowing for optionality between the three strategies listed above, and Darfense, a variety from the South, allowing for optionality only between wh-fronting and a clause-internal whP. The article will focus on these phenomena by taking into consideration the relationship with standard Italian, as well as the variation and optionality characterizing Camuno varieties. The next section will take a look at the relationship with standard Italian; section 3 will introduce the data and the main properties associated with Camuno interrogatives, while section 4 will provide an analysis of these phenomena; section 5 is dedicated to a tentative sociolinguistic account of optionality; finally, section 6 concludes. 2 Relationship between standard Italian and local languages Standard Italian can be defined as the national language of Italy and is based on the Romance vernacular spoken around the Florence area during the Middle Ages (Serianni, 2015). In short terms, at the basis of standard Italian we find the Florentine variety employed in the XIII and XIV centuries by authors such as Dante, Petrarch, and Boccaccio, among others. This vernacular has remained the language of culture in the subsequent centuries, with certain works, notably the Divine Comedy, being considered the benchmark of the Italian language and literature (Migliorini, 2001). Because of the historical and literary importance of 346 ANDA NEAGU this Romance vernacular over the others, the Florentine dialect was chosen as the national language after the Italian unification in 1871. Nevertheless, standard Italian has only started being employed in a more widespread manner around mid-XX century (Berruto, 2017). Despite previous efforts to unite Italy not only politically, but also linguistically, it is only after WWII that those efforts saw some results. Until then, the vast majority of the population would employ their local languages, a tendency that has been reverted in the past decades, with standard Italian substituting the other vernaculars (Berruto, 2017). This process has brought to what has been analyzed as either a diglossic or a multilingual relationship between standard Italian (high code) and local languages (low code) (Tamburelli, 2010). Today, most speakers employ regional standards, i.e. standard Italian with heavy influences from local vernaculars. This is also the case in the areas in which the data for this project was collected, meaning not only that I had to carefully select the participants for the elicitation task, but also that the data may be affected by the linguistic landscape that was just described. In other words, I believe that part of the data, namely the optionality observed for A-bar phenomena, can be explained not only by recurring to a formal analysis, like the one provided in the following sections, but also by observing the sociolinguistic situation of the area under consideration, in this case Valle Camonica. The following sections, as well as most of the article, will however focus on the formal syntactic analysis of Camuno interrogatives. 3 Data and main properties The following examples show the two main tendencies in Camuno interrogative patterns: example (1) illustrates interrogatives in Monnese, which allows for wh-doubling (1c), as well as clause-internal whPs (1b) and wh-fronting (1a); while example (2) illustrates interrogatives in Darfense, which only allows the latter two strategies, i.e. wh-fronting (2a) and a clause-internal whP (2b). (1) a. ‘Ngo l’ e-t vist? where CL.3SG have.PRES=CL.2SG seen b. L’ e-t vist ngont ? CL.3SG have.PRES=CL.2SG seen where c. ‘Ngo l’ e-t vist ‘ngont ? where CL.3SG have.PRES=CL.2SG seen where ‘Where did you see him?’ (2) a. Quando l’ e-t ciapada? when CL.3SG have.PRES=CL.2SG got (Monnese) Optionality and variation in Camuno interrogatives 347 b. L’ e-t ciapada quando? CL.3SG have.PRES=CL.2SG got when ‘When did you get it?’ c. Te mete-t do ke? you put.PAST=CL.2SG down what ‘What did you plant?’ Let us start by focusing on the word order in these interrogatives, which is a first indicator of the syntactic distribution of the wh-elements. Examples (1a) and (2a) pattern together in that the whP is fronted, thus appearing at the beginning of the interrogative. This is followed by an object clitic, which in turn precedes the auxiliary verb, to which the subject clitic is cliticized in an enclitic configuration. Finally, the subject is followed by the past participle. The difference with (1b) and (2b; 2c)1 resides in the fact that, in this case, the whP is clause-internal, instead of fronted, thus appearing after the past participle in terms of linear order. As far as doubling constructions (1c) are concerned, these behave similarly, with the peculiarity of spelling-out two wh-items: one fronted and one clause-internal. In sum, in Camuno interrogatives, the wh-items may occupy a fronted position, a clause-internal one, or both, depending on the variety and whether or not it allows for the doubling strategy altogether. It is also interesting to notice, however, that the inflected verb also raises, with the clitic subject attaching to it enclitically. The examples in (1) and (2) also illustrate the general tendencies in interrogative patterns that can be found in Camuno. In particular, example (1) shows that the doubling strategy also presupposes wh-fronting and a clause-internal whP; example (2), on the other hand, indicates that a clause-internal whP also presupposes wh-fronting. This, however, tends to apply to adjunct whPs (2a; 2b), while argument ones (2c) tend to favour the clause-internal distribution with a hypothetically fronted one being perceived both by the linguistic consultants and the author as a more italianized version. Lastly, only certain wh-categories allow for doubling, even in varieties where this strategy occurs. In particular, in the case of Monnese, doubling is only possible with the argument whP ke/kwé ‘what’, and the adjunct whP ‘ngo/’ngont ‘where’. Now that the main distributional patterns of wh-items have been looked at more attentively, I will switch the focus on the main properties of these constructions. First of all, there is a rigid distribution of the wh-elements when there are two forms, such as in doubling structures. This can be observed below in example (3): (3) a. ‘Ngo se-t nasciut ‘ngont ? where be.PRS=CL.2SG born where (Monnese) 1 Example (2b) also includes a fronted, phonologically independent subject, in addition to the cliticized one. 348 ANDA NEAGU ‘Where were you born?’ b. *’Ngont se-t nasciut ‘ngo? where be.PRS =CL.2SG born where ‘Where were you born?’ Example (3) shows that the two wh-items in a doubling construction cannot be freely ordered, so that the shorter form appears clause-internally and the longer one appears fronted, as in (3b). The only acceptable word-order involves a fronted shorter form, and a clause-internal longer form. I take this to mean that the two wh-elements each play a different syntactic role within the sentence, as it will be better shown in section §4. Another property worth mentioning has already been introduced earlier in this section, and it involves the presence of subject-clitic inversion. This can be observed in all the direct interrogatives in the previous examples: the subject clitic attaches enclitically to the inflected verb. What may appear, at a first glance, as an inflectional morpheme, has been extensively analyzed as a subject clitic attaching at the end of the verb (Benincà et al., 2016). Finally, as far as the clause-internal wh-element is involved, this is required to be right-adjacent to the lexical verb, as observable in the following interrogatives involving ditransitives (4): (4) a. K’ e-t dat kwé a Paolo al sera? what have.PRS=CL.2SG given what to Paolo the evening (Monnese) b. *K’ e-t dat a Paolo kwé al sera? what have.PRS =CL.2SG given to Paolo what the evening ‘What did you give to Paolo yesterday evening?’ c. E-t te scrit a ki esta letera se bé ? have.PRS =CL.2SG you written to whom this letter so well d. *E-t te scrit esta letera a ki se bé ? have.PRS=CL.2SG you written this letter to whom so well ‘To whom did you write this letter so carefully?’ Example (4) shows that interrogatives are grammatical only when the clause-internal wh-item surfaces right-adjacent to the lexical verb, as in (4a) and (4c). It does not matter whether the whP is the direct or the indirect object: whenever their final landing site is clause-internal, it must follow the lexical verb in a right-adjacent position. This becomes even clearer when looking at ditransitive constructions in declarative clauses, where the direct and indirect object can be freely reordered, as observable in (5): (5) a. L’ a dat a Paolo esta letera. (Monnese) Optionality and variation in Camuno interrogatives 349 CL.3SG have.PRS given to Paolo this letter b. L’ a dat esta letera a Paolo. CL.3SG have.PRS given this letter to Paolo ‘S/he gave Paolo a letter.’ Having introduced the syntactic patterns of Camuno interrogatives and their main properties, the next sections will be dedicated to the analysis of these structures and the evidence supporting it. 4 Analysis The analysis I propose for these constructions relies on the concept of partial copying, as introduced by Barbiers et al. (2008; 2010), who apply this analysis to non-identical doubling in Dutch dialects. A first difference between my analysis and this is the fact that, while the authors apply this analysis to long-distance dependencies, I will mostly focus on canonical interrogatives. That is, the author is aware of and acknowledges the main differences between Camuno interrogatives and the same constructions in Dutch. Furthermore, the main idea I adopt from Barbiers et al. (2008; 2010) is that partial copying targets a specific node of a layered nominal projection. Thus, given the structures observed so far for the doubling configuration, I suggest the following steps in the derivation: 1. the two wh elements enter the derivation as a layered XP: [whP kwé [QP ke]]. 2. partial copying targets only the operator portion [QP ke], which raises to the CP. 3. the restriction [whP kwé] is realized as an intermediate copy stranded at the edge of vP. This analysis, however, does not only apply to the doubling constructions, but also to the ones involving wh-fronting and a clause-internal whP. In particular, I suggest that optionality arises when only one element is pronounced at PF, while the other stays silent, being only interpreted at LF. In other words, I suggest that a unified account of the three strategies introduced thus far is possible, and that both the fronted and the clause-internal wh-elements are always interpreted by the semantic component. My analysis relies on the concepts of successive-cyclicity and phasal boundaries. In other words, Camuno wh-items occupy phasal boundaries after undergoing successive-cyclic movement to these areas. Phases, in this case CP and vP, are cyclic points of transfer to the interfaces (Chomsky, 2008). As such, the elements constituting the phasal domain that do not want to be transferred to the SEM and PHON components, employ the phasal boundary as an escape hatch (Chomsky, 2001). These phasal boundaries are often occupied by operators, such as whPs, among other elements, and have been analyzed as being more or less 350 ANDA NEAGU detailed depending on the theoretical framework one adopts. According to Cartographic approaches (Rizzi, 1997; Cinque, 1999; Belletti, 2004), these domains are highly articulated, with different functional projections hosting syntactic material. Minimalist approaches, on the other hand, tend to rely on the concept of multiple specifiers (Chomsky, 2008; Boeckx, 2012). Whichever framework one adopts, a core concept stands: the object of the phasal projection needs to be transferred to the interfaces, and the wh-items whose final landing side is outside of the domain of the phase in which they e-merge, need to make use of the escape hatch offered by the phasal boundary. These are the main theoretical premises I rely on for the analysis outlined previously in this section, and whose derivation follows the following steps: (6) CP[Ke fe-t fa vP[whP kwé [QP ke]] te [whP kwé [QP ke]]]? I analyze both ke and kwé as being part of the same A-bar chain, with the fronted element being analyzed as the operator, and the clause-internal one as the restriction of the chain2. I thus suggest that a unified analysis is possible, and that, in doubling constructions, both wh-elements are phonetically realized, in wh-fronting interrogatives only the operator is spelled out, while in “clause-internal structures” it is only the restriction that is overtly realized. The elements lacking phonological content are interpreted at LF. The following section focuses on the evidence supporting this analysis. 4.1 Evidence The first piece of evidence confirming the syntactic position of the fronted element comes from subject-clitic inversion structures. In these constructions, the inflected verb has been argued to raise into C (see Poletto, 1993 for Northern Italian dialects and Rizzi & Roberts, 1989 for French). This can be observed in example (7), where the operator is followed, in terms of linear order, by the verb to which the subject-clitic is attached. Given this premise, if the inflected verb occupies C, then the whP also occupies a position within this functional field, i.e. spec,CP. (7) Ke fe-t fa kwé te? what do.PRS=CL.2SG do what you ‘What are you doing?’ (Monnese) Another piece of evidence comes from the syntactic distribution of the restriction, which has been shown to occupy a position right-adjacent to the lexical verb. The latter, which raises from V into C when finite, raises as far as the IP domain when A similar approach to doubling has been adopted by Beck (1996) and Miyagawa (2004) to describe doubling in German interrogatives. 2 Optionality and variation in Camuno interrogatives 351 a past-participle. This can be especially observed in the presence of an aspectual adverb, which has been argued to be low in the IP domain (Cinque, 1999), as in (8). Once again, if the verb occupies a position within the IP domain, and the restriction must be right-adjacent to it3, the restriction has then raised at the edge of vP. (8) Al beker l’ ha dat amò ke a la hʧèta? the butcher CL.3SG have.PRS given again what to the girl ‘What did the butcher give again to the girl?’ (Darfense) A third behavior that constitutes evidence for my analysis is sensitivity to islands, as in (9). In particular, I take the impossibility to form canonical negative interrogatives as an indication that the NEG operator constitutes a barrier (Cinque, 1990; Rizzi, 1990) that blocks “communication” between the operator and its restriction, thus indicating that these two elements are part of the same A-bar chain. (9) a. *Ke fe-t miga majà kwé? what do.3SG.PRS=CL.3SG NEG eaten what ‘What hasn’t he eaten?’ b. *Maje-la mia ke la ho hpuda? eat.PRS=CL.3SG NEG what the his wife ‘What doesn’t his wife eat?’ (Monnese) (Darfense) Finally, the element analyzed as the operator portion functions as an operator in other A-bar dependencies, as shown in the free relative clause in (10). I take this to indicate that this is indeed the operator portion of the A-bar chain in interrogative constructions as well, and that the remaining wh-item does not possess this quality. (10) Varda ‘ngo te mete-t i pé. (Monnese) look.IMP where you put.PRS=CL.2SG the feet ‘Watch where you’re walking.’ (lit. ‘Look where you put the feet.’) To sum up the analysis, it was shown that wh-elements occupy phasal boundaries in Camuno interrogatives. In particular, they undergo successive-cyclic movement through the vP and the CP phasal domains, with the restriction being stranded at the edge of vP, and the operator raising into spec,CP. 5 A note on optionality Only certain adverbials and discourse particles can intervene between the lexical verb and the whP. 3 352 ANDA NEAGU The two varieties of Camuno analyzed so far, Darfense and Monnese, were said to be representative of the general patterns observable in Camuno interrogatives. In particular, optionality between doubling, fronting and clause-internal whPs seems to be internal to the grammar of Monnese. In other words, there seems to be no semantico-pragmatic or any other trigger that would make one strategie more appropriate than another in a given context. At the same time, at least some of the phenomena concerning optionality could be ascribed to the relationship with standard Italian that was the focus of section §2. In particular, I believe that optionality between clause-internal whP and wh-fronting in both Monnese (11) and Darfense (12) can also be explained by socio-linguistic factors: (11) a. Come fa-l tʃamà al to pader? how do.PRS=CL.3SG call the your father (Monnese) b. Fa-l tʃamà com al to pader? do.PRS=CL.3SG call how the your father ‘What’s your father’s name? (12) a. L’ e-t tʃapada CL.3SG be.PRS=CL.2SG gotten ‘When did you get the car?’ quando la machina? (Darfense) when the car b. Quando l’ e-t tʃapada la patente? when CL.3SG be.PRS=CL.2SG gotten the driving licence ‘When did you get the driving licence?’ In particular, the contrast between (a) and (b) in (11) shows that, when fronted, the wh-word come ‘how’ is phonologically the same as its standard Italian counterpart, and it also occupies the canonical position of whPs in Italian interrogatives, i.e. spec,CP. At the same time, the clause-internal strategy in (11b) is in line with the phonological form found in Monnese. In example (12), on the other hand, it is the topic of the conversation that I believe to be the trigger for the fronted distribution: whenever the topic of the conversation was more informal, e.g. concerned the car of the informant as in (12a), the strategy employed was clause-internal, while when the topic became more formal, e.g. getting a driver’s license as in (12b), the strategy employed was wh-fronting like in standard Italian. I take all of this to mean that the high degree of optionality observed in Camuno interrogatives can also be explained sociolinguistically, in particular through the contact with standard Italian, which is the official language of Italy and the language learned in school by the informants. These observations, however, require a dedicated and a more in-depth research, that is currently out of the scope of the current project. 6 Conclusions Optionality and variation in Camuno interrogatives 353 Successive-cyclicity has been argued to concern not only long-distance dependencies, but also less complex constructions (van Urk, 2020), such as the structures under consideration. In particular, if on the right track, this analysis, together with the Camuno data constitute further evidence supporting theories of A-bar movement as phasal and successive-cyclic. References Barbiers, Sjef, Olaf Koeneman, Marika Lekakou & Margreet van der Ham (eds.). 2008. Microvariation in syntactic doubling. Leiden: Brill. Barbiers, Sjef, Olaf Koeneman & Marika Lekakou. 2010. Syntactic doubling and the structure of wh-chains. Journal of Linguistics, 46(1). 1-46. Beck, Sigrid. 1996. Quantified Structures as Barriers for LF Movement. Natural Language Semantics 4. 1-56. Belletti, A. 2004. Aspects of the low IP area. In Rizzi, L. (ed.) The Structure of CP and IP: The Cartography of Syntactic Structures, vol. 2, pp. 16-51. Oxford University Press. Benincà, P., Parry, M., & Pescarini, D. 2016. The dialects of northern Italy. In The Oxford Guide to the Romance Languages. Oxford University Press. Oxford, pp. 185–205. Berruto, G. 2017. What is changing in Italian today? Phenomena of restandardization in syntax and morphology: An overview. In M. Cerruti, C. Crocco, & S. Marzo (Eds.), Towards a new standard: Theoretical and empirical studies on the restandardization of Italian (pp. 31-60). Walter de Gruyter. Biondelli, B. 1853. Saggio sui Dialetti Gallo-Italici. Bernardoni. Boeckx, C. 2012. Syntactic islands. Cambridge University Press. New York. Chomsky, N. 2001. Derivation by Phase. In Kenstowicz, M. (ed.) Ken Hale: A Life in Language. MIT Press. Cambridge, Mass. Chomsky, N. 2008. On Phases. In Freidin, R.; Otero, C. P.; Zubizarreta, M. L. (eds.) Foundational Issues in Linguistic Theory: Essays in Honor of Jean-Roget Vergnaud. MIT Press. Cambridge, Mass. (First appeared in 2005). Cinque, G. 1990. Types of A’-Dependencies. MIT Press. Cambridge, Mass. Cinque, G. 1999. Adverbs and functional heads. Oxford University Press. Migliorini, B. 2001. Storia della lingua italiana. IX ed. Tascabili Bompiani. Miyagawa, S. 2004. The nature of weak islands. MIT MS. Moseley, C. (ed.). 2010. Atlas of the World’s Languages in Danger. UNESCO Publishing. Poletto, Cecilia. 1993. Subject Clitic/Verb Inversion in North Eastern Italian Dialects. University of Venice Working Papers in Linguistics 3(1). 95-137. Rizzi, Luigi & Ian Roberts. 1989. Complex Inversion in French. Probus 1(1). 1-30. Rizzi, L. 1990. Relativized Minimality. MIT Press. Cambridge, Mass. Rizzi, L. 1997. The Fine Structure of the Left Periphery. In Haegeman, L. (ed.) Elements of Grammar. Springer. Dordrecht, pp. 281-337. Serianni, Luca. 2015. Prima lezione di storia della lingua italiana. Editori Laterza. Tamburelli, M. 2010. The vanishing languages of Italy: Diglossia, bilingualism, and language shift. In Languages in contact 2010 Piotr P. Chruszczewski & Zdzisław Wąsik (eds.). Philological School of Higher Education in Wroclaw Publishing. Tamburelli, M., & Brasca, L. 2018. Revisiting the classification of Gallo-Italic: A dialectometric approach. In Digital Scholarship in the Humanities, 33(2), pp. 442–455. van Urk, C. 2020. Successive Cyclicity and the Syntax of Long-Distance Dependencies. In Annual Review of Linguistics 6, pp. 111-130. Proceedings of CLS 58 (2022), 355-373 © Chicago Linguistic Society 2023. All rights reserved. 355 Representing Irish Mutations in Distributed Morphology and Optimality Theory* Jack Pruett Georgetown University 1 Abstract This research develops a novel approach to Irish Consonant Mutation (ICM) by framing these mutations under the theories of Distributed Morphology (Halle & Marantz 1993) and analyzing their effects using Optimality Theory (Prince & Smolensky 1993/2002/2004). Irish Gaelic (henceforth Irish), as well as the other Modern Celtic languages, have a variety of initial phonological mutations which occur in morphosyntactic (and not necessarily phonological) contexts (Fife 2010, Sims-Williams 1998, i.a.). The primary focus of this research is on the mutations that affect word initial consonants. In this paper I show that these mutations are morphemes whose vocabulary items are comprised of autosegmental phonological feature bundles (similar to what has been argued for by Ní Chiosáin 1991, Pyatt 1996, a.o.). These phonological feature bundles are realizations of various inflectional and possibly derivational morphemes which surface as mutation when they interact with other segments in their phonological environments. This analysis permits a simpler system for storing morphemes in the lexicon—one that does not require mutation specifications to be a property of the words that serve as ‘triggers’ for mutation; nor is there a need to appeal to syntactic structure to derive mutation. 2 Introduction Irish Mutation occurs in both the verbal and nominal domains. In most cases these mutations are realizations of morphological inflection, which historically took the form of clitics or affixes. These affixes have since lost their phonological segments and presently only phonological features remain to express those morphemes (Sims-Williams 1998, Fife 2010, Ó Baoill 2010, Conroy 2008, a.o.). Synchronically, these morphemes surface as mutation (Lenition or Eclipsis). The following research strives to reinforce this claim and further explain how these phonological alternations can be represented in the lexicon as their own morphemes/Vocabulary * I would like to thank and acknowledge Matthew Barros and Brett Hyde of Washington University in St. Louis for their help in advising me on this research, which is based on my undergraduate thesis (Pruett 2021). Additionally, I would like to thank Ruth Kramer and Michael Obiri-Yeboah of Georgetown University for their continued support and feedback on this research, as well as the Syntax Reading Group of Georgetown University Linguistics Department and especially Zhousi Luo for the feedback given to me on this paper. Lastly, I would like to thank my Irish teachers Sarah Johnson and Nóirín Ní Shuilleabháin for their input and for their classes from which I took some of the data presented here. 356 JACK PRUETT Items rather than some phonological property associated with lexical items (SimsWilliams 1998, Ó Baoill 2010, Fife 2010, a.o.) or a process driven by the syntax (Duffield 1997, Iosad 2017, a.o.). Irish has two main ICM. The first is Lenition, a process in which a consonant is weakened. The primary result of Lenition is that stops become their corresponding fricatives at their given place of articulation or a fricative is debuccalized or deleted. The second mutation is Eclipsis, a process by which a voiceless consonant becomes voiced and/or a voiced consonant becomes its corresponding nasal at its place of articulation. These changes are shown in (1) for each phoneme both orthographically and in IPA. (1) These mutations, though, are not purely phonological in nature. Rather, these changes occur in specific morphosyntactic environments (Ó Baoill 2010, Stenson 2020, Duffield 1997, Iosad 2017, a.o.), like specific Tense/Aspect/Mood values in the verbal domain and specific combinations of Case/Number/Gender/Definiteness in the nominal domain.1 . (2) a. No mutation in the present tense tuig-eann sí Gaeilge understand-PRES 3. FEM . NOM . SG Irish. FEM . ACC . SG [t”G Ié@n”G Si: ge:lj é@] ‘She understands Irish’ 1 Data in this paper comes from a variety of sources. Cited data comes from and is inspired by Stenson 2020 and Ó Baoill 2010. In addition, there is data which comes from examples provided in Irish Language courses I have taken. Said data comes from two native speakers of Munster Irish and Connacht Irish. Representing Irish mutations 357 b. No mutation in the feminine nominative singular without the definite article bean [bj æn”G ] woman. FEM . NOM . SG ‘a woman’ c. No mutation in the masculine nominative singular fear [fj æRG ] man. MASC . NOM . SG ‘a man’ d. Lenition in the past tense thuig sí Gaeilge [hIé Si: ge:lj é@] understand. PAST 3. FEM . NOM . SG Irish. FEM . ACC . SG ‘She understood Irish’ e. Lenition after the definite article for feminine nouns in the nominative singular an bhean [@n”G Bj æn”G ] the.DEF woman. FEM . NOM . SG ‘the woman’ f. Eclipsis in the genitive plural na bhf ear [n”G @ vj æRG ] the.DEF. PL man. MASC . GEN . PL ‘the men’s’ There are many other situations similar to those given above in (2) which will be further discussed in sections §2 and §3. The remainder of this paper proceeds as follows: sections §2 and §3 I will discuss the most common situations in which Lenition and Eclipsis apply. Then, in section §4, I will lay out my current analysis and I briefly address previous approaches in describing these phenomena. . I will build off of the previous literature on ICM and construct a framework that unifies some previous approaches under the theories of Distributed Morphology and Optimality Theory. Section §5 concludes. 3 Lenition Lenition involves two phonological processes. The first process changes the continuancy feature of a consonant (generally this takes the form [-cont.] becomes [+cont.]). This is the change observed in the labial, palatal, and velar stops2 . The second process is one of debuccalization. This primarily targets coronal consonants and the phoneme /f/. The result is that voiceless coronals become /h/, /f/ undergoes a process of deletion, and /d/ becomes a dorsal fricative (see the examples in 2 Coronal stops also undergo this change in continuancy. However, they undergo debuccalization as well. As such, they are more complex and do not fit well into this generalization. 358 JACK PRUETT (1) for a correspondence between the lenited and unlenited forms). For now, the reader should just be aware that the following generalizations broadly capture the Irish data, but do not account for all the possible alternations. So, Lenition can be captured in the following SPE rules (Chomsky & Halle 1968). (3) Lenition a. [-cont.] [+cont.] / Lenition b. [+cor.] [-cor.] / Lenition Now let us examine some contexts in which Lenition arises in Irish. Lenition in Irish occurs in both the nominal and verbal domains, as stated previously. Generally speaking, Lenition is used in the nominal domain to realize Gender, Number, and Case on nouns and adjectives. In the verbal domain, it is used to denote Tense and possibly Mood3 . Consider the examples in (4) for instances of Lenition in the Irish Verbal Domain (recall that an h after the initial consonant indicates Lenition). (4) Lenition in the verbal domain a. Past Tense: Perfective chuaigh sé siar [xu@ Se: Si@RG ] go.PAST 3. MASC . NOM . SG west ‘he went westward’4 b. Past Tense: Imperfective cheann-aíodh sé bláth-anna [çæn”G i: Se: bG ëA:h@n”G @] buy.PAST-IMPFV 3. MASC . NOM . SG flower-FEM . ACC . PL ‘he would buy flowers’ c. Conditional Mood fhan-fadh sí [An”G @ Si:] wait.PAST-COND 3. FEM . NOM . SG ‘she would wait’ d. After Negation ní thuig-im Gaeilge [nj i: hIé@mj ge:lj é@] NEG understand- PRES .1. SG Irish. FEM . ACC . SG ‘I don’t understand Irish’ e. Relative Clauses an múinteoir a fheic-im [@n”G mG u:nj tj oRj @ ec@mj ] DEF teacher. MASC . NOM . SG REL see- PRES .1. SG ‘the teacher (whom) I see...’5 3 The morphosyntactic distinctions between Tense, Aspect, and Mood in Irish are too complex to discuss in this paper. For the purposes of this paper, I will focus primarily on Tense distinctions and leaving open the question of how Tense interacts with Aspect, Mood, and mutation in Irish. 4 Data from Ó Baoill 2010. 5 Data in examples b, c, and e are from Stenson 2020. Representing Irish mutations 359 As shown in the examples in (4), Lenition occurs predominantly in the Past Tense, after negation, and in relative clauses. More specifically, regardless of aspect or mood, all verbs in the Past Tense, whether in a main clause or an embedded clause, occur in the lenited form6 . Given this generalization, it is reasonable to conclude that the realization of the Past Tense in Irish is, in fact, Lenition and nothing more (aspect and sometimes mood are realized by suffixes). As such, there must be some Past Tense morpheme that results in the surface realization of Lenition (see section §4.2). Furthermore, this analysis can be extended to the negation and relative clause data given in (4) as well. Having discussed Lenition in the verbal domain, I turn now to Lenition in the nominal domain. Lenition in the nominal domain is much more diverse and expansive than in the verbal domain. That said, all instances of Lenition in the nominal domain are the result of Gender, Number, Case, and Definiteness interactions. All nouns (and pronouns) in Irish have Grammatical Gender (masculine or feminine)7 , Number (singular or plural), and Case (nominative/accusative, genitive, and dative/prepositional)8 . Take the examples in (5) of Lenition with feminine nouns. (5) Lenition with feminine nouns a. After the definite article in the nominative/accusative singular an bhróg [@n”G BG RG o:g] DEF shoe. FEM . NOM / ACC . SG ‘the shoe’ b. Post-nominal adjectives/other attributives in the nominative/accusative singular bróg shoe.FEM . NOM / ACC . SG bheag/chapaill[bG RG o:g Bj æg/xApG I´] small.FEM . NOM / ACC . SG/horse.MASC . GEN . SG 6 This applies only for consonant initial verbs. For vowel initial verbs, the Past Tense surfaces as a prefix of the form d’. Interestingly enough, this prefix also occurs for lenited verbs that begin with /f/. Presumably, since lenited /f/ is silent, the d’ prefix is added because the deletion of /f/ renders the verb vowel initial in the Past Tense. This interaction between /f/ and the prefix will not be investigated in this paper given space limitations, but this phenomenon should be at least mentioned. 7 I use Gender in this paper as defined by Kramer 2015. Specifically, the classification of nouns as either masculine or feminine (or any other gender that the language utilizes) is grounded in a semantic core (which for human denoting nouns in many cases can correspond to social gender identity). It is possible that for some human denoting nouns the grammatical gender assigned to that noun does not necessarily correspond with the social gender identity of the person. For example, the word cailín ’girl’ in Irish is grammatically masculine (Kane et al. 2016). Outside of this core, in some languages, nouns can be assigned grammatical gender arbitrarily. The reader should thus keep in mind that Gender is discussed in its grammatical sense throughout this paper and it does not necessarily correspond to the social gender identity of any human or animate nominal. 8 For nouns, the nominative and accusative cases have collapsed into one and there is no longer any distinction between the two. However, for pronouns there is still a nominative/accusative split. Additionally, the dative case has been referred to elsewhere in the literature as the prepositional case (Fife, Ó Baoill 2010). There is no formal reason for this differentiation. In this paper, I will use the dative to describe this case. 360 JACK PRUETT ‘a (small/horse) shoe’ As can be seen in (5), feminine nouns in the nominative/accusative singular lenite after the definite article. Additionally, all post-nominal modifiers also lenite if the noun is feminine in the nominative/accusative singular. In addition to these two cases of Lenition with feminine nouns, it should be noted that all nouns, regardless of Gender, lenite after the vocative particle a (6). This also goes for any indefinite noun after certain prepositions. (6) a. a Mháire/Sheán [@ mG A:Rj @/ça:n”G ] VOC Mary/Sean. MASC . NOM . SG ‘O, Mary/Sean’ b. do bhean/chailín [d”G @ Bj æn”G /xAlj i:nj ] for woman/girl.FEM / MASC . ACC . SG ‘with a woman/girl’ With respect to masculine nouns, Lenition surfaces in the following situations. (7) Lenition with masculine nouns a. After the definite article in the genitive singular an fhir [@n”G IRj ] DEF man. MASC . GEN . SG ‘of the man’ b. Post-nominal adjectives in the genitive singular and the nominative/accusative plural báid bheag-a [bG A:dj Bj æg@] boat.MASC . NOM / ACC . PL small-PL ‘small boats’ an fhir mhóir [@n”G IRj BG o:Rj ] DEF man. MASC . GEN . SG big. MASC . GEN . SG ‘of the big man’ As is evident from (7), masculine nouns undergo Lenition in the genitive singular after the definite article and cause Lenition on the following modifier in the genitive singular and nominative/accusative plural. It is clear from the data in (5) - (7) that Gender, Number, Case, and Definiteness are distinguished by Lenition in some instances in Irish. For example, in the nominative singular, feminine nouns undergo Lenition but masculine nouns do not. Similarly, in the genitive singular masculine nouns undergo Lenition after the definite article but feminine nouns do not. Finally, masculine nouns in the plural trigger Lenition on the following modifier, whereas feminine nouns do not. It should begin to become clear that Gender, Number, Case, and Definiteness morphemes in Irish can be taken to be Lenition under certain circumstances. And, much like the Past Tense discussed earlier, it should be possible to create morphemes that encode these morphosyntactic distinctions while still triggering Lenition in the phonology. Representing Irish mutations 361 Finally, before moving on to the other mutation in Irish (Eclipsis) pronouns and occurrences of Lenition must be discussed. Lenition occurs primarily in a single situation for pronouns. Specifically, the singular possessive pronouns trigger Lenition on the following noun (8). (8) a. mo mhadra [mG @ BG Ad”G RG @] 1. GEN . SG dog.MASC . NOM / ACC . SG ‘my dog’ b. do chat [d”G @ xAt”G ] 2. GEN . SG cat.MASC . NOM / ACC . SG ‘your cat’ c. a bhád [@ BG A:d”G ] 3. MASC . GEN . SG boat.MASC . NOM / ACC . SG ‘his boat’ BUT d. a madra [@ mG Ad”G RG @] 3. FEM . GEN . SG dog.MASC . NOM / ACC . SG ‘her dog’ A crucial observation to note at this point concerns the examples in (8c) and (8d). Interestingly enough, even the third person pronoun exhibits the same pattern with respect to Lenition as nouns do (first and second person pronouns make no morphological distinction for masculine and feminine); that is, the masculine pronoun triggers Lenition (8c) but the feminine pronoun does not (8d). This difference provides further evidence that a singular morpheme for the masculine genitive singular can be formulated such that it always results in triggering Lenition. For the plural pronouns, they trigger Eclipsis and they will be discussed in the next section. And so concludes this discussion on the occurrence of Lenition in Irish. Now let us transition to the other Mutation in Irish–Eclipsis. 4 Eclipsis Eclipsis also involves two phonological processes. The first process changes the nasal feature of a consonant (taking the form [-nasal] becomes [+nasal]). This is the change observed in voiced stops. The second process is one of voicing. This primarily targets voiceless consonants. The result is that a voiceless consonant becomes voiced (see the examples in (1) for a correspondence between the eclipsed and uneclipsed forms). Eclipsis can be captured in the following SPE rules. (9) Eclipsis a. [-voi.] [+voi.] / Eclipsis b. [-nas.] [+nas.]/[+voi.]Eclipsis Now that the phonological description of Eclipsis has been established, I turn to examining some contexts in which Eclipsis occurs. Eclipsis in Irish occurs in both 362 JACK PRUETT the nominal and verbal domains, as stated previously. Unlike Lenition, however, Eclipsis is much more limited in its distribution. In the verbal domain, Eclipsis solely occurs after complementizers and when the tense is not Past. Otherwise, in the nominal domain, Eclipsis happens in the dative singular9 and genitive plural (regardless of Gender). The examples in (10) show instances of Eclipsis in Irish after varying complementizers and for nouns of both Genders in the dative singular and genitive plural. (10) a. na bhf ear/mban DEF man/woman. MASC / FEM . GEN . PL [n”G @ vj æRG /mG An”G ] ‘of the men/women’ b. an bhf uil tú ag teacht COMP. Q COP 2. NOM . SG at come. NMLZ [@n”G wIlj ”tG u: Eg tj æxt”G ] ‘are you coming10 ?’ c. ceap-aim nach dté-ann sé think-PRES .1. SG COMP. NEG go-PRES 3. NOM . SG [cæpG @mj n”G Ax dj e:n”G Se:] ‘I think that he doesn’t go’ d. ar an mbord/gcapaill on DEF table/horse.FEM / MASC . DAT. SG [ARG @n”G mG ORG d”G /gApG I´] ‘on the table/horse’ As can be seen in the examples in (10), Eclipsis occurs on nouns of both Genders after the definite article in the genitive plural (10a) and dative singular (10d). Similarly, Eclipsis surfaces on the verb after the question (10b) and negative subordinate complementizers (10c). This shows that polarity (negative vs. affirmative) and force (declarative vs. interrogative) do not play a role in whether or not the verb that follows is eclipsed or not. That is to say, regardless of the type of complementizer, Eclipsis always occurs on a verb that follows a complementizer in non-past situations (this also holds true for other complementizers like go, dá, etc.). And so, much like in the discussion of Lenition, morphemes can be constructed such that when they are spelled out they result in Eclipsis. For example, a genitive plural morpheme could have a realization that results in Eclipsis (more on this in section §4). Further evidence that Eclipsis is the universal realization of the genitive plural comes from the plural possessive pronouns, which all trigger Eclipsis on the following noun (11). 9 Eclipsis only appears after the definite article, otherwise Lenition or no mutation applies depending on the preposition. 10 In the gloss, NMLZ stands for nominalized. This is because Irish does not have a participle equivalent to ’coming’ (this is true for all verbs). Instead, Irish employs a structure whereby the copula is followed by the subject and then a predicate of the form ag + verbal noun (the preposition ’at’ + nominalized verb). Representing Irish mutations (11) 363 a. ár gcat 1. GEN . PL cat.MASC . NOM / ACC . SG [A:RG gAt”G ] ‘our cat’ b. bhur dteach 2. GEN . PL house.FEM . NOM / ACC . SG [BG URG dj æx] ‘your house’ c. a mbád 3. GEN . PL boat.MASC . NOM / ACC . SG [@ mG A:d”G ] ‘their boat’ Now, having outlined the distribution of Lenition and Eclipsis in Irish, I will briefly discuss previous approaches to analyzing these facts and outline this novel approach. Given the length and purview of this paper a full literature review is not possible. As such, the discussion in the next section is not extensive, nor does it touch upon all analyses that have been done. For these reasons, I have provided a few analyses that I find are particularly relevant to the research at hand and leave discussion of further approaches and a lengthy comparison of my approach to said approaches up to future research. 5 Current Approach This is far from the first work on Irish mutations. A vast literature dating back centuries has grappled with and discussed these aspects of the Irish language. Many have worked on Irish mutations including analyses put forth by Pyatt 1996 & 2003, Ní Chiosáin 1991, Iosad 2017, Green 2006, Duffield 1997, Conroy 2008, among many others11 . That said, the point of this research is to further the body of literature on this topic and provide a new analysis under a DM and OT framework. Previous approaches to Irish mutation fall into two main groups: syntactic approaches and morphophonological approaches. Syntactic approaches, like those of Duffield 1997, Green 2006, and Conroy 2008, primarily focus on the relationship between functional heads in the syntax (like C, D, T, etc.) and mutation domains. It has been argued that the presence of these functional heads (usually overtly present–Duffield 1997) causes a syntactic process that results in mutation (Lenition for D and T, Eclipsis for C) provided that the functional head trigger c-commands the target of mutation. There is not enough space to address all of these analyses in turn, nor can I give rebuttal to each of these previous approaches to discuss the merits and drawbacks of each approach. That said, this analysis would have trouble accounting for data like that presented in (6) where mutation seems to occur without the overt presence of a functional head C, D, or T. 11 This literature also includes some of the more descriptive work done on Irish mutations like that of Hannahs 2011, Ó Baoill 2010, Stenson 2020, Pruett 2021, a.o. 364 JACK PRUETT The other type of approach usually taken for analyzing ICM is a morphophonological one. This type of analysis can take many forms depending on the theory chosen for analysis. For example, some have argued that mutation is a lexical property of the words that serve as triggers for mutation (Iosad 2017, Ó Baoill 2010, etc.). Yet, others have argued that mutation is a realization of autosegmental morphemes comprised of phonological feature bundles (Ní Chiosáin 1991, Pyatt 1996, etc.). In the remainder of this paper I will lay out an approach that builds off of these earlier autosegmental approaches to ICM. I will show that phonological constraint interaction results in mutation when certain phonological autosegmental feature bundles are the Vocabulary Items that are spelled out for a given set of morphosyntactic features. The remainder of this section lays out my novel approach to Irish ICM. This research takes the approach that mutations in Irish are the surface realization of Vocabulary Items which correspond to various morphemes/morphosyntactic features (as is consistent with DM). Primarily, these mutations are the result of Vocabulary Items that are spelled out as autosegmental feature bundles, specifically phonological features, which trigger the surface phonological alternations. This research works under the DM framework for morphology and the OT framework for phonology to explain how these morphemes surface. Given the relationship between mutation and morphosyntax in Irish, a natural question arises concerning how we should represent this information in both the phonology and morphosyntax. If syntax is blind to phonology, then mutation must be a phonological realization of Vocabulary Items inserted after the syntactic derivation is complete (or complete at a given phase12 ). This means that certain nodes in the syntax must correspond to slots where Vocabulary Items are inserted such that the interaction of consecutive Vocabulary Items results in mutation once phonology has applied. This research proposes a set of Vocabulary Items that correspond to the morphemes which trigger mutation. This is akin to the proposals of Ní Chiosáin 1991 and Pyatt 1996 in their work on Irish Phonology and Celtic Mutation respectively. However, unlike those analyses, this research combines the theories of OT and DM in order to account for the surface realizations of these morphemes as opposed to an autosegmental analysis of the data. These morphemes are autosegments in that the phonological features that are the realization of these morphemes can operate independently of segmental phonology. But I assume the phonological computation proceeds in a constraint-based OT model. 5.1 Analysis In this section I lay out how the patterns described in sections §2 and §3 fit into the theories of DM and OT. First I show how certain morphemes are stored in the lexicon as particular Vocabulary Items in a DM framework. Then, I proceed to show that OT constraint rankings can model how these morphemes interact with the segments that surround them yielding the surface realization of either Lenition or Eclipsis. In this sense I argue that mutation is an epiphenomenon that arises as the 12 The intricacies of the syntactic model are not of the upmost importance here. Assuming a Minimalist approach to Syntax (Chomsky 1995, 2000, 2001) Vocabulary Insertion would occur after each syntactic phase is spelled out. That said, I am not sure that this assumption is necessary for this analysis to work. Representing Irish mutations 365 result of phonology and is not an intrinsic piece of Irish syntax or morphology. This, therefore, allows for the conclusion to be drawn that neither the Irish lexicon nor Irish syntax require the need to reference/access phonological processes in order to explain mutation. This means that there is no need to posit syntactic operations beyond the widely accepted ones of AGREE and MERGE. Likewise, morphology has no need to access phonology in order for mutation to arise. Mutation is a purely phonological phenomenon. 5.2 Irish Mutation in Distributed Morphology In section §2 I demonstrated how certain features in the morphosyntax always corresponded to a mutation in the phonological output. For example, Past Tense always surfaces as Lenition or the genitive plural always surfaces as Eclipsis. By mapping morphosyntactic features to phonological feature bundles, it becomes clear how mutation is able to realize such complex morphosyntactic structures. I also showed how Eclipsis and Lenition can be captured by a change in one or more phonological features. As such, it should be possible to map these phonological features to the morphosyntactic features that they realize. Using DM this can be done by formulating Vocabulary Items whose Spell Out is the phonological feature that surfaces as mutation. For example, the genitive plural could be modeled as a morpheme whose Spell Out is [VOICE , NASAL] like in (12). (12) " DEFINITE GENITIVE PLURAL #  VOICE NASAL What (12) shows is that wherever the morphosyntactic features for the genitive plural are present the Spell Out is a morpheme that is [VOICE , NASAL]. And so, when this feature bundle interacts with the next phonological segment that follows it the result will be Eclipsis. A similar type of morpheme triggering Lenition can be formulated like that in (13). (13) [ PAST ] [ CONTINUANT ] Here, whenever the Past Tense morpheme is present, the Spell Out is [CONTIN UANT ]. This will result in Lenition on the following segment. In this way, it is possible to formulate morphemes for all of the instances of mutation given in sections §1 through §3 above. For example, take a few more formulations given in (14), like the feminine nominative singular (14a), the masculine genitive singular (14b), the definite dative singular (14c), and the subjunctive (14d).13 13 For simplicity’s sake, I use subjunctive as the morphosyntactic feature that is realized as Eclipsis for any verb that follows a complementizer. This would imply that complementizers select for verbs in the subjunctive. It is outside the purview of this paper to investigate whether or not this generalization holds. For the purposes of this analysis, however, this does not matter so subjunctive could 366 (14) JACK PRUETT a. FEMININE NOMINATIVE / ACCUSATIVE SINGULAR " # [ CONTINUANT ] b. " MASCULINE GENITIVE SINGULAR " DEFINITE DATIVE SINGULAR # [ CONTINUANT ] c. #  VOICE NASAL d. [ SUBJUNCTIVE ]  VOICE NASAL Now that morphemes and their phonological content have been laid out (both the morphosyntactic feature bundles and phonological feature bundles to which they correspond are stored in the lexicon), it must be shown how OT constraint interaction predicts the outcomes that surface as mutation in Irish. In the next section I will analyze these morphemes under an OT framework to show that basic constraint rankings predict how these morphemes would interact with other segments and how they surface as mutation. 5.3 Irish Mutation in Optimality Theory In the previous section it was shown that initial mutation in Irish can be captured morphologically using DM. These morphemes are spelled out as autosegmental phonological feature bundles. In this section, I will demonstrate using OT that the surface realization of mutation can be captured by constraint rankings. As stated previously, Lenition can be captured by the presence of a preceding [CONTINUANT] feature and Eclipsis by [VOICE , NASAL]. Since these features are what change under the context of mutation it is reasonable to assume that there should be faithfulness constraints related to these features. Thus, I propose the following IDENTITY constraints (McCarthy 2008) in example (15). be replaced by any post-complementizer feature. Furthermore, this analysis hinges on the complementary distribution of this subjunctive and the Past Tense. The true nature of the morphosyntax of the Past Tense and the subjunctive is still an open question for research. Regardless of the stance taken, as long as the subjunctive and the past tense are realized such that Past Tense is more crucial for realization than the subjunctive, the predicted result will arise. Representing Irish mutations (15) 367 IDENTITY [ CONTINUANT ] ( IDENT [ C / V / N ]) 14 : Assign a violation for every segment in the output whose CONTINUANT / VOICE / NASAL feature specifi- cation is different from the input. Furthermore, in theory, Lenition or Eclipsis could be realized by the insertion of some segment to realize the features in the autosegmental bundle. Similarly, it is just as possible that the grammar simply could delete all non-segmented phonological features resulting in the deletion of said morpheme. Given these possibilities, two other constraints must be employed–DEP and MAX15 . These constraints give a good foundation for beginning to examine the phonological effects of the morphosyntactic structure laid out in the preceding sections of this paper. To begin let’s examine two straightforward derivations using these constraints. In order to compare the results, it makes sense to look at the interaction of the third person possessive pronouns since they all have the same surface realization, a, but they differ in which mutation they trigger. The following are the Vocabulary Items to derive ’his/her/their cat’ (16). (16)  a [ 3 RD PERSON ] [@]  cat [ CAT ] [kAt”G ] " # MASCULINE GENITIVE SINGULAR  " GENITIVE PLURAL FEMININE GENITIVE SINGULAR # [  CONTINUANT ] VOICE NASAL [ ø ] Using these Vocabulary Items and the constraints outlined above the following derivations can be done16 . Table 1 @ [CONTINUANT] kAt”G DEP MAX IDENT [ C ] ⌘ a. @ xAt”G b. @ kAt”G c. @ h kAt”G ∗ ∗! ∗! 14 Given that continuancy is only at play for Lenition and nasality and voicing for Eclipsis, in the following derivations the IDENTITY constraints used will only correspond to those which target the feature(s) of interest for any given mutation (i.e., IDENT [ C ] will not appear in derivations focusing on Eclipsis). 15 DEP : Assign a violation for every segment in the output which is not present in the input. MAX : Assign a violation for every feature bundle in the input that does not have a corresponding realization in the output. 16 For insertion candidates, n will be inserted for [ VOICE NASAL ] and h for [ CONTINUANT ]. 368 JACK PRUETT As should be clear from Table 1, in order for Lenition to surface in a chat ’his cat’, IDENTITY [ CONTINUANT ] must be ranked below both DEP and MAX . Now take the tableau in Table 2 for the phrase a gcat ’their cat’. Table 2 @ [VOICE NASAL] kAt”G DEP ⌘ a. @ gAt” b. @ kAt”G c. @ n kAt”G d. @ NAt”G MAX IDENT [ N ] IDENT [ V ] G ∗ ∗! ∗! ∗! ∗ From Table 2 it is clear that in order for voiceless consonants to become voiced under Eclipsis as opposed to also becoming nasal (a possibility given the features in the autosegmental bundle) IDENT [ N ] must be ranked above IDENT [ V ]. While this explains the behavior of the voiceless consonants, this does not give us any information about the behavior of voiced consonants. Nor does this ranking provide any insight into the ranking of the other constraints DEP and MAX with respect to IDENT [ N ]. For this, it is necessary to look at another example of Eclipsis, albeit with the initial consonant voiced. Consider the derivation for the example in (17) below with the relevant DM Vocabulary Items (I am only using the VIs that make up the words needed to derive the mutation, all other words of the sentence are not relevant to the points made here) as well as the OT tableau. (17) an nglan-ann tú do sheomra COMP. Q CLEAN - PRES 2. NOM . SG 2. GEN . SG room. ACC . SG ‘do you clean your room?’   an COMPLEMENTIZER [@n”G ] QUESTION [ SUBJUNCTIVE / POST- COMPLEMENTIZER [ CLEAN [ PRESENT  ] ] Table 3 ] VOICE NASAL glan [gëAn”G ]  ann [@n”G ] @n”G [ VOICE NASAL ] gëAn”G -@n”G ⌘ a. @n”G NëAn”G -@n”G b. @n”G gëAn”G -@n”G c. @n”G n gëAn”G -@n”G  DEP MAX IDENT [ N ] ∗ ∗! ∗! From this example we can establish an overall ranking for Lenition and Eclipsis, so far, with respect to the realization of mutation over insertion or deletion. This ranking is given in (18). Representing Irish mutations (18) a. b. 369 DEP , MAX » IDENTITY [ CONTINUANT ] = Lenition DEP , MAX » IDENTITY [ NASAL ] » IDENTITY [ VOICE ] = Eclipsis If only it were as simple as this to explain Irish ICM. While these constraints cover the difference between voicing and nasality in Eclipsis and accounts for the change in Lenition, one must ask if this is the only way to account for the data. Take (19) for example. (19) bróg bheag shoe small ‘a small shoe’ [ " [ Table 4 SHOE  ] bróg [bG RG o:g] FEMININE NOMINATIVE / ACCUSATIVE SINGULAR SMALL ]  # [ CONTINUANT ] beag [bj æg] bG RG o:g [ CONTINUANT ] bj æg a. bG RG o:g Bj æg b. bG RG o:G bj æg c. bG RG o:g bj æg d. bG RG o:g h bj æg DEP MAX IDENT [ C ] ∗ ∗ ∗! ∗! As can be seen in Table 4, if the word before the mutation morpheme ends in a phoneme that can undergo Lenition and the word following the mutation morpheme begins with a phoneme that can undergo Lenition, then the constraints that we have at the present moment cannot determine an optimal candidate. As such, another constraint must be introduced to account for the fact that mutation is realized at the start of the following word and not the end of the preceding one. To find such a constraint it is important to consider in which direction the autosegmental features are docking (McCarthy & Prince 1995). It is clear in all instances of mutation that the features dock to the left edge of the word directly to its right. In other words, it docks to the left edge of the following word. This behavior can be captured in the following featural ALIGNMENT constraint (McCarthy & Prince 1995, McCarthy 2008). (20) ALIGN [Feature, Left]( ALIGN [ L ]): Assign a violation for every instance of an autosegmental feature in the input that is not aligned at the left edge of a word in the output. With this new constraint it is possible to achieve the desired optimal candidate for Table 4 seen below in Table 5 with the new constraint. Table 5 370 JACK PRUETT bG RG o:g [ CONTINUANT ] bj æg ⌘ a. b R o:g B æg b. bG RG o:G bj æg c. bG RG o:g bj æg d. bG RG o:g h bj æg G G DEP MAX ALIGN [ L ] IDENT [ C ] ∗! ∗ ∗ j ∗! ∗! With this additional constraint it is possible to derive a final overall ranking to capture the phonological behavior of mutation in Irish (see example (21)). I will note that this section has left out discussion of the coronal phonemes and /f/ mentioned in section §2. The reason for this is that the coronals require phoneme specific constraints that target just the set of coronal consonants (e.g. d”G becomes G) and for lack of space these cannot be discussed here. That said, since those constraints only affect the outcome of a single segment and is not a property of mutational behavior per se, it can be assumed that they have no bearing on the overall ranking established here to account for Lenition and Eclipsis and should be investigated further in subsequent research. The ranking for Lenition and Eclipsis in Irish can be captured as follows17 . (21) a. b. DEP, MAX , ALIGN [ LEFT ] » IDENT [ CONTINUANT ] = Lenition DEP, MAX , ALIGN [ LEFT ] » IDENT [ NASAL ] » IDENT [ VOICE ] = Eclipsis 5.4 Results This research has argued for an analysis of Irish Initial Consonant Mutation (Lenition and Eclipsis) in which morphemes are spelled out as bundles of phonological features which cause phonological mutation under a DM and OT framework. Specifically, Lenition is triggered by morphemes that are spelled out as [CONTINU ANT ] and Eclipsis is spelled out as [ NASAL VOICE ]. An OT analysis of the phonology of these mutations shows that Irish has IDENTITY constraints as the lowest ranked in any derivation. I argue that mutation in Irish is a purely phonological phenomenon and happens independent of the morphosyntax. That is, mutation in Irish is no different than any other morphophonological alternation whereby morphemes have different realization because of their phonological environment. In this sense mutation is epiphenomenal in that it only surfaces as a result of morphophonological alternation and not as a result of some intrinsic property of Irish morphosyntax. This is different than other theories that argue for a more complex system whereby mutation cannot be explained by phonology alone and must require some sort of morphosyntactic property in order to explain the distribution of mutations in Irish. Furthermore, to the best of my knowledge, this research is first in analyzing Irish ICM under a DM and OT framework. 