CMY CY MY CM K Y M C Centre deLingüística Teòrica delaUniversitat Autònoma deBarcelona 27 9 5 Index 53 101 67 121 145 in CentralCatalanandtheProblemofMissingBase. Constraints. in Maximum Entropy Grammar with Lexically Specific Harmonic Uniformity. Mascaró, Joan. Mascaró, Baković, Eric Bonet, Eulàlia; Torres-Tamarit, Francesc.Introduction. Moore-Cantwell, Claire;Pater,Joe.GradientExceptionality Rebrus, Péter;Szigetvári,Péter. Diminutives: Exceptions to Piñeros, Carlos-Eduardo. Rysling, Amanda. Zuraw,Polarized Variation. Kie. Institut Interuniversitari deFilologia Valenciana ISSN 1695-6885(inpress);2014-9719(online) Catalan Journal ofLinguistics Catalan Journal . ExceptionalityinSpanishStress. http://revistes.uab.cat/catJL
Exceptions inPhonology Morphological Exceptions to Vowel Reduction Reduction Vowel to Exceptions Morphological Polish yersrevisited. Vol. 15,2016 Exceptional nasal-stopinventories.
Exceptions in Phonology CCATATALAN JJOURNAL OFL LINGUISTICS Vol. 15 Centre deLingüística Teòrica delaUniversitat Autònoma deBarcelona OF Institut Interuniversitari deFilologia Valenciana ISSN 1695-6885(inpress);2014-9719(online) Eulàlia Bonet&FrancescTorres-Tamarit C Exceptions inPhonology C J J L L OURNAL INGUISTICS AT ATAL http://revistes.uab.cat/catJL V olume Vol. 15,2016 Edited by 2016 15 AN Coberta CJL 15.pdf 2 24/10/16 12:21
Catalan Journal of Linguistics Editors Centre de Lingüística Teòrica de la Universitat Autònoma de Barcelona Institut Interuniversitari de Filologia Valenciana Catalan Journal of Linguistics style sheet
Informació general i subscripcions General information and subscriptions • Document format. Manuscripts should be should be listed chronologically, with the CATALAN JOURNAL OF LINGUISTICS és una revista de lingüís- CATALAN JOURNAL OF LINGUISTICS is a journal of theoretical written in English. Font 12 point Times New oldest publication at the top and the newest at tica teòrica, publicada pel Centre de Lingüística Teòrica linguistics published by the Centre de Lingüística Roman, interline at 1.5, use margins 2,5 cm the bottom. Please, provide the DOI of articles de la Universitat Autònoma de Barcelona (format per Teòrica de la Universitat Autònoma de Barcelona (1 inch). Phonetic symbols must be set in if available. lingüistes dels departaments de Filologia Catalana i de (comprising linguists from the departments of Catalan Doulos SIL or Charis SIL. a) Book: Filologia Espanyola) i per l’Institut Interuniversitari Philology and Spanish Philology) and by the Institut • Heading. Title, with title capitalisation, in Borer, Hagit. 1983. Parametric Syntax. de Filologia Valenciana (que integra lingüistes de les Interuniversitari de Filologia Valenciana (which boldface. Name of the author(s), affiliation(s) Dordrecht: Foris. universitats de València, d’Alacant i Jaume I de Castelló). comprises linguists from the universities of València, and e-mail address(es). Acknowledgements, DOI: http://dx.doi.org/10.1515/97831 CATALAN JOURNAL OF LINGUISTICS publica articles d’inves- Alacant and Jaume I at Castelló). as a footnote, called by a flying asterisk. 10808506 tigació en lingüística teòrica que desenvolupin teories CATALAN JOURNAL OF LINGUISTICS publishes research papers sobre l’estructura de les llengües particulars dins la in theoretical linguistics developing theories on the • Abstracts. A 150 word abstract is required. b) Journal article: perspectiva d’una teoria general del llenguatge humà. structure of particular languages within the perspective A Catalan version of the abstract should be Kayne, Richard S. 1993. Toward a modu- Els volums seran sovint monogràfics, editats per un of a general theory of the human language. Volumes will included. Authors are also asked to provide lar theory of auxiliary selection. Studia especialista en el tema. often be monographic, edited by a specialist in the topic. the keywords of the article. Linguistica 47: 3-31. € € DOI: http://dx.doi.org/10.1111/j.1467- El preu de subscripció anual és de 18 . The yearly subscription price is 18 . • Sections and paragraphs. Headings of 9582.1993.tb00837.x main sections: in boldface. Headings of first Comitè assessor Redacció order paragraphs: in italics. Headings of c) Book article: Josep Maria Brucart Marraco Universitat Autònoma de Barcelona second order paragraphs or others: in italics Jaeggli, Osvaldo & Safir, Ken. 1989. The (Universitat Autònoma de Barcelona) Departament de Filologia Catalana with text beginning at the same line. Null Subject Parameter and Parametric Maria Teresa Espinal Farré 08193 Bellaterra (Barcelona). Spain Theory. In Jaeggli, Osvaldo & Safir, Phone: +34 93 581 23 71 (Universitat Autònoma de Barcelona) • Text. The beginning of a paragraph should Ken (eds.). The Null Subject Parameter, C +34 93 581 23 72 Anna Gavarró Algueró be left indented, except for the first paragraph 1-44. Dordrecht: Kluwer. (Universitat Autònoma de Barcelona) Fax: +34 93 581 27 82 of a section. Words, phrases or sentences in a DOI: http://dx.doi.org/10.1007/978-94- M Maria Lluïsa Hernanz Carbó E-mail: [email protected] language other than English appearing in the 009-2540-3_1 Y (Universitat Autònoma de Barcelona) Intercanvi body of the text should be followed by a Sabine Iatridou (Massachusetts Institute of Technology) Universitat Autònoma de Barcelona translation in simple quotation marks. d) Dissertation, manuscript, or work in CM Michael J. Kenstowicz Servei de Biblioteques References should follow this format: preparation: (Massachusetts Institute of Technology) Secció d’Intercanvi de Publicacions Pesetsky, David. 1982. Paths and MY Chomsky (1981), Kayne (1981, 1982), Belletti Joan Mascaró Altimiras Plaça Cívica (edifici N) & Rizzi (forthcoming), Jackendoff (p.c.). If Categories. PhD dissertation, MIT. (Universitat Autònoma de Barcelona) 08193 Bellaterra (Barcelona). Spain Chomsky, Noam. 1994. Categories and CY the page number or an example is expressed: Gemma Rigau Oliver Phone: +34 93 581 11 93 Rizzi (1990: 20), Rizzi (1990: ex. (24)). When Transformations. Unpublished manus- CMY (Universitat Autònoma de Barcelona) Fax: +34 93 581 32 19 the reference is included between parentheses, cript, MIT. Henriette de Swart (Universiteit Utrecht) E-mail: [email protected] K • Publication rights. CATALAN JOURNAL OF Rafaella Zanuttini (Yale University) drop the ones enclosing the year: (see Rizzi Subscripció, administració, edició i impressió 1990: 20). LINGUISTICS is published under the licence Consell editor Universitat Autònoma de Barcelona • Examples. Example numbers should go system Creative Commons, according to the Teresa Cabré Monné Servei de Publicacions modality «Attribution-Noncommercial (by-nc): (Universitat Autònoma de Barcelona) 08193 Bellaterra (Barcelona). Spain between parentheses, left aligned (Arabic digits for examples in the text; lower case derivative work is allowed under the condition Ángel Gallego Bartolomé Phone: +34 93 58110 22. of non making a commercial use. The original (Universitat Autònoma de Barcelona) Fax: +34 93 5813239 Roman for examples in footnotes). Example work cannot be used with commercial Jesús Jiménez Martínez (Universitat de València) E-mail: [email protected] texts should start at the first tab. If there is Sandra Montserrat Buendía (Universitat d’Alacant) http://publicacions.uab.cat more than one example within a number, purposes». Therefore, everyone who sends a Xavier Villalba Nicolás ISSN 1695-6885 (in press); ISSN 2014-9719 (online) lower case alphabet letters should be used manuscript is explicitly accepting this publica- (Universitat Autònoma de Barcelona) Dipòsit legal: B-4.112-2003 followed by a stop and a tab. Examples must tion and edition cession. In the same way, Imprès a Espanya. Printed in Spain be referred to as (1), (5a,b), (6b-d), (ii), (iii he/she is authorizing CATALAN JOURNAL OF Imprès en paper ecològic c-e), (iv a,b), etc. LINGUISTICS to include his/her work in a journal’s issue for its distribution and sale. The Bases de dades en què CATALAN JOURNAL OF LINGUISTICS està referenciada • Glosses. Examples from languages other cession allows CATALAN JOURNAL OF LINGUISTICS — Bibliographie Linguistique — Linguistic Bibliography than English should be accompanied by a to publish the work in a maximum period of two — CARHUS Plus+ — Linguistic Working Papers Directory gloss and, if necessary, by a translation into years. With the aim of favouring the diffusion — CiteFactor — Linguistics Abstracts English. Align glosses with examples word of knowledge, CATALAN JOURNAL OF LINGUISTICS — Dialnet — LLBA (Linguistic and Language Behavior Abstracts) by word. Translations should be in simple joins the Open Access journal movement — Francis — Scopus quotation marks. — Índice Español de Ciencias Sociales (DOAJ), and delivers all its contents to different y Humanidades (ISOC) • References. References should be according repositories under this protocol; therefore, to the following pattern, with full names of sending a manuscript to the journal also entails CATALAN JOURNAL OF LINGUISTICS es publica sota el sistema de llicències Creative Commons segons la modalitat: authors included. Reference entries for the explicit acceptation by its author/s of this Reconeixement - NoComercial (by-nc): Es permet la generació d’obres derivades multiple works by the same author(s)/editor(s) distribution method. sempre que no se’n faci un ús comercial. Tampoc es pot utilitzar l’obra original amb finalitats comercials. Catalan Journal of Linguistics 15, 2016 1-3
Summary Catalan Journal of Linguistics = CatJL Volume 15, pp. 1-171, 2016 ISSN 1695-6885 (in press); ISSN 2014-9719 (online) http://revistes.uab.cat/catJL
5-7 Bonet, Eulàlia (Universitat Autònoma de Barcelona); Torres-Tamarit, Francesc (UMR 7023 [Paris 8/CNRS]) Introduction. Catalan Journal of Linguistics, 2016, vol. 15, pp. 5-7.
Articles 9-25 Baković, Eric (UC San Diego) Exceptionality in Spanish Stress. Catalan Journal of Linguistics, 2016, vol. 15, pp. 9-25. Stress in vowel-final non-verbs in Spanish regularly falls on the penultimate syl- lable, while stress in consonant-final words regularly falls on the final syllable. There are two main classes of exceptions to this regularity: stress on the sylla- ble preceding the regular one, and stress on the syllable following the regular one. Harris (1983) provides arguments that the second class of exceptions is morphologically systematic, but falls short of the stronger claim that this pattern is simply a subcase of the regular stress pattern. I argue here that there is much to be gained from this stronger claim, including a simple and elegant analysis of the first class of exceptions. Keywords: Spanish; stress; exceptions; derivational stem; word
27-51 Mascaró, Joan (Universitat Autònoma de Barcelona. Centre de Lingüística Teòrica) Morphological Exceptions to Vowel Reduction in Central Catalan and the Problem of the Missing Base. Catalan Journal of Linguistics, 2016, vol. 15, pp. 27-51. A certain class of Central Catalan compounds characterized by a first component that lacks a related output word is discussed and analyzed in connection with Vowel Reduction and Destressing. The first component of these compounds contains a vowel [a, ɛ, ɔ] that undergoes Destressing before the stressed vowel of the second component, but does not reduce. This causes an opacity problem because the generalization that there are no unstressed [a, ɛ, ɔ] is not surface-true for these cases. An analysis in the framework of parallel OT with output-to- output constraints is examined in detail and the function that returns the base of the relevant constituent in the candidates being evaluated is made precise. It is 2 CatJL 15, 2016 Summary
shown that such an analysis is not feasible. After showing that compounds have internal constituent structure even under noncompositional semantics, a Stratal OT analysis is presented that can handle such cases. Keywords: compounds; Central Catalan; Vowel Reduction; Destressing; opacity; output-to-output constraints; Stratal OT; compositionality
53-66 Moore-Cantwell, Claire (Yale University); Pater, Joe (University of Massachusetts Amherst) Gradient Exceptionality in Maximum Entropy Grammar with Lexically Specific Constraints. Catalan Journal of Linguistics, 2016, vol. 15, pp. 53-66. The number of exceptions to a phonological generalization appears to gradiently affect its productivity. Generalizations with relatively few exceptions are relatively productive, as measured in tendencies to regularization, as well as in nonce word productions and other psycholinguistic tasks. Gradient productivity has been previously modeled with probabilistic grammars, including Maximum Entropy Grammar, but they often fail to capture the fixed pronunciations of the existing words in a language, as opposed to nonce words. Lexically specific constraints allow existing words to be produced faithfully, while permitting variation in novel words that are not subject to those constraints. When each word has its own lexically specific version of a constraint, an inverse correlation between the number of exceptions and the degree of productivity is straightforwardly predicted. Keywords: exceptions; variation; computational phonology; Maximum Entropy Grammar; indexed constraints
67-100 Piñeros, Carlos-Eduardo (University of Auckland) Exceptional nasal-stop inventories. Catalan Journal of Linguistics, 2016, vol. 15, pp. 67-100. This article explores the topic of exceptionality in phonology focusing on nasal-stop inventories. A meticulous survey shows that it is normal for such systems to include at least two anterior units: /m/ and /n/. The finding that the introduction of /ŋ/, the first posterior unit to appear, normally follows that of both anterior units suggests that anteriority is somehow more compatible with this consonant class; however, this hypothesis is challenged by exceptions: /n/ is occasionally superseded by /ŋ/. The proposed analysis overturns the view that there is a single universal place hierarchy. It demonstrates that languages assess the cost of place features on multiple dimensions and that nasal-stop inventories are shaped by the conflict between three evaluation measures: one for articulatory cost, one for perceptual cost, and one for dispersion. A theory of far greater explanatory power emerges when each evaluation measure is empirically substantiated and their universality is strictly respected. Keywords: sound inventories; nasal stops; place features; exceptions; place hierarchies; universal rankings; articulatory cost; perceptual cost; dispersion Summary CatJL 15, 2016 3
101-119 Rebrus, Péter (Hungarian Academy of Sciences); Szigetvári, Péter (Eötvös Loránd University) Diminutives: Exceptions to Harmonic Uniformity. Catalan Journal of Linguistics, 2016, vol. 15, pp. 101-119. Front-back harmony in Hungarian is governed by the root of words: it disregards suffixes in most cases. That is, a back-vowelled root (B) followed by any number of neutral-vowelled suffixes (N) will take a back-vowelled suffix (B+N+N+B), but a root with a back vowel followed by several neutral vowels is possibly followed by a front-vowelled suffix (F): BNN+F/B. We call this Harmonic Uniformity. This is respected even in truncated stems: NB→N+N+B, although NN stems practically never take a back suffix (NN+F). Diminutive forms are the only exceptions to this pattern. We claim that this is so, because diminutive forms are much more loosely related to their “base” than is the case with any other types of suffixation. Keywords: morphology; vowel harmony; Hungarian; diminutives; truncation
121-143 Rysling, Amanda (University of Massachusetts Amherst) Polish yers revisited. Catalan Journal of Linguistics, 2016, vol. 15, pp. 121-143. It is common in linguistic research to attempt a unified analysis for similar patterns in related languages. In this paper, I argue that to do so for Polish and Russian vowel alternations would be a mistake. Although they share some notable phonological properties, they differ in their prevalence and their extensibility. I present an account of Polish under which vowel alternations are unexceptional, and exceptional blocking of alternation is achieved with lexically indexed constraints. This is the complement of Gouskova’s (2012) account of Russian, which I argue to be desirable on the basis of novel corpus statistics from the Polish lexicon and their divergences from the trends for analogous words in Russian. Keywords: Polish; Slavic; yers; vowel alternations; lexical indexation
145-171 Zuraw, Kie (University of California, Los Angeles) Polarized Variation. Catalan Journal of Linguistics, 2016, vol. 15, pp. 145-171. In cases of exceptionality, there are usually many words that behave regularly, a smaller number that behave irregularly (the exceptions), and perhaps an even smaller number whose behavior varies. This paper presents several examples of exceptionality and variation that are polarized in this way: most items exhibit one behavior or the other consistently, with only a minority of items showing variation. The result is a U-shaped histogram of behavior rates. In some cases, this requires listing of surprisingly long units. There are, however, some cases of bell-shaped histograms, where most items show variation, and only a minority are consistent. Some simple simulations are presented to show how polarized variation can result when variation is between two categorical outcomes, and both types of variation can result when variation is along a phonetic continuum. Keywords: exceptions; variation; corpus; diachronic change
ISSN 1695-6885 (in press); 2014-9719 (online) Catalan Journal of Linguistics 15, 2016 5-7
Introduction
Eulàlia Bonet Universitat Autònoma de Barcelona [email protected] Francesc Torres-Tamarit UMR 7023 (Paris 8/CNRS) [email protected]
The origin of this volume is the Workshop on Exceptionality in Phonology, which was held at the Universitat Autònoma de Barcelona the 27th of January 2015 as part of the 12th Old World Conference in Phonology, and which featured Kie Zuraw (UCLA) as keynote speaker. From the seven talks that were presented at the Workshop, six appear in this volume, and one more paper that was not presented at the Workshop has also been included. The papers presented in this monographic volume are not meant to cover the totality of theoretical approaches and research activities on the topic at hand, which is certainly a very broad topic, but to present up-to-date work by leading phonologists on a topic that is a long-debated problem for phonological theory. The issue of exceptionality in phonology generally refers to the situation in which an otherwise robust linguistic generalization fails to apply to the entire lexi- con. This kind of exceptionality is lexical and categorical, in that a pre-defined set of lexical items never undergo certain phonological processes that apply regularly to most items. Exceptional behavior can also be sensitive to the distinction between underived and derived environments, whether these are derived by virtue of the application of a phonological process or a morphological operation. It is a well- known fact that certain phonological processes only apply in derived environments, which implies that those processes are blocked in underived monomorphemic words (see, among others, Kiparsky 1993, Anttila 2009, Burzio 2011). However, it is also true that some processes fail to apply precisely in (at least phonologically) derived environments, meaning that only derived environments are instances of an exceptional behavior (see Hall 2006, for instance). A different kind of exceptional- ity is revealed by phonological variation. It is not always the case that exceptional behavior is categorical, but sometimes a number of words behaves both regularly and exceptionally. Finally, exceptionality in phonology can also refer to excep- tional, in the sense of typologically rare, sound inventories. Since the early days of generative phonology, exceptions in phonology have basically been accounted for in two different ways that we can call indexical and representational. The first type of approach makes use of diacritics that index a particular exceptional item to a phonological rule or constraint. The second type 6 CatJL 15, 2016 Eulàlia Bonet; Francesc Torres-Tamarit of approach, less often explored, encodes exceptionality directly into the lexical representation of the item by means of phonological structure that is different in some respect in regular items. Within Optimality Theory (OT), the indexical approach has taken one of two forms. One of them employs diacritics that specify exceptional items for a specific ranking of constraints that is different from the general ranking of the language, that is, co-phonologies selected by specific morphemes (Inkelas et al. 1997, Itô & Mester 1995, see Anttila 2002 for lexical items specified to select a particular ranking of unranked constraints). The other indexical approach has relied on lexically indexed constraints, which occupy a different position in the hierarchy with respect to the general, non-lexically indexed constraints from which they derive (Pater 2000). For a more extended discussion of different approaches to exceptionality in generative phonology, see Wolf (2011). In this volume, the lexically indexed constraint approach is taken by Baković, Moore-Cantwell & Pater, and Rysling. Baković’s paper on “Exceptionality in Spanish stress” shows that taking the stem as the domain for stress assignment in Spanish allows for a straighforward analysis of exceptional stress patterns by invoking the lexically indexed constraint Non-Finality. In Moore-Cantwell & Pater’s paper on “Gradient exceptionality in Maximum Entropy Grammar with lexically specific constraints”, it is shown that only a probabilistic grammar that incorporates lexically indexed constraints can model successfully both gradient productivity in nonce words, which depends on the number of exceptions to a phonological generalization, and the stable pronunciation of existing words in a language. Finally, in “Polish yers revisited”, Rysling relies on lexically indexed constraints to block vowel-zero alternations in Polish and supports her formalization on the basis of corpus statistics. Zuraw’s paper is an investigation of “Polarized variation”, that is, scenarios of exceptionality in which there is a small number of items that behave variably as both regular and exceptional items, as opposed to scenarios in which most of the items behave variably. By means of computer simulations, Zuraw shows that polarized variation in fact results from the existence of variation between two categorical outcomes. Exceptionality as evidence for serial versus parallel models of OT is addressed in Mascaró’s paper on “Morphological exceptions to vowel reduction in Central Catalan and the problem of the missing base”. By exploring the underapplication pattern of vowel reduction in Catalan in compound structures whose first com- ponent lacks a related base, Mascaró shows that internal constituent structure is necessary even under noncompositional semantics, and that only Stratal OT, as opposed to output-output constraints, can handle such cases of underapplication of vowel reduction. Rebrus & Szigetvári’s paper on “Exceptions to harmonic uni- formity” also deals with the relation between morphological constituent structure and phonological exceptionality. In this paper, Rebrus & Szigetvári describe a case of blocking of vowel harmony in Hungarian that is only found in diminutive forms, giving empirical evidence for the existence of morphologically derived environment blocking. In a different vein, Piñeros’s paper on “Exceptional nasal- stop inventories” shows that the shape of nasal inventories in languages of the Introduction CatJL 15, 2016 7 world is the result of conflicting forces on place features: ease of production, ease of perception, and dispersion. The papers presented in this volume range from accurate descriptions of phenomena of particular languages to technically accurate OT analyses that build on theoretically sound concepts, as well as work on computational mod- eling of gradient productivity and patterns of variation. We are confident that this volume provides a state-of-the-art update of studies on exceptionality in phonology and hope that it inspires others to further explore the nature of excep- tions in phonology.
References Anttila, Arto. 2002. Morphologically conditioned phonological alternations. Natural Language & Linguistic Theory 20: 1-42. Anttila, Arto. 2009. Derived environment effects in colloquial Helsinki Finnish. In Kristin Hanson & Sharon Inkelas (eds.). The nature of the word: Studies in honor of Paul Kiparsky, 433-460. Cambridge & London: The MIT Press. Burzio, Luigi. 2011. Derived environment effects. In van Oostendorp, Marc; Ewen, Colin J.; Hume, Elizabeth & Rice, Keren (eds.). The Blackwell Companion to Phonology, 2089-2114. Wiley-Blackwell. Hall, Tracy A. 2006. Derived environment blocking effects in Optimality Theory. Natural Language & Linguistic Theory 24: 803-856. Inkelas, Sharon; Orgun, Orhan & Zoll, Cheryl. 1997. The implications of lexical excep- tions for the nature of grammar. In Iggy Roca (ed.). Derivations and Constraints in Phonology, 393- 418. New York: Oxford University Press. Itô, Junko & Mester, Armin. 1995. The core-periphery structure in the lexicon and con- straints on re-ranking. In Beckman, Jill; W. Dickey, Laura & Urbanczyk, Suzanne (eds.). Papers in Optimality Theory, 181-210. GLSA: Amherst. Kiparsky, Paul. 1993. Blocking in non-derived environments. In Hargus, Sharon & Kaisse, Ellen M. (eds.). Studies in Lexical Phonology, 277-313. San Diego: Academic Press. Pater, Joe. 1994. Against the underlying specification of an ‘exceptional’ English stress pattern. In Dyck, Carrie (ed.). Toronto Working Papers in Linguistics 13: 95-121. Pater, Joe. 2000. Nonuniformity in English stress: The role of ranked and lexically specific constraints. Phonology 17: 237-274. Wolf, Matthew. 2011. Exceptionality. In van Oostendorp, Marc; Ewen, Colin J.; Hume, Elizabeth & Rice, Keren (eds.). The Blackwell Companion to Phonology, 2538-2559. Wiley-Blackwell.
ISSN 1695-6885 (in press); 2014-9719 (online) Catalan Journal of Linguistics 15, 2016 9-25
Exceptionality in Spanish Stress*
Eric Baković UC San Diego [email protected]
Received: February 25, 2016 Accepted: July 8, 2016
Abstract
Stress in vowel-final non-verbs in Spanish regularly falls on the penultimate syllable, while stress in consonant-final words regularly falls on the final syllable. There are two main classes of excep- tions to this regularity: stress on the syllable preceding the regular one, and stress on the syllable following the regular one. Harris (1983) provides arguments that the second class of exceptions is morphologically systematic, but falls short of the stronger claim that this pattern is simply a subcase of the regular stress pattern. I argue here that there is much to be gained from this stronger claim, including a simple and elegant analysis of the first class of exceptions. Keywords: Spanish; stress; exceptions; derivational stem; word
Resum. Excepcionalitat en l’accent de l’espanyol
L’accent en les formes no verbals de l’espanyol acabades en vocal cau regularment sobre la penúltima síl·laba, mentre que l’accent en paraules acabades en consonant cau regularment sobre l’última síl·laba. Hi ha dues classes principals d’excepcions a aquesta regularitat: accent en la síl·laba que precedeix la que tindria l’accent regularment i accent en la síl·laba que segueix la que el tindria regularment. Harris (1983) argumenta que la segona classe d’excepcions és morfològi- cament sistemàtica, però no arriba a defensar la posició més forta que aquest patró és simplement un subcas del patró regular. Aquí argumento que s’hi pot guanyar molt amb aquesta afirmació més forta, incloent-hi una anàlisi més senzilla i elegant de la primera classe d’excepcions. Paraules clau: espanyol; accent; excepcions; radical derivatiu; mot
* For comments on, suggestions for, and critiques of this work, I thank Brett Hyde, Ricardo Bermúdez-Otero, Rafael Nuñez Cedeño, the audience at the OCP 12 Workshop on Exceptionality in Phonology (especially Kie Zuraw and Joe Pater), and four anonymous reviewers. Remaining errors are mine. Harris (1995: 867) notes that the topic of stress had “all but overwhelmed the field of Spanish phonology” in the preceding decade or so, and only a small sliver of that work, and of work before and since, is cited here. I also take full responsibility for this failing. 10 CatJL 15, 2016 Eric Baković
Table of Contents 1. Introduction 4. Class (1) exceptions 2. Regular stress and Class (2) 5. Constraint indexation analysis ‘exceptions’ 6. Concluding remarks 3. The domain of regular stress References
1. Introduction This paper focuses on the placement of main stress in a significant subset of Spanish non-verbs, henceforth ‘words’. Stress in vowel-final words in Spanish regularly falls on the penultimate syllable (e.g. cuádro ‘frame’), while stress in consonant-final words regularly falls on the final syllable (e.g. ladrón ‘burglar’).1 There are two main classes of exceptions to this regularity:
(1) stress on the syllable preceding the regular one (i.e., antepenultimate stress in vowel-final words, penultimate stress in consonant-final words), and (2) stress on the syllable following the regular one (i.e., final stress in vowel-final words).
A third class of exceptions consists of a handful of relatively uncommon words (e.g. régimen, espécimen, ínterin), plus some loanwords and proper/place names discussed in §3.2, that distinguish themselves further by ending in a consonant but having antepenultimate stress. These words are thus exceptions to the class of exceptions noted in (1), and are identified as ‘Class X’ in the summary table in (3). I’ll have nothing substantive to say about Class X words in this paper; see instead Roca (1988, 2005, 2006) and Oltra-Massuet & Arregi (2005).
(3) Exceptional Regular Class (1) Class (2) Class X penultimate antepenultimate final V-final …σ′σ# …σ′σσ# …σ′# words sabána ‘savannah’ sábana ‘sheet’ Panamá ‘Panama’ final penultimate antepenultimate C-final …σ′# …σ′σ# …σ′σσ# words animál ‘animal’ caníbal ‘cannibal’ régimen ‘diet’
1. Spanish orthography is used throughout; main stress is always indicated. Words ending in the plural suffix -(e)s are systematically stressed on the same syllable as the corresponding singular: cuádro ~ cuádros, ladrón ~ ladrónes. Notable exceptions are a few examples with ‘moveable stress’ like régimen ~ regímenes ‘diet(s)’ and carácter ~ caractéres ‘character(s)’; see the Appendix of Harris (1983) as well as Roca (1988, 2005, 2006) for discussion. Exceptionality in Spanish Stress CatJL 15, 2016 11
Harris (1983) provides compelling arguments for his claim that Class (2) excep- tions are morphologically systematic, but falls short of the stronger claim that the pattern exhibited by this class is simply a subcase of the regular stress pattern. This stronger claim is summarized in the table in (4), in which Class (2) ‘exceptions’ are now listed as a subcase of the regular pattern.
(4) Exceptional Regular Class (1) Class X penultimate final antepenultimate V-final …σ′σ# …σ′# …σ′σσ# words sabána ‘savannah’ Panamá ‘Panama’ sábana ‘sheet’ final penultimate antepenultimate C-final …σ′# …σ′σ# …σ′σσ# words animál ‘animal’ caníbal ‘cannibal’ régimen ‘diet’
I argue here that there is much to be gained from the stronger claim sum- marized in (4), including a simple and elegant analysis of Class (1) exceptions. The argument is built up without particular theoretical commitments in sections 2 through 4, establishing in particular that Spanish stress assignment is quantity insensitive (see also Ohannesian 2004; Piñeros, forthcoming) and that its domain is the derivational stem rather than the word (Hooper & Terrell 1976; Roca 1988, 2005, 2006). This is followed in section 5 by an analysis of Class (1) exceptions in terms of lexically-indexed constraints (Pater 2000, 2010) within Optimality Theory (Prince & Smolensky 2004). The commitment to the stronger claim in (4) is shared by Roca (2006), who also employs lexically-indexed constraints in his account of Class (1) exceptions and further extends their use to account for Class X exceptions. This extension, as well as other differences in assumptions and expositional choices, lead Roca (2006) to a very different analysis than the one presented here. The present work is distinguished by its focus on the details of the consequences of the stronger claim in (4) for the analysis of Class (1) exceptions without the possibly confounding burden of aiming in addition for broader empirical coverage by tackling Class X exceptions as well.2
2. Regular stress and Class (2) ‘exceptions’ 2.1. Regular stress: vowel-final words with penultimate stress We start with Harris’s (1983: 91) key observation that vowel-final words with penultimate stress end in a terminal element (TE), generally a vowel -o, -a, or -e
2. An anonymous reviewer also points me to the indexed constraint analysis of Meinschaefer (2015), which differs in crucial respects from the analysis of Roca (2006) but aims for similarly broad empirical coverage. 12 CatJL 15, 2016 Eric Baković
(but sometimes a -Vs sequence) and typically associated with gender marking. The TE is always the last morpheme before the (inflectional) plural suffix. In the morphological analysis of Harris (1983), a TE is affixed directly to a derivational stem (henceforth referred to simply as ‘stem’). In the examples below, square brackets indicate word boundaries (preceded by the TE) and parentheses indicate stem boundaries.3 Note that the form of the TE associated with a given lexical root may vary in different morphological contexts; e.g. there is an -o in máno ‘hand’ but an -a in maníta ‘hand (dim.)’.4
(5) a. [(páp)-a] ‘potato’ f. [(pap+ít)-a] ‘potato (dim.)’ b. [(mán)-o] ‘hand’ g. [(man+ít)-a] ‘hand (dim.)’ c. [(bandíd)-o] ‘bandit’ h. [(bandid+ít)-o] ‘bandit (dim.)’ d. [(candidát)-o] ‘candidate’ i. [(candidat+ít)-o] ‘candidate (dim.)’ e. [(léj)-os] ‘far’ j. [(lej+án)-o] ‘distant’
2.2. Regular stress: consonant-final words with final stress Stress on consonant-final words is regularly final, as shown in (6a-c). Harris (1983: 91) argues that these words don’t have a TE; under derivational affixation (6d-f), a TE is added and stress once again follows the penultimate pattern.
(6) a. [(paréd)] ‘wall’ d. [(pared+cít)-a] ‘wall (dim.)’ b. [(salón)] ‘lounge’ e. [(salon+cít)-o] ‘lounge (dim.)’ c. [(animál)] ‘animal’ f. [(animal+ít)-o] ‘animal (dim.)’
2.3. Class (2) ‘exceptions’: vowel-final words with final stress Harris (1983: 116-119) further claims that vowel-final words with final stress (7a-c) also lack a TE, just like the consonant-final words in (6a-c). Support for this view comes from the fact that the final vowel of unaffixed words remains under derivational affixation (7d-f), after which other TEs are added — again, just like the consonant-final words in (6d-f).5
3. I represent the boundary between stems and derivational suffixes with ‘+’ rather than with the recursive stem structure given in Harris (1983: 92); I am agnostic on this point of morphological structure. 4. Several anonymous reviewers remind me that diminutive formation is not as straightforward a diagnostic of the relation between TEs and the derivational stem as is implied in the text; see e.g. Smith (2011) and references therein. One reviewer cites examples with -e such as cálle ‘street’ ~ callecíta ‘street (dim.)’, which Colina (2003) takes as evidence that the -e is (or has been reanalyzed as) epenthetic, not a TE. Still, the reviewers’ point is well taken. 5. The invariable presence of the stem-final vowel in this class of forms is not exceptionless, however; compare (7c,f) with [(Panamá)] ‘Panama’ ~ [(Panam+éñ)-o] ‘Panamanian’. This can be reasonably attributed to vowel deletion under hiatus (cf. [(habl+a)-r] ‘to talk’ ~ [(habl)-o] ‘I talk’); in the case Exceptionality in Spanish Stress CatJL 15, 2016 13
(7) a. [(café)] ‘coffee’ d. [(cafe+cít)-o] ‘coffee (dim.)’ b. [(dominó)] ‘domino’ e. [(domino+cít)-o] ‘domino (dim.)’ c. [(Perú)] ‘Peru’ f. [(Peru+án)-o] ‘Peruvian’
3. The domain of regular stress 3.1. Stem vs. word All of the examples illustrated in (5), (6), and (7) can be taken to be subcases of a regular stress pattern in Spanish, the generalization for which is stated in (8). This generalization can be credited to Hopper & Terrell (1976); see also Roca (1988, 2005, 2006).
(8) Stress is final in the stem.
Note that (8) accounts for penultimate stress in TE-final words (5), (6d-f), and (7d-f) as well as for final stress in words without a TE, whether consonant-final (6a-c) or vowel-final (7a-c). Despite the generality of this analysis, Harris (1983: 94-95) rejects (8) in favor of (9).
(9) Stress is penultimate in the word.
The key difference here is that (8) claims the domain of stress to be the stem, while (9) claims it to be the word. (9) must thus state that stress is penultimate, to ignore the TE that separates the right edge of the stem from the right edge of the word. This accounts for regular penultimate stress in the TE-final words in (5), (6d-f), and (7d-f).
3.2. Quantity sensitivity But of course an analysis in terms of (9) must be supplemented in order to account for final stress in words without a TE. Let’s start with the consonant-final words in (6a-c). In order to account for these, the following further condition is introduced, paraphrased from Harris (1983: 111).
(10) Branching rhymes must be stressed.
In Harris’s (1983: 22-31) analysis of the syllable structure of Spanish, closed syllables (syllables with VC rhymes) have a branching rhyme. Consonant-final words (6a-c) must thus be stressed on the final syllable by (10), overriding the penultimate stress demanded by (9).
of (7f), the final u of the derivational stem is high and can thus be glided instead. An anonymous reviewer helpfully points me to Bermúdez-Otero (2006, 2013) for a different morphophonological analysis of stem-final vowel deletion, based on a very different morphological analysis of deriva- tional affixation; see now also Myler (2015) for arguments against this morphological analysis. 14 CatJL 15, 2016 Eric Baković
On the face of it, the generalization in (8) is more straightforward than the one in (9), which requires an overriding condition (10). However, Harris argues that (10) is independently required to explain the placement of stress in Spanish. In addition to syllables with VC rhymes, there are two other types of syllables with branching rhymes in Spanish: syllables with rising-sonority diphthongs (V̯ V rhymes, as in cuádro ‘frame’) and syllables with falling-sonority diphthongs (VV̯ rhymes, as in cáusa ‘cause’). Harris (1983: 88-90) asserts that all three of these branching-rhyme types, when in penultimate position, absolutely preclude antepenultimate stress (e.g. molésto, *mólesto ‘annoying’; manióbra, *mániobra ‘maneuver’; dinosáurio, *dinósaurio ‘dinosaur’) — that is, that vowel-final words with penultimate branching rhymes cannot be Class (1) exceptions. This fact is the independent evidence adduced for the necessity of (10). There are several well-known difficulties with (10), however, which I briefly note here along with comments on their varying degrees of potential surmount- ability.
(i) As pointed out by Roca (1988: 416-418), there do in fact exist some words with penultimate branching rhymes but antepenultimate stress (e.g. Frómista, Róbinson, Wáshington, Mánchester). But these are rare, largely loanwords, and primarily proper/place names.6 (ii) Word-final VV̯ rhymes, but not V̯ V or VC, also absolutely preclude penul- timate stress (e.g. convóy, *cónvoy ‘convoy’) — that is, there are no Class (1)-style exceptions for VV̯ -final words. This requires a separate override for (9), to be discussed in §3.3 below. (iii) Word-final V̯ V rhymes also appear to absolutely preclude antepenultimate stress, even if they themselves are not stressed (e.g. princípio, *príncipio ‘beginning’), which (10) cannot account for. Harris (1983: 131-135) also notes that VC-final words with antepenultimate stress (e.g. régimen ‘regi- men’, espécimen ‘specimen’, ínterin ‘interim’) are rare, and on the basis of their “movable stress” (e.g. especímenes ‘specimens’) he concludes that these forms are “extragrammatical” (Harris 1983: 132); this is also not accounted for by (10).7 (iv) Antepenultimate stress is also (judged to be) impossible in words the final onset of which is a trill (e.g. chamárra, *chámarra ‘jacket’). Harris (1969, 1983, 2001, 2002) analyzes the trill as (underlyingly) geminate for a variety of
6. Harris (1983: 148, note 10) notes the word alícuota ‘aliquot’, dismissing it as a unique form that “the overwhelming majority of speakers have never encountered” but also speculating that “we are [perhaps] dealing with the rounded velar segment kw rather than the sequence ku̯ .” There is no such speculation available for the words in (i). 7. In this case, Harris (1983: 148, note 10) points to the unique (but not necessarily infrequent) word ventrílocuo ‘ventriloquist’, again dismissing it but speculating that it might end in kwo, not ku̯ o. An anonymous reviewer points out that words such as análisis ‘analysis’ and páncreas ‘pan- creas’ also fall into the class of VC-final words with antepenultimate stress. The final VC in these cases is a TE (cf. [(analis+áz)-o] ‘analysis (aug.)’, [(pancrea+tít)-is] ‘pancreatitis’), and these thus represent straightforward Class (1) exceptions. Exceptionality in Spanish Stress CatJL 15, 2016 15
reasons, including this one. But this analysis does not (independently) extend to palatals, all of which also appear to preclude antepenultimate stress from the same position (e.g. campáña, *cámpaña ‘campaign’; cabállo, *cáballo ‘horse’; cobáyo, *cóbayo ‘guinea pig’; estúche, *éstuche ‘case’).8 (v) Syllables with branching rhymes in antepenultimate position are expected by (10) to be stressed — especially when both the penultimate and final syllable rhymes are not branching — but alongside náufrago ‘castaway’ (antepenul- timate VV̯ rhyme), periódico ‘newspaper’ (antepenultimate V̯ V rhyme), and báscula ‘scales’ (antepenultimate VC rhyme), we have bautízo ‘christening’, bibliotéca ‘library’, and castígo ‘punishment’.9 These latter examples are (gen- erally taken to be) regular, the former Class (1) exceptions.
In sum, while the evidence for quantity sensitivity in Spanish stress is highly suggestive, it is not as solid (nor as internally consistent) as it is frequently made out to be. The argument that (10) is independently motivated, then, is without much force; the stem-domain analysis in (8) thus has more to recommend it than the word-domain (9)+(10) analysis.
3.3. Coextensive domains Even with the help of (10), (9) does not account for final stress in vowel-final words (7a-c). So why does Harris reject (8) in favor of (9)? There are two main arguments. The first argument has to do with the cyclic assignment of Spanish stress, which I will not address here. The second, stronger argument has to do with the consist- ently penultimate stress of (bisyllabic) prepositions (as pronounced in isolation, (11)) and productive truncations (12). Stress in these two sets of forms is clearly penultimate in the word and clearly not final in the stem, hence they are consistent with (9) (with or without (10)) and not with (8).
(11) a. [pára] ‘for’ c. [désde] ‘since’ b. [hásta] ‘until’ d. [sóbre] ‘over’
(12) a. [(múñe)] ‘doll’ (< muñéca) c. [(prófe)] ‘professor’ (< profesór) b. [(árqui)] ‘architect’ (< arquitécto) d. [(Máuri)] ‘Maurice’ (< Maurício)
Note that I follow Harris (1983: 94) in assuming that prepositions “have no derivational stem [because] there is no deprepositional or adprepositional deriva-
8. See Baković (2009) for a thorough review and critique of the various arguments that Harris has pre- sented in favor of the analysis of the trill as underlyingly geminate. Lipski (1990: 168) and Hualde (2004: 391) have pointed out that all of these apparently stress-attracting singleton consonants are historically derived from geminates and heterosyllabic clusters in Latin, which would have been expected to attract stress; see also Roca (1988: 417). 9. The only branching antepenultimate rhymes that appear to allow final stress are VC rhymes (e.g. bisturí ‘scalpel’). 16 CatJL 15, 2016 Eric Baković tional morphology”.10 Truncations, on the other hand, are assumed here to have such structure because they can be further derived; e.g., [(profe+cít)-o] ‘professor (dim.)’. (Harris (1983) says nothing about the possible internal structure of trunca- tions.) The relevance of this point will be clear in a moment. In order to bring final stress on the vowel-final forms in (7a-c) into the fold, then, Harris (1983: 118) proposes the rule paraphrased in (13) which, like (10), is in addition to and overrides (9).
(13) A rhyme ending in a vocoid (= a [–consonantal] segment) that is final both in the stem and the word must be stressed.
Because prepositions (11) have no stem, (13) is irrelevant to them and so they are correctly assigned penultimate stress (when pronounced in isolation) by (9). But truncations (12) do have a stem, and so they would be incorrectly assigned final stress by (13). In other words, something additional must be said about truncations under either the stem-domain analysis in (8) or Harris’s (9)+(13) word-domain alternative, both of which incorrectly predict that truncations should be stressed on the final syllable. A likely candidate for that additional something is that trun- cation fits a(n unmarked) bisyllabic trochee template (Prieto 1992; Lipski 1995; Colina 1996; Piñeros 1998, 2000a, 2000b; Feliú 2001), but see the end of §5.4 for an alternative possibility. This leaves prepositions as pronounced in isolation, about which Harris (1983: 94) has the following to say: “Normally — that is, in nonmetalinguistic discourse — prepositions are stressless in Spanish; they are proclitics to their object. […] It might be countered that penultimate main stress of prepositions in isolation is (somehow) a reflection of their usual nonprimary stress in phrases”. I accept this counterargument and claim, without further discussion, that the stress of preposi- tions as pronounced in isolation (11) is irrelevant to the choice between (8) and (9)+(13). Given that truncations (12) require something additional under either analysis, this leaves the fact that (8) accounts directly for final stress in vowel-final forms (7a-c) while Harris’s alternative has to be complicated with (13) in addition to (9) in order to account for these forms. But Harris (1983: 118-119) points out “an unexpected bonus” provided by (13): the fact noted in (ii) above, that VV̯ -final words have no Class (1)-style excep- tions (i.e., that word-final VV̯ rhymes, such as in convóy ‘convoy’, are stressed without exception). Recall from §2.1 that TEs are either vowels (-o, -a, or -e) or, sometimes, -Vs sequences. Words ending in VV̯ rhymes thus end in sequences that cannot be (and demonstrably are not) TEs, which means that the final syllable of such words — being both word-final and stem-final — will be correctly stressed
10. An anonymous reviewer notes that “it seems that (at least complex etymologically denominal) prepositions like delante de or frente a may, at least in certain varieties, undergo diminutive for- mation: delantito de, frentito a.” As the reviewer notes, this supports the analysis pursued here. While I agree with the reviewer that “[this] paper might profit from a more thorough discussion of Spanish morphological structure”, I defer to the limitations of space. Exceptionality in Spanish Stress CatJL 15, 2016 17 by (13). Even if the final V̯ of any of these words is marked as extrametrical (as in Harris’s analysis of Class (1) exceptionality; see §4 below), the preceding V satis- fies the structural description of (13) and stress is still assigned on the final syllable. I submit the following alternative analysis of VV̯ -final words, consistent with (8). First, I assume that — for the purposes of stress assignment, at least — VV̯ rhymes are treated as VV sequences. (Note that this is also a covert assumption behind (13), which refers to [–consonantal] vocoids, not vowels more specifically.) If the stress pattern of a VV̯ -final word is regular, stress is assigned by (8) to the final second vocoid, but because the sequence is ultimately realized monosyllabi- cally as VV̯ , the stress is realized on the more sonorous first vocoid. If the stress pattern of the word is Class (1) exceptional, stress is assigned to the penultimate first vocoid, where it is also realized. Antepenultimate stress when a word ends in a VV̯ rhyme is thus precluded.11
3.4. Verdict: the domain of stress in Spanish is the stem I conclude this section by accepting the stem-domain analysis in (8) as the cor- rect analysis for all of the forms in (5), (6), and (7). The ‘exceptions’ in (7a-c) heretofore classified as Class (2) exceptions are not exceptions at all, not even under the ‘systematic exceptions’ interpretation that is granted to these forms under Harris’s (9)+(13) word-domain alternative — not to mention the fact that the word-domain alternative also has to be complicated by (10) to account for final stress in consonant-final forms (6a-c), something that (8) also accounts for without further ado. As pointed out in this section, however, this analysis is not without its own potential challenges, the most obvious being that it diminishes — if not rejects — the role of quantity sensitivity in the analysis of Spanish stess. If final stress in consonant-final forms (6a-c) is due to the fact that the final syllable in such forms is stem-final and not to the fact that such syllables have branching rhymes, then (10) is unnecessary to account for them. In §3.2 I noted five independent difficulties faced by the assumption that Spanish stress is quantity sensitive; following Roca (1988), my view is that these challenges outweigh those faced by the quantity insensitive alternative advanced here.12 We arrive at the same conclusion, though via different paths: in Roca (1988), the main path is a reassessment of Harris’s (1983) theoretical assumptions about extrametricality (about which see §4 below), whereas here the main path is a reassessment of Harris’s (1983) empirical arguments and their consequences for the basic elements of the analysis.
11. It must of course be the case that final V̯ V rhymes are not also (reliably) treated as VV sequences for the purposes of stress assignment (pace Roca 2006: 252). As mentioned in both notes 6 and 7 above, at least some V̯ V sequences have an alternative segmental (and thus syllabic) analysis that merits more attention than I am in a position to give here, one in which the V̯ is not a vocoid but rather a secondary articulation on the preceding onset consonant. 12. See now also Piñeros, forthcoming. Thanks to an anonymous reviewer for pointing me to this work. 18 CatJL 15, 2016 Eric Baković
4. Class (1) exceptions Acceptance of the stem-domain analysis in (8) leaves the exceptions classified under (1): stress on the syllable preceding the regular one, now understood as stress on the penultimate syllable of the stem. Near-minimal contrasts are given in (14): regularly stressed forms in (14a-e) and Class (1) exceptionally stressed forms in (14f-j).
(14) a. [(molín)-o] ‘windmill’ f. [(cómic)-o] ‘comic(al)’ b. [(sabán)-a] ‘savannah’ g. [(sában)-a] ‘sheet’ c. [(pistól)-a] ‘pistol’ h. [(epístol)-a] ‘epistle’ d. [(animál)] ‘animal’ i. [(caníbal)] ‘cannibal’ e. [(paréd)] ‘wall’ j. [(huésped)] ‘guest’
I take it to be uncontroversial that the contrasting pairs in (14) must be arbitrari- ly distinguished somehow in the lexicon; that is, that Class (1) stress as exemplified by (14f-j) is in fact an exceptional pattern as compared to the regular pattern exem- plified by (14a-e). Harris (1983), armed with the (9)+(10)+(13) word-domain analy- sis, proposes to mark the rightmost segment of the stem of Class (1) exceptions as extrametrical, the ultimate result of this marking being that stress is assigned by (9) one syllable to the left of the regular syllable in (14f-j). There is much more to the (frought) issue of extrametricality as it has been employed in the analysis of Spanish stress than space allows for here; the interested reader should consult, in chronological order (and among others), Harris (1983), Den Os & Kager (1986), Roca (1988), and Harris (1988).13 Most relevant to the present paper is that the (lexically arbitrary) assignment of extrametricality must be restricted to single, stem-peripheral segments. Harris (1983) proposes a set of principles restricting the assignment of extrametricality in Spanish in just this way; the analysis presented in §5 avoids the need to propose any such principles.