6 Conclusion In this paper I have established a new analysis of Irish Initial Consonant Muta17 Note there are two rankings here because there are two separate phonological processes. They are not separate rankings in the sense of Cophonologies (Inkelas & Zoll 2007). Representing Irish mutations 371 tions using a Distributed Morphology and Optimality Theory framework (the first of such approaches to the best of my knowledge). This research combines different aspects of previous approaches to Irish mutation, especially those by Ní Chiosáin 1991, Pyatt 1996 & 2003, i.a., who have examined the autosegmental nature of mutation morphophonology in Irish, in order to establish a new way of looking at Irish mutation (and possibly Celtic mutation in general). This research has shown that initial mutation in Irish can be captured by a set of morphemes whose spelled out Vocabulary Items correspond to autosegmental phonological feature bundles. Furthermore, this research demonstrates that these feature bundles interact with the surrounding segments in their phonological environments resulting in overt mutation. These alternations are the product of the interactions as subject to the constraint rankings outlined in (21) above. By ranking ALIGNMENT, DEP, & MAX above IDENTITY constraints the correct surface forms are achievable. Future work would ideally capture more aspects of Irish mutation using this approach. Preliminary investigation suggests that further OT constraints can be used to explain the behavior of the coronal phonemes and /f/ when undergoing mutation. Additionally, it is possible that this approach is compatible with a Phase Theory of Syntax (Chomsky 2000 & 2001) and as such mutation domains can be established through phases. Open questions surrounding how this approach extends outside of Irish Gaelic are possible as well. For instance, it would be interesting to extend this approach to the other Celtic Languages. Finally, this research makes the prediction that Lenition and Eclipsis are productive in Irish; something that is worth exploring in an experimental capacity. References Chomsky, N. 1995. The Minimalist Program. Cambridge, MA: The MIT Press. Chomsky, N. 2000. Minimalist Inquiries: The Framework. In Step by Step: Essays in Minimalist Syntax in Honor of Howard Lasnik, 89–155. Chomsky, N. 2001. Derivation by Phase. In MIT Occasional Papers in Linguistics 18. Chomsky, N., & M. Halle. 1968. The Sound Pattern of English. New York: Harper & Row. Conroy, K. 2008. Celtic Initial Consonant Mutations - nghath and bhfuil? MA: Boston College. Duffield, N. 1997. Configuring Mutation in Irish. Canadian Journal of Linguistics 42. 75–109. Fife, J. 2010. Typological Aspects of the Celtic Languages. In The Celtic Languages, ed. by M. Ball & N. Müller, 3–27. Abingdon, UK: Routledge Taylor & Francis Group. Green, A. 2006. The Independence of Phonology and Morphology: The Celtic Mutations. Lingua 116. 1946–1985. Halle, M., & A. Marantz. 1993. Distributed Morphology and the Pieces of Inflection. In The View from Building 20, 111–76. Cambridge, MA: MIT Press. Hannahs, S. 2011. Celtic Mutations. In The Blackwell Companion to Phonology, 2807–2830. Blackwell Publishers. Inkelas, S., & C. Zoll. 2007. Is Grammar Depence Real? A comparison between cophonological and indexed constraint approaches to morphologically conditioned phonology. Linguistics 45. 133–171. Iosad, P. 2017. Celtic Mutations. In Oxford Bibliographies Online Datasets. Kane, F., R. Folli, & C. Sevdali. 2016. Irish Genitive Phrases The Pseudo-Construct State. Rivista di Grammatica Generativa 38. 173–185. 372 JACK PRUETT Kramer, R. 2015. The Morphosyntax of Gender. Oxford Academic Publishers. McCarthy, J. 2008. Doing Optimality Theory: Applying Theory to Data. Blackwell Publishers. McCarthy, J., & A. Prince. 1995. Faithfulness and Reduplicative Identity. Ní Chiosáin, M. 1991. Topics in the Phonology of Irish. Amherst, MA: University of Massachusetts. Prince, A., & P. Smolensky. 1993/2002/2004. Optimality Theory: Constraint Interaction in Generative Grammar. ed. R.U.C.f.C. Science & U.o.C.a.B.C.S. Department, Blackwell Publishers. Pyatt, E. 1996. An Integrated Model of the Phonology and Syntax of Celtic Mutation. Cambridge, MA: Harvard University. Pyatt, E. 2003. Relativized Mutation Domains in the Celtic Languages. In University of Pennsylvania Working Papers in Linguistics 9. Sims-Williams, P. 1998. The Celtic Languages. In The Indo-European Languages. Routledge Taylor & Francis Group. Stenson, N. 2020. Modern Irish A Comprehensive Grammar. New York: Routledge. Ó Baoill, D. 2010. Irish. In The Celtic Languages, 163–229. Abingdon, UK: Routledge Taylor & Francis Group. Proceedings of CLS 58 (2022), 373-384 © Chicago Linguistic Society 2023. All rights reserved. 373 Text-Setting Tradeoffs in English-Language Popular Music: Incorporating Musical Constraints Into OT Analysis* Jacob Reed University of Chicago 1 Introduction Since nearly the beginning of generative approaches to poetry—and indeed, to linguistic stress and meter more generally—musical settings have played a prominent role, serving as key illustrations of the performance of accent, rhythm, and grouping (Kiparsky 1973; Liberman 1975). At the same time, it’s always been clear that sung and recited texts tend to follow different, albeit related, logics. Thus, starting especially with the work of John Halle and Bruce Hayes, the alignment of music and text (text-setting) began to be studied in its own (Hayes & Kaun 1996).1 And this theory has had numerous great successes, accounting for, among other facts, the quatrain structures of English folksongs, differences between French and English traditional songs, and how we put the words to sea shanties like “What Shall We Do with a Drunken Sailor?” (Hayes & MacEachern 1998; Kiparsky 2006; Dell & Halle 2009; Halle & Lerdahl 1993). But, as this list of examples show, the range of musics that this theory has been applied to is somewhat narrow.2 I want to build on this work, and try to extend it to other repertoires. However, even small steps in this direction—from English folk songs to English-language popular music—require some modifications to existing work on text-setting. In particular, I will argue that a more in-depth consideration of the relevant musical factors can help clarify some of the problems that arise with text-setting in these repertoires. Moreover, I argue that putting the music first, so to speak, prompts us to reconsider even the genres that are already familiar within the literature on text-setting. In Section 2, I’ll begin by discussing some of the ways in which musical factors can induce changes in sung texts: in particular, I will argue for the existence of musically-induced instances of epenthesis and compensatory reduplication. Then, in Section 3, I will discuss some of the issues in accounting for metrical mismatches—words where the stress is on the wrong sylLAble—in pop music. Finally, in concluding (Section 4), I’ll point to the broader relevance these ideas may have for work on text and music more generally. * Audio files for the musical examples can be found on the author’s personal website: http://jacobreedorgan.com/cls58. 1 See also foundational work by Matthew Y. Chen (1983). 2 This is not to discount the geographic range of François Dell’s work in particular; c.f. Dell and Elmedlaoui 2008. 374 JACOB REED 2 Textual Changes Due to Musical Constraints Here’s a familiar example of a text-setting problem, studied in depth by Halle and Lerdahl (1993): (1) a. b. c. Put him in the scuppers with a hosepipe on him.. . . Keep him there and make ’im bail ’er.. . . (1a) gives the first verse of the song, with the ten syllables of its first line distributed into the familiar “long short-short” rhythmic pattern. Later verses, however, can have fit more syllables into this line, like (1b)’s twelve, or fewer, like (1c)’s eight. In a typical performance, the song’s surface rhythm changes to match the words. However, this is not the only possible solution to this kind of mismatch in syllable count. In other musical styles and contexts, for instance, it may be precisely musical characteristics like the “long short-short” rhythm that must be preserved. If the music cannot change, something else must give way, and thus, in such cases, the text itself can undergo changes in order to fit the music.3 Take, for instance, these lines from Jay-Z’s “Empire State of Mind”: I can make it anywhere / Yeah, they love me everywhere. (2) Since Jay-Z, like most rappers, wants to deliver his couplet in a consistent flow, the second line needs the extra word yeah to fill it out. I will call this musically-driven epenthesis. Note that such epenthesis doesn’t occur in our sea shanty example: (3) *Oh, keep him there and-a make ’im bail ’er. We could hypothesize that, as is typically the case, epenthesis is being blocked in the sea shanty by the classic faithfulness constraint D EP: (4) D EP Assess one violation for every syllable in the music that is not in the source text. Conversely, Jay-Z seems to be privileging identity of the musical rhythm between lines of couplets, which we can also formulate as a constraint: 3 In situations like (1b)—-too many syllables, not enough notes—that would entail deleting syllables or words, or finding shorter synonyms. While such processes are certainly attested (for instance in Timorese music as described in Yampolsky 2022), I will focus on the other scenario, in which syllables are added to fill out a line like (2b). Text-setting tradeoffs in English-language popular music (5) 375 I DENT ( RHYTHM ) Assess one violation for every rhythmic deviation of a musical phrase from its repetition. To a first approximation, then, the difference between these two examples can be described via variation in constraint rankings: (6) a. Keep him there and make ’im bail ’er. a. Oh, keep him there and-a make ’im bail ’er. ⌘ b. Keep him there and make ’im bail ’er. D EP !⇤⇤ I DENT ( RHYTHM ) ⇤⇤ b. I can make it anywhere / They love me everywhere. ⌘ a. I can make it anywhere / Yeah, they love me everywhere. b. I can make it anywhere / They love me everywhere. I DENT D EP ⇤ !⇤ Indeed, epenthetic insertion of either profanity or words like “uh” and “yeah” is fairly common in hip-hop and R&B. A few more examples from Jay-Z: (7) a. b. c. I been wilding since-a juvi’ / She was a good girl till she knew me (“Part II (On The Run)”) Can’t none of y’all mirror me back / Yeah, hearin’ me rap. . . (“Encore”) Got a project chick that plays her part / And if it goes down, y’all, that’s my heart (“Girls, Girls, Girls”) To be clear, I’m certainly not trying to say that all instances of “oh yeah” in rap are epenthetic: for these examples, I’ve limited myself to cases of clear rhythmic parallelism where the added word is (a) performed on a short note and (b) doesn’t add clear semantic information. Sticking to short insertions rules out the many instances in hip-hop of MCs using these syllables to fill out empty (but parallel) metrical positions in a flow.4 Playboi Carti is a prolific user of this strategy: (8) a. b. c. Ooh, yeah / Lil’ black jean jacket, yeah / Got VLONE on it, yeah . . . (“Long Time (Intro)”) That bag loaded, yeah, got it all the time . . . / The Dream Mode, yeah, tats showin’, my teams (“Lean 4 Real”) She said she need a pair, uh, spend it on myself / Got like three phones, ooh, Apple, not a pear, ooh (“Old Money”) But I should say that Carti also uses clearly epenthetic syllables quite a bit too: (9) a. Forty-five hunnid for the drink, yeah, we taxin’ / I got gold money and I bought myself a Aston (“Rockstar Made”) 4 In other words, these insertions can’t be motivated by I DENT (R HYTHM ) alone; it’s possible that they can be explained by constraints on empty strong positions like Hayes’s M AX B EAT (Hayes 2009). 376 JACOB REED b. c. She my best friend, yeah, we not a couple / She a rockstar, she a sex symbol (“Sky”) Oh, I think they like me, yeah, they like me / Diamonds on me, ice cream, ho, that slightly (“wokeuplikethis”) So, the more “structural” use of filler words doesn’t exclude their epenthetic addition by the same artist. Much the opposite, it may be that using vocables more intentionally may make them more available for insertion to fill in gaps. Something similar can be seen with a different text-setting strategy in the work of Nicki Minaj. Minaj is famous for her use of exuberant reduplications like “m-mm-monster” and “wo-wo-wo-worried” for pure effect, but she also frequently uses reduplication to fill out lines: (10) a. b. c. Th-these girls are my sons / Jo-Jon & Kate Plus Eight / When I walk in, sit up straight (“Only”) Yeah-yeah, it’s gooder than Meagan / You look good when you’re beggin’ (“Get On Your Knees”) P-P-P-Playboi Barbie / Cover of Playboy, Barbie (feat. on Playboi Carti, “Poke It Out”) Here again, we can say that Minaj is privileging I DENT (R HYTHM ) over a constraint that blocks reduplication for Carti, Jay-Z and sea shanty singers: (11) I NTEG(rity) Assess one violation for every syllable in the source text that has multiple correspondents in the music. And we can again illustrate the difference in text-setting strategies with tableaux: (12) a. Keep him there and make ’im bail ’er. a. There, keep him there, make him make ’im bail ’er. ⌘ b. Keep him there and make ’im bail ’er. I NTEG !⇤⇤ I DENT ⇤⇤ b. Playboi Barbie / Cover of Playboy, Barbie. ⌘ a. P-P-P-Playboi Barbie / Cover of Playboy, Barbie. b. Playboi Barbie / Cover of Playboy, Barbie. I DENT I NTEG ⇤⇤⇤ !⇤⇤⇤ Now, Minaj is not alone in her use of reduplication in hip-hop: (13) a. b. Own his own microphone, bring it everywhere he go / So he can bring it to you live in stere-ere-o (MF DOOM on Madvillain, “Money Folder”) I’ma do this by myself, uh, yeah / On the jet by, by myself, uh, yeah (Playboi Carti [!], “FlatBed Freestyle”) Text-setting tradeoffs in English-language popular music c. 377 Let me tell you pe-people why / Came from the bottom of the block I (Jay-Z, “My 1st Song”) However, these examples are extremely rare, and often prompted by unusual musical contexts like the use of 68 meter in the Jay-Z example (13c).5 Certainly, Minaj’s extensive use of this strategy is, if not unique, extremely marked. I would thus lean toward saying that hip-hop has a default ranking of I NTEG over D EP, and thus tends toward epenthesis over reduplication, with room for deviations by individual artists or for unusual songs and musical moments.6 Having asserted that epenthesis is the norm for hip-hop, it might be helpful to take a glance at a style that prefers reduplication. In English, a great example comes from precisely the style that Hayes, Kaun, and MacEachern studied: AngloAmerican folksongs and ballads (Hayes & Kaun 1996; Hayes & MacEachern 1998; see also Kiparsky 2006). This repertoire is so full of repetitions that it would be foolish to attribute them all to the same kinds of local musical constraints used above. The repetition of whole lines, for instance, occurs in fixed positions within the quatrains of many of these songs; the operative constraint, then, seems to be something like a matching of repeating units in music and verse (Dell 2003; see also Dell & Halle 2009). Still, there are also numerous lines like the following from “Barbara Allen,” in which the repetitions seem clearly motivated by the need for rhythmic parallelism (version from Ritchie 1997 [1965]):7 (14) a. b. c. All in the merry month of May (1st verse) So slowly, slowly she got up (3rd) Adieu, adieu, you good neighbors all (8th) These examples don’t tell the entire story: in other lines like “Oh yes I’m low, I’m very low,” epenthesis is combined with reduplication.8 If we again take this to be a deviation from the default ranking, further work will have to uncover the precise 5 For a discussion of this example—including the atypicality of reduplications like these—see Komaniecki 2019. 6 There are also MCs, especially old-school rappers like Rakim, who practically never epenthesize or repeat: D EP,I NTEG I DENT. For further nuance, we can note that Minaj also uses epenthetic words on occasion, even in the same verse as reduplications: (i) 50K for a verse, no album out / Yeah, my money’s so tall that my Barbies got to climb it / Hotter than a Middle Eastern climate, violent / Tony Matterhorn, dutty wine it, wine it (feat. on Kanye West, “Monster”) A more complete account of these phenomena would likely have to be formulated in terms of stochastic OT, partial rankings, or weighted constraints. Then we could say, informally at least, that Playboi Carti has a much higher weight for I NTEG than Nicki Minaj. 7 Note that the types of reduplication here are quite different from those in the hip-hop examples above: in future work, I plan to give a typology of musically-induced reduplications, which—like text-setting styles more generally—I expect to associate with musical styles and artists. 8 In this particular example, it’s possible that the epenthetic “oh yes” is added because of an independent constraint against compensatory retriplication, which, as far as I’m aware, does not appear in songs in this style. But such a constraint will not cover every case of epenthesis. 378 JACOB REED nature of such deviations. To summarize this section, let me sketch the beginnings of a typology of textsetting strategies, based on the constraints used so far.9 First, we have the rhythmic modifications induced by leaving the text untouched: (15) Keep him there and make ’im bail ’er. a. Oh, keep him there and-a make ’im bail ’er. b. There, keep him there, make him make ’im bail ’er. ⌘ c. Keep him there and make ’im bail ’er. D EP !⇤⇤ I NTEG I DENT !⇤⇤ ⇤⇤ This is the strategy used in the original sea shanty, and it’s the one assumed by prior studies on text-setting. Then, among strategies that preserve rhythm, we have epenthesis:10 (16) Keep him there and make ’im bail ’er. ⌘ a. Oh, keep him there and-a make ’im bail ’er. b. There, keep him there, make him make ’im bail ’er. c. Keep him there and make ’im bail ’er. I DENT I NTEG D EP ⇤⇤ !⇤⇤ !⇤⇤ In addition to hip-hop, we find epenthetic “filler words” (and not reduplication) in a variety of repertoires, including Timorese music and the “padding syllables” oWand other added words of Chinese opera (Yampolsky 2022, Yung 1983c). Finally, the tableau for reduplication: (17) Keep him there and make ’im bail ’er. a. Oh, keep him there and-a make ’im bail ’er. ⌘ b. There, keep him there, make ’im make ’im bail ’er. c. Keep him there and make ’im bail ’er. I DENT D EP !⇤⇤ I NTEG ⇤⇤ !⇤⇤ Again, there are plenty of examples worldwide, including some Ashekenazi Jewish nusachim and Black Sea-region folk music in Turkey.11 In this section, I have outlined an approach to musically-induced modifications of sung texts. In the next section, I will address a different kind of problematic mapping from text to music: stress mismatch. 9 Note the similarities with the typology of compensatory reduplication in Yu 2005; indeed, we could say that musically-induced reduplication is simply a species of compensatory reduplication to fit well-formedness constraints on rhythmic parallelism and metrical structure. 10 Not pictured, of course are violations of C ONTIG and M AX by candidate B. This is not to rule their relevance out a priori, but rather to focus on what I see as the most salient constraints. 11 Thanks to Kutay Serova for referring me to this style. Text-setting tradeoffs in English-language popular music 379 3 Approaching Stress Mismatches in Popular Music As in the previous section, I only have space here to sketch the outlines of a few ways of thinking about stress mismatch (“lexical inversions” in the terminology of Hayes 2009) in popular musics. The issue is especially pressing because stress mismatch is much more common in commercial popular musics than in the folk traditions discussed by Dell, Halle, Hayes, et al. I will argue that such text-setting does not have just one cause, but can be seen to appear from the intersections of a variety of musical and phonological constraints. But first, discussing stress mismatch in this kind of music requires a slight change in method. Here is the constraint that is violated under mismatch, according to Bruce Hayes (2009): (18) M ATCH S TRESS Assess a violation if σi and σj (in either order) are linked to grid positions Gi and Gj , respectively; σi has stronger stress than σj ; Gj is stronger than Gi ; and σi and σj occupy the same simplex word. where each σi is a syllable, and the grid is a version of standard phonological grid, grouped into binary metrical divisions (Liberman 1975; see also Liberman & Prince 1977; Prince 1983; Hayes 1984; for bracketing, see Halle & Vergnaud 1987, Hayes 1995): (19) But for Hayes and others, the grid is doing double duty as a metrical representation of the music, following Lerdahl & Jackendoff (1983). This works well for musics with relatively simple rhythmic structures, like sea shanties and folk songs, but David Temperley (1999, 2001) has pointed out how it produces unfortunate results with syncopated musics: (20) While Temperley has a straightforward way to realign syncopations to the beats they anticipate, other complex rhythmic structures like “3-generated” rhythms (Cohn 2016) imply a rhythmic structure not easily mappable to the underlying grid (example from Biamonte 2014): 380 JACOB REED (21) It would be possible to construct a local metrical grid for each of these instances, but it is not always obvious which choice of grid to make, even when it is clear which beats are stressed.12 Given this ambiguity and the fleeting nature of many of these rhythmic complexities, it seems safer to me instead to have M ATCH S TRESS refer directly to our perception of the musical surface: (22) M ATCH S TRESS (revised) Assess a violation if σi and σj (in either order) are linked to musical events ei and ej , respectively; σi has stronger stress than σj ; ej is perceived as stronger than ei ; and σi and σj occupy the same simplex word. With this in mind, let’s examine a few common scenarios in which mismatch crops up. Briefly, almost every example of mismatch I am aware of is due to one of (a) rhyme13 (b) musical/text-setting parallelism (c) intentional distortions by the artists. I will focus on the first two here. For rhyme-induced mismatches, we can take the following examples by Katy Perry:14 (23) a. b. c. d. All your insecurities / All the dirTY laundRY (“Unconditionally”) And got kicked out of the bar, so we hit the bouleVARD (“Last Friday Night”) It’s supernatural / ExTRAterrestrial (“E.T.”) Now look at me, I’m sparkLING / A firework, a dancing flame (“Part of Me”) The lines in (23d) show that rhyme can induce stress mismatches in the backward as well as the forward direction (although the latter is much more common). So we therefore need a bidirectional constraint: (24) I DENT-R HYME ( MUSICAL STRESS ) Assess a violation if the strongest perceived musical stress of a rhymephrase does not fall in analogous rhythmic position to the strongest stress of its rhyming correspondent(s). In some cases, like (23b) above, this has the same effect as I DENT ( RHYTHM ). But 12 See, for example, the disagreement between Cohn and Temperley on how to interpret the tresillo rhythm (3+3+2) in the articles cited above. 13 A possibility entertained but ultimately rejected by Hayes & Kaun (1996). 14 Even more so than in the hip-hop examples above, I am here using artists’ names as a synecdoche for the production and songwriting team as a whole (and to avoid giving all of the credit to, e.g. Max Martin and Dr. Luke). Text-setting tradeoffs in English-language popular music 381 we need the separate constraint for the other cases—and, conversely, I DENT ( RHYTHM ) explains cases that rhyme does not. Here are a few by Olivia Rodrigo: (25) a. b. c. And I don’t stick up for myself / I’m anxious and noTHING can help (“Brutal”) Yeah, I played dumb but I always knew . . . 15 / Show her off like she’s a new troPHY (“Traitor”) So find someone great, but don’t find no one better / I hope you’re happy, but don’t be hapPIer (“Happier”) (25c) is particularly interesting, since an extremely minimal modification would correct the stress mismatch: something like “I hope you’re happy, but don’t you be HAPpier.” Clearly, exact identity—of both musical rhythm and syllable count—can be extremely important to these artists, ranking above stress matching.16 Similar, but slightly distinct constraints seem to be active in producing Perry’s most infamous stress mismatch:17 (26) UncondiTIONal— / UncondiTIONally / I will love you / unCONdiTIONally (“Unconditionally”) I will propose two further, high-ranking constraints to account for this chorus: (27) (28) I DENT-R EPEAT ( MUSICAL STRESS ) Assess a violation if the musical setting aligns stress maxima with different syllables in repetitions of the same polysyllabic word.18 R HYME TO S TRESS Assess a violation if a rhyme group does not contain a local stress maximum.19 We can then see what has happened: the syllables “con” and ‘tion” have been made to rhyme with each other, and the resulting stress pattern propagates backward through the other instances of the same word. In this section, I have sketched out the beginnings of an approach to stress mismatch in popular musics. Having identified the need for a more nuanced approach to musical stress in text-setting, I hope also to have shown the value musical constraints can have in accounting for phenomena like mismatch in this kind of music. 15 This is the corresponding line in the previous verse On producer/songwriter Max Martin’s obsession with precisely matching syllable counts, see Seabrook (2015: 264–5). 17 Many thanks to Laura McPherson for getting me thinking about Perry in general and this song in particular. 18 Based on how rarely it is violated, I would posit that this constraint is quite high-ranking in most English repertoires. But it is certainly not inviolable: see, for instance, the purposeful manipulations of stress on the title words of Arcade Fire’s “Funeral” and Mitski’s “Nobody.” 19 On the other hand, I am not sure if this constraint is violable: it may be better rephrased as a (partial) definition of rhyme itself, at least in this context. 16 382 JACOB REED 4 Conclusion This paper has given a preliminary outline of some ways of extending the existing literature on text-setting to popular musics, with an emphasis on musical constraints. Section 2 described how musical factors can induce epenthesis and reduplication in song texts, while Section 3 showed how musical factors can change the pronunciation of words by causing stress mismatches. Beyond this specific project, expanding the toolkit for analyzing text-setting may be especially relevant now, given the recent explosion of interest in using text-setting (and music–text interactions more generally) as evidence in a variety of kinds of phonological analysis (e.g. Starr & Shih 2017; McPherson & Ryan 2018). If nothing else, I hope that bringing musical factors more to the fore, while retaining the enormous analytic gains already made by linguists, may contribute to bringing the two disciplines closer together in the study of song. References Biamonte, N. 2014. Formal functions of metric dissonance in rock music. Music Theory Online 20. Chen, M. Y. 1983. Toward a Grammar of Singing: Tune-Text Association in Gregorian Chant. Music Perception: An Interdisciplinary Journal 1. 84–122. Cohn, R. 2016. A platonic model of funky rhythms. MTO 22. Dell, F. 2003. Répétitions Parallèles Dans Les Paroles Et Dans La Musique Des Chansons. 499– 522. Paris. Dell, F., & J. Halle. 2009. Comparing musical textsetting in French and in English songs. In Towards a Typology of Poetic Forms: From Language to Metrics and Beyond, ed. by J.-L. Aroui & A. Arleo. Amsterdam: John Benjamins. Halle, J., & F. Lerdahl. 1993. A generative text-setting model. Current Musicology 55. 3–26. Halle, M., & J.-R. Vergnaud. 1987. An Essay on Stress. Cambridge: MIT Press. Hayes, B. 1984. The Phonology of Rhythm in English. Linguistic Inquiry 15. 33–74. Hayes, B. 1995. Metrical Stress Theory: Principles and Case Studies. Chicago: University of Chicago Press. Hayes, B. 2009. Faithfulness and componentiality in metrics. In The Nature of the Word: Studies in Honor of Paul Kiparsky, ed. by K. Hanson & S. Inkelas. MIT Press. Hayes, B., & A. Kaun. 1996. The role of phonological phrasing in sung and chanted verse. The Linguistic Review 13. Hayes, B., & M. MacEachern. 1998. Quatrain Form in English Folk Verse. Language 74. 473–507. Kiparsky, P. 1973. The Role of Linguistics in a Theory of Poetry. Daedalus 102. 231–244. Kiparsky, P. 2006. A modular metrics for folk verse. In Formal Approaches to Poetry, ed. by B. E. Dresher & N. Friedberg. DeGruyter Mouton. Lerdahl, F., & R. Jackendoff. 1983. A Generative Theory of Tonal Music. MIT Press. Liberman, M., & A. Prince. 1977. On Stress and Linguistic Rhythm. Linguistic Inquiry 8. 249–336. Liberman, M. Y. 1975. The Intonational System of English. Massachusetts Institute of Technology dissertation. McPherson, L., & K. M. Ryan. 2018. Tone-tune association in Tommo So (Dogon) folk songs. Language 94. Prince, A. 1983. Relating to the grid. Linguistic Inquiry p. 19–100. Ritchie, J. 1997 [1965]. Folk Songs of the Southern Appalachians as Sung by Jean Ritchie. Louisville: University Press of Kentucky. Text-setting tradeoffs in English-language popular music 383 Starr, R. L., & S. S. Shih. 2017. The syllable as a prosodic unit in Japanese lexical strata: Evidence from text-setting. Glossa 2. Temperley, D. 1999. Syncopation in rock: A perceptual perspective. Popular Music 18. 19–40. Temperley, D. 2001. The Cognition of Basic Musical Structures. MIT Press. Yu, A. C. L. 2005. Toward a typology of compensatory reduplication. Proceedings of the 24th West Coast Conference on Formal Linguistics . Proceedings of CLS 58 (2022), 385-402 © Chicago Linguistic Society 2023. All rights reserved. 385 On the article el heading finite clauses in Spanish: a semantic proposal and its contribution to the clause Cristina Ruiz Alonso Universitat Autònoma de Barcelona 1 Introduction This paper deals with a widespread phenomenon in the Spanish language that has not received enough attention in the scientific literature: the non-mandatory definite article determiner el (‘the’ in its masculine singular form) heading a finite clause introduced by the declarative complementizer que (‘that’): (1) (2) El que digas estas cosas me molesta [subject] the that say.2SG.SUB those things me.DAT bothers ‘[The fact] That you say those things bothers me’ Lamento el que digas estas cosas [object] Regret.1SG the that say.2SG.SUB those things ‘I’m sorry (that) you say those things’ The absence of the article does not lead to an ungrammatical clause, as it remains grammatical without it: (3) (4) Que digas estas cosas me molesta that say.2SG.SUB those things me.DAT bothers ‘That you say those things bothers me’ Lamento que digas estas cosas Regret.1SG that say.2SG.SUB those things ‘I’m sorry (that) you say those things’ Contrary to this, I argue that the article has a contribution, a syntactic and a semantic one, and its distribution is not free. This is endorsed by several consequences and properties that this particle triggers in the structure. The first proposals claimed that this construction was related to factivity (Demonte 1977), since only factive verbs accepted el, and it was argued that there was a null nominal fact between the article and the complementizer. However, this construction is not limited to factive verbs (Serrano Pardo 2015) and, as I will explore (as an extention of Picallo 2002’s ideas), the noun fact is not always possible in el-que clauses. The paper is structured as follows. In section 2, I will explore the main properties of the phenomenon and I will briefly list the verbs that can have the article, concluding that its presence does not depend on the verb, but on the properties of the clause itself; in section 3, I will present the semantic and discursive properties in order to show that the distribution of the article is not free, although it appears 386 CRISTINA RUIZ ALONSO within specific contexts where the embedded clause is not the main point of utterance nor asserted; in section 4, I will briefly present the theoretical framework where we will include our theory, this being the conceptualization of CPs as nominal due to the status of the complementizer, and I will briefly explain the previous analysis; in section 5, I will sketch a tentative analysis based on the semantics of the clause and the article. To be more specific, I will argue that el is a referential mark, which triggers presupposition in a parallel fashion to that of definite DPs. Following that, I will present some empirical data so as to endorse the proposal. Finally, in section 6, some conclusions will be presented. In this work, and due to space limitations, I will limit myself to the analysis of the semantic contribution of the construction, leaving syntax aside for further works1 . 2 Main properties El-que clauses can appear in the role of the subject (1) or the object (2). The former one is wider, either extraposed (as the specifier, cf. (5)) or postverbal (6). Regarding the DO position, this is also possible, but the examples are less attested (7)2 : (5) El que el ayuntamiento de Cabrillanes haya querido obstaculizar esa labor hubiera horrorizado a Guzmán (CORPES 21/03/2022) 3 ‘The fact that the local government of Cabrilanes has.SUB wanted to block that task would have scared Guzman’ No me sorprendió el que se manifestara tan seguro de adivinar mis pensamientos ‘It doesn’t surprise me that he was so sure about guessing my feelings’ Destacamos el respeto de las minorías y el que este proceso no lleve tensiones de tipo democrático (CREA 19/03/2022) ‘We highlight the respect for minorities and the fact that this process does not trigger democratic issues’ (6) (7) Regarding moods, in all the examples, the embedded verb is in the subjunctive4 . This verbal mood is a direct consequence of the article. First, factive verbs take their embedded clauses in this mood in Spanish, and so it happens with el-que: (8) El que ella actuara de ese modo les encantaba a todas sus amigas ‘That she behaved.SUB that way was loved by all her friends’ Even with non-factive verbs, the embedded verb is also in this mood with el-que: (9) Las centrales subrayan el que uno de cada cuatro trabajadores tenga empleo (CREA 16/03/2022) 1 Nonetheless, we will turn to syntax when required to get the best understanding of the proposal. All the examples, unless not mentioned, are taken from corpora CREA (Corpus de referencia del español actual) or CORPES 3 We use the construction the fact that in some translations into English with understanding purposes, but there is not a nominal fact-like in the structures that we are accounting for. 4 Some exceptional cases have been found in the indicative mood, although we are leaving them aside for ease of exposition. 2 El heading finite clauses in Spanish 387 ‘The centrals highlight that one out of four workers is.SUB employed’ If we argue that the subjunctive is a consequence of the article, the indicative should appear when this is not present (if the embedded properties allow it). Interestingly, this is what happens: if we remove the article in the sentence above (9), indicative is possible - in fact, preferred-: (10) Las centrales subrayan que uno de cada cuatro trabajadores tiene empleo ‘The centrals highlight that one out of four workers is.IND employed’ There is still one additional piece of evidence with verbs that can select both indicative and subjunctive, like reprochar (‘to reproach’): (11) a. b. Le ha reprochado que come con la boca abierta DAT has.AUX reproached that eats.IND with his/her mouth open boca abierta con la reprochado que coma Le ha DAT has.AUX reproached that eats.SUB with his/her mouth open ‘He has reproached him for eating with the mouth open’ As (11) shows, the verb can have embedded clauses in both moods and both are correct5 . However, when the article is inserted, indicative is ruled out and only subjunctive is available: (12) a. *Le ha reprochado el que come con la boca DAT has.AUX reproached the that eats.IND with his/her mouth abierta open boca con la reprochado el que coma b. Le ha DAT has.AUX reproached the that eats.SUB with his/her mouth abierta open ‘He has reproached him for eating with the mouth open’ Even with verbs that do not freely accept the subjunctive in their dependent contexts, the article boosts this mood. Note that for (13), the subjunctive is not accepted. Nonetheless, the example is way more adequate if the article is added (14) and, again, the indicative is ruled out: (13) (14) 5 Los hallazgos reflejan que este invierno {es ?sea} el más The findings show that this winter {is.IND ?is.SUB} the most cálido del siglo hot of-the century ‘The findings show that this winter is the hottest of the century’ /sea} el más Los hallazgos reflejan el que este invierno {*es The findings show the that this winter {*is.IND /is.SUB} the most Some semantic differences could be argued between both moods, but they are not relevant for our discussion. 388 CRISTINA RUIZ ALONSO cálido del siglo hot of-the century ‘The findings show that this winter is the hottest of the century’ This fact sheds light on the properties of the construction. Subjunctive encloses factivity in Spanish (Quer 2001; Demonte 2010; Fábregas 2014; Hooper & Thompson 1973). Moreover, Hooper & Thompson’s (1973) seminal works claim that the subjunctive cannot be asserted since it is used in unreal worlds or presuppositional contexts, where the information is already in the Common Ground (Quer 2008). Therefore, the traditional model realis-irrealis between moods does not hold for all the cases and the subjinctive is also used for presupposition in the context selected by a factive verb, which means that the information is not asserted (Hooper & Thompson 1973; Sheehan & Hinzen 2011). If, as shown, el boosts the subjunctive, there is a reason to believe that el can trigger factivity in the clauses. This claim will be elaborated on throughout the paper. However, even though el seems related to factivity, it cannot be stated that there is a null nominal fact (‘hecho’) in el-que structures. Demonte (1977 among others) argued that el-que was directly linked to factivity because the data showed that it was restricted to factive verbs. However, Serrano (2015: chap. 2) confirms that the construction is extended to other verbs. The existence of this null nominal forming a complex NP was suggested because an overt fact noun can be incorporated: (15) Me molesta el hecho de que siempre llegues tarde me.DAT bothers the fact of that always arrive.SUB late ‘It bothers me the fact that you’re always late’ However, as (15) shows, a preposition must be inserted between the noun hecho and the complementizer. This allows us to argue that, despite some semantic parallelisms, el-que structures and el hecho structures are not syntactically equivalent. First, a noun cannot take a sentence as its complement without a preposition in Spanish: (16) *Me molesta el hecho que siempre llegues tarde me.DAT bothers the fact that always arrive.SUB late ‘It bothers me the fact that you’re always late’ In fact, even when a noun is elided, the preposition must remain, as Picallo (2002) demonstrates: (17) Consideró varios hechos independientemente. El [t] de que hubieran apoyado tal propuesta era el más conspicuo ‘She considered several facts independently. The [one] of them supporting that proposal was the most famous’ On the other hand, a preposition is ruled out in an el-que clause: (18) *Me molesta el de que siempre llegues tarde me.DAT bothers the of that always arrive.SUB late El heading finite clauses in Spanish 389 ‘Intended: It bothers me the [fact of] that you’re always late’ Moreover, in ellipsis, the non-overt noun must agree with the determiner. However, with clauses, only el (the default one) can appear6 , hence the anaphoric conditions found in ellipsis do not take place with el-que (Leonetti 1999: 823). Some additional evidence can be given. First, beside de (‘of’), el hecho de constructions clauses can be headed by a preposition in Spanish: (19) Confío en el hecho de que haya venido Trust.1SG in the fact of that have.SUB come.PART ‘I trust in the fact that you came’ However, el-que clauses cannot be introduced by a preposition: (20) *Confío en el que hayas venido Trust.1SG in the that have.SUB come.PART ‘Intended: *I trust in the that you have come’ Plus, the nominal hecho cannot be added in all el-que examples: (21) Esto causó un gran revuelo en la colonia y ocasionó el (#hecho de) que se enviase a España representantes de los bos bandos para exponer sus respectivas posiciones ante el Rey ‘This caused a great stir in the colony and caused the [#the fact that] some representant from both sides to be sent to Spain in order to explain their reason to the King’ Moreover, diachronically our claim is also supported. Octavio de Toledo (2014: 949, n. 25) shows that el hecho de appeared later in Spanish, around the XIX century, whereas el-que appeared around the XVII as an evolution of the determiner heading infinitives. Thus, the nominal ellipsis is ruled out. Yet, we are dealing with a clause that is headed by a nominal element7 6 El (masculine singular article) is the only form that can appear heading clauses. We follow the hypothesis of Picallo (2002) on considering that it is a neuter-like element, instead of a true masculine since it is the defaulted gender form and, therefore, its features are negatives. We are aware of the controversy on the (non-)existence of features in clauses. We adopt Picallo’s analysis as negatives by default. Moreover, it must be said that Spanish have yet a neuter element lo that can appear with clauses: (i) Lo de que venga Juan me parace buena idea (‘That Juan comes it is a good idea for me’) Note that the preposition de here is mandatory (*Lo que venga Juan me parecede buena idea), which, according to what has been outlined above, suggests a non-overt noun. In fact, some anaphoric conditions are required since “lo de” is meaningless. Rather, a noun like idea, topic or so can be glimpsed (see Picallo 2001, 2002; Moulton 2020 and references therein for a discussion). 7 In fact, this construction is not restricted to Spanish. Apart from this fact, the claim that embedded clauses depend on a nominal element is not new and it has been argued from different perspectives (Kayne 2008, Davidson 1968, Torrego & Uriagereka 1982, Kornfilt & Whitman 2011, Panagiotidis & Grohmann 2005, Manzini & Savoia 2011 among many others). 390 CRISTINA RUIZ ALONSO Moreover, when el is overt, the extraction of the complement is blocked, being strong islands, as the impossibility to extract arguments shows: (22) a. Lamentamos que diga eso Regret.1PL that says.SUB that b. ¿Qué lamentamos que diga t? What regret.1PL that says.SUB c. Lamentamos el que diga eso Regret.1PL the that says.SUB that d. *¿Qué lamentamos el que diga t? *what regret.1PL the that says.SUB ‘What do we sorry that s/he is saying’ The impossibility of extraction suggests that the article is in an extra layer that makes the clause duller due to its semantics. The details will be presented in section 4. However, this islandhood status is not directly related to factivity since factive islands are weak islands and argument extraction is possible, contrary to what happens with el-que, as (22) shows. 2.1 Verbs that accept el-que clauses Giving an elaborated classification of verbs that accept the article heading clauses is beyond the scope of this paper. However, some notes about the data must be remarked8 . Though classifications only based on semantic meanings are not enough to define a phenomenon, Hooper & Thompson’s (1973) work about assertion and factivity on Main Clause Phenomena (MCP) sheds light on this question. The authors classify the verbs among five classes based on whether they are assertive or not. Nonetheless, Sheehan & Hinzen (2011) notice that the classification could be more subtle and fine-grained, and they propose a sixth class in order to have a tripartite classification based on assertion and factivity9 . This is as follows below: Figure 1: Classification of verbs by Hooper & Thompson (1973) amended by Sheehan & Hinzen (2011) 8 We do not split the data according to the position where it arises. Rather, we explain subject and object positions together due to space limitations. The subject position shows more cases and the distribution is wider than in the cases of the object. 9 Also, Sheehan & Hinzen introduce the “semifactive” concept since they realize that not all “factive” verbs behave the same way. This is also noticed by Kiparsky & Kiparsky’s (1970) seminal work and Hooper & Thompson themselves. El heading finite clauses in Spanish 391 According to the data collected from corpora, factive (23) and “indefinite” verbs (24) are the classes that most accept the article10 . The former appears in both subject and DO position while the latter is typically in DO: (23) (24) No nos molesta el que sean ya tres los libros que no us.DAT bothers the that are.3PL.SUB already three the books that estarán publicados en menos de tres años (CREA 20/04/2022). be.will published in less of three years ‘It doesn’t bother us the fact that three books will be published in less than three years’ Pero que quede claro que no me podrás negar el que yo But that leave.SUB clear that no me.DAT can.will deny the that I lo vea cuando yo quiera (CORPES 16/03/2022) him.ACC sees.SUB when I want ‘But it must be clear that you can’t stop me from seeing him whenever I want’ Those data bring a crucial property: though factive and indefinite verbs have different semantics, they coincide in both being non-assertive. This property is, indeed, important as will be presented in the next section. In fact, non-factive assertive verbs are incompatible with el-que, as both communication and cognitive verbs show: (25) (26) *Dijo/Comunicó/Afirmó el que venía Said/expressed/claimed the that come ‘S/he said/communicated/stated that s/he would come’ *Pienso/Creo el que viene Think.1SG/believe.1SG the that come ‘I think/I believe that he comes’ Regarding semifactive and assertive verbs, the data are less clarifying. Communication verbs can have the clausal article, but the data is more limited: (27) Su única reacción era achacarme el que yo no hubiera incluido en mi contrato el costo de un tratamiento (CORPES 17/03/2022). ‘His only reaction was to blame me for not having included the cost of a treatment in the contrast.’ As for cognitive verbs, having or not the article depends on whether the verb is assertive or not. First, some verbs that are included in this group could be emotive factive themselves. Those verbs accept the article: (28) 10 Porque respetamos el que uno haga lo que le apetezca Because respect.1SG the that one does.SUB it.CL that DAT want.SUB However, “indefinite” verbs that are classified as volitive verbs or intention like to ask for, to want, etc. do not accept the article. 392 CRISTINA RUIZ ALONSO sin molestar a los demás without disturbing.INF to the rest ‘Because we respect that every person does whatever they want without disturbing the others’ However, some verbs considered semifactives, like to see, to forget, and to know, are asserted and select their dependents in indicative in Spanish. Interestingly, those verbs cannot have the article: (29) *Sabemos/Olvidamos el que venías Know/forget.1PL the that came.2SG ‘We know/we forget [the fact] that you came’ To sum up, factive and semifactive (especially of “cognition”) verbs can generally have the article. Furthermore, indefinite verbs also accept it (see footnote 7). On the other hand, non-factive verbs show reluctance to co-occur with the article. However, despite this distribution that shows interesting patterns, the discursive context where el-que arises suggests that the presence of the article is related to the properties of the clause itself, instead of being related to the matrix verb, as we will see in the next section. 3 Semantic and discursive properties of el-que It was argued in the introduction that the distribution of el-que was not free nor optional, but it was related to some conditions. In this section, we will address this issue based on real data from corpora. As the examples will show, el-que is related to the semantic notions of Common Ground, conversational background, and the main point of utterance. To begin with, observe some examples from corpora (the underlined information is the Common Ground): (30) (31) La marroquización es la pesadilla de los canarios. Y ello está justificado porque una constante histórica de las islas ha sido “la de estar siempre a merced de la potencia y de los intereses hegemónicos en su zona atlántica”. El que esta marroquización no ocurra viene condicionado por factores externos al archipiélago. ‘The “influence of Maroc” is the nightmare of the Canarians. And that is justified because a constant historic situation of the island has been “always being at the mercy of the power and hegemonic interests in its Atlantic zone. The “possibility” that this “influence of Maroc” doesn’t happen is conditioned by external factors of the archipielago” Su persistencia es mayor debido a los adherentes que incluyen por lo que su capacidad fitotóxica también lo es. Se recomienda este tipo de azufres para el invierno y los otros para los períodos más cálidos. Además este tipo de azufres permite el que se les pueda mezclar con cobres, algas, bentonita, etc. Its persistence is greater due to the adherents that are included, and so it is its fytotoxic capacity. This kind of sulfur is recommended for the winter, El heading finite clauses in Spanish (32) 393 and the others are for the warmer periods. Moreover, this kind of sulfur allows the mixture with copper, algae, etc.’ Con todo, el alcohol se ha ido haciendo cada vez más asequible al consumo femenino y, consecuentemente, hay más mujeres alcohólicas. Aunque, si bien ya se acepta el que la mujer beba incluso fuera de las comidas, sigue existiendo una marcada intolerancia hacia la que se emborracha ‘Even so, alcohol has become more accessible to the female consumption over time, and, therefore, there are more alcoholic women. Yet, although it is accepted the fact that a woman drinks, even outside of meals, it still exists an important rejection towards the woman that gets drunk’ As the examples above show, the clause headed by el-que is related to some previous information that has been mentioned in the conversation. The information itself can be new, but it is related to information already presupposed in the conversation. To understand the semantic and discursive properties of these structures, we must address some semantic concepts that explain the different mechanisms that arise in a conversation. To specify this, the distribution of el-que suggests that these clauses belong to the Common Ground (CG. Stalnaker 1978). According to the author, the Common Ground of a conversation is the set of propositions considered to be true. When a speaker introduces new information, (s)he asserts it, and once this assertion has been accepted by the members of the conversation, it becomes part of the Common Ground. Therefore, all the possible worlds in which the information of the CG is true are candidates to be the real world. In Stalnaker’s words (2002: 716), the CG’s mechanism consists in: “that φ in a group if all members accept (for the purpose of the conversation) that φ, and all believe that all accept that φ, and all believe that all believe that all accept that φ,etc”. So far, we know that el-que clauses are related to some information that is in the CG, which means that must be previously accepted in the context. This property is directly connected to the conversational background proposed by Levinson (1983). As the author argues, the information of a clause can appear in the foreground (the main point of utterance and the main claim) and the background, where the known and presupposed information is. Kadmon (2000: 12) uses direct or partial questions as a test to identify the foreground; the background cannot be the answer nor a question because it is information that has been already accepted. Following this division of the parts of a conversation, Roberts (!996) introduces the notion of content at-issue to explain the discourse behavior. More specifically, the foreground is at-issue. This content at-issue is an assertion that takes place when the Speaker wants to make a move in the conversation; thus, the content that does not “make a move” (because the information is already in the conversation, even if it is new) is not at-issue. This is relevant for the distribution of the construction we are accounting for, since the possibility to have or not the article could depend on this property. First, el-que cannot be the foreground since its information appears in the CG. This explains that el-que cannot appear as an answer to a question: (33) A: ¿Dónde estuvo Luis anoche? (Where was Luis last night?) B: *Sus redes sociales sugieren el que estuvo en la ópera 394 CRISTINA RUIZ ALONSO His net social suggest the that was in the opera ‘His social networks suggest [the fact that] he was in the opera” Moreover, el-que cannot appear with foci adverbs. Those elements typically appear with new information that is highlighted: (34) a. #Lamento solo el que venga Juan Regret.1SG only the that comes.SUB Juan ‘Int: I’m sorry only about [the fact] Juan coming’ Juan b. #Me molesta también el que venga the that comes.SUB Juan me bothers too ‘Int: It bothers me [the fact] that also Juan is coming’ Whether el-que clauses are at-issue or not seems a crucial factor: when the verb has some evidential contribution, the embedded clause is the main point of utterance and, therefore, cannot have the article. This fact explains that semifactive verbs like saber, descubrir, ver, etc. (‘to know’, ‘to find out’, ‘to see’) cannot co-occur with the article, as well as communication, perception, or influence verbs. When the main verb encloses the main point of utterance itself, the article can be inserted if the semantic properties argued above meet. To conclude, el-que clauses cannot be assertive since their content belongs to the Common Ground, the background, and it cannot be at-issue. Rather, the content is already accepted in the set of the conversation, and it is related to information previously presupposed. Serrano (2015: 170) builds some contexts where we can see that, when the information is new and it has not appeared in the previous context, the article cannot be inserted (35). However, when the el-que clause is linked to previous information (i.e., information in the CG) it can appear (36): (35) (36) A: Estoy un poco perdido con el funcionamiento de la promoción de curso en esta país. ¿Podrías orientarme un poco? ‘I’m a little bit lost with how academic promotion works in this country. Could you help me?’ B: Pues mira, por ejemplo, (#el) que un alumno suspenda más de dos asignaturas impide automáticamente (#el) que pase de curso ‘Look, for instance, [the fact] that a student fails more than two courses automatically prevents him from promoting’ [Serrano 2015: 170 (47)] A: Mi hijo no va a pasar de curso porque ha suspendido una asignatura. ‘My son is not promoting because he has failed a course’ B: Bueno, (el) que tu hijo haya suspendido una asignatura no impide (el) que pase de curso ‘Well, [the fact] that your child has failed one course does not prevent him from promoting’ [Serrano 2015: 169(45)] Therefore, the possibility to have el depends on the properties of the clause itself and not only on the type of verb that selects the embedded clause; its content must belong to the CG, it must be in the background of the conversation and it cannot be at-issue. In the next section, we will sketch a semantic proposal of the contribution of the El heading finite clauses in Spanish 395 article in clauses in which it appears. However, first, we will make some notes on previous analyses and the theoretical framework that we will be adopting. 4 Some notes on the theoretical framework and previous analyses Although we cannot go into detail about this matter, the theoretical framework that we will be adopting in our approach can be summed up as follows: CPs are DPs due to the properties that they share, an assertion that has been defended from different perspectives (cf. footnote 7; Abney 1987; Rosenbaum 1967; Manzini & Savoia 2011; Kastner 2015; Kornfilt & Whitman 2011 and Panagiotidis & Grohmann 2005, Poletto & Sanfelici 2022 among many others). Among all of them, Manzini’s & Savoia’s (2011) and Poletto’s & Sanfelici’s (2022) will be the proposals we will use as reference, as they state that the complementizers are nominals; this fact allows them to have some special distribution and to co-occur with some elements (see references for details). One of the reasons to support this claim is that they can be headed by a determiner, as the constructions we are discussing show. This complementizer selects the CP as its complement, allowing it to be an argument in a parallel fashion to that of the determiner with the noun in DPs. Therefore, in both projections, a functional element takes the clause and the noun as its complement. What those theories predict is all CPs arguments are DPs, since they both have similar syntactic and semantic patterns (Abney 1987, Rosenbaum 1967 among many others). Thus, the article does not convert the clauses into a sort of nominalization, but most assuredly, we are facing a consequence of the nominal status of the clause. At this point, and assuming this theory, we still deal with a remaining question: if the article does not make that a clause nominal, what does it enclose? We will try to answer this in the next section. 