5. Constraint indexation analysis 5.1. The constraints I propose here an analysis of the key distinction between the regular and Class (1) exceptional stress patterns in terms of Optimality Theory (OT; Prince & Smolensky 2004) with lexically-indexed constraints (Pater 2000, 2010). A distinguishing fea- ture of this analysis is that it makes no reference to metrical structure, as most if not all work following Harris’s (1983) lead does; it only makes reference to the
13. Both Den Os & Kager (1986) and Roca (1988) aim to fix what they take to be weaknesses in the theory of extrametricality of Harris (1983) by, among other things, setting stress and extrametri- cality to the same domain. For Den Os & Kager (1986), this domain is the word; for Roca (1988), it is the stem. Exceptionality in Spanish Stress CatJL 15, 2016 19 placement of main stress. This may or may not ultimately be a desirable feature of the analysis, but given its relative simplicity, I judge it to be so. The constraint responsible for the regular stress pattern will be referred to here as FinalStress, defined informally in (15) given our acceptance of the stem- domain analysis in (8).
(15) FinalStress (informal version): Stress is final in the stem.
As will become clear in the presentation of the analysis that follows, FinalStress is assumed to be violated more the further away stress is from the end of the stem, with distance measured in terms of syllables. One way to achieve this is by formalizing FinalStress as a ‘horizontally gradient’ (McCarthy 2003: 82) alignment constraint of the kind first discussed in great detail by McCarthy & Prince (1993). Another way is by formalizing FinalStress as a categorical align- ment constraint with the same essential distance-measuring effect, following Hyde (2012). I adopt here the latter type of formalization, given in (16).
(16) FinalStress (formal version): * stem, σ´, σ˘ /(…σ´…σ˘…) ‘Assess a violation mark for each tuple stem, stressed syllable, unstressed syllable such that the stressed syllable⟨ precedes⟩ the unstressed syllable within the stem.’ ⟨ ⟩ The other constraint to be used in the proposed analysis is the OT-constraint counterpart of extrametricality, Nonfinality (Prince & Smolensky 2004; see also now Hyde 2003, 2011). This constraint is defined in (17) such that it penalizes both stem-final and word-final stress, and the latter worse than the former, the need for which will become clear in what follows.14
(17) Nonfinality: Stress is not final in the stem (= 1 violation) or word (= 2 viola- tions).
5.2. The regular pattern
The ranking FinalStress >> Nonfinality accounts for regular penultimate stress in words with a TE, like those in (5), (6d-f), and (7d-f), as illustrated in (18). ⟦ ⟧
14. This is similar but not identical to Prince & Smolensky’s (2004: 62) definition of Nonfinality, which is violated once by a word-final head foot and once more if the head of that head foot is also final. Theirs is thus a ‘vertically gradient’ constraint (McCarthy 2003: 82), but as Prince & Smolensky (2004: 63, note 33) and McCarthy (2003: 84) both note, the intended effect could also be achieved with two categorical constraints, one for the head foot (violated equally by all word-final head feet) and the other for the head of the head foot (= the main-stressed syllable, vio- lated only by word-final main stress). A similar sort of split can be performed on the definition of Nonfinality in (17): one constraint penalizing both stem-final and word-final (= “domain-final”) stress and the other less stringently penalizing only word-final stress. For the sake of simplicity, I don’t pursue this idea further here. (See also note 16.) 20 CatJL 15, 2016 Eric Baković
(18) [(candidat)-o] FinalStress Nonfinality a. ☞ [(candidát)-o] * b. [(candídat)-o] * ! c. [(cándidat)-o] ** ! d. [(candidat)-ó] ** !
The same ranking of course accounts for regular stem-final stress in words without a TE, as illustrated with a consonant-final word (like those in (6a-c)) in (19)…
(19) [(animal)] FinalStress Nonfinality a. ☞ [(animál)] ** b. [(anímal)] * ! c. [(ánimal)] ** !
…and with a vowel-final word (like those in (7a-c)) in (20).
(20) [(domino)] FinalStress Nonfinality a. ☞ [(dominó)] ** b. [(domíno)] * ! c. [(dómino)] ** !
Because there is no TE in both (19) and (20), stem-final stress and word-final stress coincide in the desired optimal candidate (a). I assume that this violates Nonfinality only twice, because stress is word-final, and not once more because stress is also stem-final — but note that nothing hinges on this assumption here, since all competitors fare worse on FinalStress. The difference between candidates (b) and (c) in all three of these tableaux illustrates the distance-measuring effect of FinalStress: (b) has stem-penultimate stress and thus one violation, while (c) has stem-antepenultimate stress and thus two violations. This distinction is not crucial for the analysis of regular stress just presented, since FinalStress is undominated, but it is crucial for the analysis of the Class (1) exceptional pattern to be analyzed in §5.3 below. The distinction made by Nonfinality between (18a) and (18d), however, is crucial here. The formal version of FinalStress in (16) does not distinguish these candidates, even though stress is stem-final in (18a) but word-final (and not stem- final) in (18d). Because a distance-sensitive demand for stem-finality of stress requires reference to a precedence relation between stressed and unstressed syl- lables within a stem, FinalStress only penalizes candidates with stress that is not stem-final if the stress is elsewhere (i.e., to the left) in the stem. Stricter adherence Exceptionality in Spanish Stress CatJL 15, 2016 21 to the claim embodied in (8), reflected in the informal version of FinalStress in (15), would also correctly distinguish (18a) from (18d), but it appears that an appropriately formalized version of the constraint would have to state two techni- cally separate requirements: that stress be assigned somewhere in the stem and that it minimize the distance from the right edge of the stem (i.e., the requirement stated by (16)). Since the first of these requirements is effectively guaranteed by Nonfinality here, I opt for the simpler formalization of FinalStress in (16).
5.3. The Class (1) exceptional pattern In order to account for the contrast between the regular pattern and the Class (1) exceptional pattern that was illustrated by the examples in (14f-j), we require a lexically-indexed version of Nonfinality that applies to all and only Class (1) exceptional words, Nonfinality(1), ranked above and forcing minimal violation of FinalStress.
(21) Nonfinality(1) >> FinalStress >> Nonfinality
The⟦ ranking in (21) achieves the desired contrasts in ⟧(14) as follows. The regu- larly stressed words in (14a-e) are not subject to Nonfinality(1), so the optimal candidate is selected by FinalStress >> Nonfinality as already illustrated in (18) and (19). But the exceptionally stressed words in (14f-j) are subject to ⟦ ⟧ Nonfinality(1), which prefers candidates without stem- or word-final stress. This means that stress cannot be stem- or word-final, but because FinalStress is better- satisfied by stress nearer to the right edge of the stem, stress on exceptionally stressed words optimally falls immediately to the left of the stem-final syllable. This is illustrated for a vowel-final word (that is, a TE-final word), like those in (14f-h), in (22)…
(22) [(epistol)-a](1) Nonfinality(1) FinalStress Nonfinality a. [(epistól)-a] * ! * b. ☞ [(epístol)-a] * c. [(épistol)-a] ** ! d. [(epistol)-á] ** ! **
…and for a consonant-final word, like those in (14i,j), in (23).
(23) [(canibal)](1) Nonfinality(1) FinalStress Nonfinality a. [(canibál)] ** ! ** b. ☞ [(caníbal)] * c. [(cánibal)] ** ! 22 CatJL 15, 2016 Eric Baković
5.4. Additional Class (1) exceptions Note that it is possible under this analysis for a vowel-final word without a TE to be lexically marked as a Class (1) exception, and thus to surface with penultimate stress just like a Class (1) consonant-final word. As it happens, examples of vowel- final words with penultimate stress the final vowel of which is demonstrably not a TE do exist, as shown in (24).
(17) a. [(whísky)] ‘whiskey’ c. [(whisky+cít)-o] ‘whiskey (dim.)’ b. [(táxi)] ‘taxi’ d. [(táxi+cít)-o] ‘taxi (dim.)’
It should be noted that Harris’s (1983) analysis predicts the same result here: his penultimate-stress-overriding rule (13) would be expected to (incorrectly) assign final stress to the words in (24a,b), but if their stem-final vowel is marked as extrametrical — as expected, given the hypothesis that these words are Class (1) exceptions — they are not subject to (13) and stress would fall on the penult, the only other available syllable. Interestingly, the two analyses make different predictions with respect to poten- tial trisyllabic or longer words of this type (vowel-final, no TE, Class (1) excep- tions). The analysis proposed here predicts that stress is still penultimate in such words, given that this satisfies Nonfinality(1) and minimally violates FinalStress. Harris’s analysis, on the other hand, predicts that stress in such words is antepen- ultimate, given that penultimate word-stress assignment ignores the extrametical stem-final vowel.15 Both predicted patterns are in fact attested in Spanish, neither particularly robustly, but they nevertheless appear to be qualitatively different: the words in (25a-c), with antepenultimate stress as predicted by Harris’s analysis, are common (though not necessarily frequent) words; those in (25d-f), with pen- ultimate stress as predicted by the present analysis, are clearly (borrowed) proper/ place names.
(18) a. [(metrópoli)] ‘metropolis’ d. [(Arrégui)] ‘Arregui’ b. [(espíritu)] ‘spirit’ e. [(Nagasáki)] ‘Nagasaki’ c. [(ímpetu)] ‘impetus’ f. [(Fujítsu)] ‘Fujitsu’
In any event, either analysis will have to say something about the stress pat- tern that it doesn’t predict. Under the present analysis, the words in (25a-c) would have the status of régimen-type words (mentioned in note 1 as well as in §3.2, (iv)), which Harris (1983: 132) dismisses as “extragrammatical”.16 Under Harris’s
15. If the penultimate syllable has a branching rhyme, however, Harris’s analysis predicts that it will be stressed given the branching condition in (10). Because the predictions made by the two analyses converge in such cases (e.g. [(mariáchi)] ‘mariachi’, [(Viváldi)] ‘Vivaldi’), I put them aside. 16. An anonymous reviewer suggests that the contrast between (25a-c) and (25d-f) could be analyzed with lexically-indexed versions of the categorical constraints corresponding to the two parts of Prince & Smolensky’s (2004) foot-based, vertically gradient Nonfinality constraint (see note 14). I do not pursue this suggestion in the interests of space. Exceptionality in Spanish Stress CatJL 15, 2016 23 analysis, the words in (25d-f) would presumably have to be bona fide exceptions to the power of (13) to override penultimate stress by (9). Recall now the prepositions and truncations discussed in §3.3. An account of the penultimate stress of these words presents itself in the present context: that these words are (predictably and consistently) marked as Class (1) lexical items. Because all of these prepositions and truncations are bisyllabic, this claim is agnostic between Harris’s analysis and the analysis proposed here: either the final vowel of these words is extrametrical (Harris’s analysis) or these words are subject to Nonfinality(1) (the analysis proposed here). But, as noted by an anony- mous reviewer, there is no independent motivation for the claim that prepositions and truncations are systematically marked as Class (1) lexical items; I leave the consequences of the claim for future research.
6. Concluding remarks The analysis of the Class (1) exceptional stress pattern just presented in §5 clear- ly depends on the stem-domain analysis of Spanish stress defended earlier in §3. FinalStress accounts for the fact that stress regularly falls on the final syllable of the stem, and its interaction with Nonfinality(1) accounts for the fact that stress exceptionally falls on the immediately adjacent syllable to the left of that stem-final syllable. Because Harris’s (9)+(10)+(13) word-domain analysis does not take the stem as the domain of stress assignment in Spanish, it cannot so simply and elegantly account for the difference between the regular and Class (1) exceptional stress pat- terns; it requires instead a lexical marking device (extrametricality) accompanied by a set of principles (albeit purportedly universal ones) to appropriately constrain it. But recall that the motivation for the stem-domain analysis was not necessarily to pave the way for this account of the Class (1) exceptional stress pattern; it was to treat the Class (2) “exceptional” stress pattern as a subcase of the regular stress pattern. I submit that this treatment on its own is a significant advance over Harris’s (1983) path-charting account of this rich area of study, and that the possibility it affords for a simple and elegant account of the Class (1) exceptional stress pattern is simply an additional and compelling argument in its favor.
References Baković, Eric. 2009. Abstractness and motivation in phonological theory. Studies in Hispanic and Lusophone Linguistics 2: 183-198.
Colina, Sonia. 1996. Spanish truncation processes: The emergence of the unmarked. Linguistics 34: 1199-1218. Colina, Sonia. 2003. Diminutives in Spanish: A Morpho-Phonological Account. Southwest Journal of Linguistics 22.2: 45-88. Den Os, Els & Kager, René. 1986. Extrametricality and stress in Spanish and Italian. Lingua 69: 23-48.
McCarthy, John J. & Prince, Alan. 1993. Generalized alignment. In Booij, Geert & van Marle, Jaap (eds.). Yearbook of Morphology 1993, 79-153. Dordrecht: Kluwer.
ISSN 1695-6885 (in press); 2014-9719 (online) Catalan Journal of Linguistics 15, 2016 27-51
Morphological Exceptions to Vowel Reduction in Central Catalan and the Problem of the Missing Base*
Joan Mascaró Universitat Autònoma de Barcelona. Centre de Lingüística Teòrica [email protected]
Received: February 25, 2016 Accepted: July 8, 2016
Abstract
A certain class of Central Catalan compounds characterized by a first component that lacks a related output word is discussed and analyzed in connection with Vowel Reduction and Destressing. The first component of these compounds contains a vowel [a, ɛ, ɔ] that undergoes Destressing before the stressed vowel of the second component, but does not reduce. This causes an opacity problem because the generalization that there are no unstressed [a, ɛ, ɔ] is not surface-true for these cases. An analysis in the framework of parallel OT with output-to-output constraints is examined in detail and the function that returns the base of the relevant constituent in the candidates being evaluated is made precise. It is shown that such an analysis is not feasible. After showing that compounds have internal constituent structure even under noncompositional semantics, a Stratal OT analysis is presented that can handle such cases. Keywords: compounds; Central Catalan; Vowel Reduction; Destressing; opacity; output-to-output constraints; Stratal OT; compositionality
Resum. Les excepcions morfològiques a la reducció vocàlica i el problema de la base absent
S’examina, en relació amb la reducció vocàlica i la desaccentuació, una classe de compostos en català central caracteritzats per un primer component que no està relacionat amb cap paraula en l’output. El primer component d’aquests compostos conté una vocal [a, ɛ, ɔ] que sofreix desac- centuació davant de la vocal accentuada del segon component, però que no es redueix. Això causa un problema d’opacitat ja que la generalització segons la qual no hi ha [a, ɛ, ɔ] àtones no és certa superficialment per a aquests casos. S’examina en detall una anàlisi en el marc de la TO paral·lela i es defineix de forma precisa la funció que assigna la base al constituent rellevant dels candidats avaluats. Es mostra que aquesta anàlisi no és satisfactòria. Després de demostrar que els compostos poden tenir estructura de constituents interna encara que no tinguin semàntica compo- sicional, es presenta una anàlisi en el marc de la TO amb estrats que dóna compte d’aquests casos. Paraules clau: compostos; català central; reducció vocàlica; desaccentuació; opacitat; restricció d’output a output; OT amb estrats; composicionalitat
* This work was supported by grant FFI2013-46987-C3-2-P of the Spanish Ministry of Economy and Competitivity. I am indebted to Ricardo Bermúdez-Otero and Eulàlia Bonet, who read previous versions and made comments that greatly improved the paper, and to two CJL reviewers. 28 CatJL 15, 2016 Joan Mascaró
Table of Contents 1. Introduction 5. Parallel and cyclic analyses 2. Compound structures: stress, 6. The identification of the base vowel reduction and missing bases 7. The problem of the missing base 3. Different kinds of compounds 8. The stratal analysis with a first member lacking a base 9. Conclusions 4. Isolates and complex structure References
1. Introduction The existence of analogical relations and groups of related words (paradigms) has been recognized for a long time. As put in general terms by Hermann Paul (1968 [1880]: 106) a long time ago, “individual words attract each other in the soul, and as a result a set of bigger or smaller groups arise.”1 The analysis of the role of para- digms in optimality theory has brought forward a range of different mechanisms to account for paradigm uniformity effects, most significantly the Base-Priority model of output-output correspondence, the theory of Optimal Paradigms, and cyclic evalua- tion in Stratal OT. When we face a given phenomenon, we may ask whether it is best analyzable, or analyzable at all, within a given model. Depending on the answers we get for different phenomena, we can either conclude that some of the models should be rejected or that linguistic theory must incorporate all or several of them in order to account for the facts. In this paper I analyze a set of compounds in Central Catalan with the intention to make some progress in answering such questions. It has been noticed that in Catalan the stem of derivatives, the first component of compounds, and stressed prefixes lose their stress because they are followed by another stress. However, they differ in that the destressed vowel of derivatives reduces, but in the case of compounds and prefixed words the destressed vowel does not reduce. Before we examine such cases, some background in vowel reduc- tion in Central Catalan is necessary (Mascaró 1976, 2002, Wheeler 2005). Central Catalan has seven underlying vowels, /a, ɛ, ɔ, e, o, i, u/, and the derived vowel [ə]. Vowel reduction is governed by the initial, approximate generalizations in (1):
(1) a. All and only a, ɛ, ɔ, e, o, i, u appear in stressed position. b. All and only i, u, ə appear in unstressed position.
In the case of alternations, vowel reduction follows the mappings a, ɛ, e → ə, and ɔ, o → u, as shown in (2a,b) below. (2a) shows stressed stems and (2b) the same stems destressed by a following stressed suffix, and (2c) presents cases of
1. “… attrahieren sich die einzelnen Wörter in der Seele, und es entstehen dadurch eine Menge grös- serer oder kleinerer Gruppen.” Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 29 nonalternating unstressed vowels. They all follow the generalizations in (1). The examples in (2d) will be discussed directly.
(2) a. b. c. d. p[á]ra p[ə]r[ɛ́]m frar[ə] fr[á]s[e] ‘s/he stops’ ‘we stop’ ‘monk’ ‘sentence’ tr[ɛ́]nta tr[ə]nt[ɛ́] bíg[ə]m [í]t[e]m ‘thirty’ ‘thirtieth’ ‘bigamist’ ‘item’ r[é]nta r[ə]nt[é]s [ə]l[ə]ment [e]v[e]r[ɛ́]st ‘s/he washes’ ‘s/he washes-pres.subj’ ‘element’ ‘Everest’ [ɔ́]bre [u]brir[á] tràng[u]l pl[á]nct[o]n ‘s/he opens’ ‘s/he will open’ ‘turmoil’ ‘plankton’ c[ó]ntra c[u]ntr[á]ri b[u]coi b[o]ns[á]i ‘against’ ‘contrary’ ‘cask’ ‘bonsai’
A set of cases that have been termed contextual, lexical, and morphological exceptions to vowel reduction in the literature disobey the generalizations in (1). In this paper we will be interested in morphological exceptions and will ignore contextual exceptions (Mascaró 1976: §1.6, 2002: 107-110, Wheeler 2005: 61-70), since they are not relevant to the central discussion. I will refer incidentally to lexical exceptions later; they affect a numerous set of lexical items in which some vowels appear as [e] or [o] even if unstressed. Unstressed [a], [ɛ], [ɔ], however, never appear as unstressed vowels in lexical exceptions.2 The corresponding gen- eralization (3a) is exemplified in (2d) above; (3b) is discussed directly.
(3) a. Lexical exceptions. A set of lexically marked items have underlying /e/, /o/ which do not undergo vowel reduction, even if unstressed. b. Morphological exceptions. In compounds and some prefixed words a destressed vowel in the first component does not undergo vowel reduction.
Let us examine morphological exceptions. Beyond the (simple) word level, several processes of phrase phonology rearrange the stress contour of the con- catenated words in different ways, and the generalizations in (1) are rendered opaque (Mascaró 2002: 93-95). The sentence Empenyeu ara! ‘Push-pl now!’ can be pronounced among other configurations with the stress pattern [ə̀mpəɲɛw áɾə], although the word stress pattern of the constituent words is [əmpəɲɛ́w] and [áɾə]. The generalization (1a) is not obeyed because [ɛ] in the first word has undergone destressing but is not reduced, and (1b) is not obeyed because the initial [ə] has received stress. I will concentrate my attention on one of these processes, namely
2. In northern subdialects [a] can appear also in lexical exceptions. 30 CatJL 15, 2016 Joan Mascaró destressing in the first component of compound structures,3 but there is a wide range of processes that have the same opacity effects.4 The different effects of a stressed vowel followed by another stressed vowel can be observed in the examples in (4a-b) repeated from (2a-b). (4a) shows a stressed vowel that is not followed by another stress. In (4b) the same root of (4a) is fol- lowed by a derivational or inflectional stressed suffix that causes destressing and vowel reduction. The examples in (4c) are compounds; the stress in the second component of the compound (4c) causes destressing but no vowel reduction in the first component. Finally, in (4d) the same word in (4a) is followed by another word within a phrase; in this case the stress weakens, but does not disappear, and the vowel does not reduce.5
(4) a. b. c. d. p[á]ra p[ə]r[ɛ́]m p[a]racaig[ú]des p[à]ra caig[ú]des ‘s/he stops’ ‘we stop’ ‘parachute’ ‘stops falls’ tr[ɛ́]nta tr[ə]nt[ɛ́] tr[ɛ]nta-c[í]nc tr[ɛ̀]nta c[í]ncs ‘thirty’ ‘thirtieth’ ‘thirty-five’ ‘thirty fives’ r[é]nta r[ə]nt[é]s r[e]ntapl[á]ts r[è]nta pl[á]ts ‘s/he washes’ ‘s/he washes-pres.subj’ ‘dishwasher’ ‘s/he washes dishes’ [ɔ́]bre [u]brir[á] [ɔ]brell[á]unes [ɔ̀]bre ll[á]unes ‘s/he opens’ ‘s/he will open’ ‘can opener’ ‘s/he opens cans’ c[ó]ntra c[u]ntr[á]ri c[o]ntracultur[á]l c[ò]ntra cult[ú]res ‘against’ ‘contrary’ ‘countercultural’ ‘against cultures’
(5) shows the structures of the examples in (4b-d), illustrated with one of them, and (6) states the corresponding generalizations.
(5) a. Derivative b. Compound c. Phrase
[Wd[Sttɾənt] [Sfɛ́]] [Wd[Wd[Sttɾɛnt-ə]] [Wd[Stsíŋ]]] [NP [Wd[Sttɾɛ̀nt-ə]] [Wd[Stsíŋ] s] ] ‘thirtieth’ ‘thirty-five’ ‘thirty fives’
3. I will use the term compound structure to refer to both compounds and prefixed words. 4. There is a tradition, mainly in the more prescriptively-oriented literature (but also elsewhere, Oliva 1992, Wheeler 2005), that considers that the first element of compound structures retains a secondary stress. This assumption derives from a confusion between word stress and emphatic and rhythmic stress. Those authors who have based their conclusions on experimental results (Mascaró 1976, Prieto 2003, Nadeu 2016) assume destressing, as does Recasens (1993). Destressing also fol- lows from the transcriptions of accurate traditional historical linguists (e.g. Moll 1931, Coromines 1989-1997). 5. Destressing of the first constituent of compounds also obtains in Spanish (Hualde 2007). Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 31
(6) a. (Simple) word stress. The stressed vowel of a root/stem of a derivative or inflected form followed by a stressed suffix loses its stress and reduces. b. Sentence stress. A word followed by another word in a phrase usually weakens its stress but does not lose it. The vowel does not reduce.6 c. Underapplication of vowel reduction in compound structure. The stressed vowel of the first component of a compound structure loses its stress but does not reduce.