5 Our proposal: the semantic contribution of el in clauses In the previous serction, we have seen the discursive properties that should arise in el-que clauses. Now, we turn to the contribution of the article when it appears. Due to its properties and the semantic and discursive contexts, it could be argued that the article encloses factivity. Considering the work of Kiparsky & Kiparsky (1970), it can be argued that factive verbs have a more complex deep structure since they have a null noun fact incorporated into their structures. This assumption, even though it does not elude controversy is followed in current terms bu Kastner (2015). As clauses with el are headed by the determiner, which is a nominal category, we could consider that the structure is akin to that proposed by Kiparsky & Kiparsky. This was defended by authors like Demonte (1977) when el-que was thought to be restricted to factive verbs. The main consequence of el-que being factive is that it would trigger the presupposition, and thus, the cancellation of the content would be impossible or, at least, difficult. Indeed, this is what happens when el appears: (37) a. El resultado refleja el que hayan ganado la liga, #aunque todavía no se sabe 396 CRISTINA RUIZ ALONSO b. The result shows the fact that they have won the championship, #although we don’t know yet El resultado refleja que hayan ganado la liga, #aunque todavía no se sabe The result shows that they have won the championship, although we don’t know yet As (37) shows, cancellation is more difficult with el-que due to the semantics of the article. However, current works point out that clauses cannot be factive since it is a property of verbs, but not of clauses. Haegeman & Urogdi (2010) and de Cuba & Urogdi (2010) assert that the property that is on track in clauses is that of referentiality. The evidence that they use to suggest their proposal is the fact that some syntactic phenomena occur, and they do not depend on whether the clause is factive or not. It is the case of the expletive azt in Hungarian. Before tackling referentially in el-que, it is necessary to explain what referential means when we talk about clauses. On the one hand, for De Cuba and Urogdi, a referential CP is “ [...] a referential entity that denotes a proposition without illocutionary force. Since referentiality does not implicate truth-conditional presupposition, both factive and non-factive predicates are compatible with this clause type. On this definition, a CP is simply used to refer to a proposition, hence (just as in the case of referring expressions in general) contextual givenness is also not a necessary requirement”. [de Cuba & Urogdi 2010: 45(12a)] Furthermore, de Cuba & MacDonald (2013: 129) claim that a referential CP is “a proposition that refers back to a resolved proposition, where a resolved proposition is a proposition that forms part of the common conversational ground, i.e. the ground shared by the speakers”. A combination of both definitions could be applied to el-que clauses. To begin with, factive and non-factive verbs can be referential and both types of verbs can accept el. In addition, we have seen that el-que is a resolved proposition that forms part of the Common Ground. Sheehan & Hinzen (2011: 47) claim that reference is an edge phenomenon. The element that is on the edge of a phase is the most referential. Plus, they argue that the edger a complement is located, the more difficult it is to extract from. In the syntactic structure, el takes the CP as its complement as it is located in a higher position (and there is not a null nominal, as was confirmed in the first section). If el-que is a phase the same way a bare CP is therefore, following their proposal, an el-que clause is more referential that a que clause and it is more difficult to extract from. This hypothesis is borne out. As was shown in the introduction, el-que blocks the extraction of the complement that otherwise could be extracted: (38) a. ¿Qué sugieren los datos que hagamos? What suggest the data that do.1PL.SUB b. *¿Qué sugieren los datos el que hagamos? What suggest the data the that do.1PL.SUB ‘What do the data suggest us to do?’ In the next section, we will show some empirical predictions that endorse the pro- El heading finite clauses in Spanish 397 posal that el is a referentiality mark. Specifically, some of them draw a parallelism between definite DPs and “definite” CPs headed by el. 5.1 Predictions of our analysis In the previous section, we have argued that el is a referentiality mark. Some predictions must be attached to prove that our proposal is on the right path. Firstly, as said, some non-factive verbs can accept el-que, which confirms that factivity is not the property involved: (39) ...Influirá en ese aspecto el que él mismo encarne la figura Influence.WILL in that aspect the that he himself embodies the figure del pastor of-the shepherd ‘Him embodying the figure of the shepherd will be of influence in that aspect’ Moreover, some examples appear in coordinated structures: in coordination, both elements should be semantically compatible. El-que examples can be coordinated with referential DPs: (40) (41) ...Todo esto tiene que ser tenido muy en cuenta para impedir su prosecución y el que sus hijos los sustituyan en su status social (CREA 16/03/2022) ‘All of this must be considered in order to prevent his prosecution and [the fact] that his children substitute them in his social status’ En una nota sobre Ortega en mi diario reconozco dos cosas: mi deuda impagable con él, y el que su desaparición dejaba a nuestro país sin un implícito censor de la tontería intelectual ‘In a note about Ortega in my diary, I admit two things: my inestimable debt with him and [the fact] that his passing left our country without a censor of the intellectual stupidity’ If el-que clauses are referential, they should be incompatible with indefinite DPs. This is borne out, as the examples below show: (42) #Reconozco dos cosas: una deuda y el que su desaparición dejaba a nuestro país sin un implícito censor de la tontería intelectual ‘I admit two things: a debt and that his passing left our country without a censor of the intellectual stupidity’ Besides, as was explained in the previous section, el-que clauses must be instantiated in the real world. This is confirmed by the fact that those clauses cannot have a conditional interpretation, being el-que an instantiation in a real situation, instead of in a potential world (Serrano 2015: 8). (43) Me molesta el que venga Juan (“The” that Juan comes bothers me) a. Declarative: “Whenever he comes/When he is coming, it bothers me” b. #Conditional: “#If he comes it will bother me” 398 (44) CRISTINA RUIZ ALONSO Me molesta que venga Juan (Juan coming bothers me) a. Declarative: “Whenever he comes/When he is coming, it bothers me” b. Conditional: “If he comes it will bother me” In (43), el-que cannot be inserted in an indefinite context, but must quantify a real and identifiable situation. The article triggers the presupposition of the existence of the content of the clause, in a similar fashion to that of the article in DPs: (45) a. b. La RAE no contempla el que haya verbos de influencia The RAE no consider the that there-are.SUB verbs of influence La RAE no contempla que haya verbos de influencia the RAE no consider that there-are.SUB verbs of influence ‘RAE (‘The Real Academy of Spanish language’) does not consider that there are verbs of influence’ In (45a), the article triggers that the verbs of influence do exist, although the RAE does not take them into account. On the other hand, in (45b) without the article, it is not clear whether those verbs exist or not and the clause solely claims that the RAE does not take them into account. Turning to the parallelism between CPs with el and definite DPs, some similar properties arise. The most relevant one concerns extraction. It was argued in the introduction that el-que blocks extraction of the complement forming strong islands (repeated in (46) below for ease of exposition). Definite DPs also block extraction (Roussou 1994; Leonetti 1999): (46) (47) a. Me molesta que digan eso ! ¿Qué te molesta Me.DAT bothers that say.SUB.3PL that ! what you.DAT bothers que digan? that say.SUB.3PL b. Me molesta el que digan eso ! *¿Qué te Me.DAT bothers the that say.SUB.3PL that ! what you.DAT molesta el que digan? bothers the that say.SUB.3PL ‘It bothers me that they say that’ ! ‘What bothers you that they say?’ a. ¿De qué materia han aprobado los exámenes varios Of what subject have.AUX passed the exams several estudiantes? students ‘Of which subject have several students passed the exam?’ b. #¿De qué materia ha aprobado los exámenes el/este Of what subject has.AUX passed the exams the/this estudiante? student ‘Of which subject has the student passed the exam?’ Furthermore, as Serrano (2015: 187 et seq.) suggests, verbs that can have el-que El heading finite clauses in Spanish 399 typically select definite DPs as their complements as well, being indefinite DPs marginal in those contexts: (48) a. b. Me fascina el que su cultural milenaria se palpe en cada uno de sus habitantes ‘I love that its millenary culture can be felt in each of its inhabitants’ Me fascina el vestido/su vestido/este vestido/*un vestido ‘I love the dress/her dress/this dress/*a dress’ This pattern is not due to factivity since some factive verbs that cannot have el-que, can have indefinite DPs as their complement: (49) Conozco/Veo/Sé una noticia/noticias ‘I know/I see/I’m aware of news’ In addition, non-factive verbs that accept el-que cannot co-occur with indefinite DPs as their complements either: (50) Su actitud accareó/facilitó {#una situación/esta situación} His aptitude triggered/facilitated {#a situation / this situation} Therefore, it can be claimed that the main contribution of the article in clauses is referentiality in a similar way to what happens in definite DPs. This could be formalized as follows: (51) Among the set of possible situations quantified in a clause hs, ti, the article (D) selects the unique salient situation hsi. 6 Conclusion In this paper, we have dealt with a phenomenon that has not received enough attention in the scientific literature of the Spanish language: the definite article el heading finite que clauses. Even though the article is not compulsory, its distribution is not free, and it is subject to some semantic and discursive properties. To be more specific, the clause must belong to the Common Ground and the Background by being related to some content present in the previous discourse. In other words, it cannot be the main point of utterance since it cannot be at issue, as it shows the impossibility of el-que appearing as an answer or preceded by foci adverbs. Due to space limitations, we have solely addressed the semantic properties of the phenomenon. We have concluded by stating that the contribution of the article in the clause is the referentiality of the clause in a similar manner to that of DPs. To endorse our proposal, we have shown some empirical predictions related to the instantiation of the el-que in the real world, unlike what happens in bare CPs. In fact, this is the main contribution of the article. Vendler (1967) or Zucchi (1993) address the different linguistic entities that can exist according to the “object” they refer to. DPs are determined physical objects, while CPs are abstract entities, even though they can be located in some space and time since they typically are events, states of affairs, etc. 400 CRISTINA RUIZ ALONSO Facing this classification, the construction we are accounting for could be an instance of a hybrid element, as has been proposed, in some ways for nominalizations (Kornfilt & Whitman 2011; Alexiadou 2010; Chomsky 2019). El-que clauses are not facts themselves, but the article forces the clause to be instantiated in the real world. The function of the article is “closing” all the possibilities that the clause headed by que has opened. Since we are defending that the clause acts as a noun, el turns the clause into a definite fixed object, being a unique salient situation belonging to the real world and being parallel to the behavior in DPs. This is proposed by Kastner (2015) who claims that complements of factive verbs are DPs due to the presupposition of existence11 . This semantic proposal is formalized in the syntax as well. Although we cannot address this topic in this work, we will briefly state that a definite operator which prevents the extraction and other movements and interventions due to the fact that the landing site is already filled by this null Operator in Spec, CP12 . We can be dealing with a cross-linguistic pattern where the nominal status of the clause and the referentiality (mentioned here as a general term) are related. References Abney, S. 1987. The English noun phrase and its sentential aspect. MIT dissertation. Alexiadou, A. 2010. The Syntax of Nominalizations across Languages and Frameworks (Interface Explorations. De Gruyter Mouton. Chomsky, N. 2019. Remarks on nominalization. In Studies on Semantics in Generative Grammar, 11–61. De Gruyter. Davidson, D. 1968. On saying that. Synthese 19. 130–146. de Cuba, C., & J. MacDonald. 2013. Referentiality in Spanish CPs. John Benjamins. de Cuba, C., & B. Urogdi. 2010. Clearing up the ‘facts’ on complementation. In University of Pennsylvania Working Papers in Linguistics 16(1), 41–50. Demonte, V. 1977. La Subordinación Sustantiva. Madrid: Cátedra. Demonte, V. 2012. Las completivas en subjuntivo y la no-vericidad en la periferia izquierda oracional. In Studium grammaticum. Homenaje al profesor José A. Martínez. Universidad de Oviedo. Fábregas, A. 2014. A guide to subjunctive and modals in spanish: questions and analyses. Borealis - An International Journal of Hispanic Linguistics 3. 1. Haegeman, L., & B. Urogdi. 2010. Referential cps and dps: an operator movement account. Theoretical Linguistics 36. 111–152. Hooper, J., & S. Thompson. 1973. On the applicability of root transformations. Linguistic Inquiry 4. 495–497. Kadmon, N. 2001. Formal Pragmatics. Oxford: Blackwell. Kastner, I. 2015. Factivity mirrors interpretation: the selectional requirements of presuppositional verbs. Lingua 164. 156–188. Kayne, R. 2008. Why isn’t this a complementizer. New York University. 11 Moreover, the author proposes the existence of a null nominal functional layer (a definite determiner) heading the clause to endorse his proposal. 12 Although we could not go deeper into this question, we can claim that the construction formed by a determiner heading a clause is not restricted to Spanish and it is also present in Greek, Hebrew or Persian. The main properties of those structures in those languages are akin to that of Spanish sketched in this paper (cf. Roussou 1991, 2010; Kastner 2015; Öhl & Lofti 2006). El heading finite clauses in Spanish 401 Kiparsky, P., & C. Kiparsky. 1970. Fact. In Progress in Linguistics. De Gruyter. Kornfilt, J., & J. Whitman. 2011. Afterword: nominalizations in syntactic theory. Lingua 121. 1297–1313. Leonetti, M. 1999. El articulo. In Gramática Descriptiva de la Lengua Española, ed. by I. Bosque & V. Demonte, 787–890. Madrid: Espasa Calpe. Levinson, S. 2000. Pragmatics. Cambridge: Cambridge University Press. Manzini, R., & L. Savoia. 2011. Grammatical categories. Variation in Romance Languages. Cambridge Studies of Linguistics. Moulton, K. 2020. Remarks on propositional nominalization. In Nominalization: 50 Years on from Chomsky’s Remarks, 255–276. Oxford University Press. Octavio de Toledo, A. 2014. Espejismo de la frecuencia creciente: gramaticalización y difusión del artículo ante oraciones sustantivas. RILCE 3. 916–958. Panagiotidis, P., & K. Grohmann. 2005. Mixed projections: Categorial switches and prolific domains. Special Issue of Biolinguistics celebrating Biological Foundations of Language 35. Picallo, C. 2001. Nominalized clauses, clausal arguments and agreement. In Catalan Working Papers in Linguistics, volume 9, 69–84. Picallo, C. 2002. Syntax 5. 116–147. Poletto, C., & E. Sanfelici. 2022. Against complementizers. In Continuity and Variation in Germanic and Romance, ed. by S. Wolfe & C. Meklenberg. Oxford Academic. Quer, J. 2001. Interpreting mood. Probus 13. 81–111. Roberts, C. !996. Information structure in discourse: towards an integrated formal theory of pragmatics. Semantics and Pragmatics 5. 1–69. Rosenbaum, P. 1967. The Grammar of English Predicate Complement Constructions. Cambridge, MA: MIT Press. Roussou, A. 1991. Nominalized clauses in the syntax of modern greek. In UCL Working Papers in Linguistics 3, 77–100. Roussou, A. 1994. The syntax of complementizers. UK: University College London. Roussou, A. 2010. Selecting complementizers. Lingua 120. 582–603. Serrano Pardo, S. 2015. Subordinación y determinación: completivas precedidas de artículo definido en español. Sheehan, M., & W. Hinzen. 2011. Moving towards the edge. Linguistic Analysis 37. 405–458. Stalnaker, R. 1978. Assertion. In Syntax and Semantics 9: Pragmatics, ed. by P. Cole, 315–322. New York: Academic Press. Stalnaker, R. 2002. Common ground. Linguistics and Philosophy 25. 701–721. Vendler, Z. 1967. Linguistics in Philosophy. Ithaca: Cornell University Press. Zucchi, A. 1993. The Language of Propositions and Events. Dordrecht: Kluwer Academic Publishers. Öhl, P., & A. Lofti. 2006. Nominalised cps in persian: a parametric account. In Second International Conference on Iranian Linguistics. Proceedings of CLS 58 (2022), 403-415 © Chicago Linguistic Society 2023. All rights reserved. 403 Instrument terms, bare singulars, and event kinds Starr Sandoval1 , Daniel Greeson2 , & Marcin Morzycki1 1 University of British Columbia, 2 Stony Brook University 1 Introduction The definite determiner in English is, of course, not normally semantically vacuous. Yet in a class of sentences that contain singular instrument terms, exactly that seems to be the case. There doesn’t seem to be any obvious truth-conditional difference between the two forms in (1): ⇢ piano . (1) Floyd plays the piano A contrast between the two forms emerges, however, in the presence of adjectival modification: ⇢ excellent piano (2) Floyd plays # . (on kind reading of piano) the excellent piano The bare form allows an adverbial reading of the adjective (in the sense of Stump (1985) and others), yielding a meaning approximately like ‘Floyd plays piano excellently’. The modified definite form lacks this interpretation of excellent. More surprising still, it also eliminates the kind interpretation of piano and must be about a particular piano rather than the kind, meaning only something like ‘Floyd plays the particular piano that was excellent’. The aim of this paper is to account for this contrast between bare and definite instrument terms as well as their distributional and interpretive differences more generally. We will propose that bare instrument terms denote EVENT KINDS. This makes possible two varieties of modification that would otherwise not be available: adjectives with adverbial readings, like excellent in (2), and adjectives associated with subkinds of playing events rather than subkinds of instruments. Crucial to our analysis will be distinguishing a form of play that occurs with definites from the play that occurs with bare nouns. The former, more canonical form combines with its object conventionally. The latter form is a light verb on which play combines with an event description of a performance. Section 2 presents novel data regarding the distributional and interpretive differences between bare and definite instrument terms, focusing on their kind-denoting properties, their occurrence with play, and the types of modifiers they allow. Section 3 discusses prior research on kinds, weak definites and bare singulars, and nonlocal modifiers—including adjectives with adverbial interpretations—and their relation to event kinds. Section 4 presents an event kind analysis of bare instrument terms and elaborates how this interacts with the light verb play to yield the observed 404 SANDOVAL, GREESON AND MORZYCKI pattern of facts. Section 5 provides a broader view, considering some possible extensions of the analysis. 2 The facts Both bare and definite instrument terms can denote kinds. This is evidenced by their compatibility with kind-level predicates, such as emerged and widespread: ⇢ Piano (3) a. emerged in the 18th century. The piano ⇢ Piano b. is widespread in this town. The piano This kind interpretation remains in play constructions. Both the definite and bare form of piano can occur as objects of play without referring to a particular piano, as in (4): ⇢ piano (4) The students played last night. the piano Crucially, (4) is true under a reading in which each student played a different piano. It is unsurprising, given previous research, that the piano can denote a kind. Definite descriptions can in general be kind denoting, irrespective of whether they involve musical instruments (Chierchia 1998; Dayal 2004): ⇢ camera emerged in the 18th century. (5) The automobile The uses of instrument terms in play constructions more closely resemble weak definites (Schwarz 2009; Aguilar Guevara & Zwarts 2011), which are definite descriptions that lack the usual presupposition of the existence of a unique identifiable individual: ⇢ post office . (6) Voters across the country sent off their ballots at the library Aguilar Guevara & Zwarts (2011) propose that weak definites also refer to kinds, so these facts both ultimately point to a kind-referring analysis of at least some instrument terms. However, bare singulars are not generally assumed to be able to denote kinds, so the kind interpretation of bare piano in (3) and (4) is already puzzling. While both bare and definite instrument terms have kind interpretations, their readings under different types of modification reveal subtle differences between them. In general, kind expressions tend to have more restricted modification possibilities, but notably can be modified by adjectives that specify a subkind. For example, electric guitar is a subkind of guitar, as it is its own distinguished type of instrument. That contrasts with other modified forms of guitar such as small Instrument terms, bare singulars, and event kinds 405 guitar or red guitar—neither of these is normally understood as a subkind of the kind GUITAR. Therefore, unsurprisingly, guitar can maintain its kind interpretation when it undergoes modification by electric, because of course, electric guitar is a widely known and generally superior form of guitar. A similar case is bass trumpet in (7), which again references a subkind of trumpet: ⇢ electric guitar emerged in the 20th century. (7) a. (The) bass trumpet ⇢ electric guitar . b. Clyde plays (the) bass trumpet Interestingly, though, there is a difference between bare instrument terms and their definite counterparts with respect to kind reference. Only bare instrument terms support a reading involving a subkind of music rather than a subkind of instrument. For example, country is a style of music, not a kind of instrument. There exists no distinguished well-established country guitar instrument kind. So country guitar has a kind reading referring to the guitar playing associated with country music, but it does not have a reading involving a (non-existent) instrument, a country kind of guitar. Jazz trumpet is similar, in that it too is a style of trumpet playing but not a kind of trumpet, and this behaves similarly. Both are impossible with the definite determiner:1 ⇢ country guitar # (8) a. The emerged in the 20th century. jazz trumpet ⇢ country guitar . b. #Clyde plays the jazz trumpet Without one, both are possible, and crucially, about styles of music-playing: ⇢ country guitar . (9) a. Clyde plays jazz trumpet ⇢ Country guitar b. emerged in the 20th century. Jazz trumpet These subkind modification facts suggest more generally that bare instrument terms denote kinds of playing events while their definite description counterparts denote kinds of physical instruments. Though more subtle, this contrast also exists without modification given the appropriate context. For example, if, in a post-apocalyptic society, archaeologists rediscovered physical guitars as artefacts, but did not know how to play them or indeed what they were used for at all, this discovery can be described with the definite description but not with the bare nominal: 1 An alternative perspective is that the country guitar in (8) coerces country guitar to be interpreted as a kind of instrument. It conveys that there exists a kind of physical instrument called a ‘country guitar’. This still falls in line with the observation that kind-denoting definite instrument terms only modifiers that denote a subkind of the instrument. 406 SANDOVAL, GREESON AND MORZYCKI (10) ⇢ The guitar #Guitar reemerged in 2050 and puzzled everyone. In the definite case, it is the physical instrument that reemerged. In the bare case, it would have to be guitar playing. This difference also has consequences for co-occurrence with adjectives that characterize quality such as excellent. Such modifiers typically prevent kind reference in definite descriptions, as these adjectives don’t identify subkinds. Thus, predictably, such modifiers eliminate the kind reading of definite instrument terms: 9 8 good > > = < bad # piano. (11) Floyd plays the (on kind reading of piano) > ; :excellent> lousy This, however, changes with bare instrument terms, which do allow quality-characterizing modification and kind reference to co-occur. Furthermore, these modifiers, again, describe the playing of an instrument as opposed to the physical instrument itself. Thus (12a) can be paraphrased as (12b): (12) a. Floyd plays good piano. b. Floyd plays piano well. Here is a table that summarizes the types of modifiers bare and definite instrument terms allow on a kind reading: (13) type of modification individual subkind event subkind (electric guitar) (country guitar) bare Yes definite Yes Yes No evaluative (good guitar) Yes No Both definite and bare instrument terms can be kind-denoting and both forms allow modifiers that denote a subkind of the instrument. Bare instrument terms are however more promiscuous—they additionally allow modification involving a subkind of music playing (e.g. country or jazz) and quality (e.g. good or bad). 3 Analytical tools 3.1 Event kinds To lay the groundwork for our analysis, we need to articulate two general analytical assumptions. The first of these is an ontological one about how the domain of kinds and the domain of events are related. Instrument terms, bare singulars, and event kinds 407 In classical discussions of kinds such as Carlson (1977) and Chierchia (1998), kinds are viewed as cousins to non-kind (object) individuals. For Carlson, it is simply a sortal distinction. For Chierchia, there is a bit more that must be said, but kinds are still constructed, ultimately, out of object individuals. That might be taken to suggest that kinds and individuals are inextricably linked. But that doesn’t seem to be the case, metaphysically or linguistically. A priori, if particular individuals can be realizations of kinds, there is no reason to assume that particular events can’t also be realizations of kinds of events. Indeed, Chierchia’s conceptualization of the issue would seem to actually predict that there should be event kinds. For him, kinds are essentially pluralities of possible individuals, so that the kind RABBITKIND is ultimately the plurality of rabbits across possible worlds. If kinds are such pluralities, it is expected that there should also be event kinds, because it is possible to construct pluralities of possible events. Thus the event kind PIANO-PLAYING-KIND is the plurality of all possible events of playing a piano across possible worlds. In this respect, the existence of event kinds is at least the null hypothesis, and arguably perhaps actually a prediction of standard theoretical assumptions. But an increasing amount of empirical evidence has also accumulated that there are kinds in the event domain. Gehrke (2019) provides a general review, but one early argument in this direction (Landman & Morzycki 2003) comes from anaphors with both adnominal and adverbial uses, which are anaphoric, in their adnominal use, to kinds, and in their adverbial use, to manners, which can be construed as event kinds. Another empirical argument for event kinds comes from adjectival participles in German (Gehrke 2015). What is especially relevant for current purposes, though, is that event kinds have independently arisen in the empirical neighborhood we’re currently exploring. Schwarz (2014) relies on them in work on the aforementioned weak definite descriptions, which definite descriptions of instrument names clearly resemble. Another connection is in how adjectives behave with instrument terms. As discussed in 2, these often get adverbial readings—playing excellent piano is playing piano excellently. The classic example of an adjective with adverbial readings is occasional and its close relatives—an occasional sailor sails occasionally, and if an occasional sailor strolled by, a sailor strolled by occasionally. Gehrke & McNally (2015) argue that this surprising adjective-adverb correspondence should also be understood ultimately as a consequence between manner and event kinds. For our purposes here, the crucial component is just that there is a solid independent basis to assume that event kinds exist, and even to assume that they are relevant to the analysis of the particular data we seek to explain here. 3.2 Derived Kind Predication and composition with instrument terms The other major analytical assumption we need to put into place involves how kinds are composed, both in general and in cases like play the piano specifically. Grammatically, it is normally possible to combine a kind-denoting expression with a predicate that expects to combine with an ordinary object rather than a kind. As a matter of logic, though, that’s unexpected, and an additional assumption is necessary to knit these pieces together. The standard one is a type shift, Chierchia (1998)’s Derived Kind Predication (DKP). A typical way to express it is by sup- 408 SANDOVAL, GREESON AND MORZYCKI posing that an instance of applying an a predicate of objects P to a kind k should be construed as an existential claim: that P holds of an object that realizes k. This much is standard. For our purposes, though, we will need to generalize the usual formulation of Derived Kind Predication slightly to accommodate event kinds. It will be crucial here that the object that realizes the relevant kind can be either an individual or an event. To reflect this, our statement of Derived Kind Predication uses the variable o, which we take to range over both individuals and events that are objects rather than kinds. Correspondingly, k ranges over both individuals and events that are kinds. Beyond that, everything will be standard, including the use of the shifting operator [ , which maps a kind to the property of realizing the kind: (14) D ERIVED K IND P REDICATION (Crosscategorial Variant) P (k) = 9o[[ k(o)^P (o)] where k is a kind, o an object, and P a property Thus, when the object-level predicate arrive combines with the kind RABBITKIND, it amounts to saying that there was a rabbit realization that arrived: (15) arrive(RABBITKIND) = 9x[[ RABBITKIND(x) ^ arrive(x)] That’s because [ RABBITKIND is simply the property of being a rabbit. We will follow Aguilar Guevara & Zwarts (2011) in assuming that ordinary play takes an object as its argument, and that it too can combine with a kind-denoting expression via Derived Kind Predication. They observe that this is sufficient to provide a semantics for cases where ordinary play combines with definite descriptions of instruments. That’s reflected in (16), which further assumes with them that the direct object is introduced via the neo-Davidsonian thematic-role predicate theme (it also uses a superscript o to flag that xo is a variable over objects rather than kinds): (16) a. J the piano K = PIANO-KIND b. J playobject K = λxo λe . play(e) ^ theme(e, xo ) c. J playobject the piano K = J playobject K (J the piano K) (by DKP) = λe . 9xo [[ J the piano K (xo ) ^ J playobject K (xo )(e)] o [ o o = λe . 9x [ PIANO-KIND(x ) ^ play(e) ^ theme(e, x )] This helps explain the definite description portion of the instrument term facts. But what about bare instrument terms? 4 Analysis 4.1 Bare singular instrument terms A natural move, especially given the Aguilar Guevara & Zwarts (2011) insight into definite description instrument terms, is to suppose that bare cases like play piano are very similar. Indeed, one might even suppose that play piano is just an elided Instrument terms, bare singulars, and event kinds 409 form of play the piano. But taking quite so literal-minded an approach is not available to us because of the differences already noted in 2 between definite and bare instrument terms. If play piano were an elided form of play the piano, we would expect precisely the same modification possibilities in these cases—yet as we have seen, play excellent piano does not mean the same thing as play the excellent piano. For the same reason, it wouldn’t suffice to suppose that bare piano means precisely what the piano means: that they both refer to PIANO-KIND. Again, this would fail to explain the difference. A relatively conservative step forward is to suppose that bare piano differs from the piano in that it refers to a slightly different kind. Where the piano refers to a kind of instrument, piano refers to a kind of music. One might even imagine implementing this syntactically as a variety of ellipsis as well, eliding the word music. This is not quite the path we will take, but it will be a helpful intermediate hypothesis. It’s progress. It would explain why bare instrument terms can accept subkind modifiers that involve subkinds of music rather than subkinds of instrument. As noted before, country guitar is a style of playing guitar, but not a subkind of guitar. If bare instrument terms—unlike their definite counterparts—refer to kinds of music, it would follow that play country guitar is possible, because it is essentially ‘play country guitar music’, but that #play the country guitar is not, because it isn’t about a subkind of music. Rather, it is about subkinds of instruments, ‘play instruments that realize the country guitar instrument kind’, which does not exist. Better still, assuming bare instrument terms refer to music kinds would explain the evaluative modification facts. Play excellent guitar is possible because it is essentially ‘play excellent guitar music’. On the other hand, play the excellent guitar is not possible on a kind reading, because it is essentially ‘play instruments that realize the excellent guitar kind’. Here there is a slight bump in the road. Aren’t excellent guitars a kind? It is, after all, possible to refer to such a kind with bare plurals: (17) Until recently, all guitars were mediocre. Excellent guitars emerged only in the last few decades. But definite descriptions of kinds aren’t just like any other form of kind reference. To refer to a kind with a definite description, it is necessary not just that the kind exist, but that it be well established (Carlson (1977), Dayal (2004), among others). The classic example, attributed by Carlson (1977) to Barbara Partee, involves kinds of bottles: ⇢ Coke bottle has a long neck. (18) The # green Green bottles may be a kind of bottle, but they are not a well established kind of bottle. How to spell this out precisely is of course an interesting issue, but not immediately crucial. What’s crucial for us is just that this independently predicts that the definite description the excellent guitar could only refer to a kind if it was well established, and excellent guitars are not a well established subkind of guitar. 410 SANDOVAL, GREESON AND MORZYCKI Another advantage of the hypothesis that bare instrument terms refer to kinds of music is that it would explain why they—surprisingly—behave like mass terms. They occur with quantity expressions like too much, which is only possible with mass terms: ⇢ cheese . (19) a. Floyd ate too much * sandwiches ) ( piano . b. Floyd played too much *pianos *the piano Kinds bear a fundamental conceptual resemblance to masses (Chierchia 1998; Dayal 2004), so its a satisfying connection to make. It is here, however, that the kind-of-music approach begins to go wrong. Music can be individuated with classifier-like expressions such as piece or arrangement, but that isn’t the case for bare instrument terms: ⇢ ⇢ piano music a piece . of # (20) Floyd played piano an arrangement Such contrasts aren’t just about the mass-count distinction. That’s especially clear in sentences like (21). Expressions like collection are perfectly compatible with both masses and plurals—the mass term piano music and the plural recipes in (21)— but they too resist bare instrument terms: ) ( recipes (21) This book is a collection of piano music . #piano Rather, the crux of the issue is that bare instrument terms refer not just to a mass kind, but a mass event kind: the playing of piano music. One reason to suspect that this is the relevant distinction is that the overtly eventive expression piano playing is impossible in both (20) and (21), just as bare piano is: ⇢ ⇢ piano a piece . of (22) a. #Floyd played piano playing an arrangement ⇢ piano # . b. This book is a collection of piano playing Piano playing can’t naturally be substituted for piano in (19), but that’s probably due to the awkwardness of repeating play. 4.2 Building an eventive denotation To spell this out, we’ll assume that bare instrument terms involve a null number head EVENT. Because it occupies the Num head, it doesn’t co-occur with plural morphology. Its role will be to shift a property of being an instrument to the property of playing that instrument: Instrument terms, bare singulars, and event kinds (23) 411 J EVENT K = λPhe, ti . \ [λe . 9xo [e is an event of playing xo ^ P (xo )]] But as (23) reflects, it does this in a particular way that puts kinds into the picture. That’s reflected in the presence of \ type shift, which shifts properties to their corresponding kinds. What EVENT does is build the property of playing some instrument, and then nominalize that property into its kind counterpart. That’s a natural move because we now have ample evidence that this construction is fundamentally about kind reference. But, as we are now discovering, it’s also about events, so it’s fitting that it’s a property of events that it nominalizes, yielding an event kind rather than an ordinary individual kind. The effect of this is that when EVENT combines with the NP piano, it yields the event kind of playing a piano, namely, PIANO-PLAYING-KIND. This explains why bare instrument terms are, in fact, bare. The NumP EVENT piano refers to a kind. Consequently, it can’t combine with an overt determiner, because English determiners expect property-denoting complements, not kind-denoting ones.2 Another advantage of this approach is that it correctly links the possibility of an eventive reading to the absence of plural morpheme with which it competes for a structural position. Eventive readings really do seem to be restricted in this way. An eventive reading is possible for (24a), but not for the plural-marked (24b), which unambiguously describes a dangerous onstage accident involving wayward pianos: (24) a. b. c. d. Piano emerged from the orchestra. Pianos emerged from the orchestra. The piano emerged from the orchestra. Drums emerged from the orchestra. (eventive) (not eventive) (not eventive) (not eventive) Because EVENT is incompatible with an overt determiner, (24c) is also not eventive. Perhaps, one might object, the problem is that pianos are only ever played one at a time, independently ruling on an eventive reading of pianos. But that’s not the case for drums, which are normally played in groups, yet (24d) still robustly resists an eventive reading. The EVENT shift may actually be a special case of more generalized shifts. Harley (2008) and Kiparsky (1997) suggest bare singulars other than just instrument terms can involve a shift to events canonically associated with the noun. This occurs, for example, in The cow calved or We saddled the horse. We will leave this connection unexplored here because making it more fully would entail explaining what blocks, for example, sandwich from being shifted into the kinds of eating of sandwiches. 4.3 Subkind and adverbial adjectives This makes the correct predictions with respect to modification. Because the kind is an event kind, it can accept event kind modifiers, so country guitar can refer to a subkind of guitar-playing event and thereby be correctly distinguished from #the 2 An exception may be galore, which seems to be a determiner that require kinds (Morzycki 2011)—and, as this predicted, piano galore is well formed. 412 SANDOVAL, GREESON AND MORZYCKI country guitar, which would on its kind reading require the existence of an instrument subkind instead. It may also correctly predict that adjectives modifying bare instrument terms get adverbial readings, with only the fairly minimal assumption that adjectives can sometimes denote properties of events, so that excellent piano could would get the denotation in (25): (25) J [DP excellent EVENT piano ] K h h ii e is an event of playing xo ^ = \ λe . 9xo piano(xo ) ^ excellent(e) Here we must confess to perpetrating a bit of compositional magic. Because excellent denotes a property of events, it must attach to a node that also denotes a property of events to be interpreted intersectively. But on our implementation, the shift to properties of events and the shift to kinds happen at the same point in the tree, so there is no node that denotes a property of events and therefore no appropriate place for the adjective to adjoin. The simplest solution would be to split the EVENT shift into two parts—the shift to events introduced by EVENT, and the shift to kinds by Chierchia’s independent freely available nominalizing type shift applying subsequently—but space limits preclude us from pursuing this further. 4.4 Eventive readings of play A major remaining question is how all this works to yield the observed readings for VPs headed by play. To assemble the pieces, we will need to first establish an empirical point. Alongside the familiar form of play that takes an instrument term as its object, there is what we’ll call eventive play, which combines instead with an event: 8 9 the gig > > < = the big closing number (26) Clyde will play . > > :the overture ; an encore Eventive play imposes on its complement the requirement that it denote some form of performance, as reflected in (27), but beyond that it is essentially a light verb. Its principal effect is to bring about as the VP denotation the property of being that event (we continue to assume the neo-Davidsonian thematic role strategy to introduce the agent): (27) a. J playeventive K = λe : performance(e) . λe0 [e0 = e] b. J the gig K = ιe[gig(e)] c. J playeventive the gig K = λe0 [e0 = J the gig K] = λe0 [e0 = ιe[gig(e)]] With this, everything else falls into place. It is playeventive that occurs with bare instrument terms, and it can compose straightforwardly via Derived Kind Predication: (28) J playeventive K (J [DP EVENT piano ] K) = λe. J playeventive K (J [DP piano ] K)(e) Instrument terms, bare singulars, and event kinds = λe . 9e0 [ [ J [DP 413 piano ] K (e0 ) ^ J playeventive K (e0 )(e)] ] (by DKP) = λe . 9e0 [[ PIANO-PLAYING-KIND(e0 ) ^ e = e0 ] = λe . [ PIANO-PLAYING-KIND(e) EVENT The resulting VP denotation is simply the property of being a realization of the piano-playing event kind—or, more simply, the property of being a piano-playing. Thus the generic flavor associated with the kind is defused, correctly predicting that there is nothing generic about Floyd played piano yesterday, and that it simply means that Floyd was the agent of an event of playing some piano. Of course, the generic flavor reemerges in Floyd plays piano, but that’s due entirely to the habitual interpretation of English tense morphology. With the other assumptions above, this also correctly predicts that play excellent piano should wind up with the same truth conditions as play piano excellently because it will simply involve the property of realizing the kind of excellent piano playings. Such an interpretation is blocked for play the excellent piano because the precludes an eventive reading. 5 Concluding remark and a look beyond instruments This paper proposed a means of interpreting bare singular instrument terms that explains their eventive interpretations and the adverbial and subkind interpretations of adjectives that modify them. The crucial ingredient is a shift to kinds of events present only in the singular, which is syntactically incompatible with plural morphology and semantically incompatible with overt determiners—and conceptually incompatible with non-instrument terms because it lexically encodes that the relevant events are instrument-playings. This, coupled with independently motivated assumptions about kind reference and semantic composition, predicts the readings of bare singular instrument terms and their define counterparts. Along the way, we identified light-verb reading of play that may well be novel. Having done this, it’s natural to wonder just how specific or general this phenomenon is. Our account suggests it’s very specific indeed, the consequence of a particular morpheme in English. That would of course be interesting in itself, and interesting for what it tells us about kinds, events, semantic composition, and so on. And yet, one has the sense that bigger fish lurk in this corner of the linguistic sea. We’ve already noted one respect in which this is the case—the Kiparsky (1997) and Harley (2008) facts about e.g. calving and saddling of horses. But it might be the case that it’s possible to broaden our focus from instrument terms in a more incremental fashion. That’s what sentences like (29) suggest: (29) a. Clyde fries a good steak. b. Floyd cuts good hair. c. Bertha throws a good ball. (Clyde is good at frying steak.) (Floyd is good at cutting hair.) (Bertha is good at ball-throwing.) These readings are a bit more unstable and variable across speakers than those of instrument terms, but clearly something must be going on here. After all, even 414 SANDOVAL, GREESON AND MORZYCKI if the readings of adjectives in (29) are marginal, what’s remarkable is that these readings are available at all. Moreover, there doesn’t seem to be any variation with other judgments in this neighborhood. All speakers reject counterparts of (29) with a definite description, except as a characterization of some particular good hair or steak or ball. Equally robust is the judgment that the reading the objects in (29) receive seem not to be available in arbitrary syntactic positions, which distinguishes them from instrument terms: (30) a. There was piano in that performance. b. There was hair in that cosmetics class. (piano = piano playing) (hair 6= hair-cutting) Nevertheless, there does seem to be some sort of connection. There may also be a connection to cognate object constructions (as in dream a nice little dream) and semantic incorporation. Also an obvious issue for future research: to what extent do readings like the ones observed for bare instrument terms here occur crosslinguistically, and what compositional mechanisms give rise to them? References Aguilar Guevara, A., & J. Zwarts. 2011. Weak definites and reference to kinds. In Proceedings of Semantics and Linguistic Theory (SALT) 20, ed. by N. Li & D. Lutz. eLanguage. Carlson, G. 1977. Reference to Kinds in English. University of Massachusetts Amherst dissertation. Published in 1980 by Garland. Chierchia, G. 1998. Reference to kinds across languages. Natural Language Semantics 6. 339– 405. Dayal, V. 2004. Number marking and (in)definiteness in kind terms. Linguistics and Philosophy 27. 393–450. Gehrke, B. 2015. Adjectival participles, event kind modification and pseudo-incorporation. Natural Language and Linguistic Theory 33. 897–938. Gehrke, B. 2019. Event kinds. In The Oxford Handbook on Event Structure, ed. by R. Truswell, 205–233. Oxford University Press. Gehrke, B., & L. McNally. 2015. Distributional modification: The case of frequency adjectives. Language 91. 837–870. Harley, H. 2008. Bare roots, conflation and the canonical use constraint. Handout for a talk presented at the NORMS Workshop on Argument Structure, University of Lund. Kiparsky, P. 1997. Remarks on denominal verbs. Complex predicates 64. 473–499. Landman, M., & M. Morzycki. 2003. Event-kinds and the representation of manner. In Proceedings of the Western Conference on Linguistics (WECOL) 2002, ed. by N. M. Antrim, G. Goodall, M. Schulte-Nafeh, & V. Samiian, volume 14, 136–147, Fresno. California State University. Morzycki, M. 2011. Quantification galore. Linguistic Inquiry 42. 671–682. Schwarz, F. 2009. Two Types of Definites in Natural Language. University of Massachusetts Amherst dissertation. Schwarz, F. 2014. How weak and how definite are weak definites? In Weak Referentiality, Linguistik Aktuell/Linguistics Today, 213–235. Ana Aguilar-Guevara and Bert Le Bruyn and Joost Zwarts. Stump, G. 1985. The Semantic Variability of Absolute Constructions. Dordrecht: D. Reidel. Proceedings of CLS 58 (2022), 415-427 © Chicago Linguistic Society 2023. All rights reserved. 415 PF-LF Domain Mismatches under Ellipsis and the Non-Simultaneous Transfer Hypothesis Yosuke Sato Tsuda University 1 Introduction In this paper, I will develop a new generalization concerning ellipsis, shown in (1), which holds that PF-deletion domains are fundamentally misaligned with LFidentity calculation domains in natural language syntax. 1 (1) The PF-LF Mismatch Generalization under Ellipsis PF-deletion of XP requires LF-identity within YP, sister to X. I will show that this generalization is supported by a wide range of possible and impossible mismatches in various languages discovered in the literature, marshalled in (2). (2) a. b. c. d. e. f. g. Possible voice mismatch under English VP-ellipsis and pseudogapping Impossible voice mismatch under modal complement ellipsis Impossible voice mismatch under Indonesian “VP-ellipsis” Polarity/finiteness/illocutionary force/tense mismatch under sluicing Antecedent-ellipsis size mismatch under antecedent-contained sluicing Causative-inchoative argument structure mismatch under VP-ellipsis Case particle and focus mismatch under argument ellipsis in Japanese I will then strive to derive the generalization (1) from a version of the NonSimultaneous Transfer Model (Marušić 2005) in tandem with computational efficiency imposed on memory buffers storing identity information at LF. 2 The PF-LF mismatch generalization under ellipsis in action In this section, I will present evidence that the generalization in (1) is motivated by a wide range of mismatches attested across languages, catalogued in (2a−g). 2.1 Voice mismatch under VP-ellipsis and pseudogapping 1 This work is supported by JSPS KAKENHI JP19K00560. I thank Mike Barrie, Jeroen van Craenenbroeck, Yoshi Dobashi, Shin Kitada, Hisa Kitahara, Si Kai Lee, Masako Maeda, Taichi Nakamura, Hajime Ono, Matthew Reeve, Yuta Sakamoto, Kenji Sugimoto, Ken Takita, Hansel Tan, Yusuke Yagi, and Dwi Hesti Yuliani for helpful comments and/or references. All errors are my own. 416 YOSUKE SATO It is well-known that both VP-ellipsis (hereafter, VPE) and pseudogapping accept voice mismatch in English as long as certain discourse/information structural factors are properly controlled for (e.g., Miller 1991; Kehler 2000, 2002; Coppock 2001; Kertz 2010; Tanaka 2011a, b, pace Merchant 2008, 2013a, b). Some representative examples are given in (3) and (4) to illustrate this observation. (3) a. The janitor must remove the trash whenever it is apparent that it should be. b. The system can be used by anyone who wants to. (Merchant 2013a:78, 79) (4) a.?Actually, I have implemented it (= a computer system) with a manager, but it should have been by a computer technician. b.?A new system can be used by anyone who could the older version. (Tanaka 2011a:476) Suppose that both VPE and pseudogapping involve PF-deletion of the VoiceP layer. As per (1), then, syntactic/semantic identity must be, and indeed is, met over the vP-sister of the Voice head, as schematically illustrated in (5). In the rest of this paper, I will annotate the PF-deletion domain with [Voi and the LF-identity calculation domain with [ v . (5) Antecedent Clause: Elliptical Clause: [VoiceP … Voice [−ACT] … [vP … ] ] [VoiceP … Voice [−ACT] … [vP … ] ] This analysis predicts that voice mismatch, which is available under VPE/pseudogapping, should be impossible under sluicing, a case of TP-ellipsis, because the VoiceP complement of the T head contains conflicting voice values, a prediction that is indeed borne out by the ungrammaticality of (6a, b). (6) a. * Someone murdered Joe, but we don’t know why by. b. * Joe was murdered, but we don’t know who. (Merchant 2013a:81) This analysis also predicts that mismatches occurring above the VoiceP layer should, in principle, be tolerated under sluicing because they lie outside the LF-identity domain. As reported by Rudin (2019), this prediction is also borne out by the availability of various forms of mismatch involving finiteness, tense, polarity, and illocutionary force, shown in (7a−d), respectively. 2 2 One might argue that (7c) is problematic for my analysis, for the NegP complement of the ellipsis site (=TP) is negative, but that of its antecedent is positive. Not necessarily. For why that is, see Kroll (2019), Sakamoto (2022), Sato (2022a), and Yagi et al. (2022), among others, who argue that polarity reversal under clausal ellipsis is obtained through the so-called Excluded Middle Assumption (Bartsch 1976; Gajewski 2005, 2007, among others) so that the seeming presence of negation in the syntactic structure underlying the ellipsis site in (7c) is actually an illusion. PF-LF domain mismatches under ellipsis 417 (7) a. The baseball player went public with his desire to be traded. He doesn’t care where [TP he will be traded]. b. Your favorite plant is alive, but you can never be sure for how long [TP it will be alive]. c. Either turn in your final paper by midnight or explain why [TP you didn’t turn it in by midnight]. d. Always save a little from each paycheck. Once you’re older, you’ll understand why [TP you should always save a little from each paycheck]. (Rudin 2019:266, 267) d Using (1) as a new heuristic too, we can go a step further and tell which language has English-type VPE or not. For example, Dagnac (2010) points out that Modal Complement Ellipsis in French, Spanish and Italian block voice mismatch, as attested by the ungrammaticality of the French examples in (8a, b). (8) a.* Il faut remplacer l’ampoule de l’escalier, mais elle ne it needs replace the.bulb of the.staircase but it NEG peut pas [XP …]. − elle est coincée. can NEG it is jammed ‘Someone should replace the bulb in the staircase, but it can’t. − it’s jammed.’ b.* Ce problême aurait dû être résolu, mais visiblement this problem must be-PST-COND solved but obviously personne n’a pu [XP …]. nobody can.PST ‘This problem should have been solved, but obviously nobody did.’ (Dagnac 2010:165) Consistent with this observation, Dagnac argues that Modal Complement Ellipsis in these three languages involves TP-ellipsis. To take another case, Fortin (2007) observes that “VPE” in Indonesian disallows voice mismatch, as illustrated in (9). (9) * Rumah+ku belum di-bersihkan selama dua minggu, jadi house+1SG not.yet PV-clean one.long two week therefore saya mesti [XP membersihkan+nya] akhir minggu ini. I must AV-clean+3SG end week DEM ‘My house hasn’t been cleaned for two weeks, so I must clean it this weekend.’ (Fortin 2007:51) Fortin proposes that this pattern is accounted for if voice specifications reside in v in Indonesian, not in Voice, as proposed for English VPE (Merchant 2008, 2013a, b). However, this proposal is arbitrary in that no explanation is given for why voice features are realized under v or under Voice depending on languages under investigation. My current approach not only avoids this pitfall, but also directly accounts for (9) from an independent observation that Indonesian “VPE” is licensed 418 YOSUKE SATO not only by modal auxiliaries, but also by temporal/aspectual makers and negation, or certain functional heads whose specifier are occupied by these expressions. Voice mismatch is blocked in (9) if the LF-identity domain there, being at least as big as TP (selected by one of these functional heads), contains Voice heads. 2.2 Size mismatch under antecedent-contained sluicing Yoshida (2010) observes that the antecedent of the elided TP in Antecedent-Contained Sluicing (ACS) is the matrix VP, not the matrix TP, and argues that this size mismatch is licensed by the mutual entailment condition on ellipsis (Merchant 2001). This analysis is supported in (10a, b): the elliptical TP in the adjunct clause cannot contain negation or modal force in the matrix clause, which it could were its antecedent the matrix TP. (10) a. John isn’t inviting anyone/someone without saying who [=without saying who he is inviting]. b. John must select a color without knowing which one [=without knowing which color he selects]. (Yoshida 2010:350) However, Takita (2013) shows that voice mismatch is blocked in German and Russian ACS. A Russian example of ACS shown in (11) illustrates this observation. If semantic identity were relevant, as argued by Yoshida, (11) should be grammatical, for the antecedent VP and the elliptical VP mutually entail each other. Then, ACS requires syntactic identity between the antecedent clause and the elided TP. (11) Hans wurde von jemandem guküßt, ohne zu wissen, Hans was by someone kissed without to know wer *(ihn küßte). who.NOM him kissed ‘lit. Hans was kissed by someone without knowing who (kissed him).’ (Takita 2013:667) My current approach captures the inconsequential behavior of ACS. That is, the TP in (10a, b) may undergo PF-deletion because the VoiceP sister to the T in the ACS site is identical to the antecedent VP, as depicted in (12), thereby correctly ensuring the impossibility of case mismatch under ACS, as shown in (11), as well as the obligatory absence of negation and modal force, as shown in in (10a, b). (12) [TP John1 [VoiceP t1 kissed someone] without knowing [CP2 who2 C [TP he1 [VoiceP t1 kissed t2 ] ] ] ] 2.3 Causative-inchoative mismatch under English VP-ellipsis PF-LF domain mismatches under ellipsis 419 Sugimoto (2018) points out that VPE in English allows a causative/inchoative mismatch. Importantly, the causative variant licenses the VPE of the inchoative variant in the elliptical clause, but not vice versa. This directional asymmetry is illustrated by the contrast in grammaticality between (13a) and (13b). (13) a. John believed that the sunshine would melt the big snowballs, but they didn’t. b.*John believed that the big snowballs would not melt, but the sunshine did. (Sugimoto 2018:146, 147) Sugimoto’s observation above runs counter to the commonly held observation in the literature (Sag 1976; Johnson 2004; Hauser et al. 2007; Merchant 2013a; Rudin 2019) that argument structure alternation is prohibited under ellipsis. We can, however, capture this one-way directional mismatch observed under VPE if we adopt a fine-grained decompositional structure for causative and inchoative variants as put forth by neo-constructionist works (Alexiadou et al. 2006; Harley 2009, 2013; Pylkkänen 2002). Suppose that the relevant parts of the structures for the causative and inchoative VPs in (13a) and (13b) are as in (14) and (15), respectively. (14) Causative VPA → Inchoative VPE Causative VPA: [VoiceP EA Voice [vP vCAUSE [√P √break IA ] ] ] Inchoative VPE: [vP IA1 vBECOME [√P √break IA1 ] ] (15) Inchoative VPA → Causative VPE Inchoative VPA: [vP IA1 vBECOME [√P √break IA1 ] ] Causative VPE: [VoiceP EA Voice [vP vCAUSE [√P √break IA ] ] ] In (14), the √P sister to the to-be-elided vP in the inchoative VPE has an identical structure-matching correlate in the antecedent causative VPA. In (15), by contrast, there is no antecedent in the inchoative VPA which is identical to the vP sister to the to-be-elided Voice P in the causative VPE. Note that my analysis also predicts that the argument structure alternation under VPE should still be blocked in the causative-inchoative antecedent-ellipsis pair like (13a, b) if √-materials within the √P layer are mismatched. Sugimoto shows that this prediction is also borne out by the contrast in grammaticality between (16a) and (16b). (16) a. John believed that the wa would raise oil prices, but they1 didn’t [vP t1 vBECOME [√P √rise t1] ]. b.*John believed that the virus would kill the patient, but he1 didn’t [vP vBECOME [√P √die t1] ]. (Sugimoto 2018:149, 150) 2.4 Case particle mismatch under Japanese argument ellipsis Argument ellipsis (AE) in Japanese is known to tolerate case particle mismatches (Takahashi 2006, 2012; Saito 2007). This mismatch has resisted a principled 420 YOSUKE SATO account in the literature (see Saito 2007 for an LF-copy analysis of the mismatch, but see Sato 2022b for some problems with the analysis). Examples in (17−18) illustrate this case particle mismatch in Japanese AE. (17) Taroo-wa zibun-no hahaoya-o/*ni oikaesita. Taro-TOP self-GEN mother-ACC/DAT chased.away ‘Taro chased his mother away.’ (18) Taroo-wa zibun-no hahaoya-ni atta-ga, Hanako-wa e oikasesita. Taro-TOP self-GEN mother-DAT met-but Hanako-TOP chased.away ‘lit. Taro met his other, but Hanako chased e (=her mother) away.’ (Saito 2007:217) (17) shows that the verb oikaesu ‘to chase away’ selects an accusative direct object. Given this restriction in mind, the grammaticality of the object ellipsis example in (18) shows that the elliptic site tolerates the case particle mismatch between the dative and accusative cases. Now, if (1) is right, we can provide a rather straightforward solution to the case particle mismatch puzzle under the so-called K(ase)P hypothesis (Travis and Lamontagne 1992; Bittner and Hale 1996), whereby a case-marked object is housed within the KP. Under this hypothesis, the relevant parts of the noun phrases (TNPs) occupying the object positions in the antecedent and elliptical clauses in (18) will be represented as in (19). (19) Antecedent TNP: [KP [NP hahaoya] ni] ] Elliptical TNP: [KP [NP hahaoya] o] ] The PF-deletion domain in the elliptical clause is the KP. (1) requires that the NP sister to the K head should find an identical expression in the antecedent clause. Since this requirement is met in (19), the case particle mismatch is accounted for. This approach can be extended to yield an illuminating account of an otherwise mysterious observation first reported by Akiyama (2014) (see also Moriyama 2017 and Sato 2022b) that AE prevents focus-marked expressions like dake ‘only’ and bakari ‘only’ from being contained in its interpretation, as illustrated by the impossibility of the focus-inclusive reading for the elliptic object position in (20). (20) Taroo-wa zibun-no tukutta ryoori-dake tabeta. Taro-TOP self-GEN cooked food-only ate Hanako-mo e tabeta. [*focus-inclusive reading: Hanako also ate only Hanako-also ate foods that she cooked on her own.] ‘lit. Taro ate only foods that he cooked on his own. Hanako also ate e.’ This observation is exactly what we predict under our KP analysis of TNPs coupled with (1). Since the to-be-elided KP for the elliptic object position requires an identical NP sister to the K head in the antecedent clause, it follows that the relevant position must be interpreted as the NP minus the K head occupied by dake ‘only’. Note that the current analysis makes another interesting prediction, namely, that if a dake-marked PF-LF domain mismatches under ellipsis 421 phrase in (20) is further embedded under the KP by addition of another particle such as the accusative −o, then such a phrase should be able to have the focus-inclusive reading because the dakeP is now shared between the TNPA and TNPE. This prediction is indeed borne out in (21), which quite easily allows the null object to be interpreted as only foods that Hanako cooked on her own, unlike what we have seen in (20). 3 (21) Taroo-wa zibun-no tukutta ryoori-dake-o tabeta. Taro-TOP self-GEN cooked food-only-ACC ate Hanako-mo e tabeta. [✔focus-inclusive reading: Hanako also ate only Hanako-also ate foods that she cooked on her own.] ‘lit. Taro ate only foods that he cooked on his own. Hanako also ate e.’ 2.5 Other cases of the PF-LF mismatch generalization under ellipsis In this subsection, I would like to point our three other miscellaneous cases from English and Japanese which may, upon closer look, support the generalization in (1), though I will leave an in-depth investigation of these individual cases for future research for reasons of space. One case is concerned with category mismatch under VPE, illustrated in (22−23) (Hardt 1993; Johnson 2001; Kehler 2002; Tanaka 2011b; Merchant 2013b; Nakamura 2013; Tan 2018; Sato 2019). (22) People say that Harry is an excessive drinker at social gatherings. Which is strange, because he never does [VP drink] at any parties. (Hardt 1993:55) (23) Allow us treat you like a graduate before you do [ VP graduate]. (Tan 2018, as cited in Sato 2019:7) Here, the ellipsis of the VP headed by a particular verb, drink and graduate, is seemingly licensed by the deverbal nominals, drinker and graduate, respectively. The acceptability of this category mismatch falls out from my approach if the presumed antecedent nominals have a verbal structure within the nominal superstructure (say, nP): the VPE is licensed in these examples because the √P sister to the V head has an identical structure-matching expression within the deverbal nominals. In this connection, Japanese examples like (24), which exhibit a category mismatch between the antecedent VP and the elliptical NP, may be amenable to the same analysis. 4 (24)? Taroo-wa Taro-TOP 3 4 2000-nin-no tyooshuu-no-mae-de [VP ut]-ta. Kare-no 2000-CL-GEN audience-GEN-front-in sing-PST he-GEN Thanks to Ken Takita (personal communication) for pointing out this observation to me. Thanks to Yusuke Yagi (personal communication) for directing my attention to this example. 422 YOSUKE SATO [NP e]-wa itukiitemo iyasareru-yo-ne. whenever.I.listen.to heal-SFP-SFP ‘intended: Taro sang in front of the 2000 audience members. His songs always heal me.’ TOP The second case worthy of further investigations is one pointed out by Anand et al. (2022) (see also Stockwell 2021) where a small antecedent is paired with a bigger clausal ellipsis, as shown in (25−26). (25) The bodies were discovered just before 1 a.m. when an employee of the shop happened to drive by, noticed [SC lights on] almost three hours after closing time and went inside to see why [TP lights were one]. (Anand et al. 2022:3) (26) With [SC the campaign on hold] − and who knows for how long [TP the campaign will be on hold] − Biden is left without any regular way to make his case to the electorate. (Anand et al. 2022:4) These examples involve TP-ellipsis whose presumed antecedent has the small clause structure, which is smaller than TP. This apparent mismatch is also possible, per hypothesis, since the small clause antecedent and the full-fledged TP-ellipsis site share some identical sub-constituent, most likely, at the PredP (Bowers 1993). The final case worth mentioning is an argument structure mismatch in other ellipsis contexts different from VPE in English. Nakamura (2022) notes that the causative/inchoative mismatch is available in Japanese gapping, as illustrated in (27). (27) Tikazukitutuaru taihu-no seide, approaching typhoon-GEN due.to a.? Kaze-wa ikioi-o ØVtran, sosite kawa-wa mizukasa-ga wind-TOP momentum-ACC and river-TOP water.level-NOM masiteiruVintr. has.increased ‘Winds have built in intensity, and rivers have swollen.’ b.? Kaze-wa ikioi-ga ØVintr, sosite kawa-wa mizukasa-o wind-TOP momentum-NOM and river-TOP water.level-ACC masiteiruVintr. ‘Winds have built in intensity, and rivers have swollen.’ (Nakamura 2022:366) In (27a), the causative verb masu ‘increase’ is gapped with its inchoative variant in the second full-fledged clause. In (27b), the inchoative verb is gapped with its causative variant in the relevant clause. These examples then show that the causative/inchoative mismatch is tolerated under Japanese gapping. PF-LF domain mismatches under ellipsis 423 3 Non-Simultaneous Transfer and Computationally Efficient Interfaces Having now presented a wide range of possible and impossible cases of mismatch under ellipsis from various languages to support (1), let us now consider why such a generalization holds, a question that forms part of the more fundamental question why there is mismatch under ellipsis in natural language syntax in the first place. It is beyond doubt that PF-deletion requires identity (syntactic, semantic, discourse, information structure) over a certain domain, so PF must be able to access identity information obtained at LF, the other wing of the grammar within the standard generative architecture. However, it remains unclear how this intermodular “information trafficking” is actually enabled under a dynamically bifurcated theory of syntax as laid out in Chomsky’s (2000 et seq.) Phase Theory, where the same mid-derivational objects (vP, CP and DP) are transferred synchronously to the two interface levels. Now, following a version of the Non-Simultaneous Transfer Model (Megerdoomian 2002; Felser 2004; Marušić 2005; Matushansky 2005), let us assume that transfer always ships a mid-derivational object to the LF-wing of grammar first and that the identity information obtained from this object is then stored in a temporary buffer so that the PF-wing of grammar can now use it to license PFdeletion. Suppose further that cumulatively updated memory loads on the buffer as syntactic derivation proceeds must be kept to a minimum, as expected under some conception of computational efficiency on interface computations. I conjecture that these two design specifications above − the asynchronous transfer requirement and the minimization of memory buffers’ loads as the information keeping device − can be satisfied in the best possible way if LF vs. PF-transfers are timed differently by one minimal computational step, precisely what the generalization in (1) amounts to. The present deduction of (1) has a couple of important theoretical ramifications. First, the proposed deduction implies that we can ever have PF-deletion as the only silencing operation for ellipsis in natural language; see also Hornstein (2016) and Fujiwara (2022) for arguments supporting the same conclusion from different perspectives. Second, the reduction also implies that there is no need to have recourse to such things as E-feature in Merchant’s (2001, 2008, 2013a, b) sense, which imposes a hybrid intermodular function on an identical constituent, XP, to the effect that it may only under ellipsis if it stands in a mutual identity relationship with some salient antecedent YP. Note that this feature is a hyper-lexical panacea, a sign of featuritis (Boeckx 2015), with no morphosyntactic reflex elsewhere in any known language. 4 Conclusion In this paper, I have argued that the ellipsis-identity domains are fundamentally misaligned in natural language syntax between the two linguistic interfaces by marshalling a wide range of (im)possible mismatches found in the literature on ellipsis. I have further suggested that this misalignment is actually virtually forced under some version of the Non-Simultaneous Transfer Model coupled with certain conceptions of the computationally efficient interface computation. I hope the paper has been 424 YOSUKE SATO successful in proposing an entirely novel way of viewing ellipsis-identity correlations. The proposal advanced here not only sheds light on the nature and functioning of intermodular channeling between sound and meaning interfaces within the context of the Minimalist Program but also helps construct a restrictive theory of ellipsis. One important ramification of the present approach is that only PF-deletion is mechanically available. I would like to end this paper by pointing out one outstanding issue with the PF-LF Mismatch Generalization. The generalization is actually at odds with a rather widely accepted view (Rooth 1992; Fiengo and May 1994; Takahashi and Fox 2005, among others) that a semantic identity domain is sometimes bigger than a PF-deletion domain. This view is supported by examples as in (28a−c). (28) a. John knows which professor we invited, but he is not allowed to reveal [PD which one1 <EC we invited t1>]. b.*John knows which professor we invited, but he is not allowed to reveal [PD which one1 we did <EC invite t1>]. (Takahashi and Fox 2005:230) Takahashi and Fox observe that VPE in (28b) is blocked by sluicing/TP-ellipsis in (28a) due to the MaxElide condition (Merchant 2001). The idea lurking behind this observation is that the parallelism domain/PD is defined over a constituent that contains not only a variable but also a binder lying outside the elliptical constituent/EC; it is this structural parallelism holding between the binder-bindee pair in the antecedent and elliptical clauses that plays a key role in licensing the ellipsis site properly contained within the relevant domain. This idea, in turn, means that the identity calculation domain is bigger than the PF-deletion domain, the exact opposite of (1), which leads us to expect that the former must be always smaller than the latter. I can only conjecture at this point that something special is happing with binding. For example, even in a derivational theory of syntax, we know that the meaning interface must be able to check multiply transferred objects in toto for binding purposes (Boeckx and Grohmann 2007). I would like to leave the question how to reconcile the aforementioned finding with my own for another occasion. References Akiyama, M. 2014. The syntax of focus-doubling in Japanese. In Formal Approaches to Japanese Linguistics 7, ed. by S. Kawahara & M. Igarashi, 1–12. Cambridge, MA: MITWPL. Alexiadou, A., E. Anagnostopoulou, & F. Schäfer. 2006. The properties of anticausatives crosslinguistically. In Phases of Interpretation, ed. by M. Frascarelli, 187−211. Berlin: Mouton de Gruyter. Anand, P., D. Hardt, & J. McCloskey. 2022. The domain of formal matching in sluicing. Ms., University of California, Santa Cruz/Copenhagen Business School. Bartsch, R. 1976. ‘Negative Transportation’ Gibt es Nicht. Linguistische Berichte 27.1–7. Bittner, M., & K. Hale. 1996. The structural determination of case and agreement. Linguistic Inquiry 27.1−68. Boeckx, C., & K. Grohmann. 2007. Remarks: Putting phases in perspective. Syntax 10. 204−222. Bowers, John. 1993. The syntax of predication. Linguistic Inquiry 24. 591−656. PF-LF domain mismatches under ellipsis 425 Chomsky, N. 2000. Minimalist inquiries: The framework. In step by step, ed. by R. Martin, D. Michaels & J. Uriagereka, 89–155. Cambridge, MA: MIT Press. Coppock, E. 2001. Gapping: in defense of deletion. In Proceedings of the Thirty-Seventh Annual Meeting of the Chicago Linguistics Society, ed. by M. Andronis, C. Ball, H. Elston & S. Neuvel, 133–148. Chicago, IL: Chicago Linguistic Society, University of Chicago. Dagnac, A. 2010. Modal ellipsis in French, Spanish and Italian: Evidence for a TP-deletion analysis. In Romance linguistics 2008: Interactions in Romance, ed. by K. Arregi, Z. Fagyal, S. Montrul & A. Tremblay, 157−170. Amsterdam: John Benjamins. Felser, C. 2004. Wh-copying, phases, and successive cyclicity. Lingua 114, 543−574 Fiengo, R., & R. May. 1994. Indices and identity. Cambridge, MA: MIT Press. Fortin, C. 2007. Verb phrase ellipsis in Indonesian. In Proceedings of the Thirty-Seventh Western Conference on Linguistics, ed. by R. Colavin, K. Cooke, K. Davidson, S. Fukuda & A. Guidice, 44−55. La Jolla, CA: Department of Linguistics, University of California, San Diego. Fujiwara, Y. 2022. Movement approach to ellipsis. Doctoral dissertation, University of Connecticutt, Storrs. Gajewski, J. 2005. Neg-Raising: Polarity and Presupposition. Doctoral dissertation, MIT. Gajewski, J. 2007. Neg-Raising and Polarity. Linguistics and Philosophy 30.298–328. Hardt, D. 1993. Verb phrase ellipsis: Form, meaning and processing. Doctoral dissertation, University of Pennsylvania. Harley, H. 2009. The morphology of nominalizations and the syntax of vP. In Quantification, definiteness and nominalization, ed. by A. Ginnakidou & M. Rathert, 320−342. Oxford: Oxford University Press. Harley, H. 2013. External arguments and the mirror principle: On the distinctness of Voice and v. Lingua 125. 34-57 Hauser, M., L. Mikkelsen, & M. Toosanvardani. 2007. Verb phrase pronominalization in Danish: Deep or surface anaphora? In Proceedings of the Thirty-Fourth Western Conference on Linguistics, ed. by E. Brainbridge & B. Agbayani, 183–195. Fresno, CA: California State University. Hornstein, N. 2016. Crucial experiments and killer data. Blog entry from “Faculty of Language”. [available at: http://facultyoflanguage.blogspot.com/2016/05/crucial-experiments-and-killer-data.html] Johnson, K. 2001. What VP-ellipsis can do, and what it can’t, but not why. In The Handbook of Contemporary Syntactic Theory, ed. by M. Baltin & C. Collins, 439–479. Oxford: Blackwell. Johnson, K. 2004. How to be quiet. In Proceedings from the Fortieth Annual Meeting of the Chicago Linguistics Society, ed. by N. Adams, A. Cooper, F. Parrill & T. Wier, 1−20. Chicago, IL: Chicago Linguistic Society. Kehler, A. 2000. Coherence and the resolution of ellipsis. Linguistics and Philosophy 23. 533–575. Kehler, A. 2002. Coherence, reference, and the theory of grammar. Stanford, CA: CSLI Publications. Kertz, L. 2010. Ellipsis reconsidered. Doctoral dissertation. University of California, San Diego. Kroll, M. 2019. Polarity Reversals under Sluicing. Semantics and Pragmatics 12.1–49. Marušić, F. 2005. On non-simultaneous phases. Doctoral dissertation, Stony Brook University. Matushansky, Ora. 2005. Going through a phase. In Perspectives on phases, ed. by M. McGinnis & N. Richards, 157−181. Cambridge, MA: MITWPL. Megerdoomian, K. 2002. Beyond words and phrases: A unified theory of predicate composition. Doctoral dissertation, University of Southern California. Merchant, J. 2001. The syntax of silence: Sluicing, islands, and the theory of ellipsis. Oxford: Oxford University Press. Merchant, J. 2008. An asymmetry in voice mismatches in VP-ellipsis and pseudogapping. Linguistic Inquiry 39. 169−179. Merchant, J. 2013a. Voice and ellipsis. Linguistic Inquiry 44:77–108. Merchant, J. 2013b. Polarity Items under ellipsis. In Diagnosing syntax, ed. by L. Cheng & N. Corver, 441−462. Oxford: Oxford University Press. Miller, P. 1991. Clitics and constituents in phrase structure grammar. Doctoral dissertation, Rijksuniversiteit te Utrecht. Moriyama, K. 2017. Dake-no koosakuzyo [Argument ellipsis of dake ‘only’]. The Handbook of the 155th Meeting of the Linguistic Society of Japan, 240-245. 426 YOSUKE SATO Nakamura, Taichi. 2013. Semantic identity and deletion. English Linguistics 30. 643–658 Nakamura, Taichi. 2022. Nihongo no kuusyoka to kookootai [Gapping and argument structure alternation in Japanese]. In Kotoba no yoosoo: Genzai to mirai o tunagu [Aspects of language: An interface between present and future], ed. by E. Shima, N. Tomizawa, Y. Ogawa, Y. Dobashi, Y. Sato & D. Lupsa Cornelia, 358−368. Tokyo: Kaitakusha. Pylkkänen, L. 2002. Introducing arguments. Doctoral dissertation, MIT. Rooth, M. 1992. Ellipsis redundancy and reduction redundancy. In Proceedings of the Stuttgart ellipsis workshop, Arbeitspapiere des Sonderforschungsbereichs 340, ed. by S. Berman & A. Hestvik, 1−26. Stuttgart: Universitäten Stuttgart und Tübingen in Kooperation mit der IBM Deutschland. Rudin, D. 2019. Head-based syntactic identity in sluicing. Linguistic Inquiry 50. 253−283. Sag, I. 1976. Deletion and logical form. Doctoral dissertation, MIT. Saito, M. 2007. Notes on East Asian argument ellipsis. Language Research 43. 203–227. Sakamoto, Y. 2022. NEG-raising via proform. Ms., Meiji University. To appear in Linguistic Inquiry. [available at: https://lingbuzz.net/lingbuzz/006175] Sato, Y. 2019.Category mismatch under VP-ellipsis: Implications for the morphosyntax and directionality of Conversion. Ms., Seisen University. Sato, Y. 2022a. When VP-ellipsis and sluicing conspire against syntactic NEG raising. Ms., Tsuda University. To appear in Linguistic Inquiry. Sato, Y. 2022b. Focus mismatch under ellipsis, polarity and head movement in Japanese. Ms., Tsuda University. Stockwell, R. 2022. Small clause predicates and sluicing. Ms., University of Oxford. To appear in Canadian Journal of Linguistics. [available at: https://ling.auf.net/lingbuzz/006869] Sugimoto, K. 2018. An argument structure alternation of change-of-state verbs under VP-deletion. JELS 35.145–151. Takahashi, D. 2006. Apparent parasitic gaps and null arguments in Japanese. Journal of East Asian Linguistics 15. 1–35. Takahashi, D. 2012. Parallelism for elliptic arguments. Paper presented at Formal Approaches to Japanese Linguistics 6, ZAS, Berlin. Takahashi, S.& D. Fox. 2005. Takahashi, Shoichi and Danny Fox. 2005. MaxElide and the rebinding problem. In Proceedings of Semantics and Linguistic Theory 15, ed. by E. Georgala &J. Howell, 223−240. Ithaca, New York: CLC Publications. Takita, K. 2013. Antecedent-contained sluicing and identity in ellipsis. English Linguistics 30. 659−676. Tan, H. 2018. Category mismatch in VP-ellipsis. Ms., National University of Singapore Tanaka, H. 2011a. Voice mismatch and syntactic identity. Linguistic Inquiry 42:470–490. Tanaka, H. 2011b. Syntactic identity and ellipsis. The Linguistic Review 28:79–110. Travis, L., & G. Lamontagne. 1992. The Case filter and licensing of empty K. Canadian Journal of Linguistics 37. 157−174. Yagi, Y., Y. Sakamoto, & Y. Tatsumi. 2022. Against syntactic neg-raising: Evidence From polarityreversed ellipsis in Japanese. In Japanese/Korean Linguistics, volume 29, ed. by K. Horie, K. Akita, Y. Kubota, D. Oshima & A. Usugi, 351−359. Stanford, CA: CSLI Publications. Yoshida, M. 2010. “Antecedent-contained” sluicing. Linguistic Inquiry 41.348−356. Proceedings of CLS 58 (2022), 427-442 © Chicago Linguistic Society 2023. All rights reserved. 427 Segmental and suprasegmental cues in competition: A nonce word experiment* Shu-hao Shih National Taiwan University 1 Introduction Across languages, multiple acoustic measures may correlate with stress in vowels: stressed vowels can be realized with increased fundamental frequency (F0), increased intensity, and increased duration (Gordon & Roettger 2017). Stressed vowels also tend to occupy a more peripheral vowel space than their unstressed counterparts, although there is a contrary effect observed in some languages whereby stressed vowels may be lower in the acoustic space than their unstressed counterparts (e.g. Garellek & White’s 2015 work on Tongan stress). Many languages also display segmental changes that are conditioned by stress or lack of stress. The typical pattern is for segments to strengthen in stressed contexts and to weaken in unstressed positions. In particular, vowels may be affected either qualitatively or quantitatively by the presence or absence of stress. Cross-linguistically, vowels often shorten and qualitatively reduce in unstressed syllables (Crosswhite 2001). For example, most vowels in English reduce to a schwa-like vowel when unstressed. The most extreme manifestation of vowel reduction is deletion, which often targets unstressed vowels. In European Portuguese (hereafter ‘EP’), duration and vowel reduction are reported to be the main cues for word-level stress (Delgado-Martins 1977, 1986, Andrade & Viana 1989, Frota 2000, 2014, Castelo 2005). This paper refines this claim by arguing that vowel reduction outweighs duration in the production of stress, using evidence from a nonce word experiment. Specifically, I will show that the error rates for applying vowel reduction are extremely low, whereas duration is limited to certain vowels and speakers, and sometimes obscured by intonation contours. This study not only contributes to the understanding of cue weighting in EP production but sheds light on the differences between segmental and suprasegmental cues in the production of stress. It also aligns with previous perceptual findings that that vowel reduction is the most robust cue for stress in EP (Castelo 2005, Correia et al. 2013, 2015). This paper is organized as follows: Section 2 provides background on EP wordlevel stress. Section 3 discusses the methodology. Section 4 presents the results from F1 and F2. Section 5 reports the results from duration. Section 6 offers a general discussion. Finally, section 7 concludes this paper. * I wish to thank Paul de Lacy and the audiences at CLS 58 for helpful comments and suggestions. 428 SHU-HAO SHIH 2 Background 2.1 Stress assignment EP has nine oral vowels on the surface (Mateus & d’Andrade 2000). The distribution of the vowels is mainly conditioned by stress. Eight vowels [i, e, ɛ, a, ɐ, o, ɔ, u] occur in stressed syllables, whereas four vowels [i, ɨ, ɐ, u] appear in unstressed syllables (See Section 2.2 on how unstressed vowels relate to stressed ones). Even though [ɐ] appears as a stressed vowel, its appearance in this context is very restricted and predictable (Mateus & d’Andrade 2000). For that reason, Mateus & d’Andrade (2000) do not assume it to be an underlying vowel, but consider it derived. Nasal vowels are not discussed here as they are not relevant to this study. In EP, the location of word stress is unpredictable: it may fall within the last three syllables of the prosodic word, and it is lexically contrastive. However, most EP words are stressed on the penultimate syllable (Cruz-Ferreira 1995). Data from the FrePOP database (Frota et al. 2010), containing 1,486,092 prosodic words from spoken and written language, shows that 74% of words with two syllables or more are stressed on the penultimate syllable, 23% of words are stressed on the final syllable, and 2% of words have stress on the antepenultimate syllable. Portuguese verbs can be placed into three classes (referred to as conjugations) according to the theme vowels in their infinitive ending: -ar (first conjugation), -er (second conjugation), and -ir (third conjugation). Each of these conjugations defines an inflectional paradigm. The verb is formed from a root morpheme, to which a theme vowel is added to form a stem. A tense/mood/aspect morpheme (which may be null) follows the theme vowel, and a person/number morpheme (which may be null as well) is added finally. When a verb is in its uninflected infinitive form, stress always falls on the last syllable of the word: e.g. [fɐ.ˈlaɾ] ‘to speak’. In the present tense, stress always falls on the penultimate syllable of the verb form: e.g. [ˈfa.lɐ] ‘(he) speaks’. As we will see below, verbs with penultimate stress (present tense) and final stress (infinitive form) are essential in the design of the nonce word experiment. 2.2 Vowel reduction In EP, vowels undergo reduction in unstressed positions. Most recent studies report the reduction pattern shown below (Mateus & d’Andrade 2000, Vigário 2003, Castelo 2005, Correia et al. 2015). (1) Vowel reduction in EP i i u e o ɛ ɐ ɔ a That is, the mid front vowels [e, ɛ] are neutralized to [ɨ], the mid back vowels [o, ɔ] are reduced to [u], the low vowel [a] is raised to [ɐ], and the high vowels [i, u] Segmental and suprasegmental cues in competition 429 remain unreduced in unstressed positions. Unlike Brazilian Portuguese, the reduction pattern in EP is identical in pre- and post-tonic positions (Mateus & d’Andrade 2000). 2.3 Phonetic realization of stress In EP, duration is reported to be the main acoustic cue for word stress. That is, stressed syllables are longer than unstressed syllables (d’Andrade & Viana 1989). However, pitch has not been considered in the EP stress literature as a potential correlate of word stress (Correia et al. 2015: 52). In EP, pitch accents signal phrase level prominence and are used to differentiate sentence types and pragmatic or discursive meanings (Frota 2000, 2014). Due to the low co-variation between word stress and pitch accent, pitch variation seems not to be a robust cue for word stress in EP (Correia et al. 2015: 53). Note that Silva (1997: 83) asserts that stress is clearly marked by an increase in pitch and volume but no instrumental data is provided to support the claim. Finally, it is not clear whether F1 and F2 play any roles in EP stress even though unstressed vowels undergo reduction (see Section 2.2). Finally, it is not clear whether F1 and F2 play any roles in EP stress even though unstressed vowels undergo reduction. 3 Methodology 3.1 Experiment design Three frame sentences were created to manipulate stress alternation. For example, for /ʒɔɡaɾ/ ‘to play’: (2) Three frame sentences a. O David e the David and b. A Sílvia também the Silvia also c. A Júlia vai the Julia is going to a the joga play jogar play Maria jogam futebol. Maria play football futebol. football futebol. football [ˈʒɔ.ɡɐ̃w̃] [ˈʒɔ.ɡɐ] [ʒu.ˈɡaɾ] While three sentences were elicited, only two were measured. For the first sentence, the target word was expected to be produced with focus effects because it is new information. Focus can bring with it special pitch accents and prosodic phrasing. So, to avoid the phonological influences of focus, only the nonce words in the second and third frame sentences were analyzed. The sentences differed in terms of stress-influencing suffixes so that the same verb root could be observed with both stressed and unstressed vowels. Note that the verb in the first two sentences has stress on the penultimate syllable, whereas the uninflected infinitive verb in the third sentence has stress on the final syllable (see Section 2.1). Vowel reduction of /ɔ/→[u] occurred in the unstressed syllable. A set of nonce words was constructed in consultation with one native speaker 430 SHU-HAO SHIH of EP. Disyllabic nonce words with the shape /CiCa/, /CuCa/, /CɛCa/, /CɔCa/, and /CaCa/ were used and treated as a verb in the experiment. The vowel in the second syllable was restricted to /a/ so that all the stimuli belonged to the first conjugation (see Section 2.1). So, stress alternated on the penultimate and final syllable: [CV́.CV]~[CV.CV́ɾ]. Each pair allowed direct comparison of vowels in both stressed and unstressed states. Note that words with the shape /CɛCa/, /CɔCa/, and /CaCa/ were excluded in the analyses of duration because the stressed vowels and their unstressed counterparts were distinct in their quality, as shown in Table 1. /CVCa/ [Cí.Cɐ] [Cú.Cɐ] [Cɛ́.Cɐ] [Cɔ́.Cɐ] [Cá.Cɐ] /CVCaɾ/ [Ci.Cáɾ] [Cu.Cáɾ] [Cɨ.Cáɾ] [Cu.Cáɾ] [Cɐ.Cáɾ] Table 1: Vowel reduction patterns. The initial consonants were limited to [p, t, k, b, d], while the second consonants were [p, t, k, s, f]. All the consonants were unaspirated in order to reduce influence on the following vowel’s duration (van Santen 1992). The second consonants were voiceless in order to keep influence on the preceding vowel’s duration relatively constant (Peterson & Lehiste 1960). This stimulus structure not only facilitated identification of vowel boundaries but also minimized segmental effects on vowels (e.g. vowel lengthening before voiced consonants). All the stimuli were created in accordance with the EP writing system. In particular, diacritics were not used on vowels, because participants might interpret them as marking stress. In the pilot study, the letters é and ó were adopted, with the hope that without explicit instructions, participants could produce [ɛ] and [ɔ] instead of [e] and [o] when these two vowels were stressed. However, the participants reported that the acute accent somehow forced them to maintain the same vowel quality even when these vowels were unstressed (i.e. in the third frame sentence). Thus, the participants were informed that the letters e and o always corresponded to [ɛ] and [ɔ] when stressed. There were six stimuli for /CiCa/, /CuCa/, /CɛCa/, /CɔCa/, and /CaCa/ words. With two stress-alternating forms for each word type, there were 60 stimuli in the experiment. The final word list was further verified by two native speakers of EP to ensure that none of the words were native words. There were three recording sessions for each experiment. Ten common native verbs were employed as fillers to encourage the participants to speak in their vernacular speech style. Each verb had two stress-alternating forms (e.g. [tápɐ] ‘(he) covers’ and [tɐpáɾ] ‘to cover’). These forms were placed in four colloquial Segmental and suprasegmental cues in competition 431 filler sentences. Fillers were interspersed among the stimuli. Three fillers were introduced at the beginning of each session to take into account the effects of any initial nervousness the participant might have about the task. The order of the stimuli was pseudo-randomized and counter-balanced in each session. In sum, each participant produced 180 tokens for the experiment (5 vowels × 6 stimuli × 2 frame sentences × 3 repetitions). 3.2 Participants Six male and two female native EP speakers participated in the experiment. Their ages ranged from 29 to 44 years old. Four participants (three males and one female) had lived in the United States for four years on average, and four participants (three males and one female) had lived in Taiwan for four years on average. All the participants still communicated in EP with their family and friends on a daily basis. Three male participants were from Lisbon, two female participants were from Coimbra, and three male participants were from Porto. The participants were naïve as to the goal of the experiment, and none had any linguistic training or history of speech impairments. They received nominal monetary compensation for their participation. 3.3 Procedure Three male and one female participants were recorded in a recording studio at the University of California, Santa Barbara in the summer of 2019, and the remaining participants were recorded in a quiet room at National Taiwan University in early 2021. The same recording devices were used in different locations to keep the recording condition constant. The participants were recorded while wearing a Shure SM35 head-worn microphone with behind-the-neck headband in order to keep the microphone at a constant distance from the mouth (thereby limiting inadvertent intensity variation). The participants were recorded using a Zoom H4n Pro at a 44.1k Hz sampling rate and 16-bit quantizing rate in mono. All the experiments were carried out by the author. Before the recording sessions began, a training session was provided to familiarize the participants with the task. Initially, the participants were provided with three native verbs and asked to produce the frame sentences. Next, the participants were presented with three nonce words and asked to repeat the same task. The three nonce words had /i/, /e/, and /o/, respectively. All the participants performed the expected conjugation for each verb and reported that they had no issues with the task. Participants were presented with the first frame sentence (with the stimuli) on a computer screen. Participants had to generate the second and third frame sentences from memory during the experiment. The purpose was to ensure that the participants applied the vowel reduction pattern without any influences from the Latin letters. Participants received sufficient training in the training session and were able to generate the second and third frame sentences with no significant hesitation. Participants had to perform a proper verb conjugation based on the 432 SHU-HAO SHIH stimuli presented in the first frame sentence: e.g. [ˈbɔ.pɐ] and [bu.ˈpaɾ] for the nonce word bopam. The recording sessions were conducted individually. Participants read the sentences when they were ready, at a normal conversational speed. Breaks were given after each recording session. 3.4 Measurements Acoustic correlates of stressed and unstressed vowels were measured, including duration, F1, and F2. Note F0 was also measured in order to check whether the participants had distinct F0 patterns for the target words. Vowels were labeled using Praat TextGrids (Boersma & Weenink 2020). Two intervals were labeled for each file: the extent of the first vowel of the target word in the second frame sentence, and the extent of the first vowel of the target word in the third frame sentence. The left boundary of each vowel was marked at the zero crossing of the first non-deformed periodic waveform. The right boundary was identified as the end of the second formant, with the help of the third formant when the end of the second formant continued into closure (Turk et al. 2006). The segmentation was performed by the author and two well-trained undergraduate students. All audio files and TextGrids were further double-checked by the author and the undergraduate students after the segmentation was finished with the goal of minimizing human error. Finally, the author examined the TextGrids and made corrections only when (a) wrong vowels were labeled, (b) consonants were mislabeled as part of a vowel, and (c) the right boundary of vowels was noted by the undergraduate assistants as uncertain; otherwise no changes were made. The labeled sound files were then run through customized Praat scripts to obtain acoustic measures. Duration was extracted from the TextGrids. For F0, F1, and F2, the midpoint of each vowel was measured. The purpose was to identify the steady point of the vowel. The results were saved to a Microsoft Excel (.xlsx) file for subsequent analysis, though they were analyzed in R (R Development Core Team 2020). 3.5 Statistical analysis In the following sections, I determine whether duration, F1, and F2 were significant correlates of stress. The values of each measure were analyzed using linear mixedeffects models (following Garellek & White 2015 for Tongan stress, and Shih 2018 for Gujarati stress). The acoustic measures (duration, F1, F2) were the dependent variable. These were implemented in R using the lmer() function of the lme4 package (Bates et al. 2015). For the duration, F1, and F2 models, VOWEL (e.g. [í] in [Cí.Cɐ] vs. [i] in [Ci.Cáɾ]), ORDER (first repetition, second repetition, third repetition), POAC1 (the place of articulation of the first consonant: labial, coronal, dorsal), and POAC2 (the place of articulation of the second consonant: labial, coronal, dorsal) were set as fixed effects. SPEAKER and WORD were included as random effects for all the models. Interaction between the fixed effects was tested in the model by using the anova() function to compare likelihood between models (Baayen 2008). Random Segmental and suprasegmental cues in competition 433 slopes for the by-speaker and by-word effects of the fixed effects were specified for each model (Barr et al. 2013), and likelihood ratio tests were run to evaluate the models. The basis for removing factors was set at a p-value of the likelihood ratio test of p < .05. When the random slope models failed to converge, the next-best models were chosen by using the likelihood-ratio test mentioned above. Visual inspection of residual plots did not reveal any obvious deviations from homoscedasticity or normality. Crucially, I report multiple pairwise comparisons for the target vowels, obtained using the pairwise() function of the emmeans package (Lenth 2020). These estimates were based on the Tukey Honestly Significant Difference (Tukey HSD) method. I report t-values as well as p-values provided in the model output. 4 Vowel reduction This section explores whether the participants apply vowel reduction to distinguish stress. Recall that the first vowel in [CV́.CV] words should be stressed while the first vowel in [CV.CV́ɾ] should be unstressed. So, /ɛ, ɔ, a/ in [CV.CV́ɾ] words should undergo vowel reduction. Results from the linear mixed-effects models show that unstressed non-high vowels undergo reduction: /ɛ/ → [ɨ], /ɔ/ → [u], and /a/ → [ɐ], whereas unstressed high vowels maintain the same quality: /i/ → [i] and /u/ → [u]. Figure 1 shows the vowel plot for stressed vowels and their unstressed counterparts. Note that [ŭ] and [ū] are used in the plot to distinguish unstressed /u/ and /ɔ/, respectively. Figure 1: Vowel plot for the first vowel in [CV́.CV] (solid circles) and [CV.CV́ɾ] (dashed circles) words. The ellipsis delineates one standard deviation from the mean value. The F1 model for [CV́.CV] and [CV.CV́ɾ] words is summarized in (3). For stressed vowels, a three-way height distinction is established. That is, [ɛ́] is significantly lower than [í] and higher than [á] (t=19.35, p<.001 for [ɛ́] vs. [í]; t=−13.37, p<.001 for [ɛ́] vs. [á]), and [ɔ́] is significantly lower than [ú] and higher than [á] (t=14.42, p<.001 for [ɔ́] vs. [ú]; t=−8.84, p<.001 for [ɔ́] vs. [á]). 434 SHU-HAO SHIH (3) F1~VOWEL+(1|Word)+(1+VOWEL|Speaker) Vowel /i/ /u/ /ɛ/ /ɔ/ /a/ Stress 295.28 (36.45) 339.89 (33.59) 528.57 (44.48) 551.49 (51.24) 735.94 (70.26) No stress 289.19 (31.45) 325.12 (38.65) 345.98 (39.5) 343.51 (44.21) 484.45 (39.19) Table 2: Mean F1 (in Hz; standard deviations in parentheses) for vowels with stress and no stress in [CV́.CV] and [CV.CV́ɾ] words. The unstressed /a/ is raised to a mid central position, which has the same height as [ɛ́] and is significantly higher than [ɔ́] (t=−3.05, p=.154 for [ɐ] vs. [ɛ́]; t=−5.78, p<.001 for [ɐ] vs. [ɔ́]). Note that both [ɛ́] and [ɔ́] have higher F1 values than the unstressed /a/. The unstressed /ɛ/ is raised to a high central position, which is significantly higher than [ɐ] (t=−10.26, p<.001). In addition, it has the same height as [ú] and is significantly lower than [í] (t=1.18, p=.966 for [ɨ] vs. [ú]; t=4.66, p=.006 for [ɨ] vs. [í]). The unstressed /ɔ/ is raised to a high back position, which overlaps with [ú] (t=0.39, p=.999). This suggests that the unstressed /ɔ/ is realized as [u]. Finally, unstressed high vowels have the same height as their stressed counterparts (t=−0.55, p=.999 for [i] vs. [í]; t=−1.53, p=.866 for [u] vs. [ú]). For F2, it was found that a three-way interaction between VOWEL, POAC1, and POAC2 significantly improved the model fit, as summarized in (4). The interaction term suggests that the influence of the flanking consonants varies across vowels. Except for [u] (reduced from /ɔ/), the statistical distinctions between vowels persist regardless of the influence of the flanking consonants. (4) F2~VOWEL*POAC1*POAC2+(1|Word)+(1+VOWEL+POAC1|Speaker) For [ɐ], it is significantly more back than [ɛ́] (t=−7.47, p<.001 for POAC1=coronal, POAC2=coronal; t=−5.84, P<.001 for POAC1=coronal, POAC2=dorsal; t=−9.34, p<.001 for POAC1=coronal, POAC2=labial) and more fronted than [ɔ́] (t=15.21, p<.001 for POAC1=coronal, POAC2=coronal; t=13.82, P<.001 for POAC1=labial, POAC2=dorsal; t=15.07, p<.001 for POAC1=coronal, POAC2=labial; t=11.06, p<.001 for POAC1=labial, POAC2=labial). For [ɨ], it is significantly more back than [í] (t=−13.04, p<.001 for POAC1=labial, POAC2=coronal; t=−12.51, P<.001 for POAC1=coronal, POAC2=dorsal; t=−13.64, p<.001 for POAC1=coronal, POAC2=labial) and more fronted than [ú] (t=16.01, p<.001 for POAC1=coronal, POAC2=coronal; t=19.46, P<.001 for POAC1=labial, POAC2=coronal; t=28.47, p<.001 for POAC1=coronal, POAC2=dorsal). Segmental and suprasegmental cues in competition 435 For [u] (reduced from /ɔ/), it has the same backness as [ú] in two out of four conditions (t=1.14, p=.97 for POAC1=labial, POAC2=dorsal; t=0.68, p=.999 for POAC1=labial, POAC2=labial) but is significantly more fronted than [ú] in another two conditions (t=4.91, p=.004 for POAC1=coronal, POAC2=coronal; t=5.1, p=.003 for POAC1=labial, POAC2= coronal). Nonetheless, the more fronted [u] is still close to [ú], given their intersecting standard deviations in each condition (in the context where POAC1=coronal and POAC2=coronal: [u]: 1180.15 Hz (SD=159.18), [ú]: 1071.46 Hz (SD=151.25); in the context where POAC1=labial and POAC2=coronal: [u]: 984.59 Hz (SD=168.66), [ú]: 880.87 Hz (SD=105.47). Moreover, [u] is significantly further back than [ɨ] (t=−11.49, p<.001 for POAC1=coronal, POAC2=coronal; t=−15.22, p<.001 for POAC1=labial, POAC2=coronal; t=−17.37, p<.001 for POAC1=coronal, POAC2=labial). Thus, the unstressed /ɔ/ has the same backness as [ú]. For [i] and [u] (reduced /u/), they consistently have the same backness as their stressed counterparts in different conditions (for [i] vs. [í]: t=−1.26, p=.947 for POAC1=labial, POAC2=coronal; t=−0.44, p=.999 for POAC1=coronal, POAC2=dorsal; t=−0.15, p=.999 for POAC1=coronal, POAC2=labial; t=−0.97, p=.99 for POAC1=labial, POAC2=labial; for [u] vs. [ú]: t=2.39, p=.392 for POAC1=coronal, POAC2=coronal; t=1.12, p=.974 for POAC1=labial, POAC2=coronal; t=1.13, p=.971 for POAC1=coronal, POAC2=dorsal; t=0.472, p=.999 for POAC1=labial, POAC2=dorsal; t=−0.27, p=.999 for POAC1=dorsal, POAC2=labial; t=−0.39, p=.999 for POAC1=labial, POAC2=labial). One goal of this experiment is to establish whether vowel reduction is productive in EP. If it is not, then vowel reduction could well be lexicalized, putting into question its phonological and psychological reality. For high vowels, all the participants did not change the vowel quality in unstressed position except for participants M3 and M4 who had some [u]-deletions. For non-high vowels, participants F2, M3, M4, and M5 had the expected vowel reduction pattern with no errors. Similarly, participants F1, M1, M2, and M6 had the expected vowel reduction pattern but they produced some reduction errors. All errors involved using vowels which are supposed to appear in stressed syllables. That is, participants used the stressed vowel in [CV́.CV] words for the unstressed vowel in [CV.CV́ɾ] words. For example, participant F1 had [pɔ́.pɐ] for the nonce word popa in the second frame sentence but *[pɔ.páɾ] for the nonce word popar in the third frame sentence. Participant F1 made three errors out of 18 stimuli containing /ɛ/ (roughly a 17% error rate). Participant M1 made two errors out of 18 stimuli containing /ɔ/ (roughly an 11% error rate). Participant M2 made four errors out of 18 stimuli containing /ɔ/ (roughly a 22% error rate) and four errors out of 18 stimuli containing /a/ (roughly a 22% error rate). Participant M6 made three errors out of 18 stimuli containing /ɛ/ (roughly an 18% error rate). The total error rates for participants F1, M1, M2, and M6 are approximately 3%, 2%, 9%, and 3%, respectively. No errors were found for participants F2, M3, M4, and M5. To probe the cause of the errors, a post-experiment interview was conducted for each participant who produced errors. However, those participants could not offer a 436 SHU-HAO SHIH systematic explanation about their decisions; they either had no clue about the errors or expressed that some errors sounded like native words to them. Some participants were found to delete [ɨ] and [u] (reduced from /u/ and /ɔ/) in [CV.CV́ɾ] words. Specifically, participants F1 and M6 had ten and six [ɨ]-deletions, respectively. Participant F2 had one [u]-deletion (reduced from /ɔ/). Participant M1 had 13 [ɨ]-deletions and one [u]-deletion (reduced from /ɔ/). Participant M3 had 14 [ɨ]-deletions and eight [u]-deletions (two from reduced /u/ and six from reduced /ɔ/). Participant M4 had 18 [ɨ]-deletions and 13 [u]-deletions (seven from reduced /u/ and six from reduced /ɔ/). Finally, participants M2 and M5 had no vowel deletion at all. Among the two vowels that undergo deletion, [ɨ] is more likely to delete than [u] is. The deletion rates are 6% (9 out of 144 stimuli) for [u] (reduced from /u/), and 10% (14 out of 144 stimuli) for [u] (reduced from /ɔ/). The deletion rate for [ɨ] is approximately 42% (61 out of 144 stimuli), which is higher than the deletion rates for [u]. Except for participant F2, participants either had [ɨ]-deletion only or had more [ɨ]-deletions than [u]-deletions. Finally, two unstressed vowels, [i] and [ɐ], were never found to undergo deletion. Based on the deletion rates, the vowels are ordered as follows: [ɨ] > [u] > [i], [ɐ]. The deletion pattern found in this study is consistent with the EP literature (Mateus & d’Andrade 2000, Vigário 2003, Freitas 2004). In summary, three generalizations emerge from the F1/F2 results of the [CV́.CV] and [CV.CV́ɾ] words. First, all the participants have phonological vowel reduction, and some have variable deletion for some vowels. Second, the error rates are extremely low for each participant who produced errors. Third, the deletion pattern is in line with previous studies on EP vowel reduction. 5 Duration Four distinct intonation groups were established based on the F0 comparisons, as summarized in Table 3. As F0 might have an influence on the duration of the target vowels, the analyses of vowel duration are based on the four intonation groups. See Shih (manuscript) for the discussion of different intonation groups. Group 1 (1 female, 3 males) Group 2 (1 female, 1 male) Group 3 (1 male) Group 4 (1 male) [CV́.CV] Low level Falling Low level Falling [CV.CV́ɾ] Falling Falling Low level Rising Table 3: Four intonation groups. 5.1 Group 1 Recall that Group 1 has a level contour on words with penultimate stress and a falling contour on words with final stress. To determine the influence of F0 on vowel duration, the [ɐ]s in [Cá.Cɐ] and [Cɐ.Cáɾ] are crucial because they are Segmental and suprasegmental cues in competition 437 unstressed and have identical quality. However, the [ɐ] with a F0 trough (57.42 ms, SE=1.8) is found to be shorter than the [ɐ] with a F0 peak (61.57 ms, SE=1.71), the opposite of what is expected. An unpaired t-test analysis shows that the comparison between the [ɐ] with a F0 trough and the [ɐ] with a F0 peak is not significant: t(141)=1.67, p=.097. So, I assume the influence of F0 on duration, if any, is minimal for Group 1. The duration model for [CV́.CV] and [CV.CV́ɾ] words is summarized in (5). It is found that stressed high vowels are not significantly longer than their unstressed counterparts, as shown in Table 4. (5) Duration~VOWEL+ORDER+POAC1+(1|Word)+(1+VOWEL|Speaker) Vowel Stress /i/ 87.42 (2.45) /u/ 78 (2.1) No stress 70.32 (1.9) 60.40 (2.31) t-value p-value 2.45 .105 2.61 .081 Table 4: Mean duration (in ms; standard errors in parentheses) for vowels with stress and no stress in [CV́.CV] and [CV.CV́ɾ] words. For the order effect, vowels in the first repetition have the longest duration (78.05 ms (SE=2.4)), followed by vowels in the second repetition (72.71 ms (SE=2.01)) and vowels in the third repetition (71.83 ms (SE=2.02)). The durational differences between the first and second repetitions and between the first and third repetitions are found to be significant (t=3.32, p=.002 for the first repetition vs. the second repetition; t=4.11, p<.001 for the first repetition vs. the third repetition). In addition, vowels preceded by a labial have the longest duration (79.38 ms (SE=1.64)), compared with vowels preceded by a coronal (70.24 ms (SE=1.96)) and a dorsal (63.32 ms (SE=3.92)). Only the comparison between vowels preceded by a labial and vowels preceded by a coronal is found to be marginally significant (t=−2.49, p=.048). 5.2 Group 2 Recall that Group 2 has a falling contour on all the target words. The F0 peak consistently falls on the initial syllable, so the influence of F0 on duration is potentially minimized. (6) Duration~VOWEL+POAC1+(1|Word)+(1|Speaker) Vowel Stress /i/ 94.48 (3.86) /u/ 94.64 (4.9) No stress 70.35 (2.86) 59.61 (3.27) t-value p-value 3.98 <.001 6.37 <.001 Table 5: Mean duration (in ms; standard errors in parentheses) for vowels with stress and no stress in [CV́.CV] and [CV.CV́ɾ] words. 438 SHU-HAO SHIH The duration model for [CV́.CV] and [CV.CV́ɾ] words is summarized in (6). First, stressed vowels are significantly longer than their unstressed counterparts, as shown in Table 5. Moreover, vowels preceded by a labial have the longest duration (87.22 ms (SE=3.39)), followed by vowels preceded by a dorsal (74.5 ms (SE=9.16)) and vowels preceded by a coronal (70.97 ms (SE=2.72)). In particular, vowels preceded by a labial are significantly longer than vowels preceded by a coronal (t=−3.56, p=.004). 5.3 Group 3 Recall that Group 3 has a level contour on all the target words, so the influence of F0 on duration is potentially minimized. The duration model for [CV́.CV] and [CV.CV́ɾ] words is summarized in (7). As shown in Table 6, there is no significant durational difference between [í] and [i]. However, [ú] is significantly longer than [u]. Note that the male participant had seven [u]-deletions out of 18 stimuli. This suggests that [u] is already acoustically very weak so it is prone to undergo deletion. (7) Duration~VOWEL+ORDER+POAC1+(1|Word) Vowel Stress /i/ 59.78 (3.44) /u/ 51.68 (3.2) No stress 45.64 (2.53) 34.02 (4.49) t-value p-value 2.52 .076 4.24 <.001 Table 6: Mean duration (in ms; standard errors in parentheses) for vowels with stress and no stress in [CV́.CV] and [CV.CV́ɾ] words. As for the order effect, vowels in the first repetition have the longest duration (54.42 ms (SE=4.3)), followed by vowels in the second repetition (47.64 ms (SE=3.2)) and vowels in the third repetition (43.49 ms (SE=2.25)). The durational difference between the first and second repetitions reaches statistical significance (t=3.54, p=.003), so as the difference between the first and third repetitions (t=4.51, p<.001). In addition, vowels preceded by a labial have the longest duration (52.87 ms (SE=2.78)), compared with vowels preceded by a coronal (41.72 ms (SE=2.56)) and a dorsal (41.15 ms (SE=1.8)). However, no statistical distinction is found for each pair, with all p>.071. 5.4 Group 4 Recall that Group 4 has a falling contour on words with penultimate stress and a rising contour on words with final stress. It is impossible to adopt the analysis in Group 1 to determine the influence of F0 on vowel duration because the unstressed vowels in [Cá.Cɐ] and [Cɐ.Cáɾ] words all have a F0 trough. So, the results reported here might be influenced by F0. The duration model for [CV́.CV] and [CV.CV́ɾ] words is summarized in (8). It is found that stressed vowels are not significantly longer than their unstressed Segmental and suprasegmental cues in competition 439 counterparts, as shown in Table 7. In addition, vowels preceded by a labial have the longest duration (81.09 ms (SE=2.84)), compared with vowels preceded by a coronal (69.01 ms (SE=2.8)) and a dorsal (56.11 ms (SE=6.25)). However, no statistical distinction is found for each pair, with all p>.122. In short, it is not clear whether duration is a cue for stress as the F0 trough might add extra duration to the unstressed vowels in [CV.CV́ɾ]. (8) Duration~VOWEL+POAC1+(1|Word) Vowel Stress /i/ 77.99 (3.69) /u/ 72.03 (3.96) No stress 79.19 (3.14) 66.68 (5.36) t-value p-value −0.14 .999 1.37 .523 Table 7: Mean duration (in ms; standard errors in parentheses) for vowels with stress and no stress in [CV́.CV] and [CV.CV́ɾ] words. In summary, of the four intonation groups, only Group 2 clearly uses duration as a cue for stress. 6 Discussion The results suggest that EP speakers weight vowel reduction more than duration in the production of stress. The hierarchy is consistent with previous studies on stress perception. Correia et al. (2013, 2015) investigate the perception of word stress in EP, both in nuclear and post-nuclear positions, by means of three experiments. Their results show that when vowel reduction is absent from the stimuli, EP speakers show a stress “deafness” effect. That is, EP speakers show significantly higher error rates in the stress contrast condition than in the phoneme contrast condition when either duration alone, or both duration and pitch accents are present in the stimuli. The result is consistent in both the ABX task (i.e. experiment 1) and sequence recall task (i.e. experiment 2). Crucially, when vowel reduction is added in the stimuli, EP speakers are able to perceive stress contrasts. Moreover, the results suggest that segmental cues (e.g. vowel reduction) are more reliable than suprasegmental cues (e.g. duration). As a consequence, stress languages without vowel reduction might show a mismatch between the acoustic cues used for stress production and perception. For example, Gonzalez (1970) reports that Tagalog uses F0, duration, and intensity to signal lexical stress. However, Hwang et al. (2019) argue that intensity does not play a crucial role in stress perception. Beyond EP, there are cross-linguistic studies showing that vowel reduction (or vowel quality) serves as a dominant cue for stress. Chrabaszcz et al. (2014) investigate how listeners’ native language affects their weighting of stress cues (such as vowel quality, pitch, duration, and intensity) in the perception of contrastive word stress. The results show that vowel quality is the strongest cue for all English, Russian, and Mandarin listeners. Chrabaszcz et al. further argue that 440 SHU-HAO SHIH vowel quality is more independent from the influence of other incidental speech characteristics, whereas acoustic cues are generally free to vary, and their utility for determining word-level stress is likely constrained to relative change across the word given a particular speech context. For instance, a short vowel is short only in the context of a longer duration while a schwa in many languages indicates an unstressed syllable regardless of its environment. The F1 and F2 results from this study provide further support to the above discussion: stressed and unstressed vowels remain distinct in their quality regardless of the influences of various intonation contours and surrounding consonants. 7 Conclusion This paper has shown that vowel reduction outweighs duration in the production of stress in EP. The results indicate that vowel reduction is a productive phonological process to distinguish stress at the segmental level. On the other hand, stressed vowels can have greater duration relative to their unstressed counterparts. However, participants differ as to whether they use duration to signal stress, and for which vowels. In addition to its implications for our understanding of stress production, this study provides a foundation for future work on languages which realize stress at both segmental and suprasegmental levels. References Boersma, P., & D. Weenink. 2020. Praat: doing phonetics by computer (version 6.2.04). http://www.praat.org/. Castelo, A. 2005. The Perception of word primary stress by European Portuguese speakers. In Prosodies: with special reference to Iberian languages, ed. by S. Frota, M. Vigário, & M. J. Freitas, 159–176. Berlin, New York: Mouton de Gruyter. Chrabaszcz, A., M. Winn, C. Y. Lin, & W. J. Idsardi. 2014. Acoustic cues to perception of word stress by English, Mandarin, and Russian speakers. Journal of Speech Language and Hearing Research 57. 1468–1479. Correia, S., J. Butler, M. Vigário, & S. Frota. 2015. A stress “deafness” effect in European Portuguese. Language and speech 58. 48–67. Correia, S., S. Frota, J. Butler, & M. Vigário. 2013. Word stress perception in European Portuguese. Proceedings of Interspeech 2013. 267–271. Crosswhite, K. 2001. Vowel reduction in Optimality Theory. New York: Routledge. Cruz-Ferreira, M. 1995. European Portuguese. Journal of the International Phonetic Association 25. 90–94. d’Andrade, E., & M. C. Viana. 1989. Ainda sobre o ritmo e o acento em português. Actas do IV Encontro Nacional da Associação Portuguesa de Linguística. Lisboa, Portugal: Edições Colibri. 3–15. Delgado-Martins, M. R. 1977. Aspects de l’accent en portugais. Voyelles toniques et atones. PhD Dissertation, University of Strasbourg. Delgado-Martins, M. R. 1986. Sept études sur la perception: Accent et intonation du portugais. Lisboa, Portugal: Instituto Nacional de Investigacao Cientifica. Freitas, M. J. 2004. The vowel [ɨ] in the acquisition of European Portuguese. In Proc. of GALA Segmental and suprasegmental cues in competition 441 2003, ed. by J. van Kampen, & S. Baauw, 163–174. Utrecht: LOT. Frota, S. 2000. Prosody and focus in European Portuguese: phonological phrasing and intonation. New York, NY: Garland Publishing. Frota, S. 2014. The intonational phonology of European Portuguese. In Prosodic typology II, ed. by Jun S.-A, 6–42. Oxford, UK: Oxford University Press. Frota, S., M.Vigário, F. Martins, & M. Cruz. 2010. FrePOP (version 1.0). Laboratório de Fonética (CLUL). Faculdade de Letras da universidade de Lisboa. Retrieved from http://frepop.letras.ulisboa.pt. Garellek, M., & J. White. 2015. Phonetics of Tongan stress. Journal of the International Phonetic Association 45. 13–34. Gonzalez, A. 1970. Acoustic correlates of accent, rhythm, and intonation in Tagalog. Phonetica 22. 11–44. Gordon, M., & T. Roettger. 2017. Acoustic correlates of word stress: a cross-linguistic survey. Linguistic Vanguard 3. Hwang, H.- K., N. Nagaya, & J. Villegas. 2019. Cue weighting in the perception of Tagalog stress. Poster presented at the 178th Meeting of the Acoustical Society of America. Lenth, R. V. 2020. emmeans: Estimated Marginal Means, aka Least-Squares Means. R package version 1.6.2-1. https://CRAN.R-project.org/package=emmeans. Mateus, M. H. & E. d’Andrade. 