2. Compound structures: stress, vowel reduction and missing bases Let us examine in more detail the morphology and phonology of morphological exceptions. Focusing on their first component, we might distinguish three types of compounds:
(7) a. [ [LO] [C] ] LO is a lexical element with an independent output paɾə-kəjɣúðəs (4c)
b. [ [LB] [C] ] LB is a lexical element with no independent output nɛu-klásik (12) c. [ [I] [C] ] I not a lexical element (it is an isolate) bɛt-əkí (9)
In (7a) the first constituent of paɾə-kəjɣúðəs is related to the lexical element páɾə ‘s/he stops’, whereas nɛu in nɛu-klásik in (7b) is a lexical element but is a bound form that never appears as the output *[nɛ́u]. In the case of bɛt-əkí the first constituent bɛt cannot be successfully related to any independent lexical item, word or affix. I will use the term isolate to refer to such constituents, and I will refer to constituents that lack a related output (both (7b) and (7c)), the structures that are of interest in this paper, as constituents that lack a base or constituents with a missing base. Compound structures with a first component missing a base can have several origins. Some of them originate through word formation processes that use as a first component a lexical item that is not a word, as is the case with reduplicative compounds, prefixes and many neoclassical compounds (§§ 3.2, 3.4). In the case of borrowings (§ 3.3) the compound structure in the source language can be preserved with loss of the original semantic compositionality. Finally, linguistic change can result in loss of semantic compositionality with preservation of constituent struc- ture; since this case is of specific interest I will consider it briefly below. What is important to notice is that, no matter what the origin, all these cases result in the same structure: a compound structure that contains two constituents, the first being unrelated to an independent output and an exception to vowel reduction. The examples presented in this paper correspond to my own speech, but even if there
6. Under different circumstances, in particular under stress clash, the first stress can disappear, as in the third syllable of the sentence Empenyeu ára cited above. 32 CatJL 15, 2016 Joan Mascaró is individual variation with regard to particular words, for all speakers the effects described are robust. In productive compounds that show transparent compositional meaning like para-sol ‘parasol’, literally ‘stops sun’, and parallamps ‘lightning rod’, lit. ‘stops lightning’, the speaker implicitly knows that they share a common first constituent, and for para-sol and gira-sol ‘sunflower’ that they share a common second constitu- ent.7 Some compounds are not compositional semantically and keep their constituent structure, e.g. [[matə][pəɾén]], a toxic mushroom of the family Boletus, lit. ‘kills relative’. In other cases noncompositionality is accompanied by loss of constituent structure, usually through historical change. For a word like passa-port [pəsəpɔ́rt] ‘passport’ (see 8a), the speaker can only discover the fact that [pəsə] is (historically) related to the form [pásə] of the verb passar ‘to pass’ through the orthographic form and rational deduction.8 This process of lexicalization of productive compounds can result in the loss of their original internal constituent structure, the two components becoming a monomorphemic root. As should be expected, in such cases, since the compound structure has been lost, all the vowels within the original first component reduce. Some examples are presented in (8a), followed in (8b) by their form in the reconstructed, previous, nonlexicalized stage; (8c) shows a proper synchronic com- pound with a first component coinciding with the one in (8b):
(8) a. Lexicalized b. Previous form c. Compound with same 1st component [pəɾájɣwə] *[[paɾ][ájɣwə]] [[paɾə][ʎáms]] ‘umbrella’ ‘stops water’ ‘lightning rod’, lit. ‘stops lightning’ [pəsəpɔ́rt] *[[pasə][pɔ́rt]] [[pasə][puɾé]] ‘passport’ ‘passes gate’ ‘food mill’, lit. ‘passes purée’ [bərsəmblán] *[[ber][səmblán]] [[beɾə][mén]] ‘likely, credible’ ‘true looking’ ‘truly’, lit. ‘true-suffix’ [kəpfikát] *[[kap][fikát]] [[kab][ʒiɾát]] ‘concerned’ ‘head put’ ‘upside-down’, lit. ‘head turned’ [fərukəríl] *[[fɛru][kəríl]] [[fɛru][məŋnətízmə]] ‘train’ ‘iron track’ ‘ferromagnetism’, lit. ‘iron magnetism’ [mələmén] *[[malə][mén]] [[mal][pərlát]] ‘badly’ ‘(in) bad mind’ ‘foulmouthed’, lit. ‘bad spoken’ [bunumíə] *[[bɔn][umíə]] [[bɔnə][βəntúɾə]] ‘conviviality’ ‘good manhood’ ‘fortune telling’, lit. ‘good venture’
7. It is easy to elicit such relatedness by different means. Speakers can also make up expressions built on gira-sol, lit. ‘turns-flower’, like Gira més que un gira-sol ‘it turns more than a sunflower’, or El fa girar més el sol que un gira-sol ‘it is turned around more by the sun than a sunflower’. 8. Although most Catalan compounds are not hyphenated, from now on their components will be always separated by a hyphen in order to make them apparent. Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 33
But crucially, in other cases of lexicalization, in fact in the majority of cases, the compound undergoes semantic drift but the constituent structure is retained and the two components do not become a monomorphemic root. Let us exam- ine a specific case in some detail. The noun meaning ‘stone’, derived from Latin pětrām, is part of many place names. The regular evolution of the internal cluster is TR>dɾ after a stressed vowel, TR>ɾ after an unstressed vowel (Coromines 1971: 183-188),9 thus we get pětrām>péðɾə. When the stress of the preceding vowel disappears as in compounds (also in proclitic prepositions: rětrō>reɾə ‘behind’) the result of TR is [ɾ]. In compounds formed at the stage when the phonetic change took place, we got pětram>péðɾə for the simple word, but when this word is the first component of a compound, as in Pera-tallada, lit. ‘cut stone’, the evolu- tion is pětrām-talleātām>peɾə-təʎáðə. At some point after this stage the lexical relation péðɾə-péɾə is lost; at later stages transparent place name compounds of the same class are formed again on péðɾə, e.g. peðɾə-fórkə Pedra-forca, lit. ‘fork stone’. What is relevant here is that at the time that the result of Latin pětram in the noun in isolation and in the first position of the compound diverged, this did not force vowel reduction of the unstressed vowel, i.e. it is not the case that /[peɾə- təʎáðə]/ became *[pəɾətəʎáðə]. We could possibly argue, for present day Catalan, that /[peɾətəʎáðə]/ has lost its compound structure, and that its syllable-initial /e/ is marked as a lexical exception to vowel reduction (see (2d) above, and discussion), as in [sopɾán-o] ‘soprano’, [deskárt-es] ‘Descartes’, [memoɾándum] ‘memorandum’ (Mascaró 1976, 2002, 2015). As argued in the next section, this explanation is impossible for cases with unstressed [a], [ɛ], [ɔ], which can never be exceptions to vowel reduction. But it does not hold even in cases with [e], [o] like [peɾətəʎáðə], because lexical exceptions to vowel reduction are recent (they probably arose in the 19th century), and it is necessary to assume a stage after TR>ɾ in which there were no exceptions to vowel reduction and in which Pera-tallada had to have the compound structure /[[peɾə]-[təʎáðə]]/. The obvious explanation for the lack of vowel reduction is that the word did not lose its compound structure and that it remained /[[peɾə]-[təʎáðə]]/, even though /[peɾə]/ was unrelated at this point to any other lexical item. This compound structure has survived in several place names (Pera-fita, Pera-fort, Pera-tallada) for more than ten centuries.10 Summing up, we can distinguish two classes of [[C][C’]] compounds, those like trenta-cinc in (8c), for which the first component [C], trenta- (4a) can be identified as an independent synchronic lexical element, the numeral trenta, and those like Pera in Pera-tallada, for which we can only identify the first component [C] as a word through diachronic analysis. Following current practice, we will refer to this independent lexical element related to a given constituent [C] (or [C’]) of a com- pound structure as the base of [C] (or [C’]). I will use the term isolate to refer to the member of a compound structure with a missing base, like Pera in Pera-tallada. Thus the preposition contra is the base of contra- in contra-cultural (4c), but the
9. We also get ɾ when a vowel between T and R was lost, i.e. VTVR>Vr: mātěr>máɾə ‘mother’. 10. There is loss of constituent structure and reduction in the case of Peralada [pəɾəláðə] 3. Different kinds of compounds with a first member lacking a base 3.1. Proper and common nouns One group of isolates with unreduced [a], [ɛ], [ɔ] contains place or person names and some common nouns. Examples are displayed in (9). (9) 1st component, Etymon glosses Meaning11 etymon palə-mós pǎlūm > Ø marsh - liquid town name palə-furʒéʎ pǎlātĭūm > pəláw palace - pers. name town name bɛt-əkí vĭdēte > bəʒɛ́w see-2pl.imp - here ‘here it is’ kamə-miʎə Greek chamai-mēlon on the ground - apple ‘chamomile’ kɔβɾə-ʎít kɔ́βɾe (Old Cat.) > kuβɾɛ́ʃ covers - bed ‘bed cover’ kɔβɾə-kálzə kɔ́βɾe (Old Cat.) > kuβɾɛ́ʃ covers - chalice ‘chalice cover’ ʃɛrkə-βíns sɛ́rkə > ʃɛ́rkə seeks - wine-pl. family name bɛni-kásim Arabic banī sons - pers. name town name taɣə-mənén — unknown mountain name tɛri-káβɾəs — unknown family name paɾə-sɛtə-mɔ́l — novel creation ‘paracetamol’ In palə-mós the Latin etymon of [palə], pǎlūm ‘marsh’, has left no simple word descendants and as a result the first component has become an isolate; the first component in palə-furʒéʎ, from Lat. pǎlātĭūm, has resulted in pəláw, to which palə cannot be related anymore. For both cases, there is a noun pálə ‘shovel’ which can- not be put in any regular compounding relation to mós ‘bite’, nor to furʒéʎ, which is not an independent word. In bɛt-əkí the original first component, vĭdēte (regularly >bəʒɛ́ w), has been reduced diachronically to bɛt which no longer can be related synchronically to bəʒɛ́ w or to any existing word.12 In the case of kamə-miʎə the 11. Orthographic forms: Palamós, Palafrugell, vet aquí, camamilla, cobrellit, cobrecalze, Xercavins, Benicàssim, Tagamanent, Terricabres, paracetamol. 12. For many speakers it is bɛt-əkít with the second component also an isolate. Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 35 original Greek compound either kept its complex character through Latin or was reinterpreted as a compound after a monomorphemic phase. The next two examples contain the apparent verbal form kɔβɾə. In Old Catalan the verb cobrir belonged to conjugation IIIb and 3sg.pres.ind. had the form kɔ́ βɾe. Later, when the verb changed to conjugation IIIa, kɔβɾə became an isolate, because the 3sg.pres.ind. form had become kubɾ-ɛ́ ʃ. Today speakers might establish a regular compound relation of VN compounds with regular compositional semantics for items like cobre-llits (‘instrument used to cover beds’). But if the base of kɔβɾə were kubɾɛ́ ʃ, we would predict *kuβɾə-ʎít. There is another verb cobrar ‘to cash, to retrieve’ with 3sg.pres. ind. kɔ́ βɾə which could be identified as the correspondent, but then we would get the wrong semantics, i.e. *’bed casher/retriever’ instead of ‘bed cover’. In ʃɛrkə- βíns the sporadic change s>ʃ has rendered the relation to the base opaque. In the following three cases the Arabic or pre-Roman origin of the first component makes the identification of a base impossible. Finally, the last example paɾə-sɛtə-mɔ́ l is a case of two nonfinal isolate components with destressed and unreduced vowels. 3.2. Reduplicative compounds Another source of isolates is found in reduplicative expressive compounds. The members of a group of such compounds consist of a reduplicated monosyllabic or disyllabic component, in many cases of onomatopeic origin. Thus if speakers want to imitate the sound of a boiling syrup they can invent the imitation form or inter- jection [blɔbblɔp] which can be pronounced with varying degrees of stress on its syllables. But if it is turned into a noun as in the sentence Quan sentis el blop-blop, para ‘when you hear the blop-blop, stop’ it will appear invariably as [blɔbblɔ́p] with final stress and unreduced unstressed vowel in the first component. Here are some actual common cases (Cabré 1993, 2002; Riera-Eures 2002).13 (10) nyeu-nyeu ɲɛwɲɛ́w ‘hypocritical talk’ (N) mec-mec mɛgmɛ́k ‘horn blast’ (N) nyam-nyam ɲamɲám ‘eating’ (N) tau-tau tawtáw ‘on equal terms’ (Adv.) xano-xano ʃanuʃánu ‘(walking) slowly’ (Adv.) poti-poti pɔtipɔ́ti ‘disorder, chaos’ (N) taf-taf taftáf ‘sound of an engine’ (N) (fer la) gara-gara gaɾəɣáɾə ‘to try to please through flattering’ (N) (lit. ‘to make the gara-gara’) 13. We can also assign to this class a set of borrowings with identical structure: Bora-Bora, Baden- Baden, pai-pai, tse-tse, beri-beri, yo-yo. 36 CatJL 15, 2016 Joan Mascaró Of course we also find cases in which the first component of these reduplicative compounds is not an isolate: kɾɛk-kɾɛ́k ‘repeated cracking sound’ (N), kɾɛ́k ‘crack- ing sound’; mɛwmɛ́w ‘repeated meow’ (N), mɛ́w ‘meow’. But in the examples in (10) the first component is always an isolate, because there is no independent word it can be related to. 3.3. Borrowings Borrowings are another source of isolates. I show some instances in (11). Some of these examples present, in addition to the nonreduced destressed vowels [a], [ɛ], [ɔ], the nonreduced vowels [e], [o] that correspond to lexical exceptions (see §1), as in [rɔkefɛ́lər] and [lɛjdmotíf]: (11) McDonalds mag-dɔ́nəls Eng. playboy plɛj-βɔ́j Eng. Pepsi-Cola pɛpsi-kɔ́lə Eng. Quasimodo kwasi-móðo Lat. leitmotiv lɛjd-motíf Ger. Tel Aviv tɛl-əβíp Heb. Vietcong bjek-kɔ́ŋ Viet. outsider ɔwt-sájðər Eng. boy scout bɔj-əskút Eng. Rockefeller rɔke-fɛ́lər Eng. The reasons for attributing compound structure to borrowings is diverse. In some cases the orthographic form might have suggested compound constituent structure (Tel Aviv), in others the speakers who introduced the borrowing might have known the source language, where the word was a compound (playboy, leit- motiv) or a phrase (Quasimodo).14 3.4. Neoclassical compounds and prefixes Many cases of compounds with a missing base are found in neoclassical com- pounds. In this case the first component is not an isolate, because it is a lexical ele- ment, but one which has no independent output, (7b) in the previous classification. The relevant examples displayed in (12) show an unstressed, unreduced vowel in the first component. Here and in (13) I omit the gloss if the English form is suf- ficiently close to the Catalan one. 14. Quasimodo derives from the Gregorian Introit Quasi modo geniti infantes ‘in the way of newborn babies’, based on 1 Peter 2:2. Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 37 (12) [a]nglo-franc[ɛ́]s ‘Anglo-French’ [ɔ]vi-f[ó]rme m[ɛ]ta-llengu[á]tge ‘metalanguage’ c[ɔ]rtico-ester[ɔ́]ide l[a]bio-dent[á]l [ɛ]cto-pl[á]sma h[ɛ]li-oc[ɛ́]ntric p[a]ra-norm[á]l p[ɛ]tro-d[ɔ́]lar t[ɛ]tra-pl[ɛ́]gia [ɔ]rto-tipograf[í]a c[a]rdio-vascul[á]r p[ɔ]li-traumat[í]sme n[ɛ]o-cl[á]ssic The other source of words with compound constituent structure and a first component which is a lexical element with no independent output is the class of stressed prefixes.15 Notice that since prefixes are, by definition, not independent words, it is impossible to establish a precise border between clear prefixes like pre- in pr[ɛ]-nat[á]l in (13) and “root” isolates like cardi(o)- in c[a]rdio-vascul[á]r in (12). Since this distinction does not bear on the issues we are dealing with, the problem will not be addressed here. Some examples of isolates traditionally described as prefixes are shown in (13): (13) [a]nti-c[ɔ́]s ‘antibody’ ps[ɛ]udo-probl[ɛ́]ma [a]rxi-satisf[é]t ‘supersatisfied’ pr[ɛ]-nat[á]l p[ɔ]li-sil·l[á]bic p[ɔ]st-operat[ɔ́]ri tr[a]ns-sexu[á]l s[ɛ]mi-form[á]l 4. Isolates and complex structure Missing bases are bound morphemes like the examples in §3.4, or isolates like the examples in §3.1-3.3 and are crucial to the arguments I present in §5-7. Since isolates are peculiar structures, in this section I will examine the evidence in favor of their existence. Containing an isolate implies having a complex structure [Z [I] [Y]], where I is an isolate. Therefore we must rule out the possibility that Z has a flat structure ([ZI ͡ Y], X ͡ Y a monomorphemic sequence) and give evi- dence in favor of the internal constituent structure. A possible argument against the complex structure [Z [I] [Y]] is that it violates strict compositionality: the meaning of [Z [I] [Y]] cannot be a function of the meaning of its parts because I is not a lexical element and therefore it has no meaning. An expression like [Z [Ibɛt] [Advəkí]] ‘here it is’ should be monomorphemic. But there is extensive evidence against strict compositionality (Jackendoff 1997, 2010b, Jackendoff 15. There are also unstressed prefixes, like a-, des-: ə-nurmál ‘abnormal’, dəz-órjðɾə ‘disorder’. In stressed prefixes the existence of underlying stress is justified by the lack of reduction. In neoclassic compounds it can be justified also, in some cases, by the existence of surface stress in the case of components that can appear both in first and in second position: m[ɔ]rfo-gènesi, antropo-m[ɔ́ ]rf; d[a]ctilo-forme, mono-d[á]ctil. 38 CatJL 15, 2016 Joan Mascaró & Audring to appear), which can be summarized in the generalization in (14a), briefly exemplified in (14b-e): (14). a There are linguistic expressions that have complex structure and noncom- positional meaning. b. Idioms. Eng. kick the bucket; Cat. prendre el pèl ‘to fool somebody’, lit. ‘to take somebody’s hair’; Cat. dinyar-la ‘to die’ (*dinyar , obj. clitic la nonreferential) c. Inflected forms. pluralia tantum, Eng. jean-s; Cat. pantalon-s ‘trousers’ d. Derivatives. Eng. prob-able (cf. prob-abil-ity); Cat. lubr-i[k] ‘lubricious’ (cf. lubr-i[s]-itat ‘lubricity) e. Compounds. Eng. bull’s eye; Cat. mata-parent ‘Boletus satanas’, literally ‘kills relative’ In (14b) the lack of semantic compositionality is obvious, and the need for several constituents also, because the verbs can inflect. The pluralia tantum in (14c) contain a plural morpheme because it triggers agreement and because the morpheme shows the typical plural allomorphy (jean[z], pant[s], breech[ɪz]). For Catalan, the sequence pantalon-s cannot be monomorphemic because in this case when referring to a set of trousers it would have a plural *pantalonsos, as in monomorphemic descans - descans-os ‘rest/rests’. In (14d) there must be a second component, the suffix –ic, as shown by its peculiar allomorphy. In the case of compounds (14e) similar arguments apply. If [Y] in [Z[I][Y]] is a constituent, then given that [Z] is a proper bracketing of the terminal structure it dominates, [I] must also be a constituent. The second component Y is a constituent because it is identifiable as a lexical element, as in the examples in (15a), repeated from (9-13); ‘—’ marks isolates: (15) Components’ glosses Gloss a. Pera-tallada — ‘cut-part.fem-sg’ Town name vet-aquí — ‘here’ ‘here it is’ Pepsi-Cola — ‘cola drink’ ecto-plasma — ‘plasm’ ‘ectoplasm’ Pala-folls — crazy-pl Town name b. poti-poti — — ‘disorder, chaos’ (N) c Taga-manent — — Mountain name cama-milla — — ‘camomile’ ɔwt-sájðər — — ‘outsider’ teri-yaki — — ‘teriyaki’ paɾə-sɛtə-mɔ́l — — — ‘paracetamol’ Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 39 In some cases there is additional evidence for an analysis of the first compo- nent as an isolate. In the town name Pala-folls the second member is identified with the masculine plural form foll-s of the adjective foll ‘crazy’. If the sequence foll-s were monomorphemic we would expect, like in the case of pantalon-s dis- cussed above, the plural *Pala-folls-os, as in descans - descans-os ‘rest/rests’, instead of Pala-foll-s. Notice that the semantic plural of fully compositional compounds with a plural second component is also identical to the semantic sin- gular: un espanta-ocell-s ‘one scarecrow’, dos espanta-ocell-s two scarecrows’. The identifiability of the second component of Pala-foll-s is confirmed by the derived demonym pala-foll-enc. In the case of reduplicative compounds (15b), examined in §3.2, the evidence is furnished by the reduplication process itself. The examples in (15c), in which both components are isolates ([Z [I][I]]) constitute a special case. The only evidence for complex structure, at least for most cases, is the lack of vowel reduction in the first component. If we derive complex structure from lack of reduction we cannot predict the latter from the former. Therefore these cases do not constitute solid evidence for the arguments put forth in §5-8. The only plausible way to account for their phonological behavior, however, is to assume complex structure.16 5. Parallel and cyclic analyses In a parallel framework like classical OT underapplication of vowel reduction causes an opacity problem, because given a noun like para-caigudes ‘parachute’, lit. ‘stops falls’, there is no way of forcing reduction in the destressed /a/ in the derivative kəjg-úð-ə-s without forcing it also in *pəɾ-ə. In /paɾ-ə-kajg-ud-ə-s/ the constraints responsible for stress placement will favor páɾ-ə-kájg-úð-ə-s, but those responsible for destressing will favor elimination of all nonfinal stresses. Vowel reduction will then force both nonfinal vowels to reduce, yielding *pəɾ-ə-kəjg-úð-ə-s, no matter how the constraints are ordered, as shown in (16). In (16) Final Stress stands for the set of constraints that assign stress to one of the last three syllables of a word, *a,ɛ̌ ,ɔ̌ ,ě,ǒ for the constraints that force vowel reduction, and *[Wd ...ˈV...ˈV...] for constraints that favor candidates with a single rightmost stress in the simple or the compound word; ‘☐‘, the necessity operator, marks the needed, acceptable candidate: 16. In some other cases there might be some kind of “weak identification” of one of the components when there is repeated partial coincidence, as in Viet-nam, Viet-minh, Viet-cong; MacDonald, MacArthur, etc.; Amsterdam, Rotterdam. Köhnlein (2015) gives extensive evidence for complex structure in Dutch place names with noncompositional semantics like Wagening-en, Loos-drecht and Amster-dam. He also shows that such structures are widespread in Germanic and in many languages from different families. 40 CatJL 15, 2016 Joan Mascaró (16) Regular compounds, parallel analysis paɾ-ə-kajg-ud-ə-s Final Stress *[Wd ...ˈV...ˈV...] *a,ɛ̌ ,ɔ̌ ,ě,ǒ paɾ-ə-kajg-ud-ə-s *! ** páɾ-ə-kəjɣ-úð-ə-s *! ☐ paɾ-ə-kəjɣ-úð-ə-s *! F pəɾ-ə-kəjɣ-úð-ə-s Harmonic Serialism would face the same problems, since constraint ordering is fixed. The obvious solution to this problem in parallel OT is transderivational output to output correspondence (OO-correspondence), which establishes a cor- respondence relation between two elements in different derivations (Kenstowicz 1996, Benua 1997, Kager 1999, Downing 2005, Downing et al. 2005). In the case of a compound structure with two constituents, [[C][C’]], the first constituent C can be put into morphological correspondence (notated here as M) with some base, B B B B a word or a constituent C , more specifically with the output C’O of C : C MC’O. Bℜ Vowel reduction in [C] underapplies because, given C MC’O, [C] (more precisely, B ℜ C’s candidates) must be faithful to C’O with respect to some specific properties. In our example the vowel features of the output of the /a/ℜ of /paɾ-ə/ in /paɾ-ə-kajg-ud- ə-s/ must be faithful to those in the output of the independent verbal form [páɾ-ə] ‘stop-3sg.pres.ind’: (17) Regular compounds, parallel analysis with OO [N[Vpaɾ-ə][Nkajg-ud-ə-s]] OO-Id(V-features) *[Wd ...ˈV...ˈV...] *a,ɛ̌ ,ɔ̌ ,ě,ǒ Base: [Vpáɾ-ə] a. F [N[Vpaɾ-ə][Nkajg-úd-ə-s]] * b. [N[Vpáɾ-ə][Nkájg-úd-ə-s]] *!* c. [N[Vpáɾ-ə][Nkəjg-úd-ə-s]] *! d. [N[Vpəɾ-ə][Nkəjg-úd-ə-s]] *! In nonparallel Stratal OT (Kiparsky 2000, in press; Bermúdez-Otero 2003, 2011, in preparation), the idea that the first /a/ in para-caigudes does not reduce because para- is an independent element of a specific sort can be captured by dif- ferent mechanisms, namely constituent structure and stratal ordering. Since para and caigudes are independent lexical elements, they must be cyclic constituents in the compound [N[Vpaɾ-ə][Nkəjɣ-úð-ə-s]]. The constituents para and caigudes are processed first, in the word cycle, and are assigned stress. The stressed /a/ in [Vpáɾ-ə] is not reduced, but the unstressed /a/ in [Nkajɣ-úð-ə-s] undergoes reduction. The next cycle is the phrase cycle, where destressing can reapply because there is a stress to the right of [Vpáɾ-ə], namely the ú in [Nkajɣ-úð-ə-s]. But vowel reduc- tion does not apply in this cycle, because it is a word-level process, i.e. it does not Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 41 apply in compound structures and beyond. In Lexical Phonology this would mean that the rule is not present in the phrasal cycles. In Stratal OT the corresponding mechanism is a different constraint ordering in the phrasal cycle, which will prevent the candidate with reduction from surfacing. The derivation proceeds as in (18a-b). (18) Regular compounds, stratal analysis a. Word stratum *[Wd ...ˈV...ˈV...] Final Stress*, Id(V-features); a,ɛ̌ ,ɔ̌ ,ě,ǒ Id(V-features) ≫ paɾ-ə *[Wd ...ˈV...ˈV...]≫ Final Stress *a,ɛ̌ ,ɔ̌ ,ě,ǒ Id(V-features) F páɾ-ə paɾ-ə 1W 1W pə́ɾ-ə 1W kajg-ud-ə-s *[Wd ...ˈV...ˈV...] Final Stress *a,ɛ̌ ,ɔ̌ ,ě,ǒ Id(V-features) F kəjg-úd-ə-s 1 kajg-ud-ə-s 1W 1W L kájg-úd-ə-s 1W L kajg-úd-ə-s 1W L b. Phrase stratum *[Wd ...ˈV...ˈV...], Id(V-features) *a,ɛ̌ ,ɔ̌ ,ě,ǒ páɾ-ə-kəjg-úd-ə-s *[Wd ...ˈV...ˈV...] Id(V-features)≫ *a,ɛ̌ ,ɔ̌ ,ě,ǒ F paɾ-ə-kəjɣ-úð-ə-s 1 páɾ-ə-kəjɣ-úð-ə-s 1W L pəɾ-ə-kəjɣ-úð-ə-s 1W L Both analyses, parallel and stratal, make similar predictions for compositional compounds like [N[Vpaɾ-ə][Nkəjɣ-úð-ə-s]]. For the OO-correspondence analysis compositionality implies the existence of an independent lexical form, for the stratal analysis it implies cyclic structure. Predictions differ, however, when there is inter- nal constituent structure but there is a first constituent that is not an independent element, or is an element that lacks an independent output.17 6. The identification of the base It follows from the previous considerations that a crucial step in the OO-correspondence analysis, and in general in the analysis of paradigm effects, is the determination of the independent output form with which the candidate stands in correspondence, its base, and the determination of the class of elements affected 17. Similar considerations are formulated by Trommer (2013: §2.5). 