2000. The phonology of Portuguese. New York: Oxford University Press. Peterson, G. E., & I. Lehiste. 1960. Duration of syllable nuclei in English. Journal of the Acoustical Society of America 32. 693–703. R Core Team. 2020. R: a language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. https://www.R-project.org/. Shih, S.-H. 2018. On the existence of sonority-driven stress in Gujarati. Phonology 35. 327–364. Shih, S.-H. MS. Phonetics of European Portuguese stress: a nonce word experiment. Silva, D. 1997. The variable deletion of unstressed vowels in Faialense Portuguese. Language Variation and Change 9. 295–308. van Santen, J. 1992. Contextual effects on vowel duration. Speech Communication 11. 513–546. Vigário, M. 2003. The prosodic word in European Portuguese. Berlin: Mouton de Gruyter. Proceedings of CLS 58 (2022), 443-454 © Chicago Linguistic Society 2023. All rights reserved. 443 NPI licensing and intrusion effects in Japanese Nianpo Su and Helena Aparicio Cornell University 1 Introduction Negative polarity items (NPIs) (e.g., ever, any, yet in English) are expressions whose distribution is restricted to particular licensing environments, i.e., they must occur within the scope (e.g., c-command domain)1 of a downward-entailing licensor (Ladusaw 1979). (1) a. No student [that the professor liked] will ever pass the exam. b. *The student [that the professor liked] will ever pass the exam. c. *The student [that no professor liked] will ever pass the exam. In (1a), the NPI ever is grammatically licensed by the determiner no, as it is c-commanded by the negative phrase. In contrast, NPIs fail to be licensed when a licensor is not present (1b), or when the NPI is outside the licensor’s c-commanding domain. Such is the case of (1c), where the licensor in the relative clause fails to c-command the matrix NPI, resulting in ungrammaticality. However, it has been observed that sentences containing improper licensors (1c) are judged to be more acceptable than sentences with no licensor (1b). In the literature, this illusory licensing effect is referred to as intrusive NPI licensing (Drenhaus et al. 2005). Previous studies have found robust NPI intrusion effects in languages such as English and German (Xiang et al. 2009, Drenhaus et al. 2005, a.o.), where the NPIlicensor dependency is retrosrpective, i.e., NPIs appear after their licensors, as in (1a). In recent years, there has been an increasing interest in the study of intrusive NPI licensing in languages such as Korean, where the NPI-licensor dependency is prospective, i.e., the NPI precedes its licensor, as shown in (2). (2) Keni-ka amwuto mannaci-anh-ass-ta. Ken-NOM anyone meet-NEG-PST-DEC ‘Ken didn’t meet anyone.’ While intrusion effects involving prospective dependencies have been reported in Turkish and Korean (Yanilmaz & Drury 2018, Yun et al. 2018, Lee & Yun 2022), the status of intrusive NPI licensing in languages with prospective NPI-licensor dependencies remains poorly understood. Here we contribute to the empirical landscape of the phenomenon by investigating the processing profile of intrusive NPI licensing in Japanese, another language with prospective NPI licensing. In two experiments (an offline acceptability judgment task and a self-paced reading study), 1 Formally, node A c-commands node B if every node dominating A also dominates B, and neither A nor B dominates the other. 444 SU AND APARICIO we show that while Japanese gives rise to intrusion effects, these are only detectable in offline dependent measures. This paper is organized as follows. Section 2 introduces two distinct features of Japanese NPI licensing as well as previous findings on intrusion effects both in languages with prospective and retrospective NPI dependencies. Section 3 presents the results of an acceptability judgment task (Experiment 1) that examines whether Japanese exhibits NPI intrusion effects. Section 4 presents results from a self-paced reading experiment (Experiment 2) that investigates the online blueprint of the intrusion effects detected in Experiment 1. Section 5 contains the discussion of our experimental results, and Section 6 concludes the paper. 2 Background 2.1 Intrusion effects Most previous studies on intrusive NPI licensing have focused on languages such as English and German, where the NPI-licensor dependency is retrospective. For instance, in a series of experiments, Xiang et al. (2013) investigate the licensing conditions of the English NPI ever and the licensors no and only. Xiang et al. find that sentences containing intrusive licensors (3b) are rated signigicantly lower than sentences containing proper licensors (3a). Crucially, they also find evidence for intrusion effects in the form of higher acceptability ratings for sentences with intrusive licensors (3b) compared to their unlicensed counterparts (3c). (3) a. No/only documentaries [that the network TV stations have played during prime time] have ever been very controversial. [Licensed] b. *The documentaries [that no/only network TV stations have played during prime time] have ever been very controversial. [Intrusive] c. *The documentaries [that the network TV stations have played during prime time] have ever been very controversial. [Unlicensed] Xiang et al. also examine the online profile of intrusion effects through a series of self-paced reading (SPR) studies. The reading time (RT) results only partially replicated the acceptability judgment patterns: while the negative licensor no gave rise to intrusion effects, such that on average the NPI in the intrusive condition (3b) was read faster compared to the unlicesed condition (3c), no significant difference was found for the licensor only. Taken together, these results suggest that the time resolution of intrusion effects is not homogeneous across different types of licensors. Parker & Phillips (2016) also find variable intrusion effects with different types of NPIs. In an SPR study, the authors report intrusion effects for the NPI ever and the licensor no (4a), thus replicating Xiang et al.’s findings (Cf. Vasishth et al. (2008) for similar effects in German). However, they fail to detect parallel intrusion effects for the NPI any (4b), suggesting that intrusion effects are not a general property of NPIs either.2 2 The same asymmetry between ever and any was found in a speeded acceptability judgment task, such that ever but not any gave rise to intrusion effects. NPI licensing and intrusion effects in Japanese (4) 445 a. No/the authors [that no/the critics recommended] have ever received acknowledgement for a best-selling novel. b. No/the authors [that no/the critics recommended] have received any acknowledgement for a best-selling novel. Finally, Parker & Phillips point out a third axis along which intrusive effects may vary. These authors argue that intrusive effects are further modulated by the amount of time spanning between the NPI and its potential licensor. In a speeded acceptability judgment task, Parker & Phillips find that the facilitatory effects incurred by intrusive licensors (5a), compared to a control unlicensed condition, disappear when a parenthetical (italicized) intervenes between the NPI and the irrelevant licensor (5b).3 (5) a. As the editors mentioned, the authors [that no critics recommended for the assignment] have ever receved a pay raise. b. The authors [that no critics recommended for the assignment] have, as the editors mentioned, ever received a pay raise. We now turn to prospective NPI dependencies. Lee & Yun (2022) investigate intrusion effects in Korean NPI licensing. Besides displaying a prospective dependency, in Korean the NPI and its licensor must be realized in the same clause. As a result, both the intrusive sentences in (6a) and (7a) are ungrammatical. In (6a), the NPI amwuto is in the matrix clause while the negation an- is in the embedded clause; in (7a), amwuto is in the embedded clause while an- is in the matrix clause. In terms of offline acceptability, Lee & Yun detected clear intrusive effects, such that (6a) and (7a) were judged to be more acceptable compared to their unlicensed counterparts in (6b) and (7b) respectively.4 (6) a. *Amwuto [Cenguni-ka Seoul-ey an-ka-ss-tako] ha-ess-ta. anyone Cengun-NOM Seoul-LOC NEG-go-PST-that say-PST-DEC (Lit.) ‘Anyone said that Cengun didn’t go to Seoul.’ [Intrusive] b. *Amwuto [Cenguni-ka Seoul-ey ka-ss-tako] ha-ess-ta. anyone Cengun-NOM Seoul-LOC go-PST-that say-PST-DEC (Lit.) ‘Anyone said that Cengun went to Seoul.’ [Unlicensed] (7) a. *Cenguni-ka [amwuto Seoul-ey ka-ss-tako] an-ha-ess-ta. Cengun-NOM anyone Seoul-LOC go-PST-that NEG-say-PST-DEC (Lit.) ‘Cengun didn’t say that anyone went to Seoul.’ [Intrusive] b. *Cenguni-ka [amwuto Seoul-ey ka-ss-tako] ha-ess-ta. Cengun-NOM anyone Seoul-LOC go-PST-that say-PST-DEC (Lit.) ‘Cengun said that anyone went to Seoul.’ [Unlicensed] 3 It is worth noting that unlike Xiang et al. (2013), Parker & Phillips did not find intrusive effects in offline acceptability in the same configurations tested by Xiang et al., despite the fact that the relative clauses used in Xiang et al.’s study were longer in terms of number of words. 4 A similar result was found for Turkish, another language with prospective NPI dependencies. See Yanilmaz & Drury (2018) for further details. 446 SU AND APARICIO The online profile of intrusive effects in Korean seems nevertheless far less clear. Lee & Yun (2022) report results from an SPR task, where contrary to what has been previously found in English and German, there were no obvious facilitatory effects in the intrusive conditions (6a-7a) compared to the unlicensed ones (6b-7b). In fact, the intrusive conditions display slower RTs than the unlicensed conditions with both matrix and embedded NPIs (though in the latter case no significant difference was found at the matrix verb between the intrusive condition and a grammatical condition containing only negation). 2.2 NPI licensing in Japanese Just like Korean, Japanese NPI licensing differs from languages like English in that the NPI precedes its licensor in the sentential linear order. This is exemplified in (8), where the NPI daremo ‘anyone’ is licensed by the upcoming negative morpheme -nakat. (8) Taro-ga daremo awa-nakat-ta. Taro-NOM anyone meet-NEG-PST ‘Taro didn’t meet anyone.’ From a processing perspective, establishing the NPI-licensor dependency in languages such as Japanese, where the dependency is prospective, arguably involves different mechanisms compared to languages such as English, where the dependency is retrospective. In Japanese, encountering the NPI triggers the expectation of an upcoming licensor downstream. In English, on the other hand, encountering an NPI requires retrieving a previously interpreted lincensor from short term memory and checking that the proper syntactic configuration between licensor and licensee obtains. A second difference is that the licensing environment is much more restricted in Japanese compared to English. Similar to what was observed for Korean, Japanese NPIs can be licensed solely by a clause-mate negation (Sohn 1995, Shimoyama 2011). Given this configurational constraint, mismatches between the syntactic positions of the NPI and the potential licensor can give rise to different types of intrusive conditions. (9) a. Daremo [Kenta-ga Tokyo-e it-ta-to] iwa-nakat-ta. anyone Kenta-NOM Tokyo-LOC go-PST-that say-NEG-PST ‘No one said that Kenta went to Tokyo.’ [Licensed] b. *Daremo [Kenta-ga Tokyo-e ika-nakat-ta-to] it-ta. anyone Kenta-NOM Tokyo-LOC go-NEG-PST-that say-PST (Lit.) ‘Anyone said that Kenta didn’t go to Tokyo.’ [Intrusive] (10) a. Kenta-ga [daremo Tokyo-e ika-nakat-ta-to] it-ta. Kenta-NOM anyone Tokyo-LOC go-NEG-PST-that say-PST ‘Kenta said that no one went to Tokyo.’ [Licensed] iwa-nakat-ta. it-ta-to] b. *Kenta-ga [daremo Tokyo-e Kenta-NOM anyone Tokyo-LOC go-PST-that say-NEG-PST NPI licensing and intrusion effects in Japanese (Lit.) ‘Kenta didn’t say that anyone went to Tokyo.’ 447 [Intrusive] As seen in the grammatical examples in (9a)-(10a), the NPI and its licensor are clause-mates: both are either in the matrix clause (9a) or in the embedded clause (10a). In contrast, in the ungrammatical examples (9b) and (10b), the NPI and its licensor belong to two different clauses. 3 Experiment 1: Acceptability judgments The goal of Experiment 1 is to determine whether Japanese gives rise to offline intrusion effects comparable to those found in other languages with prospective NPI-licensor dependencies such as Korean and Turkish. Our results show clear intrusion effects parallel to those previously detected in languages with prospective and retrospective NPI licensing. 3.1 Materials Our experiment manipulated two factors: the position of the NPI daremo ‘anyone’ (M: matrix, E: embedded) and the position of the negative morpheme -nakat (M: matrix, E: embedded, Ø: none). The two factors were fully crossed resulting in the six conditions exemplified in Table 1. Conditions a and d consist of the grammatical licensed conditions, where the NPI and the licensor are both in matrix position (a: NPIM -NEGM ) or in embedded position (d: NPIE -NEGE ). Conditions b and e contain the intrusive conditions, since the NPI and its licensor are in mismatched syntactic positions (b: NPIM -NEGE ; e: NPIE -NEGM ). Finally, conditions c and f represent the unlicensed conditions, since no negation is present to license the NPI in the matrix (c: NPIM -NEGØ ) or the embedded clause (f : NPIE -NEGØ ). In all items the matrix clause (regions 1-5) was followed by another clause (regions 6-8).5 Thirty-six sets of items were distributed across six lists following a Latin Square design. Besides the 36 experimental trials, participants saw 54 filler sentences. A comprehension question followed each of the sentences. Participants were presented with three possible answers to the question. 3.2 Participants A total of 21 Japanese native speakers (mean age = 28.0) participated in the experiment. They were recruited from italki (https://www.italki.com/), an online language learning platform. All participants were compensated for their participation. 3.3 Procedure The experiment was run on PCIbex Farm. Participants were instructed to rate each item in terms of its acceptability on a scale from 1-7 (1: least acceptable, 7: most acceptable). Next, participants responded to a comprehension question by selecting one of three possible answers. There was no time restriction, and all participants were able to finish the experiment within 30 minutes. 5 This was done in order to avoid sentence-final wrap up effects in Experiment 2, an SPR experiment that tested the same exact set of stimuli. 448 SU AND APARICIO Table 1: A sample set of experimental stimuli 3.4 Results and Discussion Answers to comprehension questions pertaining to filler trials were used as attention checks. We determined an 80% accuracy threshold that was met by all participants. Therefore, data from the 21 participants tested were analyzed. Experiment 1 results are presented in Figure 1. A linear mixed effects model was fit to the data using the R package lme4 (Bates et al. 2012). Acceptability ratings were predicted from the fixed effects predictors NPI POSITION and NEG PO SITION and their interaction. Random intercepts and slopes for the fixed effects predictors by subjects and by items were also included. Because the interaction term did not reach significance (p > 0.05), we proceeded to collapse the data across NPI position. To examine the effect of grammaticality, we first contrasted the grammatical conditions a/d against b/e and c/f . Results revealed a significant difference such that the grammatical conditions were rated significantly higher than the ungrammatical ones (β = -3.18, SE = 0.33, t = -9.612, p < 0.001). To determine the effect of intrusive licensing, we further compared the intrusive conditions (b/e) against the unlicensed conditions (c/f ). Results revealed that the intrusive conditions were rated significantly higher than the unlicensed ones (β = -1.63, SE = 0.38, t = -4.320, p < 0.001). Our results therefore show significant effects of both grammaticality and intrusion, though these effects were not modulated by the syntactic position of the NPI. 4 Experiment 2: Self-paced reading The goal of Experiment 2 was to examine the online processing signature of the offline intrusion effects detected in Experiment 1. To this end, we conducted an SPR study that tested the same stimuli set used in Experiment 1. Intrusion effects were predicted to arise in the form of higher RTs in the unlicensed condition, relative to the intrusive condition. NPI licensing and intrusion effects in Japanese 449 Figure 1: Experiment 1 acceptability judgment results. 4.1 Materials The materials used in Experiment 2 where the same as in Experiment 1. 4.2 Participants Thirty-two Japanese native speakers (mean age = 26.2) participated in Experiment 2. All participants were compensated for their participation. 4.3 Procedure The experiment was run on PCIbex Farm, and used a moving window, self-paced reading task. Participants read each sentence on a computer screen one segment at a time. Each sentence was followed by a comprehension question. To answer the question, participants were instructed to choose one of three possible answers. On average, participants completed the experiment in approximately 30 minutes. 4.4 Results and Discussion As in Experiment 1, answers to filler trials’ comprehension questions were used as attention checks. All participants had an accuracy rate above 80%, so data from all participants were included in subsequent analyses. Data points consisting of RTs below 200 ms or 3 SD away from the region mean were removed from data analysis (3.79% of the data). Next, all RTs were log transformed. To control for within-region length differences across conditions, we fitted a mixed effects model to the log transformed RTs using number of characters and segment position as fixed effects predictors and by participant random intercepts. The residuals extracted from such model were used as the new dependent variable in all subsequent data analyses. 450 SU AND APARICIO We start with the analyses corresponding to the conditions with NPIs in matrix position. Figure 2 shows the average segment-by-segment log RT residuals for such conditions. To determine any potential grammaticality effects, we fit a linear mixed effects model to the data pertaining to the critical region, i.e., the matrix verb region (region 5), which is the point where a potential NPI-licensor dependency could be established. The model predicted the residual log RTs from the fixed effects of GRAMMATICALITY (grammatical vs. intrusive/unlicensed conditions) and NEG POSITION (matrix vs. embedded). Random intercepts for subjects and items, as well as random slopes for all the fixed effects predictors were also included. Results revealed a significant grammaticality effect, such that the grammatical condition was read faster than the ungrammatical conditions (β̂ = 0.28, SE = 0.07, t = 3.920, p < 0.001). Crucially, no further distinction was detected at region 5 between the two ungrammatical conditions (intrusive vs. unlicensed: p > 0.5), indicating that participants did not display intrusion effects. Figure 2: By-region log RT residuals corresponding to conditions with matrix NPIs. Figure 3 shows the average segment-by-segment log RT residuals for sentences with embedded NPIs. The same model fit to the conditions with matrix NPIs was fit to the data pertaining to conditions containing embedded NPIs. We first analyzed region 4, which contains the point at which the NPI-licensor dependency could be potentially established, i.e., the embedded verb. No significant differences across the three embedded NPI conditions were detected in region 4. However, the following region, containing the matrix verb (region 5), displayed interesting differences. As in the matrix NPI conditions, a significant effect of grammaticality was detected, with the grammatical condition being read faster than the ungrammatical ones (β = 0.24, SE = 0.07, t = 3.297, p < 0.01). Finally, the comparison between the NPIE NEGM and NPIE -NEGØ conditions did not reach difference (p > 0.05). NPI licensing and intrusion effects in Japanese 451 Figure 3: By-region log RT residuals corresponding to conditions with embedded NPIs. 5 General discussion Our offline acceptability judgments confirmed the existence of intrusion effects in Japanese NPI licensing. That is, ungrammatical sentences with an intrusive licensor were judged as more acceptable than ungrammatical sentences with no licensor. However, no intrusion effects were detected in Experiment 2, which consisted of an SPR task. Our SPR results therefore parallel those reported by Lee & Yun (2022) for Korean, where the intrusive conditions did not display facilitation effects during reading. In this respect, languages with prospective dependencies seem to differ from retrospective languages, for which online intrusion effects have been successfully elicited using self-paced reading. Much remains to be understood about intrusive NPI licensing in languages with prospective dependencies. For instance, compared to English, not much is known about how factors such as the distance between the NPI and the licensor affect the strength of intrusion effects, or whether intrusion effects generalize to other NPIs beyond daremo or amwuto in Japanese and Korean respectively. The space of possible explanations for the asymmetric status of intrusion effects reported in Experiments 1 and 2 is therefore quite large. Here, we would like to propose that the mechanisms supporting the online processing of prospective licensing, as opposed to retrospective licensing, might be responsible for the lack of online intrusion effects in Experiment 2. When the dependency is prospective, encountering the NPI triggers the expectation of a downstream licensor. Such expectation is never triggered in languages with retrospective dependencies, since the point at which the NPI is encountered is also the point at which the dependency is resolved. It is therefore possible that processing a prospective NPI-licensor dependency incurs higher working memory demands for the processor. More specifically, the expectation of having to establish a downstream dependency would require keeping the NPI highly 452 SU AND APARICIO activated in working memory, particularly during the maintenance phase, so that the NPI representation can be readily retrieved once negation is encountered. The expectation of an NPI licensor and/or higher working memory demands during the processing of prospective dependencies might result in increased violations when the dependency fails to be established, resulting in an increase in RTs that masks any potential early facilitation effects of intrusive licensing. 6 Conclusion To summarize, we have presented the first empirical evidence showing that Japanese gives rise to intrusive NPI licensing effects. In line with what has been reported for other languages with prospective NPI dependencies such as Korean, intrusion effects were observed in offline acceptability judgments (Experiment 1), but not in reading times (Experiment 2). The current results therefore suggest that the online processing of intrusive NPI dependencies—as gauged through reading times— differs when such dependency is prospective as opposed to retrospective. In particular, in the latter case intrusive effects have been successfully detected using selfpaced reading and other time-sensitive behavioral measures. We have proposed that this asymmetry could be a direct consequence of the fact that prospective dependencies trigger a strong expectation of an upcoming licensor as soon as the NPI is encountered. When a proper dependency cannot be established either because there is no licensor, or because the licensor is not in the proper syntactic configuration, such expectation is violated, thus incurring early processing costs. Further research will be required in order to pinpoint the exact mechanisms (e.g., effects of working memory maintenance) underlying this increased processing cost. References Bates, D., M. Maechler, & B. Bolker. 2012. lme4: Linear mixed-effects models using S4 classes. Retrieved from: https://cran.r-project.org/web/packages/lme4 . Drenhaus, H., S. Frisch, & D. Saddy. 2005. Processing negative polarity items: When negation comes through the backdoor. Linguistic evidence: Empirical, theoretical, and computational perspectives, 145–165. Ladusaw, W. A. 1979. Polarity sensitivity as inherent scope relations. University of Texas, Austin dissertation. Lee, S., & J. Yun. 2022. NPI licensing and intrusion effect in Korean. In Proc. 29th Japanese/Korean Linguistics Conference, 277–289. Parker, D., & C. Phillips. 2016. Negative polarity illusions and the format of hierarchical encodings in memory. Cognition 157. 321–339. Shimoyama, J. 2011. Japanese indeterminate negative polarity items and their scope. Journal of Semantics 28. 413–450. Sohn, K. W. 1995. Negative polarity items, scope, and economy. University of Connecticut dissertation. Vasishth, S., S. Brüussow, R. Lewis, & H. Drenhaus. 2008. Processing polarity: How the ungrammatical intrudes on the grammatical. Cognitive Science 32. 685–712. Xiang, M., B. Dillon, & C. Phillips. 2009. Illusory licensing effects across dependency types: ERP evidence. Brain & Language 108. 40–55. NPI licensing and intrusion effects in Japanese 453 Xiang, M., J. Grove, & A. Giannakidou. 2013. Dependency-dependent interference: NPI interference, agreement attraction, and global pragmatic inferences. Frontiers in Psychology 4. 708. Yanilmaz, A., & J. Drury. 2018. Prospective NPI licensing and intrusion in Turkish. Language, Cognition and Neuroscience 33(1). 111–138. Yun, J., S. Lee, & J. Drury. 2018. Negative polarity illusion in Korean. In Proc. 13th Worshop on Altaic Formal Linguistics, 1–10. Proceedings of CLS 58 (2022), 455-467 © Chicago Linguistic Society 2023. All rights reserved. 455 Island sensitivity with relativization in Japanese: The case of double relatives Maho Takahashi and Grant Goodall University of California, San Diego 1 Introduction There is a well-known typological distinction between head-initial and head-final relative clauses (RCs). Both are generally taken to be largely the same in terms of their structure, differing primarily in the order between the head noun and the clause. There may be interesting generalizations as to how this order correlates with other aspects of linear order in the language, but the internal structure of the clause and its relation to the head noun appear to be uniform across both types (see Cinque 2020 for general discussion). Given this background, it is surprising to note that the literature has suggested that there is a curious structural difference between the two types of RCs. In headinitial RCs, the clause contains an A'-gap that is sensitive to island structures. Analyses differ as to what movement creates this gap (see Bianchi 2002), but movement out of an island is prohibited, as seen in (1). (1) a. the mani [RC who I believe [__i is wearing nice clothes]] b. *the mani [RC who the clothesj [RC that __i is wearing __j] are dirty] Relativization out of an embedded clause is possible in principle, as in (1a), but not when it occurs out of another RC, a well-known island constraint (Complex Noun Phrase Constraint; Ross 1967), as in (1b). We refer to cases like (1b) as double relatives. In head-final RCs, island sensitivity is often found as well, but it is also claimed that gaps may sometimes occur within island structures without any apparent penalty, as seen in the Japanese example in (2) (see Aoun & Li 2003 and Han & Kim 2004 for analogous cases in Chinese and Korean, respectively)1. (2) 1 [RC [RC __i __j ki-tei-ru] hukuj-ga wear-PROG-PRS clothes-NOM yogore-tei-ru sinsii dirty-PROG-PRS gentleman ‘the gentlemani [RC who the clothesj [RC that__i is wearing __j ] are dirty]’ The list of abbreviations used in this paper is as follows: NOM=nominative, ACC=accusative, DAT=dative, TOP=topic, PST=past, PROG=progressive, PST=past, PRS=present, PASS=passive, ADN=adnominal, NEG=negative marker, REL=relative marker, MID=middle, COP=copula. 456 TAKAHASHI AND GOODALL In this case, relativization has occurred out of another RC, yielding a double relative, yet ill-formedness does not result, unlike what we saw in (1b). This insensitivity to the island structure suggests that the gap does not arise through A'movement or that there is some other major structural difference between (1b) and (2). If this is true, though, it is a blow to the view that head-initial and head-final RCs are essentially the same in their structure, differing only in linear order. Research on islands over the past decade, however, has shown that detecting island sensitivity is not as simple as noting the (un)acceptability of single sentences in isolation, as in (1b) or (2), since many factors can influence this acceptability (Sprouse 2007, Sprouse et al 2012, Sprouse & Villata 2021). Moreover, we now know that any instance of long-distance extraction, even out of a non-island, leads to a substantial decline in acceptability, so detecting an island effect means finding a case where extraction out of a particular structure leads to significantly greater degradation than one would have expected, given an equivalent instance of long-distance extraction without that structure. Doing this requires a formal sentence acceptability experiment, so the growth of this approach to studying islands has gone together with the development of “experimental syntax” more broadly (Sprouse & Hornstein 2013, Goodall 2021). In the following sections, we use the techniques of experimental syntax to approach the fact seen in (2) where relativization appears to be immune to islands, with particular attention to Japanese. We will see that despite the relative acceptability of structures like these in Japanese, they nonetheless bear the characteristic properties of island sensitivity when we submit them to experimental scrutiny. This suggests that the gap in head-final RCs is an A'-gap after all, and that at least in this respect, the expected symmetry in structure between head-initial and head-final RCs obtains. 2 Experiment 1 2.1 Participants 44 participants were recruited on CrowdWorks, a Japanese crowdsourcing platform. 8 of them were excluded for reasons to be specified below, leaving 36 participants (age range=20-60, mean=41.3) whose data were analyzed. All participants reported that Japanese was their first language, and their parents primarily used Japanese to communicate with them. Participants received approximately $2 (in Japanese yen) upon completion of the experiment. 2.2 Materials All stimuli were declarative sentences containing an embedded clause. A 2x2 factorial design was employed, crossing embedded CLAUSE TYPE (RC vs. kotoclause) and EXTRACTION (relativization) out of that embedded clause (+ vs. -). The embedded clause was always a part of the subject of the main clause, and a complex NP headed by the light noun koto ‘fact’ (a “koto-clause”) was used as the baseline condition, since previous studies show that extracting out of a complex Island sensitivity with relativization in Japanese 457 NP headed by koto does not lead to an island effect (Omaki et al 2020). Kotoclauses were used instead of clauses headed by the complementizer -to ‘that’ so as to facilitate lexical matching across conditions; both koto-clauses and RCs can be used naturally with the same predicates, but -to-clauses cannot. The +EXTRACTION conditions all involved relativization out of the embedded clause. In order to maximize acceptability of the double relative case, three restrictions that have been discussed in the relevant literature were observed (Inoue 1976, Hasegawa 1981, Sakai 1994, Han & Kim 2004, Ishizuka 2009). First, double relatives in our stimuli involved the extraction of a subject out of an object RC, rather than extracting an object out of a subject RC as in (3), which seems to suppress acceptability. (3) [RC [RC __i __j ki-tei-ru] sinsii-ga wear-PROG-PRS gentleman-NOM koron-da fukuj fall-PST clothes ‘the clothesj [RC2 that the gentlemani [RC1 that __i is wearing __j] fell down]’ Second, the head of the outer RC was construed as the possessor of the head of the inner RC, because when this is not true, as in (4), a significant decline in acceptability occurs. (4) [RC [RC __i __j shira-nai] syoonenj-ga know-NEG boy-NOM obore-ta sinsii drown-PST gentleman ‘the gentlemani [that the boyj [that __i does not know __j] drowned]’ Third, the predicate contained in the outer RC was limited to an unaccusative, passive, or adjectival predicate. (5) is reported to be not acceptable as the relevant predicate is a transitive kan-da ‘bit’. (5) [RC [RC __i __j kawaigatte-iru] inuj-ga love-PRS dog-NOM rinjin-o kan-da sinsii neighbor-ACC bite-PST gentleman ‘the boyi [that the dogj [that __i has taken a good care of __j] bit a neighbor]’ These three factors are worth exploring further on their own, but for our purposes here, we will simply use them as guidelines to construct double relatives that are of maximal acceptability. 458 TAKAHASHI AND GOODALL A sample set of stimuli is provided below. (6) Condition 1: CLAUSE TYPE: koto-clause; EXT: [koto gakusha-ga SF-shousetsu-o kai-ta-koto]-ga professor-NOM Sci-Fi novel-ACC write-PST-fact-NOM saikin shoten-de tokusyu-sa-re-ta. recently bookstore-at feature-do-PASS-PST ‘The fact that [koto a professor wrote a sci-fi novel] was recently featured in a bookstore.’ Condition 2: CLAUSE TYPE: koto-clause; EXT: + [RC [koto __i SF-shousetsu-o kai-ta-koto]-ga Sci-Fi novel-ACC write-PST-fact-NOM saikin shoten-de tokusyu-sa-re-ta recently bookstore-at feature-do-PASS-PST gakushai]-wa hokorashige-da. professor-TOP looks.proud-COP ‘The professori [RC who the fact that [koto __i wrote a sci-fi novel] was recently featured in a bookstore] looks proud.’ Condition 3: CLAUSE TYPE: RC; EXT: [RC gakusha-ga __j kai-ta] SF-shousetsuj-ga professor-NOM write-PST Sci-Fi novel-NOM saikin shoten-de tokusyu-sa-re-ta. recently bookstore-at feature-do-PASS-PST ‘The sci-fi novelj [RC that the professor wrote __j] was recently featured in a bookstore.’ Condition 4: CLAUSE TYPE: RC; EXT: + [RC [RC __i __j kai-ta] SF-shousetsuj-ga write-PST Sci-Fi novel-NOM saikin shoten-de tokusyu-sa-re-ta recently bookstore-at feature-do-PASS-PST gakushai]-wa hokorashige-da. professor-TOP looks.proud-COP ‘The professori [RC who the sci-fi novelj [RC that __i wrote __j] was recently featured in a bookstore] looks proud.’ 20 lexically-matched sets as in (6) were created using one of 4 passive verbs (tokusyu-sa-re-ta ‘was featured’, syuzai-sa-re-ta ‘was interviewed’, happyou-sare-ta ‘was announced’, kouhyou-sa-re-ta ‘was disclosed’). These verbs all allow a koto-clause subject (Condition 1), an animate subject (Condition 2), or an inanimate subject (Conditions 3 and 4) equally naturally. Stimuli were counterbalanced through a Latin-square procedure, resulting in 4 lists (5 items per Island sensitivity with relativization in Japanese 459 condition; 20 items per list). 40 fillers were also created, consisting of sentences with varying degrees of acceptability: 10 fillers of expected high acceptability, 10 of intermediate acceptability (e.g. sentences with center-embedding), and 20 of low acceptability (e.g. sentences violating the Coordinate Structure Constraint). Fillers were identical across lists, and each of the lists consisted of 60 items. Stimuli were pseudo-randomized within each list. 2.3 Procedures The experiment was hosted on Ibex farm (Drummond 2013). Participants were instructed to rate how natural each sentence sounded by clicking on a number on a scale from 1 (very unnatural) to 7 (very natural). Participants also completed a brief language background questionnaire. To screen out participants who were not attending to the task, responses to the 10 filler items with the highest mean acceptability scores across all participants and the 10 with the lowest scores were identified. Participants whose ratings were more than 2 standard deviations away from the mean for 5 or more of these 20 items were excluded from further analysis. Two participants were filtered out in this way. In addition, a server error resulted in over-recruitment of participants for one of the lists, which led us to exclude the last 6 participants in that list (as determined by their submission date) for maintaining counterbalancing across lists. The final dataset consisted of 9 participants in each of the 4 lists, 36 in total. 2.4 Data analysis Raw acceptability scores were converted to z-scores prior to analysis, in accord with standard practice. A linear mixed-effects regression model was created using the lmerTest package (Kuznetsova et al 2017) in R (R Core Team 2022). Since the maximal model (as recommended in Barr et al 2013) did not converge, we used random intercepts and random slopes of the two manipulated factors for participant and item, and random slopes of their interaction only for participant. 2.5 Predictions Long-distance extraction is known to cause a significant decline in acceptability (Alexopoulou & Keller 2007, Fanselow 2021, Goodall 2021), so we expect the +EXTRACTION cases (Conditions 2 and 4) to be significantly less acceptable than the -EXTRACTION cases (Conditions 1 and 2). If there is no island effect, there should be no interaction between EXTRACTION and embedded CLAUSE TYPE, and the decline should occur in parallel in both the koto-clause and RC cases. If there is an island effect, on the other hand, we would expect an interaction between the two factors, with a greater decline in acceptability with extraction out of an RC than with extraction out of a koto-clause. These two possible outcomes are illustrated in Figure 1. 460 TAKAHASHI AND GOODALL Figure 1: Predicted outcome if there is no island effect (left) vs. if there is an island effect (right). 2.6 Results Mean z-scores for the four conditions of Experiment 1 are presented in Figure 2. Figure 2: Aggregated responses to critical items in Experiment 1. The model revealed a significant main effect of EXTRACTION, such that +EXTRACTION sentences were rated lower than their non-extracted counterparts, as expected (β = -0.49, SE = 0.07, p < 0.001). There was no main effect of embedded CLAUSE TYPE (β = -0.03, SE = 0.08, p = 0.76). Crucially, the interaction between these two factors was significant (β = -0.30, SE = 0.09, p < 0.01). That is, the decline in acceptability for extraction out of an RC was significantly greater than the decline for extraction out of a koto-clause. As illustrated above, this is the pattern associated with an island effect. The mean z-scores for fillers showed that participants made the expected distinctions: 1.03 for the high-acceptability fillers, -0.45 for the intermediateacceptability fillers, and -0.86 for the low-acceptability fillers. Island sensitivity with relativization in Japanese 461 2.7 Discussion The results of Experiment 1 suggest that even when sentences are constructed to make extraction out of another RC maximally acceptable, we still find the kind of interaction associated with an island effect. This suggests that double relatives in Japanese are not as exceptional as they appear at first; they still exhibit the island effect that would be expected of extraction out of an RC. The size of the interaction effect may be calculated by means of a differencesin-differences (DD) score (Sprouse et al 2012), in which the difference between the two -EXTRACTION conditions is subtracted from the difference between the two +EXTRACTION conditions. In this experiment, the DD score is 0.302, which is within the range observed for island effects in Sprouse and Villata (2021), though at the low edge of that range. A superadditive interaction as observed here is standardly taken to be a signature of A'-movement, since A'-dependencies show this effect and other dependency types do not. If this is correct, then non-A'-dependencies should not show this same type of interaction even when they are used in the same environment and under the same experimental conditions. We test this prediction in Experiment 2. 3 Experiment 2 In Experiment 1, we saw that relativization in Japanese shows an island effect, i.e. a superadditive interaction between extraction and the type of embedded clause. This effect is characteristic of A'-dependencies, but here we explore whether a superficially similar dependency will show a similar effect. We use the anaphor zibun ‘self’ since it resembles relativization in two important ways: The dependency between zibun and its antecedent can cross clause boundaries and the antecedent can appear to the right of zibun. If this dependency, like relativization, shows an interaction when tested experimentally, it would suggest that superadditivity is not limited to A'-dependencies in Japanese. If no such interaction is observed, however, it would strengthen the conclusion that only A'dependencies show superadditivity. 3.1 Participants A new group of 37 self-identified native Japanese speakers was recruited on CrowdWorks. All were compensated approximately $2 for their participation. The same attention check procedure as in Experiment 1 was used and one participant was excluded in this way, resulting in 36 participants (age range=19-56, mean=37.5) whose responses were analyzed. 3.2 Materials The design was very similar to that of Experiment 1, but with a factor of ZIBUN (i.e. presence or absence of a cataphoric dependency between zibun and an antecedent) instead of EXTRACTION. For koto-clauses containing zibun, care was 462 TAKAHASHI AND GOODALL taken in choosing the predicate of which the clause is predicated, given restrictions that have been noted in this area (Oshima 2009, Kishida 2011). The antecedent for zibun was always singular. Sample stimuli are given in (7). (7) Condition 1: CLAUSE TYPE: koto-clause; ZIBUN: (Same as Condition 1: CLAUSE TYPE: koto-clause; EXT: - in (6)) Condition 2: CLAUSE TYPE: koto-clause; ZIBUN: + [RC [koto zibuni-ga SF-shousetsu-o self-NOM Sci-Fi novel-ACC kai-ta-koto]-ga saikin shoten-de write-PST-fact-NOM recently bookstore-at tokusyu-sa-re-ta gakushai]-wa hokorashige-da. feature-do-PASS-PST professor-TOP looks.proud-COP ‘The professori [RC who the fact that [koto selfi wrote a sci-fi novel] was recently featured in a bookstore] looks proud.’ Condition 3: CLAUSE TYPE: RC; ZIBUN: (Same as Condition 3: CLAUSE TYPE: RC ; EXT: - in (6)) Condition 4: CLAUSE TYPE: RC; ZIBUN: + [RC [RC zibuni-ga __j kai-ta] SF-shousetsuj-ga self-NOM write-PST Sci-Fi novel-NOM saikin shoten-de tokusyu-sa-re-ta recently bookstore-at feature-do-PASS-PST gakushai]-wa hokorashige-da. professor-TOP looks.proud-COP ‘The professori [RC who the sci-fi novelj [RC that selfi wrote __j] was recently featured in a bookstore] looks proud.’ Conditions without zibun (Conditions 1 and 3) are analogous to Conditions 1 and 3 in Experiment 1 (i.e., those without extraction), and a similar relationship exists between Conditions 2 and 4 in the two experiments (they contain extraction in Experiment 1 and a cataphoric dependency with zibun in Experiment 2). 3.3 Procedures The same experimental procedures were taken as in Experiment 1. 3.4 Data analysis As in Experiment 1, a linear mixed-effect model with random effects of subject and item was used to fit the data and test for significance. The model predicts the acceptability (in z-scores) of sentences as a function of cataphoric dependency across an embedded clause and whether the embedded clause is an RC or a kotoclause. As the maximal random model did not converge, the resulting model Island sensitivity with relativization in Japanese 463 includes the random intercepts and random slopes of the two factors for participant and item, and random slopes of their interaction only for participant. 3.5 Predictions If dependencies involving zibun are immune to islands, we should find no interaction between ZIBUN and embedded CLAUSE TYPE, similar to the left panel in Figure 1. If this type of dependency is island-sensitive, there should be an interaction, with zibun showing a greater decline when the embedded clause is an RC than when it is a koto-clause, on a par with the right panel in Figure 1. 3.6 Results The mean z-scores for the four conditions are presented in Figure 4. The model revealed a main effect of ZIBUN (β = -0.46, SE = 0.09, p < 0.001), i.e. sentences with a dependency involving zibun were significantly less acceptable than those without such a dependency. There was no main effect of embedded CLAUSE TYPE (β = -0.07, SE = 0.08, p = 0.35). Most notably, the interaction between embedded CLAUSE TYPE and ZIBUN was not significant (β = 0.02, SE = 0.08, p = 0.79). This is different from what we saw in Experiment 1, where there was a significant interaction between embedded CLAUSE TYPE and EXTRACTION. Figure 3: Aggregated responses to critical items in Experiment 2. As before, the mean z-scores of the fillers were as expected: 1.01 for high, -0.57 for intermediate, and -0.92 for low. 3.7 Discussion The motivation for Experiment 2 was to ask whether the superadditive interaction that we saw in Experiment 1 is found more generally in non-A'-dependencies when they are used in the same environment and under the same experimental conditions. We tested this by examining cataphoric dependencies with zibun and the results show that this dependency does not exhibit the type of interaction that we saw in Experiment 1. That is, in Experiment 1, there was a significant 464 TAKAHASHI AND GOODALL interaction between embedded CLAUSE TYPE and EXTRACTION, while in Experiment 2, there is no such interaction between the two factors. Put simply, relativization shows island-sensitivity while cataphoric dependencies with zibun do not. Given that island-sensitivity is a known property of A'-dependencies, this strongly suggests that relativization in Japanese results from A'-movement. 4 General discussion The main conclusion of our experiments is that relativization in Japanese is sensitive to islands. Specifically, we saw in Experiment 1 that relativization out of an RC (an island) leads to a greater decline in acceptability than does relativization out of a koto-clause (a non-island). Importantly, this effect does not obtain with cataphoric dependencies with zibun, as we saw in Experiment 2. This difference between relativization and zibun is what we would expect if the islandsensitivity of relativization is due to its being an A'-dependency. Our finding that relativization in Japanese exhibits a central property of A'movement has important implications for the proper analysis of RCs in this language, in that it casts doubt on analyses in which A'-movement is not involved. Many prominent analyses of Japanese RCs claim that the relation between the head and the apparent gap is that of an antecedent and a null pronoun (e.g. Kuno 1973, Hoji 1986, Murasugi, 2000), but if this is correct, it is difficult to see why the island effect that we have observed would arise. One interesting fact about our findings in Experiment 1 is that despite clear evidence of an island effect in the form of a superadditive interaction, the island “violation” is nonetheless relatively acceptable. That is, the condition in which there is extraction out of an RC has a mean z-score of 0.06, which is higher than the mean for fillers that were designed to be of intermediate acceptability (-0.45). This result is perhaps not a surprise, given that relativization itself may be less sensitive to islands than other types of A'-movement (Sprouse et al 2016, Abeillé et al 2020). If we pair this with the fact that at least in some languages, RCs are extremely mild barriers to extraction (Allwood 1982, Engdahl 1982, Maling & Zaenen 1982, Lindahl 2017, Kush et al 2019), then it may not be surprising that relativization out of an RC may produce a very small island effect and consequently, relatively high acceptability for the island violation. If this is true, however, why do we not see double relatives in English? There are many factors involved, no doubt, but one could be the fact that in English, double relatives often trigger a center-embedding structure, which is notoriously difficult for speakers to parse (Miller & Chomsky 1963, Gibson 2000). This is seen in (8), where a long embedded clause intervenes between the subject the professor and its predicate looks proud. (8) *The professori [RC who the sci-fi novelj [RC that __i wrote __j] was recently featured in a bookstore] looks proud. Island sensitivity with relativization in Japanese 465 In the Japanese equivalent (Condition 4 of Experiment 1), this does not occur; the embedded clause is to the left of the subject and the predicate is to the right. Since processing difficulty is known to depress acceptability, we would expect (8) to be substantially less acceptable than double relatives in Japanese. Whatever factors might be at play in the different levels of acceptability for double relatives in Japanese and English, our experimental results point very clearly to an important commonality between the two languages: relativization is sensitive to islands. This fact has been difficult to perceive in Japanese when single sentences are considered in isolation, but the island effect becomes apparent when sentences are considered within a full factorial design, as in our experiment. This result, in turn, suggests that head-initial and head-final RCs share profound structural similarities, despite the difference in linear order. References Abeillé, A., Hemforth, B., Winckel, E., and Gibson, E. (2020). Extraction from subjects: Differences in acceptability depend on the discourse function of the construction. Cognition. 204. 104293. Alexopoulou, T., & Keller, F. (2007). Locality, cyclicity, and resumption: At the interface between the grammar and the human sentence processor. Language. 83(1). 110–160. Allwood, J. S. (1976). The complex NP constraint as a non-universal rule and some semantic factors influencing the acceptability of Swedish sentences which violate the CNPC. University of Massachusetts occasional papers in linguistics. 2(1). 1–20. Aoshima, S., Yoshida, M., and Phillips, C. (2009). Incremental processing of coreference and binding in Japanese. Syntax. 12(2). 93–134. Aoun, J., and Li, Y. A. (2003). Essays on the Representational and Derivational Nature of Grammar: The Diversity of Wh-constructions. Cambridge, MA: MIT Press. Barr, D. J., Levy, R., Scheepers, C., and Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language. 68. 255–278. Bianchi, V. (2002). Headed relative clauses in generative syntax. Part I. Glot International. 6(7). 197–204. Cinque, G. (2020). The syntax of relative clauses: A unified analysis. Cambridge: Cambridge University Press. Drummond, A. (2013). Ibex farm. Online server: http://spellout.net/ibexfarm Engdahl, E. (1982). Restrictions on unbounded dependencies in Swedish. In Readings on unbounded dependencies in Scandinavian languages, ed. By E. Engdahl & E. Ejerhed, 151– 74. Stockholm: Almquist & Wiksell Int. Fanselow, G. 2021. Acceptability, grammar and processing. In The Cambridge Handbook of Experimental Syntax, ed. By G. Goodall, 118–153. Cambridge: Cambridge University Press. Gibson, E. (2000). The dependency locality theory: A distance-based theory of linguistic complexity. In Image, Language, Brain, ed. Y. Miyashita et al, 95–126. Cambridge: Cambridge University Press. Goodall, G. (2021). Sentence acceptability experiments: What, how, and why. In The Cambridge Handbook of Experimental Syntax, ed. By G. Goodall, 7–38. Cambridge: Cambridge University Press. Han, C.-H., and Kim, J.-B. (2004). Are there “double relative clauses” in Korean? Linguistic Inquiry. 35(2). 315–337. Hasegawa, N. (1981). A lexical interpretive theory with emphasis on the role of subject. Doctoral dissertation, University of Washington. 466 TAKAHASHI AND GOODALL Hoji, H. (1986). Logical form constraints and configurational structures in japanese. Doctoral dissertation, University of Washington. Inoue, K. (1976). Henkei-bunpou to nihongo [Transformational grammar in Japanese]. Tokyo: Taishukan. Ishizuka, T. (2009). CNPC Violations and Possessor Raising in Japanese. In Proc. 2nd ICEAL. Kishida, M. (2011). Reflexives in Japanese. Doctoral dissertation, University of Maryland. Kuno, S. (1973). The structure of the Japanese language. Cambridge, MA: MIT Press. Kush, D., Lohndal, T., & Sprouse, J. (2019). On the island sensitivity of topicalization in Norwegian: An experimental investigation. Language. 95(3). 393–420. Kuznetsova, A., Brockhoff, P. B., and Christensen, R. H. (2017). lmerTest package: tests in linear mixed effects models. Journal of statistical software. 82(13). 1–26. Lindahl, F. (2017). Extraction from relative clauses in Swedish. Doctoral dissertation, University of Gothenburg. Maling, J., & Zaenen, A. (1982). A phrase structure account of Scandinavian extraction phenomena. In The nature of syntactic representation, ed. By P. Jacobson & G. Pullum, 229– 282. Berlin: Springer, Dordrecht. Miller, G. A. & Chomsky, N. (1963). Finitary models of language users. In Handbook of Mathematical Psychology, ed. By D. Luce et al, 269–321. New York: John Wiley & Sons. Murasugi, K. (2000). An antisymmetry analysis of Japanese relative clauses. In The syntax of relative clauses, ed. By A. Alexiadou et al, 231–263. Amsterdam: John Benjamins Publishing. Omaki, A., Fukuda, S., Nakao, C., and Polinsky, M. (2020). Subextraction in Japanese and subject-object symmetry. Natural Language and Linguistic Theory. 38(2). 627–669. Oshima, D. (2009). A note on backward anaphora in japanese. Ibaraki Daigaku Ryuugaku-sentaa Kiyou [Ibaraki University international student center bulletin]. 7. 79–89. R Core Team (2022). R: A language and environment for statistical computing. (Version 4.1.2.) https://www.R-project.org/. Retrieved 23 Jan 27. Ross, J. R. (1967). Constraints on variables in syntax. Doctoral dissertation, MIT. Sakai, H. (1994). Complex NP constraint and case-conversions in Japanese. In Current topics in English and Japanese, ed. By M. Nakamura, 179–200. Tokyo: Hituzi. Sprouse, J. (2007). A program for experimental syntax: Finding the relationship between acceptability and grammatical knowledge. Doctoral dissertation, University of Maryland. Sprouse, J., Wagers, M., and Phillips, C. (2012). A test of the relation between working-memory capacity and syntactic island effects. Language. 88(1). 82–123. Sprouse, J., & Hornstein, N. (2013). Experimental syntax and island effects. Cambridge: Cambridge University Press. Sprouse, J., Caponigro, I., Greco, C., and Cecchetto, C. (2016). Experimental syntax and the variation of island effects in English and Italian. Natural Language & Linguistic Theory. 34(1). 307–344. Sprouse, J., and Villata, S. (2021). Island Effects. In The Cambridge Handbook of Experimental Syntax, ed. By G. Goodall, 227–257. Cambridge: Cambridge University Press. Proceedings of CLS 58 (2022), 467-478 © Chicago Linguistic Society 2023. All rights reserved. 467 Remarks on verb echo answers and head movement in Japanese* Tomoya Tanabe and Ryoichiro Kobayashi Hokkaido University and Tokyo University of Agriculture 1 Introduction Japanese is a strictly head-final language. For this reason, it has been hotly debated whether syntactic head movement exists in the language. While scholars such as Otani & Whitman (1991), Funakoshi (2014), and Hayashi & Fujii (2015), argue for its existence, researchers such as Hoji (1998), Fukui & Sakai (2003), and Kobayashi (2016, to appear) argue against it. Recently, Sato & Maeda (2021) have presented an argument for the existence of syntactic head movement in Japanese based on their observations of Verb-Echo Answers (VEAs). A VEA can be used as a response to a polar question. (1A), in which only the verbal complex is expressed, is interpreted as an affirmative answer to (1Q). (1) Q: Naomi-wa sara-o arai-masi-ta-ka? Naomi-top dish-acc wash-pol-pst-q ‘Did Naomi wash the dishes?’ A: Arai-masi-ta-yo. wash-pol-pst-prt lit. ‘Washed.’ (‘Yes, Naomi did.’) Sato & Maeda (2021) follow Holmberg (2016) and propose that (1A) is derived via Verb-stranding TP Ellipsis (VTPE) as in (2) (see also Sato & Hayashi 2018).1 In VTPE, V first undergoes V-to-T-to-C movement, creating a V-T-C complex in syntax. Subsequently, ellipsis applies to the TP that contains the arguments, which derives a VEA such as (1A). Based on the analysis, Sato & Maeda (2021) argue that head movement occurs in Japanese in narrow syntax. (2) [CP [TP Subject [VP Object tV ] tV−T ] V-T-C] This paper argues against this line of argument for head movement by proposing the following alternative analysis. In (3), all the heads stay in-situ in syntax and the arguments are elided via Argument Ellipsis (AE) (Oku 1998, Saito 2007, Takahashi 2006).2,3 The V-T-C complex is then formed by a morphological merger (Halle & Marantz 1993) in the post-syntactic component. * We thank the anonymous reviewers of CLS58 for their helpful comments. We also thank Satoshi Oku, Yurie Hara, and Kenta Mizutani for their valuable comments. This project is partially supported by JSPS KAKENHI Grant-in-Aid for Scientific Research (C) Grant Number JP21K00574 awarded to the second author. Needless to say, any shortcomings and errors of this paper are our own. 1 For ease of exposition, the subject is in [Spec, TP] in (2). However, we are completely agnostic about whether the subject in Japanese raises to the TP-domain. 2 See Landau (2020), who proposes that AE can apply to multiple arguments in a sentence 3 Another possibility is that VEAs involve pro-drop, as depicted in (i). 468 (3) TANABE AND KOBAYASHI [CP [TP Subject [VP Object V] T] C] This paper shows that our alternative analysis in (3) is empirically superior to the VTPE analysis in (2), and concludes that VEAs in Japanese are explained with no recourse to head movement. The rest of this paper is organized as follows. Section 2 provides novel data which cannot be explained by the VTPE analysis. Section 3 reconsiders Sato and Maeda’s (2021) observation that a voice mismatch is prohibited in VEAs and argues that a voice mismatch in a VEA results in unacceptability, rather than ungrammaticality. Section 4 discusses Sato and Maeda’s (2021) argument that the scope relation between disjunction/only and negation reverses in VEAs due to syntactic NEG-raising. We extend Tanabe & Hara (2021) and demonstrate that our AE analysis coupled with discourse-based analyses accounts for a wider range of data. Section 5 concludes the paper. 2 Adverb-inclusive reading in Verb-echo answers Sato & Maeda (2021) attempt to exclude the AE analysis of VEAs in (3) based on observations of null adjuncts (see also Sato & Hayashi 2018). Sugimura (2011) observes that not only arguments but also adjuncts can be null in Japanese VEAs, as (4) shows. As an answer to (4Q), (4A) has the adjunct-inclusive reading. (4) Q: Kazuma-wa sara-o teineini arai-masi-ta-ka? Kazuma-top dish-acc carefully wash-pol-pst-q ‘Did Kazuma wash the dishes carefully?’ A: Arai-masi-ta-yo. wash-pol-pst-prt ok:‘Yes, Kazuma washed the dishes carefully.’ (adjunct-inclusive) The adjunct-inclusive reading in (4A) is unexplained by the AE analysis because AE cannot apply to adjuncts (Oku 1998). In contrast, the adjunct-inclusive reading is explained if we assume that (4A) is derived via VTPE since it elides the adjunct as well as the arguments as shown in (5) (Hayashi & Fujii 2015, Funakoshi 2016). (5) [CP [TP Kazuma [VP dish carefully tV ] tV−T ] V-T-C] However, the rest of this section demonstrates that null adjuncts are observed in VEAs when VTPE is not an available option in syntax. It follows that adjunctinclusive readings in VEAs do not substantiate the VTPE analysis. 