42 CatJL 15, 2016 Joan Mascaró by the constraint. This is what Bachrach & Nevins (2008b) call the asymmetry question, i.e. “why […] do identity effects go from some members of the paradigm towards others, and not vice versa?” and the inclusion question, “what is the set of relevant forms that learners put together”, i.e. what constitutes a relevant paradigm for identity effects. In many cases it is evident what the base of a candidate is, but in other cases it is not so clear.18 Therefore a more explicit and careful formulation of the candidate-base relation is in order. I will follow the standard assumption that the elements standing in correspondence are words or clitic groups, or constituents they contain; in what follows the term constituent of a word W has to be understood not as denoting a proper subconstituent, but as possibly referring also to the word itself. The correspondence relation that is crucial for the OO-constraints is established B between a constituent C of an input word W and a constituent CO of a morpho- B B logically related output word W , its base; as indicated above, C and CO may also B coincide with the whole words W and WO, respectively. For para-caigudes in (17), B W is the input noun /[N[Vpaɾ-ə][Nkajg-ud-ə-s]]/ and W is the verbal from /[Vpaɾ-ə]/; the relevant constituents are the first constituent of the input noun, [Vpáɾ-ə] and [Vpáɾ-ə], the output of the independent verbal form /[Vpaɾ-ə]/ (henceforth, bases are set in boldface whenever they are not easy to distinguish from their correspondent). B This correspondence relation, C MCO is mediated by a morphological process M that relates the corresponding words: for any specific OO-constraint we should B ℜ identify the base CO of any input constituent C. We can define this identification as the function whose domain is the set of triples (19) Identification of the base for /[Vpaɾ-ə]/ in /[N[Vpaɾ-ə][Nkajg-ud-ə-s]]/ a. LEX(/[Vpaɾ-ə]/, /[N[Vpaɾ-ə][Nkajg-ud-ə-s]]/, MCpd)= /[Vpaɾ-ə]/, and b. (/[ paɾ-ə]/)=[ páɾ-ə]; hence ℑGenEval V V c. (/[ paɾ-ə]/, /[ [ paɾ-ə][ kajg-ud-ə-s]]/, M )= [ paɾ-ə] ℑ V N V N Cpd V We canℑ now formulate an OO-constraint that requires identity of feature values B between vowels in any input C and its base C’O whenever they are related through the process MCpd: 18. “The asymmetry question and the inclusion question illustrate the need for a rigorous formalization of the principles governing the formation of the paradigms and ‘mini-paradigms’ used in invoking identity effects.” (Bachrach & Nevins 2008b: 7). Bermúdez-Otero (2011: 12, 29) notes also that “The implementation of this [OO-correspondence] solution poses a number of nontrivial technical challenges, such as motivating the selection of the surface base […]; transderivational theories face other questions […]: what expressions can qualify as surface bases, and how are they selected? should OO-identity be symmetrical, base-prioritizing, or both?” See also Trommer (2006) and Downing et al. (2005). Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 43 B (20) OO-Ident(V-features, MCpd) Let C be a constituent in the input word W, CO B a constituent in the output word WO, and MCpd a compounding or prefixation B process such that (C, W, MCpd)=CO. Then for any pair of vowels V, V’ stand- B ing in correspondence, V in C and V’ in CO, assign a violation mark for every feature that does ℑnot have the same value in V and V’. When the function returns Ø no output base is identified. This happens under different circumstances. First, if the constituent C is not an independent lexical element, as in the case ℑof [taɣə], which exists only in [[taɣə][mənén]] (9), or [ɲɛw] which only appears in [[ɲɛw][ɲɛ́w]] (10), no morphological process applies. Since LEX cannot return any lexical constituent, will return Ø. In other cases there can be regular morphological relatedness, but not the one determined by the mor- phologicalℑ process M of the function (C, W,ℑ M); consider the derivative parada, [N[Vpəɾ-á][Affð-ə]] ‘stop’, derived from the input /[N[Vpaɾ-a][Affd-ə]]/ and formed on the same base as the first constituentℑ of paracaigudes, the verb parar ‘to stop’. Here the first constituent /[Vpaɾ-a]/ contains the same root that appears in paracaigudes, but, unlike [N[Vpaɾ-ə][Nkəjɣ-úð-ə-s]] it does not show OO effects, since the vowel reduces to [ə]. The reason is that the morphological process M relating /[paɾ-a]/ to the verb parar is not MCpd, but a different, derivational process, hence cannot identify in this case any form for (/[paɾ-a]/, /[N[Vpaɾ-ə][Nkəjɣ-úð-ə-s]]/, MCpd). A more interesting case of returning Ø obtains in the case of the compound ℑ structures discussed in §3.4. Consider as the product LEX × GenEval (18). When ℑ the first constituent C is a bound lexical element, LEX(C, W, MCpd) will return a ℑ ℑ ℑ lexical element L, but GenEval(L)=Ø because bound elements have no independent output. In a case like [[karðj-u][βəskulár]] in (12), thereℑ exists a related lexical element ℑ /kardi/, that appears in cardíac, cardiologia, miocardi, taquicàrdia, etc., and LEX will identify it. But it is always bound, and therefore GenEval will not be able to identify an independent output *[kárði], or inflected *[kárðj-u], *[kárðj-ə], etc.ℑ The same happens with prefixes: [pɾɛ] has the same specificℑ meaning in prenatal, pre- molar, prevocàlic, preadolescent, etc. but it never surfaces independently. In both cases there is no asymmetrical process that can derive one of the elements in the paradigm from the other. In the case of neoclassical compounds one might argue for some cases that the base is a constituent in a derivative, e.g. that genito- in g[ɛ] nitounirari derives from genit-al, or antero- in ant[e]roposterior from anter-ior, or org[a]no- in organoplàstia from òrgan. But then we would get the wrong results because the base in the derivatives is destressed and has a reduced [ə]. The identification function should be made precise in other important ways which I cannot examine in detail here. For instance, when the morphological pro- cess is not asymmetrical, as mightℑ be the case sometimes in word formation, and is typical of inflection, the asymmetry has to be built into itself, which will have to be able to select a single element according to some criterion (see Albright 2002, 2005, 2008, 2010, Steriade 2008, for evidence in favorℑ of asymmetrical OO rela- tions affecting inflection). Steriade & Yanovich (2015) have also shown that OO identity constraints must have in some cases an existential interpretation: a candi- date of a derivative satisfies the OO constraint if there exists some member in the 44 CatJL 15, 2016 Joan Mascaró paradigm of the base such that the properties in the candidate and in this member mentioned by the constraint match. In order to make the morphological base of the correspondence relation more precise, I will assume an analysis of lexical representation that follows proposals in Jackendoff (2010b) and Jackendoff & Audring (to appear), an elaboration of Jackendoff (1975)—see also Bermúdez-Otero (2012). Complex words are fully specified in the lexicon and morphological schemas (similar to the redundancy rules in Jackendoff 1975) express lexical morphological regularities and relate specific lexical entries. I will not formulate the morphological processes/schemas here; for specific proposals, see the references cited. In any case, for a compound like obre-llaunes ‘can opener’, lit. ‘opens cans’ (4c) a morphological compounding process will relate the compound structure (21a-c) to the individual words (21d-f) and (21g-i). (21) obrellaunes ‘can opener’ llaunes ‘cans’, obre ‘opens’ α a. [object ; [open1(indef, can2, α)]]3 d. [plural4(can5)]6 b. [N [V3sg X ]1[N, pl Y ]2 ]3 e. [N, pl N5 – af4]6 c. [/ɔbɾə/1 /ʎawnə-s/2 ]3 f. [/ʎawnə5-s4/ ]6 g. [open7; pres8]9 h. [V [V3sg,pres.ind.8 V7]9 i. [/ɔbɾə/7,8 ]9 The entry obrellaunes consists of semantic (21a), morphosyntactic (21b) and phonological (21c) information, related by coindexation. Semantically it refers to an object objecta that is related to the predicate open which takes three arguments, an agent, a theme and an instrument; the objecta refers to the third argument, the instru- ment a. Morphosyntactically, it has the structure (21b) with a verbal and a nominal constituent. Coindexation establishes associations among parts of these structures: the meaning open1 is linked to the morphosyntactic constituent [V3sg X ]1 and to the phonological representation /ɔbɾə/1; can2 is linked to the constituent [N, pl Y ]2 and to /ʎawn-ə-s/2. The compounding process MCpd relates (21a-c) to the pair (21d-f), (21g-i). In particular it relates /ɔbɾ-ə/1 in (21c) to /ɔbɾ-ə/9 in (21i), and since the output of /ɔbɾ-ə/9 is [ɔ́βɾ-ə], can identify the base of /ɔbɾ-ə/1 in (21c) for the process MCpd as [ɔ́βɾ-ə] in (21i). Consider now the cases in which returns Ø, illustrated in (22-23) with cobre-llit and Taga-manentℑ . Since relatedness is marked by coindexation, the structure of the lexical entries allows for different kindsℑ of unrelatedness, expressed by lack of coindexation: (22) a. [bed cover]3 d. [bed4]5 b. [V [ X ]1[N Y ]2 ]3 e. [N N4]5 c. [/kɔβɾə/1 /ʎit/2 ]3 f. [/ʎit/4 ]5 Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 45 (23) a. [mount tagamanent]3 b. [N [ X ]1 [Y ]2 ]3 c. [/tagə/1 /manen/2 ]3 In (22a-c) the common noun meaning ‘bed cover’ (8) is connected through the subindex 3 to the global morphosyntactic and phonological structures, but the morphosyntactic and phonological subconstituents are unrelated to any semantic substructure. In the entry (22a-c) the second constituents in (22b,c) might be relat- ed through a morphological schema or process to the word (22d-f), but the first constituent cannot relate to any existing word. In (23a-c), representing the place name Taga-manent (8), the morphosyntactic and phonological subconstituents are not related to any existing, independent lexical entry. The lack of coindexation has as a consequence the impossibility for to identify a base; the morphologi- cal process MCpd cannot relate /kɔβɾə/1, or /tagə/1 to any base, hence (/kɔβɾə/1, ℑ 19 [/kɔβɾə/1 /ʎit/2 ]3, MCpd)=Ø and (/tagə/1, [/tagə/1 /manen/2 ]3 , MCpd)=Ø. ℑ ℑ 7. The problem of the missing base We now get to the crucial question. What happens when the function returns Ø? Obviously, if the identification of a constituent fails and an output base is missing, ℑ B an OO constraint like (20) is trivially satisfied, since there is no CO such that (C, B W, MCpd)=CO. An important prediction follows: there should be no misapplication (underapplication or overapplication) effects in this case, whereas we will find themℑ B when returns some CO. One instance of “missing bases” of this kind has been analyzed in detail in Trommer (2006, 2013) and is discussed in Bermúdez-Otero (2011).ℑ In Albanian nonactive verb forms like [foɾ.mó.hem] ‘form-nonact.1sg’ should have final stress, but they retain the stress that appears in the active forms, as in [foɾ.mój] ‘form-act.1sg’. There are however deponent verbs like the verb ‘to regret’, whose paradigm lacks active forms. Since the active base cannot be identi- fied, the prediction is that in the case of a missing base the nonactive forms should show regular, not misapplied stress. This prediction is not borne out; stress falls on the stem also in these cases: [pen.dó.hem], *[pen.do.hém]. Bermúdez-Otero (2011) analyzes in detail two other cases, Quito Spanish /s/-voicing and lenition of linking and intrusive [ɹ] in nonrhotic English dialects.20 In the case of Catalan compound structures, we get similar predictions for an OO analysis. Thus for the compounds in (24a), repeated from (4), for which returns an output word or a word constituent, underapplication of vowel reduction ℑ is correctly predicted. For those in (24b), repeated from (11), (12), LEX returns a ℑ 19. One possibility is that there is also an allomorph /kɔβɾ/, in addition to /kubr/, that would be select- ed in these particular instances. This would allow ℑLexa to select it, but since /kɔβɾ/ never has an independent output, GenEval would return Ø. 20. Bermúdez-Otero (forthcoming) presents three more cases, schwa epenthesis in Itelmen intransitive verbs, failure of glidingℑ of stem-final prevocalic /i/ in Bothoa Breton verbs and debuccalization of word-final prevocalic /s/ in Northern Chilean dialects of Spanish. 46 CatJL 15, 2016 Joan Mascaró lexical element (bound root, prefix), but GenEval cannot return an output, hence gives Ø and we incorrectly predict the reduced vowel. In (24c) LEX cannot identify a base, hence =Ø, and again we predictℑ the wrong vowel. Therefore, in general,ℑ for missing output bases an OO analysis makes the wrong predictions.ℑ ℑ (24) Phonetic form LEX GenEval Predicted vowel a. ɔβɾə-ʎáwnəs ℑLEX=/ɔbɾ/ ℑGenEval(/ɔbɾ/)=[ɔ́βɾə] [ɔ] tɾɛntə-síŋ ℑLEX=/tɾɛntə/ ℑGenEval(/tɾɛntə/)=[tɾɛ́ntə] [ɛ] b. laβju-ðəntál ℑLEX=/laβju/ ℑGenEval(/laβju/)=Ø *[ə] pɔst-upəɾətɔ́ɾi ℑLEX=/pɔst/ ℑGenEval(/pɔst/)=Ø *[u] c. kɔβɾə-ʎít ℑLEX=Ø ℑGenEval(Ø)=Ø *[u] peɾə-təʎáðə ℑLEX=Ø/ ℑGenEval(Ø)=Ø *[ə] bɔj-əskút ℑLEX=Ø/ ℑGenEval(Ø)=Ø *[u] There is one class ofℑ compounds, however,ℑ that yields to an alternative account under a parallel analysis. In reduplicatives (§3.2) like nyeu-nyeu (10) we could attribute the lack of reduction to the reduplicative process itself, which requires phonological identity between the two components. A constraint requiring partial identity, including identity in vowel features would penalize the candidate with reduction in the first component ([[ɲəw][ɲɛ́w]]) and favor [[ɲɛw][ɲɛ́w]].21 The predictions of an OO analysis can be illustrated with [A[Nlaβj-u][A[Nðənt] ál]] ‘labiodental’ whose second member, the derivative [A[Nðənt]ál], like our pre- vious example [Nkəjg-úð-ə-s] in [N[Vpaɾ-ə][Nkəjg-úð-ə-s]], has undergone regular vowel reduction (cf. [Ndén] ‘tooth’), but the vowel in [Nlaβj-u] should not reduce. In [N[Vpaɾ-ə][Nkəjg-úð-ə-s]], [Vpaɾ-ə] has a base, and the candidate [Vpəɾ-ə] with reduced [ə] in the first constituent (17d) above violates OO-Ident(V-features) because there is an output [Vpáɾ-ə]. But in (25) below the first constituent of the candidate (25d) [A[Nləβj-u][A[Nðənt]ál]] has a missing base, because it lacks an output base; therefore the OO constraint is trivially satisfied by all candidates and (25d) is the unwanted winner.22 21. I owe this observation to Eulàlia Bonet. 22. The first constituent of labiodental can of course be related to the first constituent of other words like labiovelar, but we do not know the output of labiovelar unless we find a base, and we run into a vicious circle. One could entertain the idea that, since /labio/ is a lexical element, it is evaluated, the optimal candidate being the null parse (McCarthy & Wolf 2010). The OO constraint would then pick as the base the next most harmonic candidate. There are obvious empirical and theoretical difficulties in such an analysis. In any case, it cannot be extended to isolates. Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 47 (25) Missing base, parallel analysis ([Nlabj-o], MCpd)=Ø dent ̌ ̌ ℑ[A[Nlabj-o][A[Ndent]al]] OO-I (V-features) *[Wd ...ˈV...ˈV...] *a,ɛ,ɔ,ě,ǒ Base: none a. ☐ [A[Nlaβbj-u][A[Nðənt]ál]] *! b. [A[Nláβbj-u][A[Nðént]ál]] *!* c. [A[Nláβbj-u][A[Nðənt]ál]] *! d. F [A[Nləβbj-u][A[Nðənt]ál]] 8. The stratal analysis A stratal analysis does not rely on the existence of an independent word [lábj-o], [péɾə] or [kɔ́βɾə], or one that contains these constituents; the existence of the constituent in the compound’s lexical representation is sufficient to predict underapplication. Since both first members of a compound with a base (para-caigudes) and first members missing a base (cobre-llit) are constituents, underapplication effects apply equally to both. Given the constituent structure [A[Nlabj-o][A[Ndent]al]], [Nlabj-o] and [A[Ndent]al] are computed at the word stratum, and the corresponding outputs are [lábj-u]N and [[dənt]Nál]A (26a). At this level, Final Stress, *a,ɛ̌ ,ɔ̌ ,ě,ǒ, and Wd*[ ...ˈV...ˈV...] must dominate Id(V-features), as shown in (26a). This causes reduction of all destressed vowels. At the phrase level (26b), the input [[lábj-u]N[[dənt]Nál]A] is submitted to evaluation by a different constraint ordering, with Id(V-features) and *[Wd ...ˈV...ˈV...] *a,ɛ̌ ,ɔ̌ ,ě,ǒ; this prevents phrase-destressed vowels from reducing. ≫ (26) Missing base, stratal analysis a. Word stratum [Wd ...ˈV...ˈV...] Final Stress,ɛ *a, ̌ ,ɔ̌ ,ě,ǒ *Id(V- features) ≫ ≫ [Nlabj-o] *[Wd ...ˈV...ˈV...] Final Stress *a,ɛ̌ ,ɔ̌ ,ě,ǒ Id(V-features) F [Nlábj-u] 1 [Nlabj-o] 1W 2W L [Nlábj-o] 1W L [A[Ndent]al] *[Wd ...ˈV...ˈV...] Final Stress *a,ɛ̌ ,ɔ̌ ,ě,ǒ Id(V-features) F [A[Ndənt]ál] 1 [A[Ndént]ál] 1W L [A[Ndént]əl] 1W 1 [A[Ndent]ál] 1W L 48 CatJL 15, 2016 Joan Mascaró b. Phrase stratum *[Wd ...ˈV...ˈV...], Id(V-features) *a,ɛ̌ ,ɔ̌ ,ě,ǒ [A[Nlábj-u][A[Ndənt]ál]] *[Wd ...ˈV...ˈV...] Id(V-features)≫ *a,ɛ̌ ,ɔ̌ ,ě,ǒ F [A[Nlaβj-u][A[Nðənt]ál]] 1 [A[Nláβj-u][A[Nðənt]ál]] 1W L [A[Nləβj-u][A[Nðənt]ál]] 1W L In neoclassical compounds like labiodental and in the case of prefixes a base can be identified which is not an output. In the case of proper and some common nouns in (9), reduplicated forms (10) and borrowings (11), usually no base can be identi- fied. For these cases we have to assume a constituent structure that is not justfied by the existence of a base. The lexical configurations proposed at the end of §6 allow for such structures, i.e. for the existence of morphosyntactic constituent structure which is not matched by strictly compositional semantics. Thus for the place name Tagamanent in (23) coindexation relates the constituents in [/tagə/1 /manen/2 ]3 to the morphosyntactic constituents in [N [ X ]1 [Y ]2 ]3, deriving /[[tagə][manen]]/, although there is a single global meaning ‘mount tagamanent’, the internal consti- tuent not being coindexed with individual semantic representations. 9. Conclusions I have shown that there is a large set of compound structures that share two proper- ties: a first component with a missing base and underapplication of vowel reduc- tion. Lack of vowel reduction in compounds with an identifiable base for the first component yields to both transderivational output-to-output and cyclic analyses. The compound structures with a first component missing a base cannot be handled through OO-constraints, but they can be derived cyclically under the reasonable assumption that lexical items can lack semantic compositionality while still retain- ing morphosyntactic constituent structure. At the same time, however, there are analogical effects that seem difficult to derive in a stratal analysis, without OO constraints. This is clearly the case whenever can identify an output as a base, but this base is not morphologically contained in the derived form, as in many cases of bases in inflectional paradigms (Steriadeℑ 2007, Steriade & Yanovich 2015, Albright 2002 and subsequent work). Thus in Romanian (Steriade 2008) the derivative stɨndʒ-íst ‘leftist’ appears with the palatalized consonant of the plural stɨ́ndʒ-i̯ ‘left(hand)-pl.’, not with the underlying velar that shows up in the singular stɨ́ŋɡ-ʌ. Obviously, the derivative does not contain morphologically the plural, i.e. *[[N,pl stɨ́ndʒ-i̯ ] íst ]. In other words, some cases derive from analogic effects caused by the influence of words on words, but others derive from the way lexi- cal representations are structured in terms of constituent structure and from cyclic effects. Whether some unification of these two mechanisms is possible must be left for future research. Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 49 References Albright, Adam. 2002. The identification of bases in morphological paradigms. Ph.D. dissertation, UCLA. Albright, Adam. 2005. The morphological basis of paradigm leveling. In Downing et al. (2005a), 17-43. Albright, Adam. 2008. Inflectional paradigms have bases too: evidence from Yiddish. In Bachrach & Nevins (2008a), 271-312. Albright, Adam. 2010. Base-driven leveling in Yiddish verb paradigms. Natural Language & Linguistic Theory 28.3: 475-537. Bachrach, Asaf & Nevins, Andrew (eds.). 2008a. Inflectional Identity. Oxford: Oxford University Press. Bachrach, Asaf & Nevins, Andrew. 2008b. Introduction: approaching inflectional iden- tity. In Bachrach & Nevins (2008a), 1-28. Benua, Laura. 1997. Transderivational identity: phonological relations between words. Ph.D. dissertation, University of Massachusetts, Amherst, MA. Available at ROA- 259-0498, Rutgers Optimality Archive, Jackendoff, Ray. 2010a. Meaning and the Lexicon. Oxford: Oxford University Press. Jackendoff, Ray. 2010b. The ecology of English noun-noun compounds. In Jackendoff (2010a), 413-45. (An expanded version of Compounding in the Parallel Architecture and Conceptual Semantics, in Rochelle Lieber & Pavol Štekauer (eds.). The Oxford Handbook of Compounding, 105-28. Oxford: Oxford University Press, 2009. Jackendoff, Ray, & Audring, Jenny. To appear. Morphology in the Parallel Architecture. In Audring, Jenny & Masini, Francesca (eds.). The Oxford Handbook of Morphological Theory. Oxford: Oxford University Press. Kager, René. 1999. Optimality Theory. Cambridge: Cambridge University Press. Kenstowicz, Michael. 1996. Base-identity and uniform exponence: alternatives to cyclicity. In Durand Jacques & Laks, Bernard (eds.). Currents trends in phonol- ogy: models and methods, vol. 1, 363-393. Salford: European Studies Research Institute, University of Salford. Kiparsky, Paul. 2000. Opacity and cyclicity. In Ritter, Nancy A. (ed.). A review of Optimality Theory. The Linguistic Review (Special issue), 17.2-4: 351-67. Kiparsky, Paul. In press. Paradigms and Opacity. Vol. I. Stanford: CSLI. Köhnlein, Björn. 2015. The morphological structure of complex place names: the case of Dutch. Journal of Comparative German Linguistics 18: 183-212. Mascaró, Joan. 1976. Catalan Phonology and the Phonological Cycle. Ph.D. disserta- tion, MIT: Published by Indiana University Linguistics Club, Bloomington, 1978. Expanded Catalan translation: La fonologia catalana i el cicle fonològic. Bellaterra: Sèrie Lingüística, UAB, 1983. Mascaró, Joan. 2002. El sistema vocàlic. Reducció vocàlica. In Solà et al. (2002), vol. I, 89-123. Mascaró, Joan. 2015: Regularitat i excepcions en fonologia: les reduccions vocàliques. In Lloret, Maria-Rosa, Pons-Moll, Clàudia & Bosch-Roura, Eva (eds.). Clàssics d’ahir i d’avui en la gramàtica del català, 59-69. Barcelona: Universitat de Barcelona. McCarthy, John J. & Wolf, Matthew. 2010. Less than zero: Correspondence and the null output. In Blaho, Sylvia & Rice, Curt (eds.). Modeling Ungrammaticality in Optimality Theory, 17-66. London: Equinox. Moll, Francesc de B. 1931. Estudi fonètic i lexical del dialecte de Ciutadella. In Miscelànea Filològica dedicada a don Antonio Ma. Alcover. Palma, 397-460. Reproduced in Randa 8, 1979: 5-48. Nadeu, Marianna. 2016. Phonetic and phonological vowel reduction in Central Catalan. Journal of the International Phonetic Association 46.1: 33-60. Oliva, Salvador. 1992. La mètrica i el ritme de la prosa. Barcelona: Quaderns Crema. Paul, Hermann. 1968 [1880]. Prinzipien der Sprachgeschichte. Tübingen: Max Niemeyer. Prieto, Pilar. 2003. Correlats acústics de l’accent secundari en català. Estudios de Fonética Experimental 12: 107-142. Recasens, Daniel. 1993. Fonètica i fonologia. Barcelona: Enciclopèdia Catalana. Riera-Eures, Manel. 2002. Diccionari d’onomatopeies i mots de creació expressiva: les paraules transparents de la llengua catalana. Barcelona: Edicions 62. Solà, Joan, Lloret, Maria-Rosa, Mascaró, Joan & Pérez Saldanya, Manel (eds.). 2002. Gramàtica del català contemporani. Barcelona: Empúries. Morphological Exceptions to Vowel Reduction in Central Catalan CatJL 15, 2016 51 Steriade, Donca. 2008. A pseudo-cyclic effect in the Romanian declension. In Bachrach & Nevins (2008), 313-361. Steriade, Donca & Yanovich, Igor. 2015. Accentual allomorphs in East Slavic: an argu- ment for inflection dependence. In Bonet, Eulàlia, Lloret, Maria-Rosa & Mascaró, Joan (eds.). Understanding Allomorphy, 254-314. Shefield and Bristol: Equinox. Trommer, Jochen. 2006. Stress uniformity in Albanian: morphological arguments for cyclicity. Paper given at the workshop Approaches to Phonological Opacity, 29th GLOW Colloquium, Barcelona. Trommer, Jochen. 2013. Stress uniformity in Albanian: morphological arguments for cyclicity. Linguistic Inquiry 44.1: 109-143. Wheeler, Max W. 2005. The phonology of Catalan. Oxford: Oxford University Press. ISSN 1695-6885 (in press); 2014-9719 (online) Catalan Journal of Linguistics 15, 2016 53-66 Gradient Exceptionality in Maximum Entropy Grammar with Lexically Specific Constraints* Claire Moore-Cantwell Yale University [email protected] Joe Pater University of Massachusetts Amherst [email protected] Received: February 25, 2016 Accepted: June 22, 2016 Abstract The number of exceptions to a phonological generalization appears to gradiently affect its productivity. Generalizations with relatively few exceptions are relatively productive, as measured in tendencies to regularization, as well as in nonce word productions and other psycholinguistic tasks. Gradient productivity has been previously modeled with probabilistic grammars, including Maximum Entropy Grammar, but they often fail to capture the fixed pronunciations of the existing words in a language, as opposed to nonce words. Lexically specific constraints allow existing words to be produced faithfully, while permitting variation in novel words that are not subject to those constraints. When each word has its own lexically specific version of a constraint, an inverse correlation between the number of exceptions and the degree of productivity is straightforwardly predicted. Keywords: exceptions; variation; computational phonology; Maximum Entropy Grammar; indexed constraints Resum. Excepcionalitat gradual i gramàtica de màxima entropia amb restriccions especificades lèxicament El nombre d’excepcions a una generalització fonològica sembla que afecta de forma gradual la seva productivitat. Les generalitzacions amb relativament poques excepcions són bastant pro- ductives, per les mesures en tendències a la regularització i per les produccions de mots sense sentit i altres tasques psicolingüístiques. La productivitat gradual s’ha modelat prèviament amb gramàtiques probabilístiques, incloent-hi la gramàtica de màxima entropia, però sovint no acon- segueixen recollir les pronunciacions fixes de paraules existents en una llengua, contràriament al que passa amb les paraules sense sentit. Les restriccions especificades lèxicament permeten produir els mots existents de manera fidel i al mateix temps permeten variació en mots nous, que * This material is based on work supported by NSF Graduate Research Fellowship DGE-0907995 to C.M.C., NSF grant BCS-424077 to the University of Massachusetts Amherst, and a City of Paris Research in Paris fellowship to J.P. We thank the participants in the OCP Exceptionality Workshop, and members of the UMass phonological community, especially Robert Staubs, for helpful discussion. 