2.1 Verbal Identity in Verb-stranding TP-ellipsis Funakoshi (2014, p. 347) proposes that verb-stranding VP ellipsis must satisfy an identity condition, which can be summarized as follows: (i) [ CP [ TP pro [ VP pro V] T] C] Although we set aside this possibility, the discussion in this paper is neutral about the AE analysis and the pro analysis. We leave the AE vs pro debate concerning Japanese VEAs for future work (see Sato and Hayashi 2018, who argue against the pro analysis of VEAs). Verb echo answers and head movement in Japanese (6) 469 The stranded verbs must be identical. If they are not identical, they must at least be contrasted in meaning. Since what is important is the identity of the stranded verb (not the category of the elided constituent), the condition should also hold for VTPE. Thus, the VTPE analysis predicts that null adjuncts are observed only in constructions which satisfy (6). With this in mind, consider (7). The verbs in question, namely humitsubusu ‘trample down’ and moyasu ‘burn’ are neither identical nor contrasted in meaning; hence (6A) does not satisfy the identity condition in (6). Accordingly, the VTPE analysis predicts that null adjuncts are not derived in (7A). However, the fact is that when (7A) follows (7Q), it has the adjunct-inclusive reading. (7) Q: Gozira-ga [ AdvP issyunnoutini] biru-o humitsubusi-ta-no? Godzilla-nom in.an.instant building-acc trample-pst-q ‘Did Godzilla trample down the building in an instant?’ A: (Iya,) Moyasi-ta-yo. no burn-pst-prt ok: ‘Godzilla burned down the building in an instant.’ (adverb-inclusive) (8) as a whole can be naturally interpreted as follows: “Did Godzilla trample down the building in an instant? – No, he burned down the building in an instant”. This poses problems for the VTPE analysis of null adjuncts. Likewise, in (8), the verbs, taosu ‘defeat’ and nomikomu ‘gulp down’ do not satisfy (6). Nonetheless, the unpronounced adjunct ikioiyoku ‘vigorously’ is interpreted in (8A); the interpretation that Kirby gulped down Yoshi vigorously is easily accessible. (8) Q: Kaabii-ga [ AdvP ikioiyoku] Yosshii-o taosi-ta-no? Kirby-nom vigorously Yoshi-acc defeat-pst-q ‘Did Kirby vigorously defeat Yoshi?’ A: (Iya,) Nomikon-da-yo. no gulp.down-pst-prt ok: ‘Kirby vigorously gulped down Yoshi.’ (adverb-inclusive) (Note: In Smash Bros., Kirby may gulp down his enemies, but he does not defeat them by doing so.) The data show that adjuncts can be null in a VEA even if (6) is not satisfied. Put in another way, the results indicate that null adjuncts can be derived without VTPE. Based on the observation, we propose that null adjuncts in VEAs are derived by Adjunct Ellipsis (Collins 2015, Collins 2017, Oku 2016, Kobayashi 2020). The analysis is precisely illustrated in (9), where adjunct ellipsis applies to the adjunct alone and AE applies to the arguments. Since adjunct ellipsis does not need to satisfy the identity condition in (6), (8) explains why the adjunct can be null in (7). (9) [ CP [ TP Subject [ VP Adjunct Object V] T] C] 2.2 Adverbs outside the scope of negation As illustrated in (10), applying VTPE in a negative sentence requires V to undergo a V-to-NEG-to-T-to-C movement. Thus, under the VTPE analysis, elements elided 470 TANABE AND KOBAYASHI via VTPE always fall within the scope of negation (indicated by shading). (10) [CP [TP Subject [TP [NegP [VP Object tV ] tV−NEG ] tV−NEG−T ] ] V-NEG-T-C] Indeed, one of Sato and Maeda’s (2021) arguments for the VTPE analysis of VEAs comes from their observation that the raised NEG in VTPE obligatorily takes scope over the elided elements in the TP, which we discuss in Section 4. It follows that the analysis predicts that elided adjuncts also fall within the scope of negation. In order to test this prediction, we consider data involving the adjunct iwaretatoorini ‘as one was told’. The adjunct allows both the wide scope and the narrow scope readings in relation to negation; hence, it provides us with an empirical testing ground. (11) illustrates the ambiguity; it is ambiguous between the manner adverb interpretation of the adjunct (i.e., the NEG > Adjunct reading) and the sentential adverb interpretation of the adjunct (i.e., the Adjunct > NEG reading). (11) Kenta-wa iwaretatoorini geemu-o si-nak-atta. Kenta-top as.he.was.told dish-acc do-neg-pst ok: ‘Kenta did not play the video game in the way he was told to.’ (NEG > Adjunct) ok: ‘Kenta did not play the video game following what he was told (he was told not to play the video game).’ (Adjunct > NEG) The VTPE analysis predicts that if the adjunct is elided, the resulting VEA has only a NEG > Adjunct reading since NEG in the V-NEG-T-C complex scopes over the adjunct contained in the elided TP. To see whether the prediction is correct, consider the following example. Interestingly, (12A) allows an adjunct-inclusive reading not only under the NEG > Adjunct but also under the Adjunct > NEG reading. (12) Q: Kenta-wa [AdvP iwaretatoorini] geemu-o si-nak-atta(-no)? Kenta-top as.one.was.told game-acc do-neg-pst-prt ‘Did Kenta not play video games as he was told?’ A: Si-nak-atta-yo. do-pol-neg-pst-prt ok: ‘Kenta did not play the video game in the way he was told to.’ (NEG > Adjunct) ok: ‘Kenta did not play the video game following what he was told (he was told not to play the video game).’ (Adjunct > NEG) For those who find the Adjunct > NEG reading difficult to get in (12A), specifying a context may be helpful. The context is such that the adjunct is unambiguous; it is understood only as a sentential adverb. Nonetheless, Kenta’s answer to Yuka’s question has the adjunct-inclusive reading. (13) Context: Yuka, Kenta’s mother, told Kenta not to play video games over the weekend, when she was away from home. When she returned home at the end of the weekend, Yuka asks Kenta: Yuka: Anata-wa [AdvP iwaretatoorini] geemu-o si-nak-atta(-no)? you-top as.one.was.told game-acc do-neg-pst-prt Verb echo answers and head movement in Japanese 471 ‘Did you not play video games as he was told?’ Kenta: Si-nak-atta-yo. do-pol-neg-pst-prt not: ‘I did not play the video game in the way you told me to.’ (NEG > Adjunct) ok: ‘I did not play the video game following what I was told (I was told not to play the video game).’ (Adjunct > NEG) The Adjunct > NEG reading in the VEA is unexplained under Sato and Maeda’s (2021) VTPE analysis, where the raised NEG obligatorily takes the wide scope with respect to the TP-internal elements. (14) illustrates this point. NEG in VTPE ends up in a higher position than the adjunct, regardless of the adjunct’s position. (14) [CP [TP Subj [TP Adj [TP [NegP [VP Obj tV ] tV−NEG ] tV−NEG−T ]]] V-NEG-T-C] Here again, the adjunct ellipsis analysis explains the result in (13).4 In (15), NEG does not raise, and hence the adjunct that is higher than NEG can be elided. (15) [ CP [ TP Subject [ TP Adjunct [ TP [NegP [VP Object V] NEG] T]]] C] This section has shown that observations regarding adjunct-inclusive readings in VEAs are better explained by our analysis, which posits that VEAs are derived by applying AE to the arguments and adjunct ellipsis to the adjuncts. 3 Voice mismatches in Verb-echo answers Sato & Maeda (2021) argue that a voice mismatch in VEAs results in ungrammaticality, as in (16A). They claim that the impossibility of voice mismatches in VEAs supports their VTPE analysis because sluicing in English, which can be analyzed as an instance of TP-ellipsis, disallows voice mismatches in (17b) (Merchant 2001). (16) (17) Q: Kenta-wa Yuka-o sikari-masi-ta-ka? Kenta-top Yuka-acc scold-pol-pst-q ‘Did Kenta scold Yuka?’ A: *Shikar-are-masi-ta-yo. scold-pass-pol-pst-prt lit. ‘Was scolded.’ (Intended: ‘Yuka was scolded by Kenta.’) a. I know someone scolded John, but I don’t know who. b. *I know someone scolded John, but I don’t know by whom. If (16A) involves VTPE, then a syntactic identity condition which prohibits voice mismatches in TP-ellipsis explains the ungrammaticality of both (16A) and (17b).5 While we agree that (16A) is “bad” to some degree, we argue that (16A) is not ungrammatical due to syntactic identity but it is unacceptable due to discourse factors. 4 5 As mentioned above, the pro-analysis of null arguments in VEAs is also a potential alternative. The line of arguments was first presented in Sato & Hayashi (2018). 472 TANABE AND KOBAYASHI First of all, the unacceptability of voice mismatches does not characterize VEAs in Japanese. (18) is a case of AE rather than TP-ellipsis, and yet a voice mismatch is unacceptable to some extent. The fact is not surprising since congruent questionanswer pairs normally do not tolerate a voice mismatch see (Weir 2017, for detailed discussion). Yet, some speakers find (18A) more acceptable than (16A), and one may claim that the contrast supports the VTPE analysis. (18) Q: Kenta-wa Yuka-o sikari-masi-ta-ka? Kenta-top Yuka-acc scold-pol-pst-q ‘Did Kenta scold Yuka?’ A: ??Yuka-wa sikar-are-masi-ta-yo. Yuka-top scold-pass-pol-pst-prt lit. ‘Yuka was scolded.’ (Intended: ‘Yuka was scolded by Kenta.’) Nonetheless, we maintain that the contrast (if exists) is also explained by a pragmatic analysis as follows: In (16A), neither of the arguments is expressed. Thus, a heavy burden is on the listener to infer ‘who was scolded by whom’ from contextual clues(, which are absent in the constructed example). (18A) may be more acceptable than (16A) because the information about ‘who was scolded’ is expressed, and the information about ‘by whom Yumi was scolded’ is easily inferred on the basis of the question being addressed by the conversation participants. The line of pragmatic analysis is supported by (19). (19A) is a VEA, and yet it is more acceptable than (16A) arguably because it is clear in the discourse that the utterer of (19A) is the one who was scolded by Kenta (as pointed out by Satoshi Oku (p.c.)). (19) Q: Kenta-wa anata-o sikari-masi-ta-ka? Kenta-top you-acc scold-pol-pst-q ‘Did Kenta scold you?’ A: Shikar-are-masi-ta-yo. scold-pass-pol-pst-prt lit. ‘Was scolded.’ (Intended: ‘I was scolded by Kenta.’) Importantly, the contrast between (16A) and (19A) is not explained if they are derived via VTPE. Sato and Maeda’s (2021) analysis based on syntactic identity on TP-ellipsis wrongly predicts (19A) to be ungrammatical on a par with (16A). In this section, we have shown that the alleged impossibility of voice mismatches in VEAs does not support Sato and Maeda’s (2021) VTPE analysis. 4 Negative scope reversal in Verb-echo answers? Now we discuss scopal interactions between disjunction/only and NEG in VEAs. We demonstrate that scope patterns support our analysis over the VTPE analysis. 4.1 Disjunction and negation In languages like English, a disjunction can take both a wide and narrow scope with respect to its clause-mate negation, as in (20). In Japanese, however, disjunction does not take scope below its clause-mate negation, as in (21) (Goro 2007). While Verb echo answers and head movement in Japanese 473 (21) means that (Hayato didn’t eat bread) or (Hayato didn’t eat rice) (or>neg), it cannot mean that Hayato didn’t eat bread or rice (neg>or). (20) (21) Ronald did not eat bread or rice. (or>neg, neg>or) Hayato-wa pan-ka-kome-o tabe-nak-atta. (or>neg, *neg>or) Hayato-top bread-or-rice-acc eat-neg-pst lit. ‘Hayato did not eat bread or rice.’ Sato & Maeda (2021) adopt Shibata (2015) and assume that the or>neg reading is derived in the structure (22), where the disjunctive phrase occurs higher than NegP.6 (22) [ CP [ TP Subject [ TP bread-or-rice [ TP [NegP [VP tbread-or-rice V] NEG] T]]]C] Based on this assumption, Sato & Maeda (2021) provide the following example to support their VTPE analysis of VEAs. They report that the otherwise unavailable neg>or reading becomes available in the VEA in (23A). Interestingly, (23A) does not seem to allow the or>neg reading (see also Funakoshi 2013, Maeda 2019). (23) Q: Hayato-wa pan-ka-kome-o tabe-ta-no? Hayato-top bread-or-rice-acc eat-pst-prt ‘Did Hayato eat bread or rice?’ A: Tabe-nak-atta-yo. (*or>neg, neg>or) eat-neg-pst-prt lit. ‘Did not eat.’ Sato & Maeda (2021) propose that the observation can be explained by the VTPE analysis in (24), in which the raised NEG takes scope over the disjunction. (24) [CP [TP Subj [TP b-or-r [NegP [VP tb-or-r tV ] tV-NEG ] tV-NEG-T ]] V-NEG-T-C] However, as Sato & Maeda (2021, p. 370) themselves acknowledge in a footnote, the VTPE analysis faces the counter-example in (25). The VEA which follows the negative question has the or>neg reading (see also Sakamoto 2016, Maeda 2019). This is unaccounted for if the VEA is derived via VTPE as illustrated in (24). (25) Q: Hayato-wa pan-ka-kome-o tabe-nak-atta-no? (or>neg, *neg>or) Hayato-top bread-or-rice-acc eat-neg-pst-prt ‘Did Hayato not eat bread or rice?’ A: Tabe-nak-atta-yo. (or>neg, *neg>or) eat-neg-pst-prt lit. ‘Did not eat.’ Our analysis straightforwardly explains the or>neg reading in (25A). As depicted in (26), the subject and the disjunctive phrase are elided by AE. Since the disjunctive phrase falls within the scope of negation, the or>neg reading in (25A) is obtained. 6 Shibata (2015) proposes that the disjunction “acyclically” merges to the argument only after it moves to a position above NegP. For ease of exposition, the analysis is not precisely illustrated in (22). 474 (26) TANABE AND KOBAYASHI [ CP [ TP Subject [ TP bread-or-rice [NegP [VP tbread-or-rice V] NEG] T] ] C] On the other hand, the AE analysis does not explain the neg>or reading in (23A). However, we point out that the judgement reported in Sato & Maeda (2021) is dubious. For us and five informants we consulted, the most likely interpretation of (23A) is that Hayato ate neither bread nor rice rather than it is not the case that Hayato ate bread or rice (i.e., neg>or). Given this, we propose that the source of the seemingly available neg>or reading in (23A) is (27A). Notice that (27A) does not involve disjunction, and yet it is truth-conditionally equivalent to neg>or.7 (27) Q: Hayato-wa pan-ka-kome-o tabe-ta-no? Hayato-top bread-or-rice-acc eat-pst-prt ‘Did Hayato eat bread or rice?’ A: (Iya,) Hayato-wa pan-mo kome-mo tabe-nak-atta-yo. no Hayato-top bread-also rice-also eat-neg-pst-prt ‘Hayato ate neither bread nor rice.’ Thus, the structure of (23A) is (28), where all the arguments are elided by AE. (28) [ CP [ TP Subject [ NegP [ VP bread-also rice-also V] NEG] T] C] While the AE analysis accounts for the observations in (23) and (25), a question remains as to why (23A) does not have the or>neg reading. AE should be able to derive a structure like (26), in which the disjunctive phrase above NEG is elided. Yet, this seemingly mysterious scope pattern can be explained by Tanabe and Hara’s (2021) analysis based on question-answer congruence. Let us reconsider the question-answer pair in (23). (23Q) asks whether there is a food among bread and rice that Hayato ate. Notice that the or>neg reading of (23A), (Hayato didn’t eat bread) or (Hayato didn’t eat rice) does not provide an answer. (29) shows that the or>neg reading is false if and only if both Hayato ate bread (p) and Hayato ate rice (q) are true. This means that answering (23Q) by (23A) under the or>neg reading amounts to providing an insufficient answer as follows: Hayato may have eaten bread or rice, but he may have eaten neither. At least, he did not eat both. This provides neither an affirmative answer nor a negative answer to (23Q). (29) p∨q p q ¬p ¬q T T T F T T F F T F T F F F T T F T F T ¬(p ∨ q) neg>or F F F T ¬p ∨ ¬q or>neg F T T T In contrast, asserting that Hayato ate neither bread nor rice provides a negative 7 Funakoshi (2013) also suggests that an apparently available neg>or reading in a null object construction is not derived from the neg>or structure but from pro. As repeatedly mentioned above, the pro analysis is also a potential alternative analysis to the VTPE analysis. Verb echo answers and head movement in Japanese 475 answer to (23Q), and thus it is available in (23A). This reveals that the otherwise available or>neg reading of (23A) can be excluded by question-answer congruence. The line of pragmatic analysis is compatible with the observation that the or>neg reading is available if the question is a negative question, as in (25). The negative question in (25Q) asks whether there is a food among bread or rice that Hayato did not eat. The or>neg reading of (25A) , which asserts that there is a food among bread or rice that Hayato did not eat gives an affirmative answer to (25Q). To summarize our discussion in this subsection, we have shown that the AE analysis coupled with a discourse-based analysis successfully explains the contrast between (23A) and (25A), which the VTPE analysis cannot explain. 4.2 -dake ‘only’ and negation This subsection demonstrates that the scope patterns of -dake ‘only’ in relation to negation is also better explained by the AE analysis. Sato & Maeda (2021) claim that the negation obligatorily takes scope over -dake ‘only’ in (30A). (30) Q: Kana-wa pan-dake tabe-ta-no? Kana-top bread-only eat-pst-prt ‘Did Kana eat only bread?’ A: Tabe-nak-atta-yo. (*only>neg, neg>only) eat-neg-pst-prt lit. ‘Did not eat.’ Sato & Maeda (2021) propose to explain the neg>only reading in (30A) by the VTPE analysis in (31), where the raised NEG scopes over -dake ‘only’. (31) [CP [TP K [TP bread-only [NegP [VP tb-only tV ] tV-NEG ] tV-NEG-T ]] V-NEG-T-C] However, the VTPE analysis again faces the following counter-example. In (32), the VEA answers a negative question, and it has only the only>neg. (32) Q: Kana-wa pan-dake tabe-nak-atta-no? (only>neg, *neg>only) Kana-top bread-only eat-neg-pst-prt ‘Did Ken not eat only bread?’ A: Tabe-nak-atta-yo. (only>neg, *neg>only) eat-neg-pst-prt The AE analysis in (33) straightforwardly explains the observation. The subject and the object suffixed with -dake ‘only’ are individually elided by AE. (33) [ CP [ TP Kana [ TP bread-only [NegP [VP tbread-only V] NEG] T] ] C] The contrast between (30A) and (32A) with regard to the availability of the only>neg reading is again accounted for by Tanabe and Hara’s (2021) discourse-based analysis. That is, (30Q) presupposes that Kana ate bread and asks whether it is the case that she did not eat anything else. On the other hand, the only>neg reading entails that Kana did not eat bread, which causes a presupposition failure. In contrast, 476 TANABE AND KOBAYASHI (32A) with the only>neg reading provides an affirmative answer to (32Q). Yet, if NEG does not raise to C, a question arises as to why the neg>only reading is derived in (30A). The neg>only reading reported in Sato & Maeda (2021), however, is arguably not real. To our ears, the most salient interpretation of (30A) is Kana did not eat bread (to begin with), where the meaning of ‘only’ is not present (see also Akiyama 2014, Moriyama 2017, Sato 2020). As mentioned above, this interpretation causes a presupposition failure; hence, the VEA sounds more natural if we add a comment that cancels the presupposition, as in (34A). (34) Q: Kana-wa pan-dake tabe-ta-no? Kana-top bread-only eat-pst-prt ‘Did Kana eat only bread?’ A: Iya, toiuka, tabe-nak-atta-yo. (*only>neg, ??neg>only) no in.fact eat-neg-pst-prt ok: ‘In fact, Kana did not eat bread to begin with.’ The alleged neg>only reading in (30A) therefore is derived from (35), in which the subject and the object (without -dake‘only’) are simply elided by AE.8 (35) [ CP [ TP Kana [ NegP [ VP bread V] NEG] T] C] Overall, we have shown that the scopal interactions between -dake ‘only’ and negation in VEAs are also explained by the AE analysis. 5 Conclusion This paper has revealed that the VTPE analysis of Japanese VEAs is inadequate. We have also shown that our AE analysis is empirically superior to the VTPE analysis of VEAs. Therefore, we conclude that the relevant observations of VEAs do not constitute evidence that syntactic head movement exists in Japanese. References Akiyama, M. 2014. The syntax of focus-doubling in Japanese. In Formal Approaches to Japanese Linguistics (FAJL) 7, ed. by S. Kawahara & M. Igarashi, 1–12. Cambridge, MA: MIT Working Papers in Linguistics. Collins, C. 2015. Adjunct deletion. Manuscript., NYU. Collins, C. 2017. Incomplete comparatives as ellipsis. Manuscript., NYU. Fukui, N., & H. Sakai. 2003. The visibility guideline for functional categories: Verb raising in Japanese and related issues. Lingua 113. 321–347. Funakoshi, K. 2013. Disjunction and object drop. Tampa Papers in Linguistics 4. 11–20. Funakoshi, K. 2014. Syntactic head movement and its consequences. College Park: University of Maryland dissertation. Funakoshi, K. 2016. Verb-stranding verb phrase ellipsis in Japanese. Journal of East Asian Linguistics 25. 113–142. 8 As repeatedly mentioned above, the pro analysis can also account for the data in (34). Verb echo answers and head movement in Japanese 477 Goro, T. 2007. Language specific constraints on scope interpretation in first language acquisition. College Park: University of Maryland: Doctoral dissertation. Halle, M., & A. Marantz. 1993. Distributed morphology and the pieces of inflection. In The view from the building 20: Essays in linguistics in honor of Sylvain Bromberger, ed. by k. Hale & S. Keyser, J, 111–176. Cambridge, MA: MIT Press. Hayashi, S., & T. Fujii. 2015. String vacuous head movement: The case of V-te in Japanese. Gengo Kenkyu (Linguistic Research) 147. 31–55. Hoji, H. 1998. Null object and sloppy identity in Japanese. Linguistic Inquiry 29. 127–152. Holmberg, A. 2016. The syntax of yes and no. Cambridge: Cambridge University Press. Kobayashi, R. 2016. Japanese V-te V compounds as post-syntactic compounds. Studies in Generative Grammar 26. 223–238. Kobayashi, R. 2020. A case against the verb-stranding VP-ellipsis analysis in Japanese. In Proceedings of Seoul International Conference on Generative Grammar 22, ed. by T. S. Kin & S.-Y. Cho, 298–304, Seoul. Korean Generative Grammar Circle. Kobayashi, R. to appear. Against V-to-T-to-C movement in Japanese: A case study of nonconstituent coordination. In Proceedings of the 11th Edition of the Workshop on Altaic Formal Linguistics, ed. by G. Tsoulas, Cambridge, MA. MIT Working Papers in Linguistics. Landau, I. 2020. On the nonexistence of verb-stranding VP-ellipsis. Linguistic Inquiry 51. 341– 365. Maeda, M. 2019. Argument ellipsis and scope economy. Syntax 22. 419–437. Merchant, J. 2001. The syntax of silence: Sluicing, islands, and the theory of ellipsis. Oxford: Oxford University Press. Moriyama, K. 2017. Dake no koosakujo [Argument ellipsis with only-XP]. Pre-proceedings of the 155th Meeting of the Linguistic Society of Japan 240–245. Oku, S. 1998. A theory of selection and reconstruction in the minimalist program. Storrs: University of Connecticut dissertation. Oku, S. 2016. A note on ellipsis-resistant constituents. Nanzan Linguistics 11. 57–70. Otani, K., & J. Whitman. 1991. V-raising and VP-ellipsis. Linguistic Inquiry 22. 345–358. Saito, M. 2007. Notes on East Asian argument ellipsis. Language Research 43. 203–227. Sakamoto, Y. 2016. Scope and disjunction feed an even more argument for argument ellipsis in Japanese. In Japanese/Korean Linguistics, ed. by M. Kenstowicz, T. Levin, & R. Masuda, volume 23, Stanford, CA. CSLI Publications. Sato, Y. 2020. Focus mismatch under ellipsis in Japanese, polarity and head movement. In Proceedings of the 22nd Seoul International Conference on Generative Grammar, ed. by T. S. Kim & S.-Y. Cho, 224–232, Seoul. Korean Generative Grammar Circle. Sato, Y., & S. Hayashi. 2018. String-vacuous head movement in Japanese: New evidence from verb-echo answers. Syntax 21. 72–90. Sato, Y., & M. Maeda. 2021. Syntactic head movement in Japanese: Evidence from verb-echo answers and negative scope reversal. Linguistic Inquiry 52. 359–376. Shibata, Y. 2015. Negative structure and object movement in Japanese. Journal of East Asian Linguistics 24. 217–269. Sugimura, M. 2011. Domain extension: A study of restructuring predicates in Japanese and the role of head movement. Montreal: McGill University dissertation. Takahashi, D. 2006. Apparent parasitic gaps and null arguments. Journal of East Asian Linguistics 15. 1–35. Tanabe, T., & Y. Hara. 2021. Question under Discussion-based analysis of Japanese ellipses. In Pre-proceedings of 162nd Meeting of the Linguistic Society of Japan, 329–335, Online. Weir, A. 2017. Cointensional questions and their implications for fragment answer. In Proceedings of the 21st annual Sinn und Bedeutung conference, ed. by R. Truswell, Edinburgh. University of Edinburgh. Proceedings of CLS 58 (2022), 479-501 © Chicago Linguistic Society 2023. All rights reserved. 479 A featural analysis of Mandarin classifiers and plural morpheme -men and its implications on the semi-lexicality of Mandarin Chinese Fanghua Zheng University of Cambridge 1 Abstract This paper revisits the syntax and semantics of the plural morpheme -men in Mandarin Chinese. In previous literature, there are mainly three formal approaches to explain the distribution of -men: 1. -Men is a plural morpheme and it is incompatible with classifiers due to the violation of Head Movement Constraint. 2. -Men is incompatible with classifiers because they compete for the same syntactic positions. 3. -Men is a group function local to nouns and its semantics is incompatible with individual classifiers. This paper shows that the first two analyses are empirically insufficient, and the third analysis is conceptually complicated. Instead, this paper is trying to defend a traditional descriptive generalization of -men that has not been given enough attention, namely -men can only co-occur with approximate quantity (Lü 1980). This paper accounts for this generalization by arguing that in Mandarin a multiplication relatino holds between Numeral Phrase and Classifier Phrase. The numeral in the specifier of Numeral Phrase is a Multiplier and the numeral implied by the classifier is a Multiplicand. The distribution of -men is sensitive to the result of the multiplication. This paper aims to account for this generalization by employing the feature theory to grammatical number proposed by Harbour (2014) and developed by Martí (2020a, 2020b). Different types of classifiers in Mandarin are realizations of different featural content: [+atomic] feature spells out individual classifiers; [-atomic, +minimal] spells dual classifiers (that is, classifiers meaning ‘couple’), and [-atomic, -minimal] spells out group classifiers. The [+atomic] feature entails the multiplicand to be ‘one’; the combination of [-atomic, +minimal] entails ‘two’; the combination of [-atomic, -minimal] leaves the multiplicand unspecified. The morpheme -men can merge into the structure only when the multiplication result of multiplier and multiplicand is an approximate quantity. 2 Introduction In the past two decades, the research on the syntax and semantics of plurality in the nominal domain has made lots of progress. There are three major developments in the syntax of plurality. The first view is put forward by Borer (2005) who argues that plural is a division function. Borer has proposed a universal DP structure and instead of regarding the plural morpheme as a function for counting in Number Phrase, she argues that the plural is a division function, which divides mass nounts into countable units, which can be further counted by Number Phrase. The second view argues that plural morphemes are distributed across the nominal spines, either merging as heads or adjuncts (Wiltschko 2008, Mathieu 2014, 480 FANGHUA ZHENG Acquaviva 2008, and Kramer 2016). They hold the view that there are many types of plural cross-linguistically and they should not be treated in the same manner. Even in a single language, there could be different types of plurals. They could occupy different layers in the functional spine of noun phrases, such as D, Numeral, Division, n, root. The third approach looks at the issue of plurality from the perspective of grammatical feature. Harbour (2014) proposed a featural analysis for grammatical number which aims to explain how grammatical number behaves cross-linguistically. Martí (2020a, 2020b) further develops Harbour’s system and uses several sets of primitive grammatical features to account for the behaviour and interpretation variations of number markings in languages such as Turkish, English, West Armenia, etc. Languages like Turkish and Hungarian are extremely interesting to the current paper for two reasons: first, there is a limited number of classifiers in these two languages; second, they both have a plural marker that cannot co-occur with numerals. These properties are, to some extent, similar to Mandarin Chinese. A major concern when looking at -men is its categorial status. The first question is whether -men is a lexical item or a grammatical item. The second question is whether Mandarin has grammatical number or not. The current paper takes the view that Mandarin has grammatical Number head in its nominal spine and -men is a semi-lexical semi functional item which merges to the functional structure when the merging conditions are satisfied. Corbett (2000) mentioned two languages that seem to have no number category: Piraha and Kawi (Old Javanese). These two languages don’t have plural nouns, not even plural pronouns. In these two languages, number could be indicated by quantifiers such as ‘many’ and ‘all’ by conjoining constructions. This is certainly not the case of Mandarin Chinese. In Mandarin Chinese, the morpheme -men is used with pronouns and animate human nouns to express plural meaning explicitly, such as xuesheng-men (the students), or pronouns such as ta-men (they). Although -men cannot be used with all common nouns, unlike inflectional morpheme -s in English, but -men is highly productive among animate human nouns and pronouns. Besides, when -men is attached to human nouns, it conveys plurality. Thus, it is problematic to treat -men as either fully functional or fully lexical. The in-between property of -men is an important starting point the current inquiry, namely, how do we capture this semi-lexical and semi-functional property in a syntactic theory which seems to embrace a fully functional structure. There are good reasons to take -men as semi-lexical items. According to Cavirani-Pots et al. (2021), semi-lexical items have unexpected or degraded morpho-syntactic behaviours compared to fully functional items. This is exactly the case for -men which has semantic constraints for nouns it combines with, and a plural meaning is implied when -men is attached. Harbour (2014) proposed a featural theory on grammatical number which is further developed by Martí (2020a, 2020b). Harbour and Martí’s systems are intended to explain number categories with purely inflectional properties. The plural morpheme -men, however, is not a fully inflectional morpheme, which means that we cannot apply Harbour’s and Martí’s systems to Mandarin Chinese directly. However, it does not mean that their theory is totally irrelevant when it comes to the Mandarin classifiers and plural morpheme -men 481 number system of Mandarin Chinese because categories like classifiers and morpheme -men do not look purely lexical. To accommodate to the semi-lexical nature of Mandarin Chinese, we need to make some modifications of the featural theory at hand. After all, we do not want to throw the baby out with the bath water. The structure of this paper will be as follows. Section 3 summarizes major properties of -men. In the meantime, I evaluate previous analyses of -men and propose a new descriptive generalization that can cover more empirical data. Then, I review some literature on the syntactic height of plural marking and argue that -men should be inserted under the Number head rather than little n. In section 4, I introduce the theoretical background of the featural analysis of grammatical number. Then, I propose a featural analysis for Mandarin individual classifiers and group classifiers and argue that a multiplication relation holds between the numeral in Spec Number Phrase and the numeral implied by the classifier to account for the insertion constraints of -men. I provide three syntactic environments where -men can be inserted to derive different interpretations and uses, including plural marking, associative plural and lexical use of -men. Section 5 concludes the main points argued in this paper, 1) A multiplication relation holds between the numeral in the Spec Number Phrase and the numeral implied by the Classifier head. 2) The plural morpheme -men can be inserted when the result of the multiplication is an imprecise quantity. 3) -Men is a semi-lexical item that can be merged into different syntactic environments to derive different interpretations as long as the merging conditions are satisfied. 3 Properties and previous analyses In this section, I will summarize the properties of -men, including its animacy requirement, definiteness effect, interactions with numerals and classifiers, associative plural use, and lexical formation with kinship terms. Then, I will propose a new descriptive generalization on the distribution of -men which can cover more data. Finally, I will evaluate literature on the syntactic height of plural marking and argue that -men should be merged under the Number head. 3.1 Animacy In Mandarin, -men can express plurality when it combines with human common nouns, as shown in (1). The translation shows that when -men combines with animate common nouns, it tends to have definite effect. This point will be elaborated later. Plants and animals like flowers and cats can take the morpheme -men to express plurality only when they are anthropomorphised, such as in fairy tales, shown in (2). Inanimate nouns do not take -men to express plurality as shown in (3). (1) a. b. xuesheng student/students xuesheng men student men the students 482 (2) (3) FANGHUA ZHENG huaer men zai weixiao flower men Prog smile The flowers are smiling *zhuozi men hen xin table men very new The tables are new. The morpheme -men is also employed to form plural pronouns, as shown in (4). When it combines with the 1st person singular pronoun, the result plural pronoun can denote a meaning either including the listener or excluding the listener, depending on the context. (4) SG PL 1st Wo Wo men (inclusive/exclusive) 2nd Ni Ni men 3rd Ta Ta men Based on the animacy resuirement mentioned above, we tentatively assume that one of the insertion conditions of -men is [+human] feature (or [+animate] feature if plants and animals are included). Note that this is a semantic feature, not a syntactic feature. It is a semantic selection restriction on -men. Typologically, the animacy requirement on plural marking is quite common. Based on the notion of Animacy Hierarchy given in (5) (see Comrie 1989: 185200), Corbett (2000) proposed a constraint of the Animacy Hierarchy on the singularplural distinction described in (6). As we can see, when it comes to the use of -men to distinguish singular and plural, the split is between the animate and the inanimate. (5) (6) Animacy Hierarchy 1st person pronoun > 2nd person pronoun > 3rd person pronoun > kin > human > animate > inanimate Constraint of the Animacy Hierarchy on the singular-plural distinction The singular-plural distinction in a given language must affect a top segment of the Animacy Hierarchy. (Corbett 2000: pp 56) 3.2 Definiteness effect When -men combines with common animate nouns, it will give rise to a definite interpretation. It presupposes the existence of what the noun denotes. The definiteness effect can be tested in the existential sentences illustrated in (7). (7) a. *you xuesheng men zai shuohua have student men Prog talk Intended: There are students talking. b. you xuesheng zai shuohua have student Prog talk There is/are some student(s) talking. Mandarin classifiers and plural morpheme -men 483 As shown in (7a), N-men cannot appear in an existential sentence. In order to express the meaning in (7a), the bare noun is used, and it denotes a general number, that is, the interpretation of the bare noun xuesheng (student) is ambiguous between a singular reading and a plural reading. 3.3 Constraints on distribution In this part, I will introduce the most significant morpho-syntactic restrictions on -men and show that these restrictions are empirically insufficient. The new generalization is in line with the traditional view of -men mentioned in descriptive grammar but might have long been overlooked in formal linguistic approaches. 3.3.1 View 1: -Men is incompatible with classifiers Mandarin is known as a classifier language. When a noun occurs with a numeral, a classifier is often needed between the numeral and the noun to assist the counting process as in (8). (8) san ge xuesheng three CL student three students Unlike English plural morpheme -s, -men is infelicitous in counting constructions, as shown in (9). It is also impossible for -men to occur in a structure with classifiers as shown in (10). (9) *san xuesheng men three student men three students (10) *san ge xuesheng men three CL student men three students In previous literature, there are mainly two ways to explain the ungrammaticality of -men in (10). First, Li (1999) argues that -men is a plural marker like English -s. The syntactic structure given by Li is demonstrated in (11). As we can see, both men and -s are realizations of plural features (Pl) base generated under the Number head. The only difference between Mandarin and English lies in the presence or absence of the Classifier projection. In English, the Classifier projection is absent, thus the plural morpheme -s in Number head can lower down to attach to the NP. In Mandarin, where the Classifier projection is present, the plural morpheme -men cannot lower to NP due to the intervention of the Classifier head. 484 FANGHUA ZHENG (11) In addition, as mentioned earlier, -men is productively attached to pronouns and proper names to express plurality. Li (1999) and C.T. Huang & Li (2009) hold the view that pronouns and proper names are base generated in D head. If -men can be attached to these D elements productively, it is expected that -men is suffixed to elements in D. In the cases where -men is attached to bare common nouns, it is the result of movement from N to D. In other words, -men is suffixed to a common noun only when it is raised from N to D. Li’s analysis has two advantages. First, by saying that -men moves to D, it accounts for the definiteness effect of -men directly. Second, it also explains why N-men is incompatible with classifiers. This is because the Classifier projection prevents N-to-D head movement of the common noun. It violates the Head Movement constraint (Travis 1984). The second approach to account for the incompatible between is Bošković & Hsieh (2013). Different from Li (1999), Bošković & Hsieh argues that -men could be a classifier. The spirit of this view could be traced back to Borer (2005) who proposed that plural marking in English and classifiers in Mandarin are realizations of the same feature, namely the division feature, the function of which is to divide the noun into portions that can be counted. This implies that plural marking and classifiers are in complementary distribution as they compete for the same position. The complimentary distribution of plural marking and classifiers has long been observed (see T’sou 1976, Doetjes 1997). Bošković & Hsieh follow this line of thinking and argue that this is why -men is incompatible with numeral classifier sequence. For them, -men is a classifier that selects non-individuals. They explain the definiteness effect of N-men by following Cheng and Sybesma. Cheng & Sybesma (1999) argue that definiteness in Mandarin is syntactically localized in the classifier projection. The definiteness effect thus follows naturally if -men is base generated in the classifier position. Now, let’s evaluate these two analyses from an empirical point of view. Their starting point largely relies on the generalization that -men is incompatible with classifiers. I will show that the generalization does not hold. Let’s start by looking at (9), repeated here as (12) where classifiers are absent. If we follow Li’s approach, would predict (10) to be grammatical because no head could intervene for N-to-D movement. If we follow Bošković & Hsieh’s approach, classifiers are not there to compete for the same position with -men, thus the struc- Mandarin classifiers and plural morpheme -men 485 ture should be felicitous, contra to the fact. (12) *san xuesheng men three student men three students In addition, Jiang (2017) mentions another two types of empirical data that challenge the two analyses mentioned above. First, -men can occur with individual classifiers when the number is approximate. In (13), ji means ‘a few’ and when it combines with nouns, it requires classifiers as normal numerals do. As we can see, in (13), -men can appear with individual classifiers. (13) ji ge haizi men a.few CL-individual child men a few children Second, -men can occur with group classifiers as shown in (14). Group classifiers like qun or zu means ‘group’ and ‘team’. Group classifiers have quite similar syntactic distribution with individual classifiers. The null hypothesis thus is to treat group classifiers to be generated in the same syntactic position as individual classifiers, namely the Classifier head, with differences in semantic meaning. If -men can co-occur with group classifiers, it is problematic to say that -men occupies the Classifier head. (14) yi qun haizi men one CL-group child men a group of children From the discussion above, we can see that -men is compatible with group classifiers and even individual classifiers when the numeral is not specified. This shows that analyses that starting out with the assumption that -men and classifiers are incompatible misses the point. These analyses are empirically insufficient. 3.3.2 View 2: -Men is incompatible with numerals In languages like Turkish and Hungarian, and plural morpheme cannot co-occur with numerals. This leads us to think whether this is also the case in Mandarin Chinese. First, it is crucial to make it clear what I mean by numeral here. Let’s define it structurally. The numeral I am referring to is the specifier of Number Phrase. It is different from quantity. The distinction between numeral and quantity is important here. The view that -men si incompatible with numerals holds true if we look at (15): (15) *san xuesheng men three student men three students However, if we consider the data including group classifiers, we find that -men can co-occur with numerals like one. Therefore, it is inaccurate to say that -men 486 FANGHUA ZHENG is incompatible with numerals. Example (16) also shows that -men does not agree with the numeral in the specifier of Number Phrase, otherwise -men cannot co-occur with numeral one, which is supposed to be compatible only with singular number. (16) yi qun xuesheng men one CL-group student men a group of students The above discussions show that -men is neither incompatible with classifiers nor numerals. In the following section, I will propose a new constraint on the distribution of -men. 3.3.3 The real constraint on -men From the data above, we can see the view that -men is incompatible with classifiers or numerals both miss the point. This paper holds that the real constraint on the (non)-occurrence of -men is whether the quantity is precise or not. Many native speakers I consulted with shared this intuition and it has been mentioned in traditional descriptive grammar in Mandarin. However, it is not treated seriously in formal approaches. It is particularly important to distinguish quantity and numerals in languages with classifiers because the semantic meaning of classifiers varies. Individual classifiers denote atoms whereas group classifiers do not. This leads to different values of quantity and numerals. In this paper, I argue that quantity is the result of the multiplication of the numeral in Spec Number Phrase and the numeral implied by certain type of classifiers. The (non)-occurrence of -men is sensitive to the result of the multiplication, namely, the quantity. When the result is precise, -men cannot appear. When the result is imprecise, -men can appear. This constraint can account for all the data mentioned so far better than the constraints summarized in previous literature. A good way to test this constraint is the dual classifier in Mandarin. In Mandarin, there is a dual classifier dui (couple) which can be used as a classifier to denote a couple, as shown below. If we follow the constraint proposed above, we could predict that the structure in (17a) would be incompatible with -men. This is born out, as shown in (17b). (17) a. san dui qinglü three CL-couple couple three couples b. *san dui qinglü men three CL-couple couple men three couples The reason that (17b) is ungrammatical is because the quantity is precise, that is, six. The multiplication of the numeral three and the numeral two implied by the dual classifier gives the result six, which is a precise quantity. Since -men can only be merged when the quantity is imprecise, (17b) is ungrammatical. Let’s refer to the numeral at the specifier of Number Phrase as a Multiplier and Mandarin classifiers and plural morpheme -men 487 the numeral implied by the classifier as Multiplicand. The value of the multiplier and the multiplicand can be specified or unspecified. This gives rise to four possibilities of the combination. If we add the possibility of bare nouns, then we have five possibilities in total, as demonstrated in (18). (18) Multiplier (Spec NumP) Specified (1,2,3...) Multiplicand (Spec CLP) Result -men Specified Specified Banned (Individual or dual classifiers) Specified (1,2,3...) Unspecified Unspecified OK (Group classifiers) Unspecified Specified Unspecified OK (ji ‘a few’) (Individual or dual classifiers) Unspecified Specified Unspecified OK (ji ‘a few’) (Group classifiers) None (bare noun) None (bare noun) Unspecified OK This chart correctly captures the distribution of -men in the data mentioned so far. It shows that -men is banned in the structure when the quantity is precise, and the quantity is calculated through a multiplication relation. In this section, we have reviewed two analyses on -men. We showed that they are empirically insufficient. Then we propose a new constraint on the distribution of -men. This constraint relies on a multiplication relation holding between the Number Phrase and the Classifier Phrase. In the next section, we will return to summarize the final property of -men. 3.4 Associative plural use of -men and multifunctionality In the previous section, we mainly looked at data with -men attaching to common nouns to express plurality. However, it can also be attached to pronouns and sometimes proper names to express associative plural meaning. Associative plural constructions usually consist of a noun X, typically a human reference as a representative, and an associative plural marker to express the meaning ‘X and other people associated with X’ (Daniel & Moravcsik 2013). For instance, in (19), when the third person plural pronoun ta-men ‘they’ is used after a proper name Xiaoqiang, it means Xiaoqiang and his associates. (19) Xiaoqiang ta-men lai le Xiaoqiang he-men arrive Perf Xiaoqiang and his associates arrived. Some people also accept (20), where -men is directly attached to the proper name to express associative plural meaning. However, there are also people who found the associative interpretation of (20) rather unnatural (including me). The morpheme -men can be attached to proper names, but a more natural interpretation of proper name -men is a group of people who share the same property. For example, the most natural interpretation of (20) is interpretation 2, where Xiaoqiang-men denotes a 488 FANGHUA ZHENG group of people who share the same property of being named as Xiaoqiang, similar to ‘Johns’ or ‘Toms’ in English. (20) Xiaoqiang-men lai le Xiaoqiang-men arrive Perf ??Interpretation 1: Xiaoqiang and his associates arrived. Interpretation 2: People whose name is Xiaoqiang have arrived. The shared property does not have to be the person’s name. It could be any salient property of the person that the proper name denotes. For instance, in (21), Leifeng is well known for his kind heart, thus we could use Leifeng-men to denote those ordinary people who share similar personalities with Leifeng. This usage is very similar to the cases when -men is attached to common nouns. It is likely that the proper name in (21) is base generated in the N position, just like common nouns, such as student, child, etc., and then being pluralized. It is hard to get the associative plural meaning for examples like (21). Therefore, I will only treat structures involving plural pronouns occurring after proper names in (19) as typical associative plural constructions. (21) Leifeng-men Leifeng (the name of a warm-hearted person)-men Ordinary people who are ready to help others without asking returns Another property of associative plural is that it can be followed by quantity phrase. This is different from N-men. In 3.3, we spent a lot of time discussing the fact that -men cannot co-occur with precise quantity phrase like san ge ren (three people). However, in associative plural constructions, where -men precedes the quantity phrase, -men is obligatory when the numeral is larger than one. (22) Xiaoqiang ta-men san ge ren Xiaoqiang he-men three Cl person Xiaoqiang and another two associates This reminds us of the generalization we reach on the distribution of -men in 3.3.3. The generalization says that -men can only co-occur with imprecise quantity. However, in (22), the quantity being three is precise and -men can appear. In fact, -men is obligatory in (22). One way out of this is to say that what we see in (22) is an appositive structure, the numeral-classifier sequence following the pronoun is a non-restrictive modifier. It provides additional quantity information to the proper name and pronoun it modifies. The syntactic structure could be the one shown in (23), which is a modifying structure. (23) In (23), the NumP2 in the right adjunction is optional. It can be left out when the quantity of the associates is not important. In the left branch, -men can be inserted in under the NumP1 . This does not violate the constraint on -men we proposed earlier because no precise quantity is specified in NumP1 . Note that -men is obligatory in (23) when the quantity denoted by NumP2 is larger one. We could say that the Mandarin classifiers and plural morpheme -men 489 obligatory appearance of -men is a semantic number agreement holding between DP and NumP2 . So far, we have shown that -men can be attached to various types of noun phrases. When it is attached to common nouns and proper names, it denotes plural meaning. When it is attached to pronouns, it denotes associative plural meaning. This is an indication of the multifunctionality of this morpheme -men. To further prove its multifunctionality, I would like to include another piece of data to elaborate on this point. The morpheme -men can be attached to some kinship terms to form new words, as shown in (24). Note it is not productive. The meaning of the formed words is not that transparent. It is more like a lexical process. A piece of evidence to show that this might be a lexical process is that these words are always pronounced with rhotacization, which might be an indication of the word boundary. (24) a. b. c. ge-men-r brother-men-Rhotacization Male best friend jie-men-r Sister-men-Rhotacization Female best friend ye-men-r Grandfather-men-Rhotacization Manhood So far, we have summarized the major properties of -men. We also reviewed two syntactic analyses of -men and pointed out their empirical starting point is to some extent problematic. In addition, we showed that -men can appear in various linguistic environments to express different meaning, such as plural, associative plural, and opaque meaning when it is involved in a lexical formation process. Before we end this section, I’d like to review another essential paper on -men by Jiang (2017), who took into the empirical weakness of the previous accounts into consideration and proposed a unified analysis of -men. After summarizing her analysis, I’ll show that we could make improvement both empirically and conceptually. There is a reason that I review Jiang’s article after reviewing the associative plural use of -men. Jiang examines four types of structures as listed in (25), the specific examples of which have been covered in the previous discussion. Jiang proposed a unified analysis of -men. She argues that -men is an associative plural morpheme generated in little n, which is local to nouns and lower than numerals 490 FANGHUA ZHENG and classifiers. The syntactic structure given by Jiang is like the one in (26b). (25) (26) (27) a. N+Men b. Numeral+CL+group+N+Men c. Numeral Approximation+CL+N+Men d. N/Pronoun/Proper Name+Men+CL a. *san ge xuesheng men three CL student men three students b. menhek,he,tii = λkλY[U khuman |Y| 2 ^ G ^ (k) = Y Jiang argues that -men is an associative plural morpheme which maps a kind to a salient group. In (27), khuman means the noun selected by -men is a human kind. Y is a set of plural individuals and G is a group function mapping a kind to a salient group. The semantic type of -men is hek , hr, tii, and it seeks kind-denoting terms. The result of N-men is thus he, ti. Jiang holds the view that common nouns are regular kind and proper names or pronouns are individual kind. When -men combines with them, the output is type he, ti, a predicate. To further turn the predicate into an argument, Jiang relies on iota operator or generic operator. In the following section, I will point out two issues with this analysis. Jiang explains the ungrammaticality of Num-CL-Noun-men in (26) in the following way. The morpheme -men first combines with the common noun xuesheng (student). The result xuesheng-men denotes a salient group of plural individuals which cannot provide the correct semantics individual classifiers look for, as individual classifiers look for atoms. However, we mentioned earlier that -men can co-occur with individual classifiers when the numeral is imprecise, such as ji (a few), repeated below in (28). In fact, Jiang took this possibility into consideration and assumed a modification structure for it as shown in (28). The numeral phrase is a right adjunction to the N-men phrase. (28) ji ge xuesheng men a.few CL-individual student men Mandarin classifiers and plural morpheme -men 491 a few students (29) This right adjunction analysis looks rather suspicious. For Jiang, when the numerical value is specific, like three, the classifier forms a constituent with the noun. When the numerical value is vague like a few, the classifier forms a constituent with the number. It is not very consistent. I think the new generalization mentioned earlier is better in this aspect. In my generalization on -men, I do not assume two different structures for numeral-classifier-noun phrase. The second point I am taking issue with about Jiang’s analysis is the syntactic position of -men. I think placing -men at little n might be too low. Jiang adopts Mathieu’s (2014) split analysis of plurality and plural typology, the main spirit of which is that there are different types of plurals and each occupies a different functional head, be it Number, Classifier (or Division head in Borer’s (2005) term), or little n. Since Mathieu’s (2014) paper is also of great importance to my analysis, I will review it in detail in the following section. In the meantime, I will show that although I am also adopting Mathieu’s splitting view of plurality, I think -men occupies a higher functional head than little n. 3.5 Syntactic height of -men and Mathieu 2014 Let’s start with the core data Mathieu’s analysis relies on. Mathieu examines the plural of singularized nouns in Arabic. Arabic has a series of collective nouns denoting groups that can be turned into individuals via singulative morphology. Once the collective has been turned into a singulative, the output can further be pluralized. Mathieu provided a list of such words and I selected some to present in (30). (30) Collective Singulative Plural of singulative šajar ‘trees’ šajara(h) ‘a tree’ šajaraat ‘trees’ èajar ‘stones’ èajara(h) ‘a stone’ èajaraat ‘stones’ burtogaal ‘oranges’ burtogaala(h) ‘an orange’ burtogaalaat ‘oranges’ Mathieu assumes that collectives denote kinds and are weakly referential. They do not refer to entities. Following Borer (2005), Mathieu argues that singulatives in Arabic are classifying plural under the Division head, the function of which is to divide and classify. The plural of the singulative is a counting plural realized higher in the structure under the Number head. Once the singulative has been realized 492 FANGHUA ZHENG under the Division head, the counting plural can target the singulative. The tree in (31) illustrates the hierarchy of different plurals Mathieu & Zareikar. (31) In (31), the plural or group meaning denoted by the collectives is realized at little n. The singulative morpheme and English plural morpheme -s are realizations of Division head. The plural of singulative is realized in the Number head (#0 ). If we recall Jiang’s analysis, we may find it is not compatible with Mathieu’s framework in three aspects. First, if -men is the realization of little n, and individual classifiers are realizations of Division head, then we may wonder why individual classifiers in Mandarin cannot divide the N-men as Arabic singulatives do to the collective. Second, Mathieu treats collective nP as kinds rather than predicates. In Jiang’s analysis, -men is a kind-seeking group function and the result of N-men is predicate, rather than kind. This is also different from Mathieu’s framework. Third, Mathieu mentions that collectives in Arabic denote kinds. They are weakly referential and number neutral. They do not refer to uniquely identifiable entities. This is extremely different from the definite effect of N-men and the plurality interpretation it adds to the structure. In the following part, I will elaborate on the third point and show that -men should be analysed higher in the nominal spine. Mathieu points out a possible connection between the hight of the plural and the strength of its referentiality. Mathieu holds the view that classifying plural nominals are weakly referential and has an inclusive reading, whereas counting plural nominals are strongly referential and has an exclusive reading. In other words, the higher the plural is, the stronger its referentiality it becomes. Before we go further, let me explain briefly about the inclusive and exclusive reading of plural nominals. It has been observed that in languages like English, the plural allows both inclusive reading and exclusive reading. The inclusive plural means one or more than one, which includes both singular and plural individuals. The exclusive reading means more than one only, which introduces only plural individuals. For example, in (32), the plural form children could only mean more than one child. However, in (33), the plural form can take the interpretation of ‘one’ or ‘more than one’, which is an instance of inclusive plural reading. (32) (33) I have children. Exclusive reading only: I have more than one child. a. Do you have children? Yes, I have one. Mandarin classifiers and plural morpheme -men b. 493 I didn’t eat apples. Mathieu has noticed a similar contrast of inclusive and exclusive reading in Arabic. He points out that collectives and singulatives are used in inclusive reading environment whereas plural of singulative is only used exclusively. For instance, in (34a), the plural of the singulative šaQraat (strands of hair) is ungrammatical in an inclusive reading environment. Instead, the collective form is used in (34b). (34) a. b. Qindah šaQraat? has-he hair FEM-PL Does he have strands of hair? Qindah š a Qar? has.he haircol Does he have hair? Mathieu provides a series of environment to the test the inclusive and exclusive contrasts and conclude that the plural of singulative (the counting plural in Number head) always get the exclusive plural reading whereas the collective (nP) and the singulative (the dividing plural) always get inclusive plural reading. He also holds the view that there is a connection between the strength of referentiality and the availability of exclusive reading. The counting plural is strongly referential and is possible for exclusive plural reading only. Now, let’s come back to -men again. I would like to argue that -men is only possible with the exclusive reading. As for the including reading environment, Mandarin usually use bare nouns. For instance, in (35), the Mandarin counterpart of (32) and (33), we find that -men is incompatible with an inclusive reading environment. Also, in Mandarin, bare nouns are regarded as number neutral, that is, bare nouns can be interpreted as ‘one’ or ‘more than one’. However, when -men appears with the noun, the noun is interpreted as plural only, namely ‘more than one’. (35) a. ni you haizi ma you have child question particle Do you have children? b. *ni you haizi-men ma you have child-men question particle Do you have children? The strong referentiality and the exclusive plural reading of -men remind us of the counting plural in Arabic. Therefore, it makes more sense to put -men higher in the structure rather than little n as Jiang argued. I would like to add a further thought on Mathieu’s work. It is interesting to think if it is possible that the counting plural in Number head is sensitive to a multiplication relation and the dividing plural is sensitive to an addition relation. If we assume that -men merges at the Number head when its semantic requirement is satisfied, that is, the result of the multiplication of Spec Numeral Phrase and the Spec of Classifier Phrase is an imprecise quantity, we may wonder whether such a multiplication relation holds between the two heads in Arabic or not. Well, it could 494 FANGHUA ZHENG be, but we cannot tell. The reason is that the Division head in Arabic is realized by singulatives, thus the numeral implied by the Division head is always one. The result would be the same regardless of whether we take a multiplication relation or an addition relation between Numeral Phrase and Division Phrase. It is possible that the English type of plural, that is, the classifying plural under Division head, is sensitive to the addition relation, whereas the plural of singulative in Arabic and Mandarin -men in Number head are sensitive to the multiplication relation. 