54 CatJL 15, 2016 Claire Moore-Cantwell; Joe Pater no estan subjectes a aquestes restriccions. Quan cada mot té la seva pròpia versió d’una restricció especificada lèxicament es prediu directament una correlació inversa entre el nombre d’excepcions i el grau de productivitat. Paraules clau: excepcions; variació; fonologia computacional; gramàtica de màxima entropia; restriccions indexades Table of Contents 1. The problem: gradient exceptionality 3. Case study: Dutch voicing alternations 2. The proposal: Maximum Entropy 4. Conclusions Grammar with Lexically Specific References Constraints 1. The problem: gradient exceptionality The most general form of the problem that we address in this paper is the inad- equacy of a two-way distinction between a regular / rule-governed / general pho- nological pattern and an exceptional / lexical / minor one. This two-way theoretical distinction is inadequate because it does not match the observed data. Phonological patterns across languages display a continuum of productivity, or conversely, of exceptionality. This is demonstrated at length in Hayes’ (2008: ch. 9) textbook chapter, which offers a number of examples at various points of this continuum. We discuss two empirical examples of gradient exceptionality in this paper. In this first section, we use the example of gradient exceptionality in lexical stress placement to elaborate on the problem, and in Section 2 we use it to exemplify our solution. In section 3, we present computational modeling results for another case, that of voicing alternations in Dutch (Ernestus and Baayen 2003). 1.1. Gradient exceptionality in lexical stress placement The problem of gradient exceptionality was perhaps first pointed out by Fidelholtz (1979: 58): It appears to be a problem for linguistic theory that there is nothing in the formal description of Polish stress which would indicate that Polish is a ‘penultimate-stress’ language, as compared with the similar rules in English, which is essentially a free-stress language. When the penultimate syllable is light, stress falls on the penultimate syllable of some English and Polish words (e.g. English banána, Polish spokójny ‘quiet’), and on the antepenult on others (e.g. English Cánada, Polish fízyka ‘physics’). In English, both patterns are well-attested (Pater 1994), and each word’s pronuncia- tion is stable; there is apparently no regularization to either penultimate or ante- penultimate stress. In Polish, there are very few antepenultimately stressed words – Peperkamp et al. (2010) note that about 0.1% of the vocabulary has exceptional stress, which also includes final stress. Antepenultimately stressed words tend to Gradient Exceptionality in Maximum Entropy Grammar CatJL 15, 2016 55 be borrowings and or learned words, and there is frequent regularization to penul- timate stress. Besides frequency of regularization, another piece of evidence that the differ- ence in lexical statistics correlates with a difference in productivity comes from Peperkamp et al. (2010) psycholinguistic research with native speakers of languag- es with varying degrees of exceptionality. They use a nonce word memory task to determine how well participants are able to encode stress differences. Participants hear a sequence of 5 bisyllables that differ only in the placement of stress. The task is to report which syllable was stressed in each of the words. Performance on this task is compared with performance on a similar task with a segmental contrast. Peperkamp et al. (2010) find evidence of three-way grouping of participants by language background. Spanish participants, whose language has the highest degree of exceptionality (closest to English), perform as well on the stress contrast as on the segmental contrast. French, Finnish and Hungarian participants, whose lan- guages have predictable stress, perform much worse on the stress contrast. Polish participants’ level of performance is in between the two other groups. The problem in linguistic theory that Fidelholtz is alluding to in the cited passage is that the number of exceptions, few or many, does not affect the grammatical status of a pattern in a standard generative grammar. The formal descriptions of both English and Polish, as well as Spanish, would require lexical marking of one of the patterns, with the other generated by rule. This is true of the SPE formalism of Fidelholtz’s time, of metrical rule approaches, and of OT accounts with lexically specific constraints, posited independently by Kraska- Szlenk (1995) for Polish and Pater (2000) for English stress. Standard generative accounts can generate the existing words of the languages, but they do not account for the varying degrees of productivity of the patterns. 1.2. Previous generative approaches to gradient productivity Gradient productivity has been the subject of considerable research in the recent generative literature, especially in the modeling of nonce word judgments or pro- ductions. Zuraw (2000), Ernestus and Baayen (2003), Hayes and Wilson (2008) and Hayes, Zuraw, Siptár and Londe (2009) and others attack the problem using stochastic grammars, either Stochastic OT (Boersma 1998) or Maximum Entropy Grammar (MaxEnt: Goldwater and Johnson 2003; referred to as a log-linear model in Ernestus and Baayen). As a simple example of this approach, the tableaux in (1) and (2) show the activity of two MaxEnt grammars, one generating penultimate stress with 0.95 probability, and another choosing between penultimate and ante- penultimate stress at chance. We consider only left-headed binary feet, and have all unstressed vowels as schwa, as in English. The constraints choosing the posi- tion of the foot are Align-R, which demands a foot at the right edge of the word, and Nonfinality, which demands that the final syllable be unfooted. Violations are indicated by negative integers, and the weights of the two constraints are shown beneath their names. The column labeled H shows the weighted sum of violations, or Harmony (Smolensky and Legendre 2006). The probability of each candidate 56 CatJL 15, 2016 Claire Moore-Cantwell; Joe Pater is shown in the last column; it is proportional to the exponential of its Harmony. When Align-R has higher weight, as in (1), penultimate stress has higher prob- ability. When the two constraints have equal weight, as in (2), penultimate and antepenultimate stress have equal probability. (1) A grammar choosing penultimate stress 95% of the time /bætækæ/ Align-R NonFin H p 4 1 bə(tǽkə) –1 –1 0.95 (bǽtə)kə –1 –4 0.05 (2) A grammar choosing penultimate stress 50% of the time /bætækæ/ Align-R NonFin H p 2 2 bə(tǽkə) –1 –2 0.50 (bǽtə)kə –1 –2 0.50 These tableaux would be appropriate for a case in which multiple pronunciations of a single word are generated by a single speaker’s grammar. For the production of a nonce word by a native speaker of English, a grammar of this type would be appropriate, since for a given word of this shape either penultimate or antepenultimate stress would be assigned with about equal probability (Moore- Cantwell 2015). For existing words in the English lexicon, however, this would not be the correct analysis (or as we will shortly claim, it’s not the complete analysis). Words with final schwa in English, like Cánada and banána do have close to a 50/50 split in penultimate vs. antepenultimate stress (Moore-Cantwell 2015), but each of the words is produced in a single way. In other words, English stress displays what’s traditionally called exceptionality, or what we might more neutrally call lexically conditioned application. An unelaborated MaxEnt grammar of this type is instead appropriate for what’s traditionally called variation (Goldwater and Johnson 2003), as displayed in English t/d-deletion (though it’s worth noting that variation is often, if not always, lexically conditioned; Coetzee and Pater 2011). The challenge in modeling gradient exceptionality is to account for how it can affect responses in nonce word production and other psycholinguistic tasks, while at the same time allowing existing real words to be produced in a non-variable way. This challenge has long been recognized (Zuraw 2000), but it is sometimes not addressed. For example, Hayes et al. (2009) provide a MaxEnt grammar for Hungarian vowel harmony that is trained on the lexicon, and that generates patterns that are compared with nonce word productions. They do not confront the problem that the grammar would generate the wrong outcome for many real Hungarian words (see also Ernestus and Baayen’s 2003 similar Stochastic OT and MaxEnt modeling of Dutch voicing alternations). To deal with gradient Gradient Exceptionality in Maximum Entropy Grammar CatJL 15, 2016 57 exceptionality, Zuraw (2000) develops a model in which morphologically complex words are stored both as wholes and in decomposed forms, with the grammar choosing between the two based on the constraint ranking. This might be applied to the Dutch or Hungarian examples, but it does not seem to be applicable when the gradient pattern is over underived words, as in the stress cases we have been discussing here. Conversely, Hayes and Wilson (2008) develop a MaxEnt model of phonotactics that defines a probability distribution over the space of possible words, and which generates well-formedness scores for underived nonce forms. The Hayes and Wilson (2008) model is not applicable to alternations (though see the extension in Allen and Becker 2015 and Gouskova and Becker to appear). To meet this challenge, we propose to combine MaxEnt grammar with lexically specific constraints, which are able to encode lexical conditioning (see Pater 2010 for an overview and comparison with alternatives, such as co-grammars). Real words have stable pronunciations because their associated lexically specific con- straints have sufficiently high weight; nonce words have no associated constraints. Lexically conditioned constraints have in fact been suggested as a means of coping with gradient exceptionality in a deterministic version of OT (Pater 2005, Becker, Nevins and Ketrez 2011), but this requires a special grammar-external calculation over the lexicon to get the influence of the degree of regularity. 2. The proposal: Maximum Entropy Grammar with Lexically Specific Constraints We assume that the grammar is composed of general constraints, and lexically specific versions of (some of) them, indexed to particular items. Continuing with the example from the last section, we show in (3) how lexically indexed constraints can stabilize the pronunciation of individual lexical items. As we saw in (2) above, equally weighted Align-R and Nonfin on their own generate equal probability for penultimate and antepenultimate stress. The tableau in (3) adds an indexed Align-R constraint for banana, whose weight leads to near-fixed penultimate stress for that word, and an indexed Nonfin constraint for Canada, whose weight leads to near-fixed antepenultimate stress. A correct grammar would presumably make the probability of misstressing either of these words vanishingly low. Scaling these weights (multiplying them by a constant) would bring the probability of the correct form arbitrarily close to 1. (3) Align-R-i Nonfin-j Align-R Nonfin H p 5 5 2 2 → bə(nǽnə)i –1 –2 0.99 (bǽnə)nəi –1 –1 –7 0.01 kə(nǽdə)j –1 –1 –7 0.01 → (kǽnə)dəj –1 –2 0.99 58 CatJL 15, 2016 Claire Moore-Cantwell; Joe Pater There is a range of possibilities for how the set of lexically specific constraints is composed. In Pater (2010), for example, it is proposed that lexically specific constraints are posited only to resolve inconsistency. For example, if we had just Align-R and Nonfin in our constraint set, and no other constraints (such as faith- fulness) to distinguish Canada from banana, then adding one of the indexed con- straints would be necessary, and sufficient, to generate the correct pronunciations. The proposal in Pater (2010) does not deal with gradient productivity. To do so, we assume that constraints have lexically specific instantiations for every lexical item. Because we have separate constraints for every lexical item, the number of items displaying one or another pattern can affect both the weight of the general constraints, and of their lexically specific versions. When a pattern is relatively general across the lexicon, the general constraint winds up with a relatively high weight. In this case there are relatively few exceptions, and the lexically specific constraints encoding them need to have relatively high weight for the words to be pronounced in a fixed fashion. On the other hand, when a pattern is relatively rare, the general constraint winds up with a relatively low weight. Exactly how lexically specific constraints are induced is a topic we leave for further research, as it will require considering cases more realistic than the small illustrations we present here (for example, to determine whether unworkably large constraint sets are created under some assumptions). One general question in previ- ous research on lexically indexed constraints is whether markedness constraints, faithfulness constraints, or both, can be indexed (see Pater 2010 for references and discussion). Choices in this regard will not only be important in terms of keeping the size of the constraint set manageable, but it will also effect how exactly learn- ers generalize. An attractive possibility is that the lexically indexed constraints are simply the constraints encoding the phonological features of the word, that is, that they are realizational constraints in the sense of Xu (2007), but a full exploration of the consequences of abandoning underlying representations and faithfulness is of course beyond the scope of this short paper. As an illustration of our proposal, we consider a set of simple stress ‘languages’ with varying degrees of regularity. Stress can fall in one of two positions, and there is a general constraint preferring each (e.g. Align-L and Align-R). There are 100 words, and thus 100 lexically specific versions of Align-L, and 100 of Align-R. The languages have stress in one position in 100 to 50 words (fully regular to fully lexical), making 51 languages. Our learning algorithm is batch Gradient Descent. It is similar to the Stochastic OT / HG Gradual Learning Algorithm (GLA; Boersma 1998; Boersma and Pater 2016), except that each epoch is a presentation of the entire dataset (see Pater and Staubs 2013 and Moreton, Pater, and Pertsova 2015). We used batch Gradient Descent for convenience: a single run closely approximates the averaging of results of multiple runs of the regular on-line GLA (the batch algorithm doesn’t have a stochas- tic component, and thus gives a single outcome for a single starting point). We ran it for 1000 epochs, with a learning rate of 0.1, and starting constraint weights of zero. The resulting grammars give probability near 1 for the correct pronunciations of the words it was trained on. To generate probabilities for nonce word productions, Gradient Exceptionality in Maximum Entropy Grammar CatJL 15, 2016 59 1.0 0.8 Pattern-observing output 0.6 bability 0.4 Pro 0.2 Exceptional output 0.0 0 10 20 30 40 50 Number of Exceptions Figure 1. Probabilities assigned by grammars with only the general constraint weights, learned for grammars with 0-50 exceptions. we simply remove the lexically indexed constraints. The resulting probabilities are shown in Figure 1. When there are no exceptions, the grammar without the lexically specific con- straints assigns stress following the general pattern with probability 1.0, and when the number of forms following each pattern is equal, nonce words are predicted to be assigned to one or the other pattern with equal probability. In between, the grammar assigns a range of probabilities, thus allowing for the modeling of gradi- ent productivity. Interestingly, the curve is not linear: adding an exception to a language with relatively few exceptions is predicted to have less of an effect on productivity than adding an exception to a language that has relatively many. We do not know of any current data that bear in this prediction. Our claim that gradient productivity exists across the stress patterns of the world’s languages was not based on nonce word data. In section 1.1, we introduced two pieces of evidence that the stress pattern of Polish, while admitting exceptions, is more productive than that of English or Spanish: exceptions are more often regularized, and speakers have more difficulty encoding lexical stress in an experimental task. In terms of our model, these differences between Polish on the one hand, and English and Spanish on the other, can be understood as follows. In Polish, the general constraint(s) demanding penultimate stress would have relatively high weight compared with English or Spanish. For a word with antepenultimate stress to be encoded faithfully, the relevant lexically specific constraint would itself need to have relatively high weight. Thus, encoding lexical stress would require more experience for a Polish learner (assuming gradual learning), and absent that experience, regularization or unfaithful encoding would be predicted to be more common. 60 CatJL 15, 2016 Claire Moore-Cantwell; Joe Pater 3. Case study: Dutch voicing alternations In this section, we turn to a case of gradient productivity within a single language, for which we have existing nonce word data to model. For this detailed case study, we analyze the classic case of Dutch voicing alternations presented by Ernestus and Baayen (2003). In Dutch, word-internal obstruents contrast in voicing, as can be seen in the minimal pair [vɛrʋɛidɛn] (‘to widen’) vs. [vɛrʋɛitɛn] (‘to reproach’). However, obstruents devoice word-finally, leading to neutralization: [vɛrʋɛit] is a homophone, meaning either widen or reproach. Final neutralization is exceptionless in Dutch, but Ernestus and Baayen (2003) show that there are gradient phonologi- cal generalizations about the likelihood that a word-final voiceless obstruent will surface as voiced under suffixation – that is, in terms of the standard analysis, about whether it is underlyingly voiced. In the lexicon of Dutch, both the place and manner of the word-final obstru- ent affects the likelihood that a word will undergo alternation, with the obstruent becoming voiced word-internally. For example, as shown in Table 1, about 70% of words with word-final labial fricatives (f/v) alternate, while only 9% of words with word-final labial stops alternate. Data from Ernestus and Baayen’s (2003) search of the CELEX corpus are given under the column headed “% voiced in lexicon”. Also shown in Table 1 are the percentages from Ernestus and Baayen’s pro- duction experiment. They found that participants could ‘guess’ whether or not a form should alternate based on its neutralized form, and their guesses followed the statistics of the lexicon. A participant would be presented with a nonce word, say kuuf, and would be asked for the past tense form of the word, formed by adding -te for final voiceless obstruents and -de for final voiced obstruents. A response of kuufte indicates that the participant has interpreted the final f of kuuf as underlyingly voiceless, while a response of kuuvde indicates that the participant has interpreted the f as voiced. As Table 1 indicates, for each place and manner category, participants roughly matched the probability of voicing in the lexicon. Participants exhibited an overall preference against voicing which is not seen in the lexicon (likely because they were given the devoiced form as a prompt), but otherwise, they roughly match the probability of voicing in the lexicon for each type of word-final obstruent. Table 1. Percent of word-final voiceless consonants that become voiced intervocalically under derivation, in Ernestus and Baayen’s 2003 CELEX corpus search and experimental results % voiced in lexicon % voiced in experiment p/b 9% 4% t/d 25% 9% s/z 33% 23% f/v 70% 49% x/ 97% 80% ɣ Gradient Exceptionality in Maximum Entropy Grammar CatJL 15, 2016 61 This is a case where participants ‘frequency match’, closely copying in nonce forms the distribution in the lexicon of their language. Some generalizations reported by Ernestus and Baayen are nearly exceptionless, while others have many exceptions. Our model represents the generalizations with fewer exceptions via higher weights on the general constraints, lower weights on the lexically specific constraints of observers, and very high weights on lexically specific constraints of violators. Generalizations with large numbers of exceptions are represented with lower weights on the general constraints, and similar weights on lexically specific constraints for observers and for violators. As even the title of Ernestus and Baayen’s paper implies (“Predicting the unpredictable”), the standard analysis in which the voicing of the derived form is simply stored as part of the Underlying Representation of the underived form does not predict that speakers should show awareness of the generalizations about which types of consonant alternate. An alternative is to treat the pattern as lexically conditioned intervocalic voicing (see also Becker, Kitrez and Nevins 2011 on Turkish), as in the following adaptation of Ernestus and Baayen’s (pg. 20) analysis, which uses a subset of their constraints. We used seven general constraints, including two very general constraints demanding that all intervocalic obstruents be voiced or voiceless. We note from the outset that these constraints may well be different from those one would use in a standard OT analysis, and that they gloss over details of the Dutch system, such as the effects of final vowel length, and additionally do not represent the morphological and prosodic environments relevant to the voicing alternation. We use these constraints for ease of comparison with Ernestus and Baayen’s analysis, and because they are sufficient to illustrate our approach. (4) The most general constraints *VTV Assign a violation mark to an intervocalic voiceless consonant *VDV Assign a violation mark to an intervocalic voiced consonant Each lexical item is associated with a lexically specific version of one of these constraints – words with a voiced obstruent, like [vɛrʋɛidɛn], are assigned a copy of *VTV, while words with a voiceless obstruent ([vɛrʋɛitɛn]) are assigned a copy of *VDV. For convenience, we omitted lexically specific constraints that would be violated in the correct output (these would simply receive a weight of zero in the learned grammar). (5) *VTVwiden Assign a violation mark to an intervocalic voiceless consonant when its underlying correspondent is contained in /vɛrʋɛid/ ‘widen’ *VDVreproach Assi gn a violation mark to an intervocalic voiced consonant when its underlying correspondent contained in /vɛrʋɛit/ ‘reproach’ 62 CatJL 15, 2016 Claire Moore-Cantwell; Joe Pater Additionally, we use the following five place-specific constraints, which are general in the sense that they do not pertain to particular lexical items. They militate against voicing for the obstruent types that tend to be voiceless intervocalically (p/b, t/d and s/z), and militate against voicelessness for the obstruent types that tend to be voiced intervocalically (f/v, x/ɣ). (6) Place and manner constraints *P[+voice]: Assign a violation mark to an intervocalic voiced labial stop *T[+voice]: Assign a violation mark to an intervocalic voiced coronal stop *S[+voice]: Assign a violation mark to an intervocalic voiced sibilant *F[-voice]: Assign a violation mark to an intervocalic voiceless labial fricative *X[-voice]: Assign a violation mark to an intervocalic voiceless velar fricative Note also that our training data consist of the pairs of derivationally related words from Ernestus and Baayen’s corpus search, and that they report that the sta- tistics for obstruents at other locations than root-final differ (pg. 7). A fuller simula- tion may well need to specify the constraints in (5) and (6) to root-final position. In this simulation, we did not allow all general markedness constraints to be indexed to specific lexical items; rather we only allowed the most general marked- ness constraints *VTV and *VDV to be lexically indexed. This choice was made primarily to save on processing time (with 1700 lexical items, allowing all seven constraints to be indexed to any relevant lexical item would result in a possible 6800 constraints). As discussed in Section 2, more investigation is needed into whether all constraints should be available for indexation, or if not, on what basis constraints are chosen for indexation. Like the above stress simulation, this simulation used Batch Gradient Descent, where a single learning epoch constitutes an update over the entire set of training data, which match the corpus data from Ernestus and Baayen (2003). The algorithm was run for 10,000 epochs with a learning rate of 0.01. Constraint weights were regularized using a Gaussian prior with a mean of zero and a variance of 100.1 After 10,000 epochs, the predicted percentages of voicing on all real words in the training data was very close to either 100% or 0%, depending on whether that word was voiced or voiceless in the lexicon. Table 2 shows the percentage of voic- ing predicted by our simulation for each type of input obstruent for ‘wug’ words. These are generated from the weights of the constraints, without any lexically spe- cific constraints. Again, our assumption is that novel forms do not have associated lexically specific constraints (or if they do, that they have very low weight). The overall pattern matches the place and manner distinctions shown in the lexical data and in the experiment. The trends are generally more extreme in the predictions of our model than in either the lexicon or in the experimental data; the percent voicing is closer to zero for the voiceless trend cases, and closer to 100 for voiced. Patterns 1. The parameter settings can affect the outcome, especially the setting of the regularization term. The need to tune the parameters to match the experimental data is a potential weakness of this approach. Gradient Exceptionality in Maximum Entropy Grammar CatJL 15, 2016 63 Table 2. Percentage of intervocalically voiced forms, in the training data, in the outcome of our learning simulation, and Ernestus and Baayen’s experiment Training Data % voiced % voiced %voiced Trend Lexicon no. forms Experiment (EB) Simulation p/b voiceless 9% 230 4% 1% t/d voiceless 25% 719 9% 9% s/z voiceless 33% 451 23% 18% f/v voiced 70% 166 49% 84% x/ voiced 97% 131 80% 99% ɣ in experimental data of this type tend to show more randomness than in lexica. To adjust for this, Hayes et al. (2009) use a temperature parameter in their MaxEnt model, and a similar approach could be adopted here. Figure 2 shows the model’s behavior on items with a velar fricative, which in the lexicon are voiced nearly exceptionlessly. On the left, the weights of the four types of constraints relevant to such a lexical item are plotted over the course of the learning period. Epochs are on the x axis, and weights on the y axis. The gen- eral constraint *VTV has a very low weight throughout the entire learning process (except for a ‘burn-in’ period over the first approximately 300 epochs). Because it applies to all types of word in the training data in the same way, it does much less work than the place/manner specific constraints. The constraint demanding voic- ing on velar fricatives *X[-voice] gets a relatively high weight, and in particular much higher than the weight of the lexically-specific *VTV constraints associated with lexical items which follow the trend. The weights of the lexically specific *VDV constraints, on the 3% of lexical items which violate the trend are very high - they must overcome *X[-voice] in evaluation for the exceptional words to be pronounced correctly. Figure 2. Velar fricatives – relevant constraint weights (left) and predicted probabilities of different word types (right) over the course of the simulation. 