4 Proposal 4.1 Theoretical background: a feature account of grammatical number Harbour’s (2014) theory postulates a small set of primitive features that derive possible number systems and explains impossible ones. It also provides explanations on the interpretation and morphological realization of these features. In what follows, I will summarize Martí’s implementation of Harbour’s grammatical number theory. Martí’s (2020a, 2020b) analysis relies on four primitive features: [±atomic] and [±minimal] to account for the typology of grammatical number and account for the properties of the numeral noun constructions. [±Atomic] is sensitive to atoms vs non-atoms. [±Minimal] is sensitive to whether there is minimal part in the semi-lattice or not. [±minimal] is relative notion, it depends on what else is in the set. These features play a role in morpho-syntax and semantics. When we apply these two sets of features independently, the result is the same. They distinguish the singular vs plural noun phrases. However, it is not the case that [±minimal] always gives the result as [±atomic]. A strong argument for this is the existence of dual. The feature combination gives rise to the following grammatical numbers: (36) a. b. c. d. [+minimal, +atomic]=singular [+minimal, -atomic]=dual [-minimal, -atomic]=plural [-minimal, +atomic]=empty set Feature repetition is also constrained to avoid over-generating grammatical numbers that are not attested. Harbour’s system only allows the feature to be repeated if the value of the feature is not the same. The syntactic structure is as follows: (37) Mandarin classifiers and plural morpheme -men 495 Martí proposed that languages, such as English, are [±atomic] systems. Let’s look at how Martí explains the number system in English. (38) a. J[+atomic][nP boy]K = λx.J[nP boy]K(x) and atom(x) b. J[ atomic][nP boy]K = λx.J[nP boy]K(x) and ¬atom(x) c. #J[+atomic][two card][nP boy]K d. J[ atomic][two card][nP boy]K = λx.J[nP boy]K(x) ^ card(x) = 2 e. J[+atomic][one card][nP boy]K = λx.J[nP boy]K(x) ^ card(x) = 1 f. #J[+atomic][one card][nP boy]K ! boy ! boys ! two boy ! two boys ! one boy ! one boys Martí assumes that numerals are intermediate between the noun and the functional projection holding the grammatical features. In other words, the sequence is grammatical number feature>numeral>NP. (38c) is ungrammatical because when the numeral two combines with the NP boy, it selects all the elements that contain two atoms {ab, ac, bc...} and when this set further combines with [+atomic] feature, the result is an empty set because there is no single atom in that set. Similarly, (38f) is ungrammatical because the numeral one has already selected the atomic elements and form a set {a, b, c...}, and when this set combines with [-atomic] feature, the result is empty set. 496 FANGHUA ZHENG Martí’s framework is important for my analysis below. Martí’s featural system implies the relationship between the number and the feature it can combine with. For instance, [+atomic] feature implies numeral one, [+minimal, -atomic] implies numeral two. In the following part, I will propose that although the numerals implied by these features are not expressed overtly in Mandarin, they can serve as the multiplicand. The numerals generated in the specifier of Number phrase are, on the other hand, multiplier. 4.2 Proposal for -men In the following section, I will propose that individual classifiers in Mandarin are realizations of [+atomic] feature, group classifiers are realizations of [-minimal, -atomic] features, and dual classifiers are realizations of [+minimal, -atomic] features. I will argue that morpheme -men is a semi-lexical item and it is specified with insertion condition [+person, +def, imprecise quantity]. Note that the feature imprecise quantity can be formalized as the following equation: n x m = ? The letter n stands for the numeral in Spec Number Phrase. The letter m stands for the multiplicand implies by the classifier. The question marker ? stands for an imprecise result. Let’s look at again the feature sets Martí uses to derive different numeral values, as shown below. When we look at the individual classifiers, dual classifiers and group classifiers, it maps well the following features. The individual classifiers are realizations of [+atomic] features which pick out sets of atoms. The dual classifiers are realizations of [-atomic, +minimal]. The group classifiers are realizations of [-minimal, -atomic] feature. The tree representation is shown in (40). (39) a. b. c. d. [+atomic]=individual classifier [+minimal, -atomic]=dual classifier [-minimal, -atomic]=plural classifier [-minimal, +atomic]=empty set (40) The number value of the multiplicand is never overtly expressed. It is only implied by the classifiers we use. The multiplier is the numerals used in the numeral noun phrase. The morpheme -men can be merged under the Number head when Mandarin classifiers and plural morpheme -men 497 three conditions are satisfied. First, the nP has semantic person feature. Second, the result of the multiplication between multiplier and multiplicand is imprecise quantity, namely n x m = ?. Third, the insertion of -men could give the structure a definite interpretation. This structure also indicates the difference between English numerals and Mandarin numerals. According to this structure, Mandarin numerals occupy the Multiplier position, whereas English numerals occupy the Multiplicand position. English plural -s is sensitive to the additive plural, whereas Mandarin -men is sensitive to the multiplication plural. In the above discussion, I have accounted for the distribution of -men with individual classifiers and group classifiers. In the following section, I will discuss the associative plural use of -men. I propose an adjunction structure for associative plural construction. There are two reasons for doing so. First, -men can only get association plural interpretation when it is attached to pronouns. Second, when -men is used as an associative plural, it can appear with precise quantity phrase. These two properties are different from the plural use of -men when it is attached to common nouns and proper names. (41) (42) (43) Zhangsan (proper name) the people whose name are Zhangsan xuesheng men (common noun) student men the students Zhangsan ta-men san ge ren (pronoun) Zhangsan he-men three CL person Zhangsan and his two associates. The structure I propose for the associative plural is demonstrated below in (44). The numeral phrase serves as modifiers of the associative plural DP, adding complementary information to the structure, namely specifying the number of people involved. Since the numeral phrase is a modifier, it can be freely omitted. In the left branch, the quantity is not specified, thus -men can be inserted into the structure. The insertion condition of -men can remain the same as previously mentioned. It is desirable to keep the insertion condition the same. (44) Finally, I will explain the last piece of data, namely the lexical use of -men. In 498 FANGHUA ZHENG previous parts, we mentioned that -men can be attached to some kinship terms to form new words although the process is highly limited. The words are repeated as follows: (45) a. b. c. ge-men-r brother-men-Rhotacization Male best friend jie-men-r Sister-men-Rhotacization Female best friend ye-men-r Grandfather-men-Rhotacization Manhood I would like to propose that -men in these words is inserted in little n and first merges with the root. These words do not necessarily convey plural meaning. They are number neutral like bare nouns. I take the little category as the derivational domain. The semantic composition of components is not as transparent as in inflectional domain above nP. Besides, the rhotacization could also be seen as an indication of the word boundary. (46) In this part, I have proposed structures for three types of constructions involving -men, they are: 1) Num CL Nouns -men; 2) Associative plural use of -men; 3) Kinship terms with -men. The morpheme -men can be inserted into the structure when certain conditions are satisfied. First, the quantity expressed by the structure should be approximate and imprecise. Second, -men s-selects for human or animate nouns. Interestingly, as we can see from the above analysis, -men can be inserted into three types of structures and derive different meanings, although the insertion condition remains the same. This is in line with literature on semi-lexical items Cavirani-Pots et al.. Semi-lexical items can be inserted in different syntactic structures to serve various functions. They are semi-lexical and semi functional. This renders the functional structure necessary because the structure will impose some insertion condition for the semi-lexical item. On the other hand, the meaning of the semi-lexical items also imposes a semantic selection constraint for the environment they merge into. These semi-lexical items are not pure functional items that can be merged under the functional head to give rise to the interpretation directly. Mandarin classifiers and plural morpheme -men 499 5 Implications and conclusion In this article, we have seen that -men can co-occur with individual classifiers, group classifiers, and numerals, contra to the constraints proposed in the previous analyses. We have defended the traditional descriptive constraint on the distribution of -men, namely -men can only co-occur with imprecise quantity. The descriptions can be captured if we assume a multiplication relation holds between the Number Phrase and the Classifier Phrase in Mandarin. The morpheme -men can only be inserted in the structure when the result of the multiplication is imprecise. Also, if we follow Mathieu’s (2014) hypothesis on the connection between the referentiality and the height of the plural, we would like to say that -men is merged under the Number head. This also explains the exclusive plural reading of -men. In the meantime, if we accept Mathieu’s view of split plurality, it seems possible that the counting plural in Number head is sensitive to a multiplication relation whereas the classifying plural is sensitive to an addition relation. This may also lead to different semantics in numerals in different languages. For instance, numerals in Mandarin can only merge in Spec Number Phrase positions and numerals in English can merge in Spec Division Phrase. However, such a proposal is highly hypothetical and needs empirical support in future research. Last but not the least, this article treats -men as a semi-lexical item. On the one hand, it has insertion conditions which need to be satisfied. On the other hand, we could see that semi-lexical items can be inserted into the different syntactic environments and various meanings can be derived. The insertion conditions are purely semantic. In the meantime, my analysis shows that a functional category like Number Head is needed in Mandarin. Although Mandarin does not have an English type of plural morpheme that seems purely syntactic, it does have a plural morpheme that can have a clear-cut form-meaning distinction when inserted into the structure. References Acquaviva, P. 2008. Lexical Plurals: A Morphosemantic Approach. Oxford: Oxford University Press. Borer, H. 2005. Structuring Sense, Volume 1: In Name Only. Oxford: Oxford University Press. Bošković, Z., & I. Hsieh. 2013. On word order, binding relations, and plurality in Chinese noun phrases. Studies in Polish Linguistics 8. 173–204. Cavirani-Pots, C., M. D. Belder, & H. Klockmann. 2021. Semi-lexicality: syntax or lexicon? In Proceedings of GLOW 44. Cheng, L., & R. Sybesma. 1999. Bare and not-so-bare nouns and the structure of NP. Linguistic Inquiry 30. 509–542. Comrie, B. 1989. Language Universals and Linguistic Typology: Syntax and Morphology. Oxford: Blackwell. Corbett, G. 2000. Number. Cambridge: Cambridge University Press. C.T. Huang, Y. L., & Y. Li. 2009. The Syntax of Chinese. Cambridge: Cambridge University Press. Daniel, M., & E. Moravcsik. 2013. The Associative Plural. In The World Atlas of Language Structures Online. Leipzig: Max Planck Institute for Evolutionary Anthropology. 500 FANGHUA ZHENG Doetjes, J. 1997. Quantifiers and selection: on the distribution of quantifying expressions in French, Dutch and English. Leiden University dissertation. Harbour, D. 2014. Paucity, abundance and the theory of number. Language 90. 185–229. Jiang, L. 2017. Mandarin associative plural -men and NPs with -men. International Journal of Chinese Linguistics 4. 191–256. Kramer, R. 2016. A split analysis of plurality. Linguistic Inquiry 47. 527–559. Li, Y. 1999. Plurality in a classifier language. Journal of East Asian Linguistics 8. 75–99. Lü, S. 1980. Xiandai Hanyu Babai Ci (800 Words in Chinese). Beijing: Shangwu Press. Martí, L. 2020a. Inclusive plurals and the theory of number. Linguistic Inquiry 51. 37–74. Martí, L. 2020b. Numerals and the theory of number. Semantics and Pragmatics 13. 1–53. Mathieu, E. 2014. Many a plural. In Weak Referentiality, 157–181. Amsterdam: John Benjamins. Mathieu, E., & G. Zareikar. 2015. Measure words, plurality, and cross-linguistic variation. Linguistic Variation 15. 169–200. Travis, L. 1984. Parameters and Effects of Word Order Variation. MIT dissertation. T’sou, B. 1976. The structure of nominal classifier systems. In Austroasiatic studies II, Oceanic Linguistics Special Publication, volume 13, 1215–1247. Honolulu: The University Press of Hawaii. Wiltschko, M. 2008. The syntax of non-inflectional plural marking. Natural Language and Linguistic Theory 26. 639–694. Proceedings of CLS 58 (2022), 501-517 © Chicago Linguistic Society 2023. All rights reserved. 501 The interaction between structure, discourse, and prosody in wh-questions in English Zihui Zhu and An Nguyen Johns Hopkins University 1 Introduction Prosody describes the melodic aspects of human speech in terms of intonation and rhythm, which involves the systematic use of suprasegmental properties such as pitch, amplitude, and duration of speech components (Cole 2015). The present study mainly focuses on prosodic stress, which signals the relative prominence of components within a prosodic unit via combinations of suprasegmental properties (e.g., Liberman & Prince 1977, Pierrehumbert & Hirschberg 1990). While prosody itself is a phonological concept, many recent studies have investigated the relationship between prosody and other aspects of human speech such as syntax and pragmatics. Our study contributes to this discussion through a production experiment testing the prominence of the wh-word in various types of English wh-questions in terms of acoustic correlates of stress. We chose English wh-questions as our target for investigation as recent studies have claimed that English information-seeking wh-questions can vary in their word order (e.g., fronted versus wh-in-situ questions), their wh-word position (front-, mid- or final-position), and discourse restrictions (e.g., prominent versus non-prominent common ground), with remaining questions about the interaction of prosody, syntax, and discourse in this empirical domain. The paper is structured as follows: we start by introducing the wh-variants in English in terms of their syntactic, pragmatic, discourse, and prosodic properties in Section 2. In Section 3, we review the literature on the syntax-discourse-prosody interaction and formulate hypotheses to explain the wh-stress patterns. We then introduce the design of our production experiment in Section 4. Finally, Section 5 concludes the paper. 2 Wh-variants in English In this paper, we consider a wh-question information-seeking if it asks for novel information that has not been previously mentioned in discourse. English allows for both fronted wh-questions (FQs) and wh-in-situ as information-seeking questions, though wh-in-situ questions are marked and typically used only in certain contexts, such as quiz shows (Comyn 2013), legal questioning (Pires & Taylor 2007), classroom settings, and child-directed speech (Nguyen & Legendre 2022). 1 Thanks to Géraldine Legendre, Jane Li, Colin Wilson, and Angela Xu for providing extremely helpful feedback on the project. We also appreciate the comments and criticisms from the anonymous reviewers and audiences at the 2022 CLS meeting. 502 ZHU AND NGUYEN We follow Nguyen & Legendre (2022) to call information-seeking wh-in-situ Probe questions (PQs) to separate them from wh-in-situ echo questions (EQs). EQs are typically used to request repetition or clarification of an immediately preceding utterance instead of new information. The difference between information-seeking and repetition-seeking is illustrated below. (1) a. (information-seeking question) [A teacher is showing the class images of technology used in the 90s.] A: (FQ) Look, what is this? / (PQ) Look, this is what? B: A computer. B’: I don’t know. b. (repetition-seeking question) A: That’s a computer. B: (EQ) (That’s) a WHAT? A: a computer. A’: #I don’t know. Among the three types of wh-questions, FQs are the unmarked variant and don’t have any context restrictions. An FQ can occur out of the blue (Biezma 2020) and have broad focus consisting of the entire sentence. PQs, on the other hand, have stricter context requirements and usually cannot be out of the blue (Pires & Taylor 2007, Biezma 2020), as illustrated by examples (2a-b) from Biezma (2020). PQs typically need to satisfy discourse givenness on the non-wh portion, which can be provided by either the discourse context or extra-linguistic information in addition to preceding linguistic utterances. In example (3) taken from Biezma (2020), the non-wh portion is considered given by the extra-linguistic context because the interlocutors share the immediate environment, and it is natural for B’s friend to assume that the pronoun these refers to the dishes that B is holding. (2) [B stops a random pedestrian on the street and says.] a. (FQ) B: Excuse me, where can I buy an Italian newspaper? b. (PQ) B’: #Excuse me, I can buy an Italian newspaper where? (3) [B is helping to tidy up after dinner at her friend’s house and enters the kitchen carrying the dishes.] B: These go where? When the discourse requirements are satisfied, PQs can typically be used interchangeably with FQs. Both variants can be used to elicit the same information and have the same set of answers. In contrast to information-seeking wh-questions, repetition-seeking EQs can only appear in response to the preceding utterance (Banfield 1982). The strict context requirement, plus the specific purpose of EQs as a request for clarification or repetition, leads to a strong presupposition that the addressee knows the answer and can provide it when asked. The response to an EQ must be the original utterance, a synonym of the original utterance, or at the very least, a description that is co-referential with the original utterance (Blakemore 1994). It is infelicitous for Structure, discourse, and prosody in wh-questions 503 the addressee to answer “I don’t know” to an EQ, as shown in (1b) (Nguyen & Legendre 2022). Finally, in terms of prosody, many studies have characterized EQs as having a distinct prosody consisting of heavy stress on the wh-word and an overall rising pitch contour (Authier 1993) characterized by an L+H∗ pitch accent and an HH% boundary tone (Bartels 1999, Artstein 2002). In contrast, PQs are said to lack such “focal stress” on their wh-words (Pires & Taylor 2007). Both PQs and FQs display a “falling declarative pattern” and the fronted wh-word does not require an accent (Bartels 1999, Nguyen & Legendre 2022). However, previous comparative work on the prosody of EQs and PQs may be lacking as they typically focus on wh-in-situ questions with the wh-word at the sentence-final position, ignoring those with the wh-word in the mid-sentence position. We will return to this point in Section 4. 3 The syntax-discourse-prosody interaction In this section, we briefly review the literature on the three-way interaction among syntax, discourse, and prosody. We start by summarizing theories on the syntaxprosody and the discourse-prosody interfaces in Sections 3.1 and 3.2, respectively. We then narrow down the focus to the interaction in wh-questions in Section 3.3 and formulate hypotheses on the relationship between various syntactic and pragmatic factors and wh-word stress. 3.1 Prosody and syntax Syntax provides the necessary structural information for prosody to realize. For example, to assign stress on the phrasal level, we first need access to syntactic constituencies to determine which words form a phrase. There are two main theoretical lines on the relationship between syntactic structure and application domains of phrasal phonological processes (see Bennett & Elfner 2019 for a review). Direct reference theories argue for the direct influence of sentence structure on prosodic patterns with no mediating factors (e.g., Rotenberg 1978, Cinque 1993, Newell & Piggott 2014). Under this view, phrasal phonological phenomena are defined solely in terms of syntactic relations such as c-command (Kaisse 1985) or sisterhoods (Bennett & Elfner 2019). Indirect reference theories, on the other hand, assume that syntax and prosody are mediated by the prosodic structure, which is an intermediate level of representation derived from syntax (e.g., Selkirk 1978, Beckman 1986). This mediating level consists of a hierarchy of prosodic categories including syllables, phonological words, intonational phrases, etc. These abstract prosodic constituents provide the basis for metrical stress theory (Liberman 1975, Liberman & Prince 1977) and the autosegmental-metrical theory of intonation (Ladd 2008) because this hierarchy defines various levels of locations to assign suprasegmental features and realize different levels of stress and pitch accents. The prosodic hierarchy remains the most widely accepted framework in the syntax-prosody interface and we will discuss our operational definition of prosodic stress under this framework in Section 3.3. Because of the foundational role of syntax, prosody is often conceptualized as a complementary factor in terms of establishing sentence structure (Dahan 2015). For 504 ZHU AND NGUYEN example, prosody can indicate syntactic phrase boundaries (Section 3.2, Cole 2015) and help reduce syntactic ambiguity (e.g., Price et al. 1991). The effect of prosody on syntactic structures, on the other hand, is less studied but has been receiving growing attention (Anttila 2016). The prosodic size of a constituent can affect its placement within an utterance and thus the surface word order. Serbo-Croatian, for example, only allows prosodically large phrases such as de Žaneiru ‘in Rio de Janeiro’ to be fronted and topicalized but not the prosodically smaller ones such as u Riju ‘in Rio’ (Zec & Inkelas 1990). Another prosodic factor with similar effects on syntax is metrical prominence, which results from placing a word in a strong metrical position. For instance, Szendrői (2003) observed that contrastively focused elements in Hungarian are fronted and receive extra metrical stress compared to the unfocused ones. Assuming leftmost main metrical stress in Hungarian, Szendrői explained that the focused unit needs to be fronted to the left periphery to pick up the main stress and satisfy the Stress–Focus Correspondence Principle (Reinhart 1995). These examples suggest that prosody can be a licensing factor for noncanonical surface word orders. To summarize, literature on the syntax-prosody interface suggests that there exist bidirectional influences between syntactic structure and prosodic properties of speech constituents. Syntax at least partially establishes the prosodic structure on which phrasal prosodic patterns realize. Prosodic properties such as prosodic size and metrical prominence can influence surface word order. 3.2 Prosody and discourse Prosody has a “leading role” in conveying the discourse meaning of an utterance, contrary to its supportive role in establishing the segmental structure (Dahan 2015). Prosody bears pragmatic functions and can mark a sentence in terms of its illocutionary force, e.g., question, confirmation, request, etc. Each type of utterance can be associated with distinct pitch contours resulting from the complex interaction between intonational features (Cole 2015). For example, (4) is a statement with the canonical intonational contour for declaratives in American English. The same declarative syntactic structure can denote a question when the phrasal tone changes from L- to H- (Pierrehumbert & Hirschberg 1990). (4) You deliberately deleted my files. H∗ H∗ H∗ L-L% Another aspect of discourse functions closely related to prosody is information structure, or the division of an utterance into discourse-new (focus) or discoursegiven (background) information. In English, a focused constituent is marked by assigned phrasal prominence and associated pitch accents rather than morphosyntactic devices (e.g., Brown 1983, Rooth 1992). Pierrehumbert & Hirschberg (1990) claimed that pitch accents are associated with specific information structures. For example, H∗ usually conveys discourse-new information and marks the focus. In (5), B introduces the new referent, Bob, to A by placing a rising pitch accent on Bob. L∗ , on the other hand, labels existing beliefs and is associated with discourse givenness. In (6), because B knows that A is already aware of his birthday wishes, B marks the desired item with L∗ indicating that it is part of their mutual beliefs. Structure, discourse, and prosody in wh-questions 505 (5) A: Whom did Sarah diss? B: Sarah dissed Bob. (Watson et al. 2008) H∗ (6) [A asks B to provide a list his birthday wishes. A remembers that B mentioned a Pavoni espresso machine before but just wants to double check.] (Pierrehumbert & Hirschberg 1990) A: What would you want for your birthday? B: Well, I’d like a Pavoni... L∗ However, such an implicit one-to-one correspondence between prosodic forms and discourse functions has been challenged by recent experimental results. For instance, Roettger et al. (2019) found “low and inconsistent performance” in such form-and-function-mapping tasks in a perception experiment when participants were only given prosodic cues without further discourse information. According to the authors, the suboptimal though often above-chance performance reflects that prosodyinformation structure mapping is “systematic but probabilistic” in nature. Overall, literature on the syntax-prosody and pragmatics-prosody interfaces suggests a complex interaction among the three. Information structure often contributes to such complexity due to its association with specific prosodic and syntactic cues. 3.3 The syntax-discourse-prosody interaction in wh-questions The present study characterizes wh-question prosody in terms of wh-phrasal prominence realized through prosodic stress. Previously in Section 3.1, we introduced several indirect reference theories, which all assume the hierarchical prosodic structure. According to the metrical stress theory, metrical prominence results from placing a word in a strong metrical position and is associated with increased duration and intensity (e.g., Fry 1955, Beckman 1986). Based on the closely related autosegmental-metrical model of intonation, intonational prominence should coincide with metrical prominence because the stressed elements are potential “anchors” for pitch accents (Beckman & Pierrehumbert 1986, see Bocci et al. 2021 for experimental evidence). Therefore, we should expect an expanded pitch range (intonational prominence) and longer duration (metrical prominence) as possible acoustic correlates of linguistic stress. In the subsections below, we formulate specific hypotheses on the relationship between three non-phonological constructs and wh-word stress. 3.3.1 Intonational prominence and movement of the wh-word Fronted wh-questions have traditionally been analyzed in terms of [+Q] and [+WH] features on C (Cheng 1991, Rizzi 1996, Adger 2003). These encode interrogative illocutionary force and operator-status of wh-phrases, respectively. In informationseeking questions, the [+Q] feature on C triggers overt T to C head movement, and the [+WH] feature triggers overt phrasal wh-movement to SpecCP. For wh-in-situ to be acceptable, it has been proposed that intonational prominence comes into play. In particular, Carnie (2006) proposed that the EQ complementizer has a special [+INT] feature, which “instructs” the phonology to render a 506 ZHU AND NGUYEN rising pitch contour. Artstein (2002) alternatively claimed that the rising pitch accent marks focus on the wh-word, which licenses wh-in-situ because wh-focusing is an alternative strategy to wh-fronting for realizing a “question denotation”. This relationship between syntactic derivation and focus prosodic stress is further supported by Bocci et al. (2021). Through their production experiments, Bocci et al. found that the main stress and Nuclear Pitch Accent both fall on the lexical verb in direct wh-questions in Italian. To account for the fact that the stressed lexical verb does not have a focus reading itself, the authors proposed that the lexical verb acquires the [focus] feature from the wh-phrase as an intervening head on the whmovement path. It is then assigned with extra stress as the rightmost element within the syntactic unit carrying the [focus] feature. According to the autosegmental-metrical model of intonation, intonational prominence entails metrical prominence because pitch accents only occur in strong metrical positions. Therefore, pitch accents on the wh-words of PQs and EQs but not on FQs imply that the former is no less prominent metrically than the latter. In summary, if prosodic stress on wh-words licenses the non-canonical in-situ word order, we would predict that 1. The in-situ variants, EQs and PQs, have comparable wh-phrasal prominence both intonationally and metrically and hence similar length and (positive) pitch change. 2. The wh-word in FQs is less stressed intonationally and thus has a smaller pitch change than those in EQs and PQs. It also receives no more metrical prominence and is thus no longer than EQ and PQ wh-words. 3.3.2 Non-canonical intonation signals non-canonical wh-question pragmatics Hedberg et al. (2010) analyzed the nuclear contour associated with each pragmatic class of wh-questions. Hedberg et al. suggested that EQs fall under the “clarifying” category for their non-information-seeking nature, which is predominantly associated with a rising nuclear contour. In contrast, the information-seeking variants exhibit a variety of falling patterns. Building on their corpus analysis results, Hedberg & Sosa (2011) further proposed that non-canonical question intonations convey non-inquisitiveness. Under the framework of Inquisitive Semantics (Groenendijk & Roelofsen 2009), an inquisitive utterance opens up the possibility of updating the discourse common ground (i.e., information-seeking, as defined in Section 2). Canonical wh-questions are inquisitive because they elicit an open set of responses that could be discoursenew. Contrarily, EQs merely ask for the repetition of the previous utterance and are thus non-inquisitive. The non-canonical pitch contour of EQs helps convey such a non-canonical pragmatic function. Note that the corpus results in Hedberg et al. (2010) do not necessarily speak to wh-word stress because nuclear pitch contour contains the final pitch accent and the boundary tone in an intonational phrase and an EQ wh-word need not be in the sentence-final position. Nonetheless, we can extend the inquisitiveness-based theory. In Section 2, we mentioned that the wh-word in canonical informationseeking FQs are unaccented, which should also apply to PQs given their similar Structure, discourse, and prosody in wh-questions 507 pragmatics. A rising pitch accent can then be put on the EQ wh-word to signal the non-canonical pragmatic function, resulting in the following predictions: 1. Because EQs are not information-seeking (i.e., EQs do not have the canonical function of wh-questions), their wh-word is intonationally stressed with a rising pitch accent and has positive a pitch change. 2. Because PQs and FQs are information-seeking, their wh-word are not intonationally stressed. As intonational prominence implies metrical prominence, wh-words on PQs and FQs are not more prominent metrically or longer than that in EQs. 3.3.3 Information structure: discourse givenness drives optimization for surface word order and main stress placement Hamlaoui (2011) observed that in Francilian French (the variety spoken in Paris and its surrounding areas), in-situ and fronted information-seeking questions differ in terms of their information structure and main (metrical) stress placement. An FQ constitutes a broad focus set, while in-situ ones have a narrow focus on the wh-word only. While both wh-words are “inherently focused”, in-situ questions require their non-wh portion to be discourse-given. Furthermore, in-situ questions tend to have the sentence main stress on the wh-word, while FQs have the main stress at the canonical sentence-final position. Adopting an Optimality Theory approach (Prince & Smolensky 2004), Hamlaoui presents a simultaneous optimization of syntax and prosody given the question information structure. When the non-wh portion is discourse-new (Tableau 3, Hamlaoui 2011), in-situ candidates fail to stress all focused elements and are eliminated. The fronted one with main stress on the wh-phrase is also ruled out for violating the rightward metrical head alignment constraint, which makes the fronted option with rightmost main stress optimal. When the non-wh portion is given (Tableau 4, Hamlaoui 2011), the fronted variants with main stress on the right edge or on the lexical verb violate the high-ranked DISTRESS - GIVEN constraint, so does the in-situ candidate with rightmost stress. This leaves the in-situ variant with main stress on the wh-word as the optimal candidate. Given that (1) English and French both allow fronted and in-situ informationseeking questions, (2) both have discourse constraints (i.e., the non-wh portion is given) for information-seeking wh-in-situ, and (3) both languages have canonical rightmost main stress, we attempt to extend Hamlaoui’s proposal on French to English. An English-adapted version of Hamlaoui’s proposal will predict: 1. The PQ wh-word is metrically more prominent and thus longer than the FQ one. Because the latter is fronted, it does not receive the rightmost main stress in FQs and is thus not as prominent metrically as the former, which does carry the strongest metrical stress within a sentence. 2. Because EQs also have to satisfy discourse-givenness, the wh-word in an EQ should be as metrically stressed and thus similar in duration as the wh-word in a PQ. 3. The wh-words in PQs and EQs are at least as prominent intonationally as that in FQs. This is because the PQ and EQ wh-words are placed in metrically 508 ZHU AND NGUYEN strong positions and are eligible for pitch accents, whereas the FQ wh-word is metrically weaker and should not be accented. 4 Experiment In Section 3.3, we formulated predictions regarding three non-phonological conditioning factors of wh-word stress based on previous theories. While the first two hypotheses make direct predictions about intonational prominence and are thus exclusive, the third one is primarily on metrical prominence and could be compatible with either intonational claim. These hypotheses all have implications for the other aspect of stress because the autosegmental-metrical theory assumes that pitch accents can only appear on strong metrical positions. We next introduce our experimental setup to test these three sets of predictions. 4.1 Methods 4.1.1 Participants 20 adult native English speakers with North American accents (M age = . years, SD = 14.61 years, range = 18−69 years) participated in the study through Prolific. One participant was excluded from the study due to their failure to pass the attention checks, leaving us with a total of 19 data points. 4.1.2 Materials and design The study included a total of 33 trials, consisting of 27 targets and 6 fillers, divided into three equal groups of what-, where-, and who-questions. Each trial included a context and a conversation presented to participants in both text and audio forms. The conversation always included at least one target or filler question. The fillers were indirect wh-questions. The 27 targets consisted of 9 FQs, 9 PQs, and 9 EQs grouped into 9 trios. The contexts within each trio were controlled to be roughly similar in length and content (see (7a-c) for an example of a trio). All audio clips were created using Amazon Polly, a natural-sounding human speech synthesizer. (7) Examples of target trials a. (PQ) [Michael and his mom are talking about his friend from childhood. She asks him whether he remembers his friend’s family.] Michael: I remember that he has a baby sister. Mom: And her name is what? b. (FQ) [Michael is telling his mom about his friend’s baby sister. The little girl was born just the day before.] Michael: My friend has a new baby sister. Mom: What is her name? c. (EQ) [Michael is telling his mom about his friend’s baby sister. But his mom is distracted and does not hear him.] Michael: My friend has a new baby sister named Katherine. Mom: Her name is what? Structure, discourse, and prosody in wh-questions 509 As previously mentioned in Section 2, most existing experimental stimuli and corpora data containing wh-in-situ have sentence-final wh-words, which made it impossible to tell whether the differences between EQs and PQs arise from the stress on the wh-word or from the sentence-final prosody. To address this concern, onethird of all in-situ stimuli had a wh-word at the rightmost position, as in (7), while the rest involved a wh-word in the middle of the sentence (e.g., Mario made what for you today?). 4.1.3 Procedure Each participant went through all 33 trials in a randomized order. For each trial, they were asked to record themselves saying the target sentences verbatim as naturally as possible. The experiment also included multiple-choice attention check questions to ensure that participants paid attention to the context of each trial, which was necessary for them to get the right prosody. Each participant saw 4 attention check questions (randomly selected from a pool of 12 questions). Participants who failed to answer at least 3 questions correctly were excluded from data analysis. The example below is an attention check question on the content in (7a) with (b) being the correct answer: (8) Who were Michael and his mom talking about? A. Michael’s teacher. B. Michael’s childhood friend. C. Michael’s sister. 4.2 Results 4.2.1 Comparing wh-in-situ against fronted wh-questions From the recordings collected, a total of 602 target sentences were analyzed. 25 additional recordings were excluded due to distorted voice (3 recordings), blank audio (10 recordings), noisy background (1 recording), and wrong content where participants failed to precisely produce the target sentence (11 recordings). The target sentences were forced-aligned using the Montreal Forced Aligner (McAuliffe et al. 2021). The phonetic features were extracted using Praat (Boersma & Weenink 2021) and the Parselmouth library in Python (Jadoul et al. 2018), including F0 characteristics and the duration of the wh-word. See examples of aligned pitch tracks in Figure 1. We used ∆ F0 to capture the variation in pitch over a specific interval, which is measured by the difference between F0 at the offset and F0 at the onset of the interval of interest. Higher ∆ F0 suggests a larger pitch range and thus more metrical prominence. Phonetic features including ∆F0 and duration across the three types of wh-questions were visualized in Figure 2. In general, results showed that the wh-word ∆F0 of FQs and PQs were not significantly different (M (SD)FQ = .(.) Hz, M (SD)PQ = .(.) Hz, t() = −., p = .). In contrast, the wh-word ∆F0 of FQs significantly differed from that of EQs (M (SD)EQ = .(.) Hz, t() = −., p < .). In terms of wh-word duration, the wh-word in FQs had the shortest duration (M = .22 s, SD = .07 s), followed by EQs (M = .33 s, SD = .10 s) and finally PQs 510 ZHU AND NGUYEN Figure 1: Example pitch tracks: (1a) an EQ with sentence-final wh-word; (1b) an EQ with sentence-medial wh-word; (1c) a PQ with sentence-final wh-word; (1d) an PQ with sentence-medial wh-word; (1e) an FQ. (M = .37 s, SD = .12 s). The duration of all three question types significantly differed from one another (ps < .01). A more detailed analysis comparing the wh-word of FQs against the wh-word of wh-in-situ separated based on its position (mid-sentence versus final-sentence) is reported in Appendix A. 4.2.2 Analysis of wh-in-situ questions To account for the potential influence of sentence-final prosody on wh-word phonetics, we compared the phonetic features of the in-situ wh-words in different positions within the question. A linear mixed effect model was built using the lme4 package in R (Bates et al. 2015). The model included the Phonetic features as dependent Structure, discourse, and prosody in wh-questions 511 Figure 2: Pitch slope and duration of the wh-word in each type of wh-question. variables, Question type (EQ, PQ) and Wh-word position (mid, end) as the fixed effects, and Participant and Trial as random effects. The R formula was: (9) Phonetic feature ∼ QuestionType + WhPosition + QuestionType:WhPosition + (1 | Participant) + (1 | Trial) The regression outputs are summarized in Table 1. Question type and Wh-word position were both significant fixed effects for both phonetic features, suggesting that the wh-word pitch slope as well as duration in a wh-in-situ question differed depending on the type of questions (information-seeking versus repetition-seeking) as well as its position in the sentence. There were no significant interaction effects for either phonetic feature. β SE p Response variable: ∆F0 Intercept .74 .21 < .001∗∗∗ QuestionType (PQ) −.86 .16 < .001∗∗∗ WhPosition (mid) −.62 .24 .01∗ QuestionType:WhPosition (PQ:mid) .28 .20 .16 Response variable: wh-word duration Intercept .25 .27 .35 QuestionType (PQ) .36 .14 .01∗ WhPosition (mid) −.60 .30 .05∗ QuestionType:WhPosition (PQ:mid) −.11 .17 .53 Table 1: Outputs of linear mixed effect model: Effects of question type (PQ vs. EQ) and wh-word position (sentence-middle vs. sentence-final) on wh-word phonetics (∆F0 and wh-word duration). A closer look at the phonetic features show that the wh-words in the sentencemiddle position also tended to have smaller ∆F0 and shorter wh-word duration than the wh-word in the sentence-final position. When the wh-word was at the 512 ZHU AND NGUYEN sentence-final position, it received a sentence-final stress that prolonged its duration and increases its pitch contour. At the same time, PQs tended to have smaller ∆F0 and longer wh-word duration than EQs regardless of the wh-word position (Figure 3). Thus, it is safe to conclude that the two types of wh-in-situ questions are different, and such difference is likely induced by the question function, i.e., information-seeking versus repetition-seeking. Figure 3: Pitch slope of the in-situ wh-word its position within a sentence: (3a) sentence-middle; (3b) sentence-final. 4.2.3 Summary of the results In general, the three wh-variants had different phonetic realizations on their whwords. Wh-words in PQs and FQs had comparable pitch contours (indicated by non-significantly different ∆F0), while wh-words in EQs had a significantly higher rising pitch pattern (indicated by larger ∆F0) compared to the other two question types. All pairwise differences in wh-word duration were significant. There was a main effect of in-situ wh-word position, with the sentence-final position being associated with longer wh-word duration and higher ∆F0. We found no interaction effects between question type and in-situ wh-word position on wh-word phonetics. The main effect of question type persisted even after separating the in-situ whwords by their positions. 4.3 Discussion In section 3.3, three hypotheses were formulated based on previous studies entertaining different aspects of the interaction between prosody, information structure, pragmatics, and surface word order. Let us revisit each hypothesis in light of the experimental data. The first hypothesis (see Section 3.3.1) predicts that in-situ wh-words carry more intonational stress than the fronted ones to license the non-canonical word Structure, discourse, and prosody in wh-questions 513 order. This contradicts our experimental data as the pitch contours of in-situ PQs were similar to canonical FQs but different from in-situ EQs. This suggests that intonational prominence is not a licensing factor for wh-in-situ, at least in English. The second hypothesis (see Section 3.3.2) makes a different claim about the role of question intonation in terms of pragmatic functions. Our results support the main intonational prediction, given that the two information-seeking variants (FQs and PQs) had comparable wh-word intonation and differed from the repetition-seeking EQs. Its secondary predictions on metrical prominence based on the autosegmentalmetrical theory of intonation only received partial support. While EQ wh-words were significantly more prominent metrically than the FQ ones, as predicted, they were metrically weaker than the PQ ones, against the prediction. This indicates that the high-rising pitch accent on a EQ wh-word primarily conveys its non-canonical question pragmatics, i.e., repetition-seeking. The third hypothesis (see Section 3.3.3) argues for the conditioning role of information structure on surface word order and metrical prominence assignment. Its primary prediction on metrical stress patterns matched our data because wh-words in narrowly focused questions (EQs and PQs) were more prominent metrically than those with broader focus (FQs). However, it is still not clear why PQs’ wh-words were metrically stronger than the EQ ones. Our experiment thus partially supports the claim that information structure drives the simultaneous optimization of metrical stress placement and question word order. The significant difference in metrical stress on the wh-words between EQs and PQs was unexpected because the in-situ wh-words occupied approximately the same metrical positions within their respective sentence, given that the experimental stimuli were near-minimal pairs. One potential explanation is that participants were surprised or felt unsure about using PQs, a marked variant, to ask for new information instead of using fronted wh-questions (the unmarked/default variant). Previous studies have identified the correlation between prolonged utterance duration and speaker’s hesitation. The specific acoustic correlates include pre-speech delay (Smith & Clark 1993), more frequent and lengthened filled pauses (e.g., um, uh) (Barr 2003, Gibbon & Tseng 2006), and lengthened phonemes (“local reduction in speech rate”, Betz & Wagner 2016). These authors claimed that the overall lengthening effect reflects the increased cognitive load in discourse planning when experiencing uncertainty. In our case, the non-default PQ targets might have disrupted participants’ original plan of using the default FQs to request new information. If this were true, we would expect longer overall durations for PQs than EQs in their non-wh portion. To test this claim, we conducted an utterance duration analysis on EQs and PQs in contexts where the target sentences contained the same set of words (e.g. We should visit who for an eye exam?) or were minimally different (e.g. We saw what in the mall? vs. You saw what in the mall?). This analysis showed that PQs had significantly longer durations than EQs, excluding the wh-word (M EQ = . s, SDEQ = . s, M PQ = . s, SDPQ = . s, t() = −., p = .). This indicates that the slightly longer wh-duration in PQs than in EQs can arise from non-linguistic factors such as hesitation in discourse planning and thus does not necessarily contradict our hypotheses. The overall results of our production experiment along with previous literature suggest a complex interaction between syntax, pragmatics, and prosody. Based on 514 ZHU AND NGUYEN indirect reference theories, the metrical structure of a sentence is derived from its syntactic structure. These two factors (metrical prominence assignment and surface word order) can be simultaneously influenced by information structure, as proposed by Hamlaoui (2011). This receives support from the metrically stronger wh-words in narrowly focused questions than the broadly focused ones in our experiment. According to the autosegmental-metrical model of intonation, the metrical structure of an utterance establishes the basis of its intonational pattern by restricting potential sites for pitch accent assignment. Pragmatic functions can determine the actual realization of pitch accent, as predicted by the inquisitiveness-based model of question intonation in Hedberg & Sosa (2011). This is also supported by our experiment because information-seeking and repitition-seeking questions are different in both their question pragmatics and intonational prominence on the wh-word. 5 Conclusion The present study investigated various claims regarding the syntax-pragmaticsprosody interface through a speech production experiment of wh-variants in English. We compared their prosodic patterns by analyzing the acoustic correlates of stress on the wh-words and were able to separate the effect of sentence-final stress. The wh-words in repetition-seeking EQs were intonationally more prominent than the wh-words in information-seeking questions (PQs and FQs). Questions with a narrower focus set (PQs and EQs) had metrically more prominent wh-words than questions with a broader focus set (FQs). The wh-word phonetics provided evidence 1) for the use of non-canonical intonation to license non-canonical pragmatic function, 2) for information structure as a conditioning factor for surface word order and metrical prominence, and 3) against the use of exaggerated prosody to license non-canonical word order. References Adger, D. 2003. Core syntax: a minimalist approach. Oxford: Oxford University Press. Anttila, A. 2016. Phonological effects on syntactic variation. Annual Review of Linguistics 2. 115–137. Artstein, R. 2002. A focus semantics for echo questions. In Workshop on information structure in context, volume 98, p. 107. IMS, University of Stuttgart. Authier, J.-M. 1993. Nonquantificational wh and weakest crossover. Linguistic Inquiry 24. 161– 168. Banfield, A. 1982. Unspeakable sentences: Narration and representation in the language of fiction. Boston, MA: Routledge & Kegan Paul., 1st edition. Barr, D. J. 2003. Paralinguistic correlates of conceptual structure. Psychonomic Bulletin & Review 10. 462–467. Bartels, C. 1999. The Intonation of English Statements and Questions. New York: Routledge, 1st edition. Bates, D., M. Mächler, B. Bolker, & S. Walker. 2015. Fitting linear mixed-effects models using lme4. Journal of Statistical Software 67. 1–48. Beckman, M. E. 1986. Stress and Non-Stress Accent. Dordrecht, the Netherlands: Foris Publications. Structure, discourse, and prosody in wh-questions 515 Beckman, M. E., & J. B. Pierrehumbert. 1986. Intonational structure in Japanese and English. Phonology 3. 255–309. Bennett, R., & E. Elfner. 2019. The syntax–prosody interface. Annual Review of Linguistics 5. 151–171. Betz, S., & P. Wagner. 2016. Disfluent lengthening in spontaneous speech. In Elektronische Sprachsignalverarbeitung (ESSV) 2016, ed. by O. Jokisch. Dresden: TUD Press. Biezma, M. 2020. Non-informative assertions: The case of non-optional wh-in-situ. Semantics and Pragmatics 13. 18. Blakemore, D. 1994. Echo questions: A pragmatic account. Lingua 94. 197–211. Bocci, G., V. Bianchi, & S. Cruschina. 2021. Focus in wh-questions. Natural Language & Linguistic Theory 39. 405–455. Boersma, P., & D. Weenink. 2021. Praat: doing phonetics by computer (Version 6.1.38). http://www.praat.org/. Retrieved 02 January 2021. Brown, G. 1983. Prosodic structure and the given/new distinction. In Prosody: Models and Measurements, ed. by A. Cutler & D. R. Ladd, 67–77. Berlin, Heidelberg: Springer. Carnie, A. 2006. Syntax: a generative introduction. Malden, MA: Wiley-Blackwell, 2nd edition. Cheng, L. L. S. 1991. On the Typology of Wh-questions. Cambridge, MA: Massachusetts Institute of Technology dissertation. Cinque, G. 1993. A null theory of phrase and compound stress. Linguistic Inquiry 24. 239–297. Cole, J. 2015. Prosody in context: a review. Language, Cognition and Neuroscience 30. 1–31. Comyn, M. 2013. Wh-in situ is accompanied by which formal features? master’s thesis, Ghent University, Ghent, Belgium. Dahan, D. 2015. Prosody and language comprehension. WIREs Cognitive Science 6. 441–452. Fry, D. B. 1955. Duration and intensity as physical correlates of linguistic stress. The Journal of the Acoustical Society of America 27. 765–768. Gibbon, D., & S.-C. Tseng. 2006. Discourse functions of duration in Mandarin: resource design and implementation. In Proc. of the Fifth International Conference on Language Resources and Evaluation (LREC’06), Genoa, Italy. European Language Resources Association (ELRA). Groenendijk, J., & F. Roelofsen. 2009. Inquisitive semantics and pragmatics. In Workshop on Language, Communication and Rational Agency. Hamlaoui, F. 2011. On the role of phonology and discourse in Francilian French wh-questions. Journal of Linguistics 47. 129–162. Hedberg, N., & J. M. Sosa. 2011. A unified account of the meaning of English questions with non-canonical intonation. In International Seminar on Prosodic Interfaces. Jawaharlal Nehru University, New Delhi, India. Hedberg, N., J. M. Sosa, E. Görgülü, & M. Mameni. 2010. The prosody and meaning of whquestions in American English. In Proc. Speech Prosody 2010, Chicago, IL. Jadoul, Y., B. Thompson, & B. de Boer. 2018. Introducing Parselmouth: A Python interface to Praat. Journal of Phonetics 71. 1–15. Kaisse, E. M. 1985. Connected speech: the interaction of syntax and phonology. Orlando: Academic Press. Ladd, D. R. 2008. Intonational phonology. Cambridge: Cambridge University Press, 2nd edition. Liberman, M., & A. Prince. 1977. On stress and linguistic rhythm. Linguistic Inquiry 8. 249–336. Liberman, M. Y. 1975. The intonational system of English. Cambridge, MA: Massachusetts Institute of Technology dissertation. McAuliffe, M., M. Socolof, S. Mihuc, M. Wagner, & M. Sonderegger. 2021. Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi (Version 2.0.0). http://montrealcorpustools.github.io/Montreal-Forced-Aligner/. Retrieved 25 September 2021. Newell, H., & G. Piggott. 2014. Interactions at the syntax–phonology interface: Evidence from Ojibwe. Lingua 150. 332–362. Nguyen, A. D., & G. Legendre. 2022. The acquisition of wh-questions: Beyond structural economy and input frequency. Language Acquisition 29. 79–104. Pierrehumbert, J., & J. Hirschberg. 1990. The meaning of intonational contours in the interpretation of discourse. In Intentions in Communication, ed. by P. Cohen, J. Morgan, & M. E. 516 ZHU AND NGUYEN Pollack, chapter 14, 271–311. Cambridge, Massachusetts: MIT Press. Pires, A., & H. Taylor. 2007. The syntax of wh-in-situ and common ground. In Proc. from the Annual Meeting of the Chicago Linguistic Society, volume 43, 201–215. Chicago Linguistic Society. Price, P. J., M. Ostendorf, S. Shattuck-Hufnagel, & C. Fong. 1991. The use of prosody in syntactic disambiguation. The Journal of the Acoustical Society of America 90. 2956–2970. Prince, A., & P. Smolensky. 2004. Optimality Theory: Constraint interaction in generative grammar. Oxford: Blackwell. Reinhart, T. 1995. Interface strategies. OTS working papers in linguistics 55–109. Rizzi, L. 1996. Residual verb second and the wh-criterion. In Parameters and functional heads: Essays in comparative syntax, ed. by A. Belletti & L. Rizzi, chapter 2, 63–90. Oxford: Oxford University Press. Roettger, T. B., T. Mahrt, & J. Cole. 2019. Mapping prosody onto meaning — the case of information structure in American English. Language, Cognition and Neuroscience 34. 841–860. Rooth, M. 1992. A theory of focus interpretation. Natural language semantics 1. 75–116. Rotenberg, J. 1978. The syntax of phonology. Cambridge, MA: Massachusetts Institute of Technology dissertation. Selkirk, E. O. 1978. On prosodic structure and its relation to syntactic structure. In Nordic Prosody II, ed. by T. Fretheim, 111–40. Trondheim: Tapir. Smith, V. L., & H. H. Clark. 1993. On the course of answering questions. Journal of Memory and Language 32. 25–38. Szendrői, K. 2003. A stress-based approach to the syntax of hungarian focus. The Linguistic Review 20. 