64 CatJL 15, 2016 Claire Moore-Cantwell; Joe Pater The right-hand side of Figure 2 shows the probability of voicing on words that are voiced in the lexicon (black) and words that are voiceless in the lexicon (grey), both of which have associated lexically specific constraints. The black line is also the predicted probability for nonce words – the data from the nonce words exactly matches the probabilities of the words that are voiced in the lexicon. Words that follow the trend toward voicing are learned early and at the end of learning are pronounced correctly 100% of the time. However, the exceptional voiceless words are learned much slower, and even at the end of learning still have some non-zero probability of error. Nonce words are also predicted to voice velar fricatives 100% of the time, in contrast to the 80% voicing on these items observed by Ernestus and Baayen – this is one instance of the general “trend exaggeration” seen in the output of our model discussed with respect to Table 2 above. Figure 3 shows the weights and probabilities associated with forms with [s/z], which have a lower lexical probability of intervocalic voicing than [x/ɣ], and fall in the middle of the range for different place/manner combinations in Table 2. As for [x/ɣ], the general *VTV constraint gets zero weight. The constraint *S[+voice] gets some weight, but this time both lexically specific constraints of observers (*VDVi in this case) and of violators (*VTVj) get higher weights than the general mark- edness constraint. The probability of correct pronunciation of an existing lexical item observing the generalization is the same as the correct pronunciation existing lexical item violating the generalization - close to 1.0. However, the probability of voicing on a nonce word, unspecified for voicing and lacking a lexically specific constraint, is 0.33. This is close to the value observed by Ernestus and Baayen for these forms, 23% voiced. Figure 3. Coronal fricatives – relevant constraint weights (left) and predicted probabilities of different word types (right) over the course of the simulation. 4. Conclusions Recent work in generative linguistics has begun to address the problem of gradient productivity by adopting probabilistic grammars. Our proposal seeks to address an outstanding problem in the modeling of lexically gradient patterns: that although the Gradient Exceptionality in Maximum Entropy Grammar CatJL 15, 2016 65 proposed grammar models successfully model the gradience seen in experimental results such as nonce word productions, they often do not deal with the fixed pro- nunciations seen in individual lexical items. We have shown that by incorporating lexically specific constraints into a Maximum Entropy model, both gradient pro- ductivity and fixed pronunciation of individual lexical items can be successfully modeled. Moreover, the influence of the lexicon on the weights of the constraints encoding the general patterns comes from the basic operation of the learning algo- rithm with the proposed constraint set, rather than through any special calculation over the lexicon. We have illustrated our proposal using a toy case of differences of lexical frequency of stress patterns across languages, as well as an attested case of gra- dience in nonce word production from Dutch voicing alternations. Much remains to be done in developing this model and in comparing it to alternatives, such as MaxEnt models that operate over sub-lexicons (Allen and Becker 2015, Gouskova and Becker to appear, Moore-Cantwell and Staubs 2014), and various forms of analogical model (see Ernestus and Baayen 2003 and Moore-Cantwell 2015, as well as the Appendix of Moreton 2015 on formal connections between analogical and MaxEnt models). We find the results thus far with our relatively simple model encouraging, and present this as a potentially useful further step in addressing a longstanding problem. References Allen, Blake & Becker, Michael. 2015. Learning alternations from surface forms with sublexical phonology. Unpublished manuscript, University of British Columbia and Stony Brook University. Available as lingbuzz/002503 Becker, Michael & Gouskova, Maria. To appear. Source-oriented generalizations as grammar inference in Russian vowel deletion. Linguistic Inquiry. Available as ling- buzz/001622. Becker, Michael, Ketrez, Nihan & Nevins, Andrew. 2011. The Surfeit of the Stimulus: Analytic biases filter lexical statistics in Turkish laryngeal alternations. Language 87(1): 84-125. Hayes, Bruce. 2008. Introductory Phonology. Malden, MA: Wiley-Blackwell. Hayes, Bruce & Wilson, Colin. 2008. A maximum entropy model of phonotactics and phonotactic learning. Linguistic Inquiry 39: 379-440. Exceptional nasal-stop inventories* Carlos-Eduardo Piñeros University of Auckland [email protected] Received: February 25, 2016 Accepted: June 22, 2016 Abstract This article explores the topic of exceptionality in phonology focusing on nasal-stop inventories. A meticulous survey shows that it is normal for such systems to include at least two anterior units: /m/ and /n/. The finding that the introduction of /ŋ/, the first posterior unit to appear, normally follows that of both anterior units suggests that anteriority is somehow more compatible with this consonant class; however, this hypothesis is challenged by exceptions: /n/ is occasionally superseded by /ŋ/. The proposed analysis overturns the view that there is a single universal place hierarchy. It demon- strates that languages assess the cost of place features on multiple dimensions and that nasal-stop inventories are shaped by the conflict between three evaluation measures: one for articulatory cost, one for perceptual cost, and one for dispersion. A theory of far greater explanatory power emerges when each evaluation measure is empirically substantiated and their universality is strictly respected. Keywords: sound inventories; nasal stops; place features; exceptions; place hierarchies; universal rankings; articulatory cost; perceptual cost; dispersion Resum. Inventaris excepcionals d’oclusives nasals Aquest article explora el concepte d’excepcionalitat en fonologia en relació amb els inventaris de les oclusives nasals. Un repàs meticulós d’aquests inventaris demostra que és normal de trobar-hi com a mínim les dues nasals anteriors /m/ i /n/. El fet que la introducció de la nasal velar /ŋ/, la primera uni- tat posterior a aparèixer, vagi precedida normalment per la introducció de les dues unitats anteriors suggereix que l’anterioritat és d’alguna manera més compatible que la posterioritat pel que fa a la classe de les consonants nasals. Tot i això, aquesta hipòtesi es veu compromesa per algunes excep- cions: /n/ és substituïda de manera ocasional per /ŋ/. L’anàlisi proposada en aquest article invalida la idea que hi ha una sola jerarquia pel que fa al punt d’articulació. S’hi demostra que les llengües avaluen el cost pel que fa als trets de punt d’articulació en diverses dimensions i que la forma dels inventaris de les consonants nasals ve determinada pel conflicte entre tres mesures d’avaluació: cost articulatori, cost perceptual i dispersió. S’assoleix una teoria amb molt més poder explicatiu quan cada mesura d’avaluació troba suport empíric i la seva universalitat es respecta estrictament. Paraules clau: inventari de sons; oclusives nasals; trets de punt d’articulació; excepcions; jerar- quia de punts d’articulació; rànquings universals; cost articulatori; cost perceptiu; dispersió * This article benefited from the comments of two anonymous reviewers, to whom I would like to express my sincere gratitude. Funding for my research on nasal-stop inventories was provided by a University of Auckland Faculty Research Development Fund. This was crucial for the development of the database on which this article is based. I also want to acknowledge the help provided by my student Melissa Irvine, who was the Research Assistant for this project. 68 CatJL 15, 2016 Carlos-Eduardo Piñeros Table of Contents 1. Introduction 5. Assessing the quality 2. Questionable explanations of phonological contrasts 3. An alternative based 6. Integration on functional hierarchies 7. Conclusion 4. A three-way conflict References “Isn’t it possible that there is something that we might call normal in the sense that if we look at the thousands of languages that now exist or have been recorded from the past, the overwhelming majority of them work in a certain way? (Ferguson 1974: 5) 1. Introduction It makes sense to begin a study on exceptions by establishing what the norm is. As concerns nasality, there are two main tendencies in its phonological use. Languages tend to have distinctive nasality in consonants, rather than in vowels, and in ante- rior, rather than in posterior articulations. These are the properties underpinning the notion of ‘normal nasality’ defended by Ferguson (1974, 1975).1 The proclivity of nasality to appear in anterior consonants is manifested by the crosslinguistic prevalence of /m/ and /n/, the labial and coronal nasal stops. Ferguson’s claim that languages normally have at least this pair of nasal phonemes is substantiated by the Auckland Nasal-Stop Inventory Database (ANSID).2 The results in Table 1 provide a sense of the extent to which this generalization is true. Here we see that 419 of the languages in the sample do indeed employ both /m/ and /n/. Since this figure amounts to 92.70%, it is sensible to regard this pattern as a language universal; however, it would not be absolute as in (1), but probabilistic as in (2). (1) All languages have at least two nasal-stop phonemes, both of which are ante- rior. 1. It is assumed that consonants articulated in front of the protuberance of the alveolar ridge are anterior, while those articulated behind it are posterior. 2. ANSID is essentially a revision of the nasal-stop systems catalogued in UPSID-451 (Maddieson and Precoda 1990). The discovery of numerous coding and analytical errors in the existing pho- nological segment inventory databases led to the conclusion that an accurate typology of nasal- stop inventories cannot be developed based on them. The nature of the problems is such that the recommendation not to use such tools for typological phonological research has repeatedly been made in the literature (Basbøll 1985, Pagliuca and Perkins 1986, Simpson 1999, Vaux 2009). The decision to emulate UPSID was motivated by its effort to ensure a reasonable level of genetic diversity: only one language from each small family grouping (Maddieson 1984: 5). Exact replica- tion of the sample was not possible, however, because there were a few languages for which reliable primary sources could not be obtained, in which case a related language was used as substitute. The substitutions made were Afitti for Nyimang, Gbaya Kara for Gbaya Bossangoa, Ipai for Kumeyaay (Diegueño), !Xoon for Jul’hoan (!Xu), and Urdu for Hindi-Urdu. The only other alteration was the addition of Palauan, a language exemplifying a pattern that would have been missed otherwise. This raised the total to 452 languages. Exceptional nasal-stop inventories CatJL 15, 2016 69 Table 1. Quantification of the norm in ANSID Language type Number Percentage a. with both /m/ and /n/ 419 92.70% b. without both /m/ and /n/ 33 7.30% TOTAL 452 100.00% Table 2. Quantification of the exceptions in ANSID Language type Number Percentage a. without phonemic nasal stops 26 5.75% b. with /n/ only 4 0.89% c. with /m/ only 2 0.44% d. with /m ŋ/ only 1 0.22% TOTAL 33 7.30% (2) Most languages have at least two nasal-stop phonemes, both of which are anterior. The present study focuses on the exceptions to normal nasality; that is, the 33 ANSID languages with fewer than two anterior nasal stops. As seen in Table 2, only four types of exception were found. In first place comes the absence of nasal stop phonemes, which evinces that this consonant class falls short of being universal. The next two types reveal that languages may also opt for a single nasal-stop pho- neme: either /n/ or /m/. The final type is the most unexpected. Dorsal /ŋ/, the most common of the posterior units, manages to oust /n/ in one language. Were it not for this particular pattern, the preference for anterior articulations would be absolute. There are several challenges that these findings pose to Phonological Theory. To begin with, the theory must provide a mechanism whereby phonologies can opt out of nasality, so that those that function without it can be accommodat- ed. Secondly, it must ensure that the selection of nasal-stop phonemes prioritizes anterior units over their posterior congeners. This is imperative to account for two facts: the absence of nasal solos consisting of a posterior unit (e.g. */ŋ/, */ɲ/, */ɳ/, etc.) and the absence of nasal duos consisting of two posterior units (e.g. */ŋ ɲ/, */ŋ ɳ/, */ɲ ɳ/, etc.). Thirdly, while both anterior units must be granted special sta- tus, there must also be a way to assess their properties separately, so that it can be explained why some languages compose their nasal solos with /n/, while others do so with /m/. Lastly, the theory must find a principled way to justify the fact that, occasionally, an anterior unit may be excluded for the sake of a posterior one. This is essential to account for the duo /m ŋ/ without invalidating the hypothesis which facilitates the analysis of all other languages; that is, that anterior nasal-stops make better phonemes than their posterior congeners. One gathers from the above that a satisfactory account of normal and excep- tional nasality must strike a balance between restrictiveness and flexibility. It needs 70 CatJL 15, 2016 Carlos-Eduardo Piñeros to be highly restrictive to prevent the creation of random nasal-stop inventories, yet flexible enough to allow a modicum of variation. 2. Questionable explanations The challenges identified above are hardly trivial. This section reviews three lines of analysis in order to demonstrate that the dichotomy between norms and excep- tions escapes historical, structural, and hierarchical explanations. 2.1. Historical approach One way to explain why nasal-stop inventories do not come in an endless variety of forms is to refer to previous evolutionary stages. The argument is that normal and exceptional patterns follow from diachrony. To provide a concrete example, one could argue that it is for historical reasons that the nasal-stop inventory of Onge― one of the languages spoken in the Adaman Islands―is /m n ɲ ŋ/. Diachronic analyses have concluded that this nasal quartet was already present in Proto-Ongan, which passed it on to its descendants: Onge and Jarawa (Blevins 2007). From this optic, it would seem that ancestry is the key to understanding why nasal-stop inven- tories are the way they are. Onge has two anterior and two posterior units because it inherited them from its ancestor. Unfortunately, this type of explanation does not get very far. It soon runs into trouble because there are numerous languages which have modified the nasal-stop inventories of their ancestors. It is well known that language evolution may involve the introduction of phonemes that were missing from the ancestor as well as the elimination of phonemes that the ancestor did have. An example of the former case is Spanish, one of the Romance languages which developed a palatal nasal-stop phoneme that Latin did not know. Thus, while the modern language has the trio /m n ɲ/, the classical language had the duo /m n/ (Penny 2002). A language whose evolution involved reduction of the nasal-stop inventory is Palauan. Foley (1975) demonstrates that a sound shift turned Proto-Austronesian /n/ into Palauan /l/, as a result of which the trio /m n ŋ/ of the parent language became the duo /m ŋ/ in the descendant. The latter example is of special significance. It tells us that, because language evolution may involve the loss of anterior nasal stops, ancestry cannot be the reason why languages are prone to having them. Another obstacle to upholding explanations based on ancestry is that they keep sending us back to previous grammatical stages without ever addressing the mat- ter. They tells us that language A has such and such sound inventory because it inherited it from language B, which inherited it in turn from language C, and so on; however, it is evident that this only serves to evade the question. At some point in this constant referral to the past, the analyst seeking to gain a deeper understanding of the facts realizes that he must face the inevitable question: how did languages come to have the nasal-stop inventories that they have? Evolutionary Phonology (Blevins 2004) resolves some of these objections by recognizing that diachrony may lead to maintenance, enrichment, or impov- Exceptional nasal-stop inventories CatJL 15, 2016 71 erishment of sound inventories. The latter two are logical consequences of sound change. But the explanation that this theory offers for the consolidation of some patterns as norms and of others as exceptions remains unsatisfactory. It is claimed that sound patterns may be frequent or rare because the sound changes which create them are frequent or rare themselves (Blevins 2004: 83). The problem with this view is that, because the ultimate cause remains unidentified, the mind is led to recast the question. Why are certain sound changes frequent whereas others are rare? A more explanatory answer is possible by drawing on phonetics. Sound changes which are phonetically motivated are natural, hence frequent, whereas those stem- ming from analogy, rule inversion, rule telescoping, or accidental convergence are unnatural, hence rare. The need to distinguish between natural and unnatural sound changes is recognized by most phonological theories, but Evolutionary Phonology sets itself apart by assuming that, even when they are natural, sound changes are non-teleological. That is to say that, rather than leading to the achievement of a goal, they are the outcome of random phonetic drift. Little insight is gained when normal nasality is seen through this lens. Evolutionary Phonology would say that this phenomenon exists because there are natural sound changes which give rise to /m/ and /n/ for no practical purpose. It is the final part of this explanation that is questionable. If the development of anterior nasal stops afforded no practical benefits, why would most languages evolve in that direction? It does not stand to reason that a regularity that holds for over 90% of world’s languages is haphazard. 2.2. Structural approach Positing structural differentials is one of the tactics that one can use to tackle the problem synchronically. The pillar of this idea is the assumption that there are vari- ous degrees of complexity among nasal stops. This opens the door for the argument that the predilection for anterior units is driven by economy: simpler structures yield savings which enable grammars to operate at a lower cost. Figure 1 illustrates an analysis along these lines proposed by Rice and Avery (1993: 143). It is hypothesized that coronal is the unmarked place of articulation and, since /n/ is a member of this category, it is held that its phonological repre- sentation includes no structure under the Place node. Other places of articulation are assumed to be more marked, from which it follows that their phonological representations should be more elaborate. Observe that the structure of /m/ includes one level under the Place node, while that of /ŋ/ includes two. This analysis yields some positive results. First of all, the economy ensuing from the simple structure attributed to /n/ ensures that this unit will be the top choice for nasal solos. Secondly, since /m/ has only one additional level, it will be able to secure one of the positions available in nasal duos. Lastly, the first posterior unit to make the cut will be /ŋ/, which is expected to seize one of the positions available in nasal trios because it represents the next increase in structural com- plexity. The patterns /n/, /m n/, and /m n ŋ/, are thus accounted for. 72 CatJL 15, 2016 Carlos-Eduardo Piñeros /n/ /m/ /ŋ/ Place Place Place Peripheral Peripheral Dorsal Figure 1. Increasing degrees of structural complexity among nasal stops. But there are some negative results as well. The solo /m/ and the duo /m ŋ/ contravene this analysis because they favor the selection of an alleged complex unit over a simpler one (i.e. /m/ over /n/ in one case, /ŋ/ over /n/ in the other). Adjusting the representation of place categories according to the choices made by individual languages would remove these obstacles; however, that strategy does not translate into real progress because it merely trades one problem for another: the theory would go from undergenerating to overgenerating. Consider that, while it is possible to concoct representations conducive to the desired results (e.g. fewer structural levels for /m/ than for /n/, more structural levels for /n/ than for /ŋ/, and so on), these would empower grammars to generate all sorts of nasal-stop inventories, when the reality is that the exceptions to normal nasality are not so diverse (Table 2). Doubts about the representations in Figure 1 are also raised by the fact that the articulators of /n/, /m/, and /ŋ/ perform essentially the same task: they occlude the oral cavity. This parallelism suggests that, rather than the amount of structure, it is the presence of a unique trait in each that makes these segments different. To be more precise, /n/ would not have the quality it has, if coronality were not part of its composition and, similarly, /m/ and /ŋ/ would not be the entities they are without the contribution of labiality to the former and of dorsality to the latter. A more sensible way to represent such distinctions is, therefore, to posit a privative feature for each place of articulation (Figure 2). When one admits that the distinctions among /n/, /m/, and /ŋ/ hinge on the com- mutation of one place feature for another, the explanation based on structural econo- my collapses. These phonemes have exactly the same degree of structural complexity /n/ /m/ /ŋ/ [coronal] [labial] [dorsal] Figure 2. Equal structural complexity among nasal stops. Exceptional nasal-stop inventories CatJL 15, 2016 73 and, therefore, it is unfounded to claim that simplicity is the source of the crosslin- guistic preference for anterior nasal stops. In no way is /n/ structurally poorer than /m/ nor is /ŋ/ structurally richer than either one of its anterior congeners. 2.3. Hierarchical approach Another scheme one can pursue is to order place features hierarchically so that they can be selected according to their rank. This is the foundation of the construct known as the Place Hierarchy. Optimality Theory offers the machinery needed to implement this vision (Prince and Smolensky 1993/2004). A family of structural constraints banning place features―say *[place]― is pitted against a family of faithfulness constraints requiring the preservation of such entities―assume it to be Ident(place). The outcome of this clash will depend on how individual grammars rank these principles with respect to one another. Ranking Ident(place) over *[place] will ensure that the segment bearing the feature in question is grammatically affordable but, because each grammar is free to decide which specific member of *[place] is to play that role, the results will vary. The most interesting part of this analysis is how the crosslinguistic preference for certain places of articulation is accounted for. It is assumed that the ranks among the members of *[place] are universal. To appreciate the merits of this approach, consider the ranking in (3). This version of the Place Hierarchy states that the features [labial] and [coronal] are equally costly, but less so than [dorsal].3 (3) Place Hierarchy (Prince and Smolensky 1993/2004) *[dorsal] >> *[labial], *[coronal] It is easy to foresee that the asymmetry that exists between anterior and poste- rior nasal stops is amenable to this interpretation. In a grammar where the ranking of Ident(place) with respect to the Place Hierarchy is set as in (4a), both anterior units will be able to emerge without their posterior congeners. This positive result is strengthened by the fact that the preference for anteriority persists, even if one makes an express effort to favor posteriority. Note, for instance, that /ŋ/ will emerge in a grammar where Ident(place) rises to the position it occupies in (4b); yet, it will not be alone, but in the company of /m/ and /n/, whose co-selection is guaranteed by the fact that the constraints against their place features hold the lower ranks of the Place Hierarchy. (4) a. *[dorsal] >> Ident(place) >> *[labial], *[coronal] b. Ident(place) >> *[dorsal] >> *[labial], *[coronal] 3. The study of place-of-articulation phenomena has led to the proposal of multiple versions of the Place Hierarchy. The version in (3) has been chosen to illustrate this line of analysis because it is the most successful in accounting for nasal-stop inventories. 