37–78. Watson, D. G., M. K. Tanenhaus, & C. A. Gunlogson. 2008. Interpreting pitch accents in online comprehension: H∗ vs. L+H∗ . Cognitive Science 32. 1232–1244. Zec, D., & S. Inkelas. 1990. Prosodically constrained syntax. In The phonology-syntax connection, ed. by S. Inkelas & D. Zec, 365–378. Chicago: University of Chicago Press. A Appendix ∆F0 Question type Mean (SD) t(df ) EQ Mid 60(73) −7.0(269) End 118(80) −11.9(216) PQ Mid 6(94) .2(271) End 38(99) −2.9(215) p Wh-word duration Mean (SD) t(df ) p < .001∗∗∗ < .001∗∗∗ .31(.1) .38(.08) −8.8(269) −14.5(216) < .001∗∗∗ < .001∗∗∗ .86 < .01∗∗ .34(.12) .42(.1) −10.8(271) −16.6(215) < .001∗∗∗ < .001∗∗∗ Table 2: Results of two-sample t-tests comparing wh-word phonetics of FQs against the in-situ questions (EQs or PQs with sentence-medial or sentence-final wh-word). Proceedings of CLS 58 (2022), 517-537 © Chicago Linguistic Society 2023. All rights reserved. 517 What is a gendered voice? Intersectional perspectives on ethics, empiricism, and epistemologies in the sociophonetics of gender Lal Zimman (he/him or they/them) University of California, Santa Barbara 1 Introduction1 The goal of this paper is to critically engage with a foundational assumption of most (socio)phonetic research: that voices–perhaps all voices–can be unproblematically characterized as either “female” or “male.” This assumption shapes methods at every level, from hypothesis formulation to acoustic settings to participant binning. It is also increasingly unviable in today’s changing genderscape, if it ever was. This paper shows that interrogating the concept of vocal sex/gender reveals that it suffers from under-theorization, significant logical shortcomings, and limited analytic usefulness, at least in its typical implementation. I explore this issue by drawing on three complementary perspectives: ethical, epistemological, and empirical. I begin with a discussion of the ethical issues that motivate this paper: the implications of researchers’ attributions of sex/gender to their participants and the way such practices work in tandem with the persistent erasure or distortion of intersectionality (Crenshaw 1989; miles-hercules 2022) in (socio)phonetic approaches to the gendered voice. The discussion then shifts to an epistemological perspective to consider how different forms of knowledge and ways of thinking are brought to bear on gender attributions based on the voice, as well as the logical circularity embedded in vocal sex/gender as a concept.2 Finally, I discuss a pilot experiment on gender attribution, the results of which raise empirical questions about the production of vocal/sex gender together with the production of race. Overall, this paper underscores the importance of reconsidering theories of sex/gender in (socio)phonetics in ways that both avoid misgendering trans people and address the legacy of racism in the characterization of the gendered voice. 2 The ethics of representation The central issues driving this investigation are fundamentally ethical ones. Research ethics include numerous components, among them how researchers interact with participants; the relationship between the researcher and the research community or population; the types and extent of agency participants are granted; and how they are represented within the research and its dissemination (Pickering 1 Thanks to members of the Trans Research in Linguistics Lab at UCSB and other colleagues whose ideas have shaped this paper, especially deandre miles-hercules, Dozandri Mendoza, Mary Bucholtz, and Anne Charity Hudley. 2 I refer to (socio)phonetics to highlight the fact that sociophoneticians often reproduce the same models of sex/gender employed in non-socially-oriented phonetics. This should be a red flag, since one function of sociophonetics is to theorize the social impacts on the voice. 518 LAL ZIMMAN & Kara 2017). Representation itself is multidimensional, including the social characteristics attributed to participants and their voices, especially those thought to be audible through the embodied voice, as sex/gender and race are. When it comes to the ethics of representation, we might consider questions like the following: What kind of humanization is afforded to participants? Are they represented individually, or only in the aggregate as representatives of a larger category? Does the analysis or interpretation contradict the way participants understand themselves or the world? Does it correspond more to the way they are seen by others, especially those with more social power? And are participants informed, consulted, or given any degree of control regarding the way they are presented in publications and talks? Focusing more specifically on the ways transgender and non-binary3 speakers are represented in (socio)phonetic research, we might ask how information about gender is collected; how many gender categories are used when binning speakers or modeling the data; and the extent to which choices about trans representation might be motivated by methodological convenience, intellectual traditionalism, or lack of awareness of non-hegemonic understandings of gender. In general, (socio)phoneticians continue to rely on the same model of gender that was in use half a century ago: a binary, fixed, primarily biological phenomenon that directly causes a great proportion of the differences that can be observed between women and men, even if some degree of social influence is accepted as a form of “elaboration” on biological difference. This view maps onto the voice: binary biological difference is a first-line explanation for many phonetic gender differences, even when evidence to the contrary is apparent (see Zimman 2018). Sometimes representation is implicit, as when participants are written about as including “24 speakers (13 male)” or as having “equal numbers of both sexes,” which deny the possibility of more than two. Indeed, sociophoneticians’ continued deployment of the word sex, rather than gender, implies that differences between women and men at the phonetic level are biological in origin, even when biology has purportedly been factored out (e.g. when explaining differences in normalized formant values). For some disciplines, even recognizing that gender is a social construct is radical, but an important contribution of trans, queer, and poststructuralist thinking is that sex is no less social than other aspects of gender (Zimman 2014). This perspective is not that embodied difference doesn’t exist or doesn’t matter, but that it is always tied up with cultural frameworks and practices (Fausto-Sterling 2000; Butler 1993; Snorton 2017). For example, the notion that there are two biological sexes is contradicted by the well-documented variability in human sexual characteristics, such as those in bodies classified as intersex (Dauphinais 2022), and this binary is best understood as a form of erasure that helps 3 Trans here refers to anyone who does not identify (or exclusively identify) with the gender assigned to them at birth. This definition includes non-binary individuals, but at times I refer to “trans and non-binary people” in recognition that not all non-binary people identify as trans. What is a gendered voice? 519 constitute a cultural framework that recognizes only two genders. That is, the “natural fact” that there are two sexes is itself a cultural construction. Some of these considerations surrounding gendered representation find their double in race, but others take a decidedly different form. For instance, speaker sex/gender is almost always reported in (socio)phonetic research, but speaker race is often unspecified unless it is an explicit focus of the study or the study includes speakers whose race differs from the way their language is generally racialized (Rosa & Flores 2017). In this way, white speakers of European languages become representative of categories that are at once ethnonational, colonial, and linguistic, such that they can be described simply as “English speakers” or “speakers of French” without reference to race. The unmarkedness of whiteness in the context of English legitimates white citizenship and linguistic repertoires, even when speakers of color are referred to in ways that seem conscious of race. Reporting race, of course, is a low bar, as even when race is reported, it may not be analyzed or even inform the analysis; in other cases, it may inform the analysis in damaging ways (about which more is said below). There are other issues that might be shared when it comes to the representation of gender and race – e.g., how many categories to count and the extent to which self-identification should be authoritative – but the answers to these questions will likely be different, as gender and race often operate through different logics. Despite these differences, this paper posits that the types of biological deterministic models of gender that misgender trans people also inhibit intersectional analyses of gender and race. One parallel that exists between race and trans status (i.e., whether a person is cis or trans) is in reporting norms. Although speaker sex/gender is indicated in (socio)phonetic research as a matter of course, trans status tends only to be mentioned if the study is focused specifically on trans people, or because the researcher learns that at least one speaker is trans. Few authors report how information about gender identity and trans status was collected, e.g., using a form, face-to-face questions as part of an interview, or simply the researcher’s own attributions. Yet even the most trans-affirming methods, such as allowing participants to self-identify their genders, may be undermined after the fact, e.g., by binning according to sex assignment at birth or through the assumption that anyone who doesn’t describe themself as trans must be cis. 4 This is a form of cisnormativity, or the framing of cisness as the normal, natural, default state of humanity. Importantly, cisnormativity is not unique to biological determinist theories of gender difference; feminist research on gender and the voice has often tied such differences to language socialization early in life (e.g., Brend 1975). Whether understood through biological or social forces, the notion that gender is fixed, binary, and in alignment with physiological traits and/or assigned sex at birth 4 A non-cis research participant may not identify as trans because they do not wish to disclose that information, because they use other terminology that distinguishes them from cisgender women and men (e.g. Two-Spirit), or because they identify simply as women or men and see their trans status as unimportant for characterizing their gender. 520 LAL ZIMMAN has a misgendering perlocution for trans people. Representation is not only about the way participants in a given study are presented; it also exists in the form of absences and implicit contrasts. That is, ethical representation demands that we think about what is said (and not said) about gender even when discussing cisgender people and men; what is said (and not said) about race when discussing white people; etc. Even in studies that seem to have nothing to do with marginalized populations, representation of these groups occurs through the characterization of dominant or normative groups. This section can be concluded with two pronouncements that are at once obviously true and yet carry radical implications: (1) If research misgenders trans people, directly or indirectly, it promotes and advances transphobia (i.e., discourses and practices that harm trans people).5 Put simply, such research is transphobic. (2) If research presents whiteness as default, fails to engage with the hierarchical nature of race, or ignores race completely, it promotes and advances white supremacy. Put simply again, such practices are racist.6 As I discuss below, the ways transphobia and white supremacy are manifested in (socio)phonetic research on gender and the voice are deeply intertwined with one another, and a point I wish to underscore is that dismantling one set of ideologies requires simultaneously working to dismantle the other (Steele 2022). 2 Epistemological frameworks for gender attribution One way to understand the problems with the ways gender attribution has functioned in (socio)phonetic research is through attention to the epistemological foundations of that process In this section, I consider the ways vocal sex/gender might be more explicitly theorized, and hence thought about, in (socio)phonetics. This exercise poses questions regarding what it means for a voice to be female or male, how many vocal sex/genders exist, whether vocal sex/gender always matches a speaker’s sex/gender, and how one knows whether a voice is female or male. By making implicit theories of identity more explicit, we can see the ways vocal sex/gender rests on undertheorized – and seemingly logically circular – ways of knowing, and how these theories work against trans-affirming and raciallyinformed approaches to the voice. I also discuss problematic interpretive practices that are often regarded as legitimate routes to knowledge about linguistic variation. 5 The harms associated with misgendering are numerous, particularly when it comes to mental health and overall well-being (e.g. Russell et al. 2018). Additionally, transphobic violence and institutionalized transphobia are typically grounded in and accompanied by acts of misgendering. 6 Of course, racism in academic research is in no way particular to linguistics, but rather part of a legacy of scientific racism mentioned later in this paper. What is a gendered voice? 521 2.1 Models of sex/gender, race, and the voice 2.1.1 Three models of vocal sex/gender Short of approaches that draw on evolutionary biology (e.g., Ohala 1984), it is rare to encounter (socio)phonetic work that overtly theorizes or provides empirical justification for any aspect of its model of sex/gender. Instead, researchers tend to categorize voices as female or male without ever explaining what they mean by this characterization or how they arrived at it. There are at least three ways of making sense of the construct of vocal sex/gender, ball of which leave something to desired. The first possible model of the gendered voice, which I call the vocal gender as speaker gender model, treats the former as a derivation of the latter. Within this framework, a female voice is simply the voice of a female person, however that category is understood. In all likelihood, this is how vocal sex/gender is most often implemented, particularly for cisgender speakers. This means that what unifies “female” or “male” voices is not their acoustic characteristics, nor even how they are heard by others, but by speaker gender alone. Even if only cisgender speakers are considered, this model would mean that some “female voices” would not be identified as female by listeners, and may have more in common, acoustically speaking, with “male voices”. Of the frameworks I describe, this is the one that is most compatible with the affirmation of trans identities and the recognition of more than two gender categories, because it allows for any type of voice to be connected to any gender. Of course, it is also entirely possible to implement this model in transphobic ways. A number of trans students of sociophonetics have told me that about faculty instructing them to classify trans speakers according to “biological sex” – by which they seem to mean sex assignment at birth – without regard for how that person’s voice actually sounds, the current embodied state of their vocal anatomy, or the ethical implications of this choice. Ultimately, however, vocal sex/gender offers minimal analytic usefulness if it is simply another way of referring to the sex/gender of the speaker and tells us nothing about the voice itself. The second and third possible models for vocal sex/gender are closely intertwined. In the second, which is an acoustic model of vocal sex/gender, a voice can be classified on the basis of its acoustic properties, such as having a fundamental frequency above or below a certain range. This approach seems to offer some empirical grounding rather than simply acting as a proxy for speaker sex/gender. However, this definition would mean that some women – including some cisgender women, even if their voices are heard as female – may not be considered to have female voices, and that some men – including cisgender men who are recognized as men by others – may have female voices, based solely on acoustic measurements. More importantly, this approach still requires an a priori definition of gender in order to determine what acoustic ranges should be used to classify voices. That is, how do we determine the F0 range (among other traits) that tells us a particular voice is female or male? Presumably by relying on research on the pitch ranges produced by women and men, taking us full circle, back to the 522 LAL ZIMMAN same issue of how we define those categories. From whom would these measurements come? How we answer this question has important implications for trans people: are trans women and men included in the characterization of possible female or male acoustic ranges? Is there a place for non-binary speakers at all? Moreover, the fact that research on gender differences in the voice has relied disproportionately on white speakers means that any definition that relies on existing findings regarding “women” and “men” would in practice reproduce a white standard of gender normativity. The third model of vocal sex/gender is an attribution model that focuses on listener categorization rather than acoustics per se. From this perspective, a female voice is a voice that is classified as female by listeners. Such an approach has some benefits over the acoustic definition just described, in that it does not require researchers to discover all of the acoustic features that correlate with listeners’ attributions or define any precise acoustic boundaries. It also appears to sidestep the question of how researchers classify speakers/voices as female or male because it relies on listeners to do so. However, it leaves us with the same counter-intuitive situation where some (cisgender) women must be considered to have male voices and some (cisgender) men to have female voices. It also carries some of the same weaknesses as the acoustic model regarding which speakers would be selected as representatives of different gender groups while also introducing additional shortcomings, such as whose attributions are used as the basis of these classifications and how their social context shapes how they classify gender. No matter the model used, the construct of a “(fe)male voice” presumes certain alignments of identification, vocal production and social attribution. The centering of speakers for whom these alignments exist (or are presumed to) has allowed transphobic and racist ideas about vocal sex/gender to circulate uncritically. These problems do not necessarily mean that the notion of vocal sex/gender could not be rehabilitated, or that no other frameworks are in use. However, they do suggest that more serious engagements with critical theories of gender, identity, and embodiment will be critical for the continued viability of the concept. 2.2 Thinking about race and the voice However implicit the theorization of sex/gender and the voice has been in (socio)phonetics, the field’s theorization of race and the voice is even more so. Not only does race frequently go uncommented upon, it might be doubted that a voice “has” a race at all. For one thing, not all racial identities could be considered equally audible; in a U.S. context, some racialized populations are strongly associated with particular phonetic features, such as African Americans, while others have less salient links–at least to community outsiders–such as Asian Americans. Furthermore, it seems clear that not all members of these groups deploy the phonetic or phonological features associated with their racial positionalities, and that listeners will not be equally aware of those features for every racialized group. What is a gendered voice? 523 If it seems obvious that some speakers of color use racially unmarked styles that result in their decontextualized voices being categorized as white, and that some listeners are better than others at accurately identifying speakers’ races, the acoustic and perceptual models discussed for gender also fail when it comes to race insofar as it would be inappropriate to say that some people have a vocal race that is different from their social race. Instead, the only model that seems to make sense is an identity-based one, where the race of a voice is based on the race of the person who produces it, regardless of how it sounds. Of course, as for gender, a concept like “vocal race” seems to offer little analytic utility, instead offering obvious potential to do harm through the legacy of scientific racism, as the next section explores in greater depth. 2.2 Interpretive frameworks Another central element in the epistemologies of vocal sex/gender concern the interpretation of findings. The emphasis on biology surfaces here as well, as (socio)phoneticians show a persistent drive to identify biological origins for gender differences in the voice. This section emphasizes how this fact also undermines the intersectional thinking required to see how race and gender are sociophonetically co-produced. Trans-affirming and intersectional lenses can together undermine hegemonic interpretations and offer alternative routes to knowledge. 2.2.1 The pull of biological determinism To understand the influence that biological determinism exerts on (socio)phonetic research, we can consider examples of texts that highlight the lengths to which cisnormative epistemologies can go in order to preserve their narratives. Although I rely on specific examples to explore these points, but it is important to note that studies like the one discussed next are not exceptional but rather rely on common frameworks for the (socio)phonetics of sex/gender. For several decades now, /s/ has been investigated as an index of gender and sexuality across a number of languages (e.g., the articles in Levon, Maegaard & Pharao’s 2017 special issue), but particularly in English. These investigations frequently, but not universally, find that women tend to produce a more fronted or higher frequency /s/, compared to men, and that men with a more fronted or higher frequency /s/ are associated with non-normative sexualities or gender expressions (see Calder 2020). Cross-linguistic similarities can intensify the existing drive for physiological explanations, and Fuchs and Toda (2010) investigate the possibility that gender differences in /s/ among German and English speakers are caused by undocumented differences in the mouths of (presumably cisgender) women and men. To this end, they performed several anatomical measures of speakers’ palates as well as articulatory and acoustic measures related to the production of /s/ for six women and six men from each of the two languages. Race is not mentioned at any point, suggesting that “German” and “English” might be deployed as ethnoracial as well as linguistic terms such that the speakers are white or else that the authors do 524 LAL ZIMMAN not consider race relevant to their analysis. Trans status is not mentioned either, nor are the criteria or method for categorizing speakers as female or male. Fuchs and Toda’s findings mirror previous work in showing a robust acoustic gender difference for /s/ among both the German and English speakers; the authors also report an indirect relationship between palate length and spectral properties of /s/ in which palate length was correlated with the size of the front cavity (i.e., the space between the tongue and teeth) and the size of the front cavity was correlated with the spectral properties of /s/. However, this indirect relationship held only for the English speakers, and Fuchs and Toda found no significant correlations between the acoustics of /s/ and any of the anatomical measurements in either language. Most importantly, they found no significant sex/gender-based differences in the anatomical measures they took; in fact, the differences between the German and English speakers were greater than those between women and men. Despite finding no sex/gender differences in their anatomical measures, Fuchs and Toda put weight on a “trend” in one oral measurement, which also held only for the English speakers, and which failed to meet the standard of significance set for their analysis. The authors interpret these results as supporting a partly biological and partly social explanation for gender differences in /s/ (p. 299), rather than as providing strong support for the influence of social factors and limited, if any, support for direct biological causes. There are some things to be appreciated about Fuchs and Toda’s analysis, such as their conclusion that it might be better to use “speakers’ morphological data instead of simply splitting data into male and female results” (ibid.), with which I am in agreement. At the same time, the choice of this interpretation – over one that acknowledges how much more evidence this study produced for a social interpretation than a biological one – highlights the distorting force that the drive for physiological explanations for gender differences can exert within (socio)phonetics. It also points to the importance of bringing trans epistemologies into (socio)phonetic research as alternatives to the cisnormative epistemological frameworks that continue to dominate the field. 2.2.2 Clashes between biological determinism & intersectionality Binary, biological determinist models of gender may be at odds not only with the affirmation of trans identities, but also with intersectional and anti-racist perspectives on gender and race. The practice of ignoring race entirely seems connected to the notion that sex is a universal, biological binary that applies equally to all of humanity, making further subclassifications unnecessary. If all women and all men share distinct bio-linguistic essences, it may be seen as unnecessary–or recognized as dangerous–to investigate whether the same differences occur across populations that are racialized in different ways. In reality, sociophoneticians examining the relationship between gender and race have found that gender differences attributed to “(American) English speakers” do not always hold among (American) English speakers of color. Continuing with sibilants, Calder and King’s (2020) analysis of /s/ among African American speakers in Rochester, NY and Bakersfield, CA revealed that the former group followed the pattern described What is a gendered voice? 525 previously on the basis of white or racially unmarked speakers, where women produce a higher frequency /s/, but the latter group showed no such difference (see also Pharao et al. 2014; Podesva 2013; Steele 2022). Faced with results like this, Calder and King’s conclusions point crisply toward a sociolinguistic explanation, in which gender differences in /s/ are learned in the context of a local community. On the other hand, if researchers refuse to let go of the expectation that gender differences will have biological origins when confronted with the fact that gender differences displayed by white speakers may not be paralleled in non-white populations, they run a risk of adopting an epistemology of scientific racism in which culturally defined “races” are understood in terms of physiological or evolutionary hierarchy (Fairchild 1991). This line of thinking is what we see in Ryalls, Zipprer, and Baldauff’s (1997) investigation of gender and Voice Onset Time among African American and white American speakers of English. This piece, which is not widely cited, reports on a gender difference among white speakers where women maintained a greater contrast between voiced and voiceless consonants than men. The study, however, found no such gender difference among its Black participants. Like Fuchs and Toda, Ryalls and coauthors look first and foremost to biological differences as the cause for this articulatory variation.7 The reliance on binary sex difference might seem to be contradicted by the absence of any gender difference for VOT among Black participants, yet the authors hold strong to the notion that gender differences are biological, and hence conclude that the racial difference, too, may be due to “fundamental physiological differences” (1997:642) between Black and white populations. They go as far as to cite a blatantly racist early-apartheid South African publication–“The anatomy of the South African Negro larynx” (Boshoff 1945)–as support for a physiological interpretation of gendered and racialized articulatory norms. In referencing a case like this, the point is not to castigate authors, nor to claim that this specific line of thinking is held by most researchers. Instead, what is most notable is how this interpretation is facilitated by the biological determinism so enthusiastically taken up with respect to gender and the voice. Indeed, there is a fundamental incompatibility between the idea that biological sex is a simple binary producing uniformly “female” or “male” bodies and voices that are essentially the same AND the recognition that sex/gender is produced in ways that are inextricable from race, among other subjectivities. The approach of Ryalls et al. exemplifies the absence of any articulable theory of race, let alone its intersection with gender, in much of the (socio)phonetic research on gender difference (see also Charity Hudley, Mallinson, & Bucholtz 2020). It is with these observations in mind that we should approach the study of the voice and its relationship to speakers’ social and embodied positionalities, which is the goal of the final section of this paper. 7 VOT is determined by the timing of articulatory gestures and vocal fold vibration, and has no anatomical correlate. Even if the gender difference Ryalls et al. find for white speakers were universal, it would be unclear what sort of sex-based physiological mechanism could cause it. 526 LAL ZIMMAN 3 Empirical insights on the attribution of gender and race This section turns to an ongoing project on the attribution of gender and race and a (flawed) pilot experiment conducted on gender attribution using digitally manipulated guises of trans speakers. In addition to talking about this experiment, I discuss the concepts of experimental refusal and resistance and how trans voices of color reveal limitations in non-intersectional approaches to gender attribution. 3.1 Pilot on gender attribution 3.1.1 Experimental research on gender attribution The most substantial body of literature on the process of gender attribution based on trans people’s voices comes from the field of speech-language pathology (SLP), which has a history of providing services to trans people seeking to change the gendered characteristics of their voices. If the goal is to shift how a speaker’s voice is gendered, it is critical to know where the perceptual boundaries are; I refer to these boundaries as crossover points. A number of authors have focused on identifying a single crossover point for fundamental frequency (e.g., Spencer 1988, Wolfe et al. 1990, Gelfer & Schofield 2000), and have proposed 150 to 165 Hz as a potential range in which voices are ambiguously gendered, or that divides voices identified as female from those judged to be male. Figure 1 visualizes this range on a continuum along with the oft-repeated–but rarely sourced–F0 ranges that are supposedly average for male and female speakers of American English: 100-120 Hz and 200-220 Hz, respectively. Figure 1: Fundamental frequency values from previous studies on F0 crossover point. However, there are important exceptions to the proposal of 150-165 Hz as a potential crossover zone, as Figure 1 also shows. Günzburger (1993) discusses a trans woman whose voice was consistently identified as female by listeners despite having a mean fundamental frequency as low as 119 Hz, while Gelfer and Schofield (2000) describe a speaker who is consistently categorized as male despite having a mean F0 as high as 181 Hz. These exceptions suggest that different speakers have different crossover points, in keeping with bricolage-based theories of sociolinguistic style (Eckert 2004; Zimman 2017). From this perspective, we would What is a gendered voice? 527 expect that where an individual speaker’s crossover point lies will be shaped by the other semiotically meaningful aspects of their self-presentation. Although an experiment cannot reflect the true social richness through which this process occurs in context, it can test the possible influence of specific phonetic variables on identity attribution. Based on the rich bodies of work on vowels and sibilants as segments displaying gender-related variation, the phonetic variables selected were the F1, F2, and duration of stressed vowels and the center of gravity (weighted mean) and duration for /s/ within the stimulus phrase. 3.1.2 Pilot methods The stimuli for this pilot, which was originally carried out in 2013, came from recordings of 15 trans speakers reading the “Rainbow Passage” (Fairbanks 1960). The recordings came from a long-term ethnographic study of transmasculine speakers in the early years of hormone replacement therapy in order to track changes in their F0 and other aspects of their voices. A single sentence was selected for use in the attribution experiment: Throughout the centuries people have explained the rainbow in various ways. From the multiple recordings made of each speaker, I selected the one in which their average F0 for this phrase came as close as possible to the range of 150-165 Hz. The average original F0 for all of the speakers was 142 Hz (see Table 1 for each speaker’s mean F0 prior to digital manipulation). Aside from F0, these speakers had quite variable gendered phonetic styles. Some had voices that were already very masculine before their F0 began to lower, while others used and embraced styles regarded as feminine; Zimman (2017) offers an analysis of the production data and descriptions of speakers’ identities. Speaker Adam Carl Dave Devin Elvis Ethan James Jeff Mean F0 131 Hz 162 Hz 113 Hz 129 Hz 138 Hz 170 Hz 131 Hz 99 Hz Speaker Joe Jordan Kam Kyle Mack Pol Tony Mean F0 125 Hz 177 Hz 129 Hz 137 Hz 187 Hz 164 Hz 134 Hz Table 1: Original mean F0 for experimental stimuli (by speaker pseudonym). The recording from each speaker was then used to create 8 guises by globally manipulating mean F0 with Praat (Boersma & Weenik 2011). The guises were shifted to have a mean F0 separated by 10 Hz (± 1 Hz) from 120 to 190 Hz. Each speaker therefore had guises with means of 120 Hz, 130 Hz […] 180 Hz, 190 Hz. The experimental task was administered via Amazon’s Mechanical Turk (MTurk). Each listener heard only one guise per speaker, and could listen to each audio clip as many times as they wished before answering whether the speaker was 528 LAL ZIMMAN female or male. Listeners also had the option to provide general comments after completing the attribution tasks. Ten participants heard a single guise from each speaker, making a total of 150 listeners. In early days of using MTurk for experiments, the simple, forced-choice task was designed to mimic the type of work users on that platform were often paid to do. Since then, much has been written on MTurk and its users (e.g., Paolacci & Chandler 2014). MTurk responses were used to identify an F0 crossover point for each speaker. A speaker’s crossover point was defined as the highest F0 guise that was categorized as male by a majority of listeners. That is, a crossover point of 140 Hz indicates that a majority of listeners rated that speaker as sounding male at 140 Hz and below.8 Some speakers were perceived as female even at 120 Hz, and their crossover point was set at 110 (Hz) for the purposes of visualization and statistical analysis, with the caveat that it is possible they might not have been classified as male by a majority of listeners even with a mean F0 of 110 Hz or lower. The acoustic measurements, which were made prior to the design of the experimental data, included F1, F2, and duration for the stressed vowels (7 tokens) and center of gravity and duration for /s/ (3 tokens). The number of tokens is small, but reasonable, both because listeners had only a small amount of speech to base their attributions and because the tokens are matched for content and context. Finally, the crossover points and acoustic data were included in a linear mixedeffects regression model using the lme4 (Bates et al. 2015) and lmerTest (Kuznetsova et al. 2017) packages in R, with the crossover point as the independent variable. The five acoustic measures – F1, F2, vowel duration, center of gravity for /s/, and /s/ duration – were set as fixed effects and word was set as a random effect. 3.1.3 Pilot results The first question up for consideration is whether these speakers support the notion of an F0 crossover point or range that would be shared by all speakers. By isolating global F0 through digital manipulation, this analysis suggests that the answer is no. Figure 2 shows the crossover points for each of the 15 speakers. Some, such as Kyle, Carl, Tony, and Jeff, crossed over at 150, 160, or 170 Hz – close to the range of 150-165 Hz proposed in other studies. But more registered crossover points closer to the periphery, particularly at the lower end. Four speakers were categorized as female a majority of the time even with an F0 as low as 120 Hz, and another four were consistently categorized as male only in the lowest F0 guise. Others were at the other end of the spectrum: Joe was heard as male by a majority of speakers even with an F0 of 190 Hz, James had a crossover point of 180 Hz, and Tony and Jeff crossed over at 170 Hz. This distribution falls out much like the other variation in this sample discussed by Zimman (2017). The speakers on the upper end of this spectrum tend to have more typically masculine voices and gender 8 There was one case in which a speaker was perceived as male by a majority of listeners at 150 Hz (as well as 110 Hz, 120 Hz, and 130 Hz), but by fewer than 50% of listeners at 140 Hz. This speaker’s crossover point was set at 150 Hz. What is a gendered voice? 529 expressions, such as Joe, a straight, white, working-class trans man in his 40s whose masculine self-presentation predated his self-identification as trans. Those toward the lower end tend to have less typically masculine gender presentations, including genderqueer-identified speakers like Devin and Elvis, and queer fem(me) maleidentified speakers, like Jordan and Dave. Figure 2: The crossover point for each speaker in this study. The methods used in this study were different from those employed in the SLP research on trans voices by isolating F0 from other variables that might influence the attribution of gender. On one hand, these findings suggest that shifts in F0 alone may not be enough to bring about a change in gender attribution. At the same time, these data are highly impoverished in that the voices of some trans speakers in this study changed not only with respect to F0, but also in formant frequencies and/or /s/ – not to mention salient, if not dramatic, changes in visual cues. Among the participants in this study, most found that they were typically identified as male by others, especially in face-to-face interactions. 9 Despite these complexities, what seems clear is that these speakers do not share a single crossover point. The second set of questions concern the role of the acoustic measurements in predicting variation in crossover point. The results of the regression model appear in Table 2. Two of the acoustic variables modeled for this analysis reached a p value smaller than 0.001: F2 and center of gravity for /s/. Both correlated negatively with crossover point such that a speaker with a higher F2 (Figure 3) or higher /s/ center of gravity (i.e., a fronter /s/; Figure 4) had to reach a lower F0 before being categorized as male. Means for each speaker’s acoustic measures are in Table 3. 9 One exception may be phone interactions with strangers, in which case trans people – and, one might venture, cis people as well – find gender misattributions to be more common than in other contexts. 530 LAL ZIMMAN Variable Intercept F1 F2 Vowel duration /s/ center of gravity /s/ duration Estimate 335.1 -0.026452 -0.08278 5.44984 -0.007296 239 Std error 43.9 0.044944 0.018271 35.434192 0.001947 200 df 38 t value 7.641 -0.589 -4.531 0.154 -3.748 1.193 p value < 0.001 n.s. < 0.001 n.s. < 0.001 n.s. Table 2: Fixed effects from linear mixed-effects model of crossover point. This analysis has identified two gendered phonetic characteristics that predict where these speakers’ individual crossovers point might be. These results support a style-based understanding of the gendered voice, in which it is the combination of variables that produces a social meaning, in this case that the voice is “female” or “male.” In other words, what constitutes a gendered voice is not a single trait but a multitude of characteristics that may be combined in more or less normative ways. Speaker Crossover Adam Carl Dave Devin Elvis Ethan James Jeff Joe Jordan Kam Kyle Mack Pol Tony 120 Hz 160 Hz 120 Hz <120 Hz <120 Hz 140 Hz 180 Hz 170 Hz 190 Hz <120 Hz 120 Hz 150 Hz <120 Hz 120 Hz 170 Hz F1 F2 475 Hz 592 Hz 545 Hz 522 Hz 497 Hz 544 Hz 512 Hz 465 Hz 529 Hz 529 Hz 562 Hz 467 Hz 508 Hz 537 Hz 518 Hz 1973 Hz 1976 Hz 1996 Hz 1961 Hz 1905 Hz 1752 Hz 1476 Hz 1894 Hz 1691 Hz 1838 Hz 1898 Hz 1824 Hz 2073 Hz 1847 Hz 1817 Hz Vowel duration 73 ms 107 ms 98 ms 88 ms 81 ms 79 ms 82 ms 63 ms 150 ms 82 ms 78 ms 111 ms 103 ms 197 ms 84 ms /s/ COG 6011 Hz 4771 Hz 9667 Hz 9989 Hz 7734 Hz 3985 Hz 7256 Hz 6071 Hz 4905 Hz 6901 Hz 7165 Hz 6508 Hz 5407 Hz 6252 Hz 4574 Hz Table 3: Means for the five acoustic measures. /s/ duration 58 ms 66 ms 87 ms 61 ms 71 ms 65 ms 60 ms 76 ms 74 ms 70 ms 49 ms 71 ms 76 ms 96 ms 68 ms What is a gendered voice? 531 Figure 3: F2 frequency by crossover point. Figure 4: Center of gravity for /s/ by crossover point. While these findings are compelling, there is still a great deal to be said about the process of gender attribution not reflected here. I address two of these in the final sections of this analysis: refusals and race. 3.2 Refusals and racialization: Problematizing gender attribution Anthropologists have examined the concept of refusal from a number of perspectives, often in relation to things social actors perform in relation to a state and colonial empire (Simpson 2007). These insights are also applicable to research encounters that involve refusal or non-compliance. McGranahan (2016) theorizes refusals as generative, social and affiliative, hopeful and willful, and yet distinct 532 LAL ZIMMAN from resistance. In an experimental context, both refusals and resistance on the part of participants is often erased. For example, it is common for online surveys to be left incomplete for unspecified reasons: perhaps participants simply got distracted or opted to abandon the task for something more interesting, or perhaps they found the design of the study objectionable in some way that led to a more conscious refusal to continue. These datapoints may then be removed. “Forced choice” tasks– like the binary gender selection required of participants in my pilot–are generally regarded as a handy tool to get desirable data, but they also limit participants’ agency to represent their own perspective. Another experiment I carried out as a graduate student involved an online survey in which participants were asked to identify the gender and/or sexual orientation of a group of speakers. I solicited participants from a variety of online communities of which I was a member, and I found that not all of my associates were happy to make judgements about these identities based on the voice alone. In the open comments portion of the survey, one participant openly asserted, “you can’t tell what gender someone is from just talking. they have to tell you.” Another shared, “I felt so uncomfortable being asked to do this. I finished the survey but I didn’t like it.” Others remarked that they felt like they were guessing at times. These comments were few and far between, but may have constituted overtly transaffirming stances in a research context that seemed to have no room for critical reflection on gender. There is an irony here, as the goal of my work on gender attribution is to problematize the role of cisnormativity and transphobia in these processes; however, my experimental design provided no clue to participants why they were being asked to complete these tasks. Listeners in this experiment and others also oriented to the difficulty of attributing gender without some reference to other modalities, especially gendered and racialized embodiment. One participant noted, “It was surprisingly hard to pick genders because I don’t know what [the speakers] look like.” In other cases, participants were more specific, as when a listener said, “one of the black voices was low pitched but sounded like my auntie,” connecting directly to their lived experience with specific individuals. This listener’s reference to Black family members also points to the possibility of different degrees of attention to race among respondents. A final comment was especially instructive: “I noted a few [of the voices] sounded non-white […], and I thought one person could have been either a woman or an African American guy.” The final point, that this listener’s gender attribution depended on a presumed racial identification as either a racially unmarked (presumably white) woman or a Black man, is part of what drives the current phase of my work on attribution. Before shifting to that project, we can look briefly at a segment of an interview with Xane, a Black trans man from Seattle whose lived experience aligns with the survey comment just discussed. He was in his 30s and working both at a café and as a security guard when I interviewed him in San Francisco in early 2011 as part of my dissertation fieldwork. Our conversation was one of a few that served as the genesis for a line of thinking that was not yet well enough developed for me to What is a gendered voice? 533 apply to my earlier experimental work. Excerpt 1 follows my observation that other Black trans men had described to me that they were only ever “clocked” – i.e., identified as trans rather cis men – by other Black people. 01 02 03 04 05 06 07 07 08 09 10 11 12 X: LZ: X: LZ: X: LZ: X: Oh yeah. Oh definitely.= =Did you find that [to be the case as well? [Definitely, definitely. Yeah. Definitely. Why do you think that is? Just tapping into different-? Um:. I think so, I think so. I mean, uh, i- i- it’s, it’s Maybe there’s more of a close, [...] recognizable thing about about certain things within your own race? Excerpt 1: Segment of interview with Xane, a Black trans man in his 30s Xane enthusiastically agrees with the reports I describe, saying repeatedly that it’s “definitely” the case that he had experienced something similar in which he was clocked as trans or misgendered primarily by other Black people. He has a few false starts before finding the words to express this experience to me, but goes on to suggest that it is a matter of “close[ness]” (line 10) and that people may recognize “certain things within [their] own race” (line 12). He does not specify what those “certain things” may be, but one possibility is that speakers are attending to racialized linguistic patterns. Another possibility, not described by Xane, is that non-Black speakers are more apt to assume he is cis male because of raciolinguistic ideologies (Rosa 2018) about Blackness and hypermasculinity. Indeed, both processes may co-exist. 3.3 An ethically-minded intersectional attribution experiment In response to the findings and other observations described above, members of the Trans Research in Linguistics Lab (TRILL) at UC Santa Barbara are developing a mixed-methods investigation on the process of gender and race attribution centered around the voices of trans people of color. We began with interviews that touch on experiences with the attribution of gender and race, along with a variety of other topics, along with more controlled tasks like reading passages and narrating the Pear Story film (Chafe 1980). During post-interview debriefing, participants are asked for consent not just to have their voices included in an experiment, but specifically that listeners in this experiment will be asked to guess their race and gender and that they may be misgendered as a result. We consider this an important part of informed consent even if speakers are not present for the misgendering. 534 LAL ZIMMAN Our experimental design departs from the pilot described in this paper in several ways. The first phase of the experiment uses an open-ended design to collect listeners’ impressions about speakers without prompting them for any specific traits. In the second phase, listeners will be asked both about gender – conceptualized both categorically and on linear scales of femininity and masculinity – and about race, presenting these questions in variable order. Finally, in the third phase we will recreate the digitally manipulated guises comparable to those described above for a gating experiment that further explores the crossover point of speakers with a much wider variety of identities and forms of embodiment. The experiment also differs in the way that listeners are engaged with. Rather than simply asking them to complete the task cold, we inform listeners up-front that they will be asked to judge people’s identities based on voice, that we know this can be difficult, touchy, and sometimes stereotype based, but that we are genuinely curious about people’s gut reactions and how they can help us understand the way we interact with each other and where stereotypes come from. While this kind of positioning is uncommon in studies that value “naïve” participants and avoid what might look like political pronouncements (albeit one that is toned down in relation to our lab’s actual political frameworks), members of TRILL hold the stance that the absence of a positioning text would have its own political ramifications, which would equally shape the results; we reject the possibility of a politically neutral context for attributing gendered and racial identities, and instead prioritize the wellbeing of participants who might be impacted negatively by being asked to do this type of categorization without knowing where the researchers are coming from. Going further into listeners’ experiences, we will conduct brief interviews with a subset of our listeners in order to learn more about their intersubjective experience with attributing race and gender to disembodied voices. This will allow us to delve more deeply into the kinds of stances that were hinted at in the comments of previous experimental participants discussed above. Our design allows participants to opt out of assigning a gender and/or race to any number of participants, and we are especially interested in hearing from those who refuse to make these assignments, rather than excluding them from our datasets. The goal here is to explore the ways different kinds of listening subjects (Inoue 2003; Flores & Rosa, 2015) perform and talk about the process of identity attribution. Our project is large and ambitious in part because there is so much groundwork to be done: examining the relationship between gender and race attribution, looking at how trans people of color’s voices are categorized, asking about what is going on for listeners, and considering what the proposal of a gender-based crossover point might look like when speakers of color are centered. 4 Conclusions The insights of intersectionality and trans experience are by no means new, but they have not yet been integrated into the analysis of human voices. In most (socio)phonetic research, gender is operationalized as a binary that is assumed to be based in embodied sex differences, while race is frequently not reported or What is a gendered voice? 535 acknowledged at all. These problems, I have argued, are related. The norms, received knowledge, and resources related to acoustic analysis of the voice are based primarily on speakers who are (assumed) white and cisgender, though their actual identities are often unknown. This reality has real impacts not only on our ability to properly analyze gendered voices, but also the application of this type of knowledge to people’s everyday lives. For instance, English-based speech recognition tools have been shown to do better for white speakers than for speakers of color (Koenecke et al. 2020). For trans people of color, the general lack of information about the gendered voice among speakers of color means that even the most trans-affirming experts offering voice support for trans people (as some speech-language pathologists do) may rely entirely on what is known about normative white voices, which are also typically marked by class privilege, lack of disability, heteronormativity, non-rurality, and so on. In the tradition of social justice-focused, community-based research methods, I have taken just a few steps toward reimagining the linguistic construction of the gendered voice. The steps have included questioning conventional knowledge, making implicit theories more explicit, addressing racialized exclusions and erasures, rejecting transphobic logics, and advocating for new ways of asking questions in (socio)phonetics. An intersectional sociophonetics requires a recentering of the field that attends not only to difference across racio-gendered lines, but the lived consequences visited upon those who live at certain intersections over others. References Bates, D., M. Mächler, B. Bolker, & S. Walker. 2015. Fitting linear mixed-effects models using lme4. Journal of Statistical Software 67. 1–4. Boersma, P., & D. Weenik. 2011. Praat (Version 5.3). http://www.praat.org. Retrieved 15 Oct 11. Boshoff, P. 1945. The anatomy of the South African Negro larynx. South African Journal of Medical Science 10. 35–50. Brend, R. M. 1975. Male-female intonation patterns in American English. In Language and Sex: Difference and Dominance, ed. by B. Thorne & N. Henley, 84–88. Rowley, MA: Newbury. Butler, J. 1993. Bodies That Matter: On the Discursive Limits of “Sex”. New York: Routledge. Calder, J., & S. King. 2020. Intersections between race, place, and gender in the production of /s/. U Penn Working Papers in Linguistics 26. Article 5. Calder, J. 2020. From ‘gay lisp’ to ‘fierce queen’: The sociophonetics of sexuality’s most iconic variable. In The Oxford Handbook of Language and Sexuality, ed. by K. Hall & R. Barrett. New York: Oxford. https://doi.org/10.1093/oxfordhb/9780190212926.013.49 Chafe, W., ed. 1980. The Pear Stories: Cognitive, Cultural, and Linguistic Aspects of Narrative Production. Norwood, NJ: Ablex. Charity Hudley, A. H., C. Mallinson, & M. Bucholtz. 2020. Toward racial justice in linguistics. Language 96. e200–e235. Crenshaw, K. 1989. Demarginalizing the intersection of race and sex. U Chicago Legal Forum 1989. 314–343. Dauphinais, A. L. 2022. Intersex bodies, language, and medical authority. In The Oxford Handbook of Language and Sexuality, ed. by R. Barrett & K. Hall. New York: Oxford. https://doi.org/ 10.1093/oxfordhb/9780190212926.013.75 Eckert, P. 2004. The meaning of style. Texas Linguistics Forum 47. 41–53. 536 LAL ZIMMAN Fairbanks, G. 1960. Voice and Articulation Drillbook. New York: Harper & Row. Fairchild, H. H. 1991. Scientific racism: The cloak of objectivity. Journal of Social Issues 47. 101– 115. Fausto-Sterling, A. 2000. Sexing the Body. New York: Basic Books. Flores, N., & J. Rosa. 2015. Undoing appropriateness. Harvard Educational Review 85. 149–171. Fuchs, S., & M. Toda. 2010. Do differences in male versus female /s/ reflect biological or sociophonetic factors? In An Interdisciplinary Guide to Turbulent Sounds, ed. by S. Fuchs, M. Toda, & M. Zygis, 281–302. Berlin: Mouton. Gelfer, M. P., & K. J. Schofield. 2000. Comparison of acoustic and perceptual measures of voice in MtF transsexuals perceived as female versus those perceived as male. Journal of Voice 14. 22– 33. Günzburger, D. 1993. An acoustic analysis and some perceptual data concerning voice change in male-female trans-sexuals. European Journal of Disorders of Communication 28. 13–21. Inoue, M. 2003. The listening subject of Japanese modernity and his auditory double: Citing, sighting, and siting the Modern Japanese woman. Cultural Anthropology 18. 156–193. Koenecke, A., A. Nam, E. Lake, J. Nudell, M. Quartey, Z. Mengesha, C. Toups, J. R. Rickford, D. Jurafsky, & S. Goel. 2020. Racial disparities in automated speech recognition. PNAS 117. 7684–7689. Kuznetsova A., P. B. Brockhoff, & R. H. B. Christensen. 2017. lmerTest package: Tests in linear mixed effects models. Journal of Statistical Software 82. 1–26. Levon, E., M. Maegaard, & N. Pharao, eds. 2017. Special issue: The sociophonetics of /s/. Linguistics 55(5). miles-hercules, d. 2022. The real tea: Language at the intersections. In The Oxford Handbook of Language and Sexuality, ed. by R. Barrett & K. Hall. New York: Oxford. https://doi.org/10.1093/ oxfordhb/9780190212926.013.70. Ohala, J. J. 1984. An ethological perspective on common cross-language utilization of F0 of voice. Phonetica 41. 1–16. Paolacci, G., & J. Chandler. 2014. Inside the Turk. Current Directions in Psychological Science 23.184–188. Pharao, N., M. Maegaard, J. P. Møller, & T. Kristiansen. 2014. Indexical meanings of [s+] among Copenhagen youth. Language in Society 43. 1–31. Pickering, L., & H. Kara. 2017. Presenting and representing others. International Journal of Social Research Methodology 20. 299–309. Podesva, R. J. 2013. Gender and the social meaning of non-modal phonation types. Proc. of the Annual Meeting of the Berkeley Linguistics Society 37. 427–448. Rosa, J. 2018. Looking like a Language, Sounding like a Race. New York: Oxford. Russell, S. T., A. M. Pollitt, G. Li, & A. H. Grossman (2018). Chosen name use is linked to reduced depressive symptoms, suicidal ideation, and suicidal behavior among transgender youth. Journal of Adolescent Health 63. 503–505. Ryalls, J., A. Zipprer, & P. Baldauff. 1997. A preliminary investigation of the effects of gender and race on Voice Onset Time. Journal of Speech, Language, and Hearing Research 40. 642–645. Simpson, A. 2007. On ethnographic refusal. Junctures 9. 67–80. Snorton, C. R. 2017. Black on Both Sides. Minneapolis, MN: University of Minnesota Press. Spencer, L. E. 1988. Speech characteristics of male-to-female transsexuals. Folia Phon 40. 31–42. Steele, A. J. 2022. Enacting new worlds of gender. In The Oxford Handbook of Language and Sexuality, ed. by R. Barrett & K. Hall. New York: Oxford. https://doi.org/10.1093/oxfordhb/ 9780190212926.013.75 Wolfe, V. I, D. L. Ratusnik, F. H. Smith, & G. Northrop. 1990. Intonation and fundamental frequency in male-to-female transsexuals. Journal of Speech & Hearing Disorders 55. 43–50. Zimman, L. 2014. The discursive construction of sex. In Queer Excursions, ed. by L Zimman, J. L. Davis, & J. Raclaw, 13–34. New York: Oxford. Zimman, L. 2017. Gender as stylistic bricolage. Language in Society 46. 339–370. Zimman, L. 2018. Transgender voices. Language & Linguistics Compass 12. e12284.

RELATED PAPERS

RELATED TOPICS

Log In

Representing Irish Mutations in Distributed Morphology and Optimality Theory

Representing Irish Mutations in Distributed Morphology and Optimality Theory

Related Papers

RELATED PAPERS

RELATED TOPICS