74 CatJL 15, 2016 Carlos-Eduardo Piñeros Nasal solos respond well to this treatment too. Since the Place Hierarchy under consideration leaves the ranking between *[labial] and *[coronal] unspecified, Ident(place) will be able to subdue each one of these constraints separately. Some grammars may therefore adopt the order in (5a), while others may opt for (5b). In either case, the only affordable nasal stop will be anterior. (5) a. *[dorsal] >> *[labial] >> Ident(place) >> *[coronal] b. *[dorsal] >> *[coronal] >> Ident(place) >> *[labial] Despite such positive results, the hierarchical approach is not without problems; it too suffers from undergeneration. The duo /m ŋ/ is its nemesis. On the assump- tion that *[dorsal] dominates *[coronal], this inventory is incomprehensible. Why should a language drop /n/ to make room for /ŋ/ if the latter segment bears a place feature which is universally more costly? One way to circumvent this impasse would be to grant languages the freedom to modify the Place Hierarchy at their discretion. One could assume that most lan- guages accept the ranking *[dorsal] >> *[coronal], while a few opt to reverse it. The problem with this solution is that any modification to the Place Hierarchy―even if it is for just one language―invalidates its universal status and, if this premise is given up, then the explanation for the proclivity towards anteriority vanishes too. Nothing would prevent posterior units from crowding nasal-stop inventories. It is worth pointing out that, in order to cope with exceptions, the hierarchical and the structural accounts are pressured to abandon their ground assumptions. One gives in by reversing dominance relations among the members of the Place Hierarchy, while the other one does so by adjusting the representation of place categories. Such repairs are to no avail. They merely turn the inability to generate a few patterns (undergeneration) into the ability to generate any conceivable pattern (overgeneration). To stay away from ranking reversals, the hierarchical approach could capital- ize on the mechanism that Optimality Theory uses to model conflict: constraint interaction. This would require the postulation of an additional principle against the feature [coronal]. Let us call it Anti-Cor. If dominant, the new player could tone down the coercive power of the faithfulness constraint. This happens in (6), a grammar where /n/ cannot emerge alongside /m/ and /ŋ/, despite the fact that the rank of Ident(place) is high enough to justify the cost of three place features. It goes without saying that this solution is ad hoc. Anti-Cor―an obvious duplicate of *[coronal]―is brought in expressly to block an otherwise felicitous unit. (6) Anti-Cor >> Ident(place) >> *[dorsal] >> *[labial], *[coronal] While blocking by a higher-ranking constraint safeguards the universality of the Place Hierarchy, it does not avoid overgeneration. This problem is palpable in (7), where the strategy used in (6) is reapplied to generate the solo */ŋ/. Observe that, when an additional constraint against [labial]―call it Anti-Lab―is ranked above Ident(place), both anterior units are ruled out while the posterior unit manages to Exceptional nasal-stop inventories CatJL 15, 2016 75 go through. This effectively reverses the dominance relations encoded in the Place Hierarchy; hence, it is no different from assuming that [dorsal] is less costly than both [labial] and [coronal] or, more straightforwardly, that there is no universal Place Hierarchy. (7) Anti-Lab, Anti-Cor >> Ident(place) >> *[dorsal] >> *[labial], *[coronal] The ineluctable conclusion is that the hierarchical approach does not measure up to the task either. This is unsurprising considering that the Place Hierarchy is simply a set of stipulations. To this date, no justification has been provided for the alleged parity between the anterior place categories or for the advantage they presumably have over their posterior congeners. Are there any actual properties behind such an appraisal? As long as this question remains unanswered, there will be no hope for the hierarchical approach. 3. An alternative based on functional hierarchies The solution advanced here is both functional and hierarchical. I draw on articula- tory and perceptual functions to validate the hypothesis that there are universal dominance relations among place features. This view is founded on the premise that different linguistic structures strain the human linguistic capacity to different extents. Accordingly, place features which are articulatorily taxing will rank high on a hierarchy guided by articulatory factors. A parallel thesis is defended for the perceptual dimension. Place features which are perceptually taxing will rank high on a hierarchy guided by perceptual factors. Also central to the proposal is the assumption that the difficulty in articulating or perceiving place features depends on the type of segment within which they appear. That is to say that the assessment of their cost is not absolute, but relative to each segmental class. The split between articulation and perception and the diversity of segmental classes join forces to overturn the view that there is a single universal place hierarchy. Phonologies must refer to multiple place hierarchies because, due to their different nature, articulatory and perceptual demands need to be assessed separately, and also because place features occur in different segmental classes, some of which are more compatible with certain places of articulation than with others. Importantly, the recognition of this plurality does not preclude the postulation of universal rankings. The condition that a hierarchy must meet in order to qualify as universal is to hold true for all languages where the relevant variables are present. For instance, if an articulatory place hierarchy appraises that feature X is more difficult to articulate than feature Y in segmental class Z, then that must be the case for all languages where X and Y are present in Z. This condition notwithstanding, it is still possible that X be less difficult to perceive than Y in the same segmental class or that it be less difficult to articulate than Y in a different segmental class. In this section I develop two universal place hierarchies which are crucial for the selection of nasal-stop phonemes. The first one is the Articulatory Consonant Place Hierarchy, which assesses the articulatory cost of place features within con- 76 CatJL 15, 2016 Carlos-Eduardo Piñeros sonantal segments. The second one is the Perceptual Nasal-Stop Place Hierarchy, which assesses the perceptual cost of place features within nasal stops. 3.1. The articulatory cost of place features in consonantal segments It is defended here that certain sound structures are articulatorily more demanding and, therefore, more costly to the phonology than others. In assessing the articu- latory cost of place features, it is essential to consider whether they are part of a consonant or a vowel. When included within a consonant, features such as [labial], [coronal], and [dorsal] are tied to the ability of the articulators to travel at high speed. This is because consonants are produced by constricting the vocal tract at a local point, an event which must be rapidly executed given that the duration of consonants is significantly shorter than that of vowels (Cooper et al. 1952, Klatt 1976, Borzone de Manrique and Signorini 1983). Such requirements suggest that reaching the constriction location (i.e. the place of articulation) is less taxing for articulators excelling at fast movement. Articulator velocity thus emerges as one of the criteria that can be used to guide the assessment of articulatory cost. On the basis of this substantive factor, I propose the constraints in (8) and the hierarchy in (9). (8) A family of articulatory structural constraints *Dor-in-C = Do not use the dorsum to articulate a consonant. *[dorsal] / [ ___, consonantal] *Lab-in-C = Do not use the lips to articulate a consonant. *[labial] / [ ___, consonantal] *Cor-in-C = Do not use the corona to articulate a consonant. *[coronal] / [ ___, consonantal] (9) Articulatory Consonant Place Hierarchy (fragment)4 *Dor-in-C *Lab-in-C *Cor-in-C 4. Languages may, of course, distinguish more than three places of articulation, which means that the Articulatory Consonant Place Hierarchy must be more elaborate than in (9). In this article I focus on the fragment that is needed to explain the preference/dispreference for anterior/posterior articulations. Exceptional nasal-stop inventories CatJL 15, 2016 77 The findings of several experimental studies support the claim that, within a consonantal segment, [dorsal] is articulatorily more costly than [labial], which is in turn articulatorily more costly than [coronal]. Hudgins and Stetson (1937: 92) conducted a series of diadochokinetic experiments with a view to determining the maximum speed of certain speech movements. The nine speakers who participated in this study were asked to repeat rhythmic groups of repeated syllables as rapidly as they could. The results indicate that the corona is the fastest speech organ. It yielded a maximum-rate average of 8.2 syllables per second, while the corresponding figures for the dorsum and the lips were 7.1 and 6.7, respectively. The decision to concede the lowest rank of the Articulatory Consonant Place Hierarchy to *Cor-in-C is thereby justified. It captures the fact that drawing on the corona for the purpose of forming a consonantal constriction is most efficient; hence, least costly. A subsequent experimental study employed lateral-view cineradiography to measure the velocity of the corona, the lower lip, and the dorsum relative to the maxillary and the mandibular systems (Kuehn and Moll 1976). The results for the maxillary system show that the corona moved the fastest (196mm/s), followed by the lower lip (166mm/s), which was in turn faster than the dorsum (129mm/s). The data for the mandibular system also show that the corona was the fastest articu- lator (142mm/s), but neither the lower lip (102mm/s) nor the dorsum (102mm/s) was faster than the other in this case. These findings corroborate the decision to place *Cor-in-C below the other members of its family and hint that *Lab-in-C should be ranked below *Dor-in-C. Roon et al. (2007) examined the same set of articulators using a different tech- nique. They employed electromagnetic articulography to calculate stiffness, a meas- urement of the velocity that an articulator can reach independently of the distance it travels (Byrd 1992, Roon et al. 2007). Given that an organ of high stiffness returns to its equilibrium position faster than one of lower stiffness, it can be inferred that the former has a superior ability for fast movement. Two speakers participated in this experiment. For one of them, the stiffness of the corona (mean=3.15, SD=2.00) was significantly higher than that of the lower lip (mean=2.61, SD=0.27), whose stiffness was in turn significantly higher than that of the dorsum (mean=1.85, SD=0.31). The same order followed from the results for the second speaker (coro- na: mean=2.21, SD=2.48, lower lip: mean=2.06, SD=0.39, dorsum: mean=1.44, SD= 0.47), although it should be noted that, in this case, the advantage of the corona over the lower lip did not prove to be statistically significant. These findings back up the decision to rank *Lab-in-C below *Dor-in-C. To synthesize, there is recurrent evidence that the corona is the articulator of highest velocity, which clinches the argument that, when it comes to producing consonants, no other articulator is more efficient. Somewhat less forceful is the evidence available for the difference between the lips and the dorsum. While the earliest study placed the dorsum ahead of the lips (Hudgins and Stetson 1937), the more recent studies, which were able to use more advanced techniques, indicate that the lips outperform the dorsum (Kuehn and Moll 1976, Roon et al. 2007). The development of new experimental techniques will most certainly afford 78 CatJL 15, 2016 Carlos-Eduardo Piñeros more accurate measurements in the future; however, the evidence that is currently available is consistent enough to assert that the corona is the fastest articulator, the lips come next, and the dorsum falls behind them. The ranking *Dor-in-C >> *Lab-in-C >> *Cor-in-C is thereby substantiated. It is important to point out that, although the same dominance relations were previously posited by some versions of the Place Hierarchy (e.g. de Lacy 2002, 2006), the present proposal breaks new ground in two ways. Firstly, the adoption of articulator velocity as the criterion that guides the ranking has the benefit of avoiding stipulation. The hierarchical organization of place features now has a justification. It stems from substantive factors; that is, from differences in the ability of the articulators to perform their functions. Secondly, the Articulatory Consonant Place Hierarchy acknowledges that the established order of prec- edence does not apply across the board. It is specific to consonantal segments and not from every angle, but specifically from an articulatory perspective. The lesson to learn from this is that the search for a place hierarchy that works for all place-of-articulation phenomena is futile. The fact that the cost of place features is conditioned by segmental classes as well as by articulatory and perceptual factors means that they can be organized in multiple hierarchies, which are none- theless universal. 3.2. The perceptual cost of place features in nasal stops A lack of symmetry among sound structures is also true of the perceptual dimen- sion. Depending on the segment within which they appear, certain place features are more difficult to perceive and, consequently, more costly to the phonology than others. I argue next that, in the context of a nasal stop, the cost of perceiving place features increases proportionally to the distance that separates the constric- tion location from the lips. That is to say that the more posterior the articulation, the more challenging the perception of the nasal consonant. The application of this generalization to /m/, /n/, and /ŋ/ gives rise to the constraints in (10) and to the hierarchy in (11), where [coronal] has been bumped out of the bottom rank by [labial]. This new assessment does not change the status of [dorsal], which remains less affordable. (10) A family of perceptual structural constraints *Dor-in-N = Do not listen for cues to [dorsal] in a nasal stop. *[dorsal] / [ ___, nasal, occlusive] *Cor-in-N = Do not listen for cues to [coronal] in a nasal stop. *[dorsal] / [ ___, nasal, occlusive] *Lab-in-N = Do not listen for cues to [labial] in a nasal stop. *[labial] / [ ___, nasal, occlusive] Exceptional nasal-stop inventories CatJL 15, 2016 79 (11) Perceptual Nasal-Stop Place Hierarchy (fragment)5 *Dor-in-N *Cor-in-N *Lab-in-N The evidence for the ranking in (11) is to be found in the nasal murmur―the sound radiated through the nose while the oral cavity is occluded.6 Acoustic stud- ies have identified several spectral traits characteristic of nasal murmurs (House 1957, Delattre 1958, Fant 1960, Fujimura 1962a, b, Recasens 1983, 1992, among many others). The most conspicuous one is a series of increasingly higher nasal formants (N1, N2, N3, N4, N5, …), among which the first one is clearly domi- nant given its much higher intensity. It will be important to keep in mind that the intensity of nasal formants decreases as their frequency rises; for instance, N5 is less intense than N4, which is less intense than N3, and so on. Of special inter- est here is the concomitant presence of a nasal antiformant or zero (NZ), whose frequency correlates with place of articulation (House 1957, Fujimura 1962a, b, Recasens 1983).7 To appreciate how the place of articulation of a nasal stop relates to its anti- formant, consider the diagrams in Figure 3, where the most anterior unit appears 5. From the existence of additional place features, it follows that this is not the full hierarchy. Only the fragment which is relevant to the account of normal and exceptional nasality is presented here. 6. Despite the fact that vowel transitions also carry place cues for nasal stops, the decision not to base the Perceptual Nasal-Stop Place Hierarchy on them was made for the following reason. Studies such as Cooper et al. (1952), Liberman et al. (1954), Malécot (1956), and Delattre (1958) have revealed that the F2 and F3 transition shapes which cue place of articulation in nasal stops are similar to those that serve the same purpose in their oral counterparts. That is to say that there is acoustic parallelism among the members of the sets /m b p/, /n d t/, and /ŋ ɡ k/. It is possible to hypothesize from this that, if vowel transitions were responsible for the difficulty in perceiving posterior articulations, /ŋ/, /ɡ/, and /k/ would pattern together, or, to put it another way, the pres- ence of a gap in the nasal-stop series would be mirrored by the presence of a similar gap in the oral-stop series. The reality is, however, that there is a tendency for languages that have /m/ and /n/ to be deprived of /ŋ/, whereas languages that have /b/ and /p/, on the one hand, and /d/ and /t/, on the other hand, normally possess /ɡ/ and /k/ as well (Maddieson 1984, Recasens 1992, Clements 2009). The fact that it is specifically the nasal-stop series which tends to lack a posterior member indicates that the factor responsible for the difficulty in perceiving posterior articulations resides in a property specific to nasal stops: the nasal murmur. The centrality granted to the place cues carried by the nasal murmur is what sets this proposal apart from that by Narayan (2006). This author attributes the difficulty in perceiving /ŋ/ to the shapes of the F2 and F3 vowel transitions. The problem with this approach is that it fails to establish a differential between /m/ and /n/. 7. Simply put, a nasal formant represents frequencies amplified by the pharyngeal and nasal cavities, whereas a nasal antiformant stands for frequencies attenuated by the occluded oral cavity. 80 CatJL 15, 2016 Carlos-Eduardo Piñeros /m/ > /n/ > /ŋ/ Figure 3. Place of articulation determines oral-cavity size8. to the left, while the most posterior one appears to the right. It is easy to see that the three diagrams are identical except in one respect: the size of the oral cavity. A quick comparison shows that this particular resonator becomes gradually smaller as the constriction retracts. This phenomenon is bound to have acoustic consequences because the frequency at which a resonator vibrates depends largely on its size. Given that a small resonator vibrates faster than a larger one, a natural consequence of reducing the size of the oral cavity is that there is a rise in NZ values. Confirmation that a reduced oral cavity leads to higher NZ frequencies is abun- dant (see House 1957, Delattre 1958, Fant 1960, Fujimura 1962a, b, Dukiewicz 1967, Recasens 1983, Gubrynowicz et al. 1985, and Ladefoged and Maddieson 1996). These studies report that the large oral cavity of /m/ leads to a low NZ (around 800 Hz), whereas the small oral cavity of /ŋ/ results in a high NZ (around 3700 Hz). It has also been verified that occlusions formed at intermediate locations between the velum and the lips give rise to intermediate NZ values. The NZ of /n/, for instance, occurs at about 1780 Hz. The presence of a formant cluster above N2 is another aspect of the nasal spec- trum that bears on place of articulation (Fujimura 1962a, b, Recasens 1992). By plotting their movements through the murmur, Fujimura discovered that nasal for- mants are involved in clustering patterns induced by the antiformant. He observed that, while the formants which are out of the immediate vicinity of the antiformant remain relatively constant, those surrounding it come closer together. In particular, the N2 and N3 of /m/ form a cluster with its low NZ, while the N3 and N4 of /n/ do the same with the significantly higher NZ that characterizes this unit. The fact that the NZ of /ŋ/ is even higher entails that the clustering formants are higher as well: N5 and N6. On the basis of this discovery, Fujimura (1962a: 1871) describes the difference between the murmurs of the three units under discussion as follows: “/ŋ/ has four formants in the main frequency range (up to about 3000 cps), /n/ is obtained when the third of the formants is replaced by a cluster consisting of two formants and one antiformant, /m/ is obtained when the second formant is replaced by a similar cluster.” In summary, the findings of the aforementioned acoustic studies indicate that posterior nasal stops have higher NZ and N-cluster values than their anterior 8. In these diagrams, the resonators of the vocal tract (i.e. the pharyngeal, oral, and nasal cavities) are represented as ducts contoured by straight lines. The horizontal duct that is occluded at the left end is the one representing the oral cavity. Arrows indicate the direction of airflow. Exceptional nasal-stop inventories CatJL 15, 2016 81 congeners. This may give the impression that /ŋ/ is acoustically superior to its competitors, but the reality is that it performs poorly. It turns out that having high frequencies is counterproductive for nasal stops because energy concentrates less intensely in the higher than in the lower regions of the nasal spectrum. As a result, it is the lower frequencies that are advantageous because they are more salient (House 1957, Ohala 1975, Ohala and Ohala 1993). Combined with the uncontroversial assumption that acoustic salience makes speech sounds easier to perceive, this means that, from a perceptual point of view, /m/ is the most affordable member of its class. It can be further inferred that /n/ and /ŋ/ strain the human auditory system to gradually greater extents because their higher NZ and N-cluster are gradually less salient. The ranking *Dor-in-N >> *Cor-in-N >> *Lab-in-N reflects precisely that. It is worth noting that the Articulatory Consonant Place Hierarchy and the Perceptual Nasal-Stop Place Hierarchy invalidate two common assumptions about the major place categories. In the face of the phonetic evidence assembled above, it can no longer be claimed that the features [coronal] and [labial] are on equal standing (contra Prince and Smolensky’s Place Hierarchy) nor that [coro- nal] is invariably less costly than [labial] (contra de Lacy’s Place Hierarchy). In actuality, the cost of place features varies across segmental classes and across speech facets (i.e. articulation vs. perception). The evidence indicates that, for consonantal segments in general, the feature [coronal] is articulatorily less costly than [labial]; but for nasal stops in particular, [labial] is perceptually less cost- ly than [coronal]. 4. A three-way conflict The stage is now ready to present the formal analysis. A cogent explanation for the crosslinguistic preference for anterior nasal-stops will emerge from the resolu- tion of a three-way conflict. The Articulatory Consonant Place Hierarchy is bound to clash with the Perceptual Nasal-Stop Place Hierarchy because, although both oppose the use of place features, they disagree as to which value is the least costly. Additionally, the structure-saving nature of the place hierarchies puts them in direct conflict with Faithfulness, the constraint family demanding structural preservation. I assume that the principle in (12) is the relevant faithfulness constraint. (12) Faith(nasal place): Nasal segments standing in correspondence must be faithful to each other in terms of place features. The size and structure of nasal-stop inventories will depend on the rank of Faith(nasal place) with respect to the place hierarchies. In general, the higher rank- ing the faithfulness constraint, the larger and more diverse the inventory will be. For expository ease, I will begin with the smallest inventory and advance towards the larger ones. 82 CatJL 15, 2016 Carlos-Eduardo Piñeros 4.1. Absence of nasal stops Although the prioritization of anterior nasal stops is a strong crosslinguistic ten- dency, there are a few languages where it is imperceptible. No signs of partiality can be found in languages deprived of phonemic nasal stops because all members of the class are left out. Such systems can be construed as the grammatical state in which both place hierarchies have absolute supremacy over Faith(nasal place). (13) Grammar of languages deprived of nasal stops *Dor-in-C *Dor-in-N *Lab-in-C *Cor-in-N *Cor-in-C *Lab-in-N Faith(nasal place) This arrangement has the effect of setting a cost threshold equal to nil, the upshot of which is that even the least costly units turn out to be unaffordable. Since nothing can be more restricted than an empty set, the extreme ranking in (13) seems appropriate for this language type. I take this to be the most primitive stage in the development of nasal-stop inventories and argue that growth from that point forth is brought about by the ascent of the faithfulness constraint with respect to the structural constraints. Attached to this proposal comes a responsibility. It is imperative that every step in the ascent of Faith(nasal place) be examined in order to verify that the theory does not overgenerate. This is not a small task. Given that the relevant fragment of each place hierarchy consists of three constraints, the number of posi- tions that Faith(nasal place) can occupy with respect to each one of them is four. It follows from this that the total number of rankings that need to be examined is sixteen (4 x 4). A factor that makes it quicker to complete this task is that many of the posi- tions that Faith(nasal place) can occupy with respect to the place hierarchies yield the same result. That is the case with the absence of nasal stops. It turns out that, in addition to (13), there are seven other rankings whereby these constraints yield an empty set. Three of those rankings have been assembled in (14) on the basis of a common property: all members of the Articulatory Consonant Place Hierarchy retain their supremacy over Faith(nasal place). Exceptional nasal-stop inventories CatJL 15, 2016 83 (14) Three additional rankings where the articulatory-cost threshold is nil9 *Dor-in-C >> *Lab-in-C >> *Cor-in-C Faith(nasal place) *Dor-in-N >> *Cor-in-N >> *Lab-in-N *Dor-in-C >> *Lab-in-C >> *Cor-in-C Faith(nasal place) *Dor-in-N >> *Cor-in-N >> *Lab-in-N *Dor-in-C >> *Lab-in-C >> *Cor-in-C Faith(nasal place) *Dor-in-N >> *Cor-in-N >> *Lab-in-N The differences among these rankings boil down to the number of members of the Perceptual Nasal-Stop Place Hierarchy that succumb to the faithfulness constraint. In an initial stage, only the lowest member of this place hierarchy gives in, but in subsequent stages two or three additional members yield as well. The development that is unfolding here is, therefore, a unilateral ascent, which is why there is no success. Given that all members of the Articulatory Consonant Place Hierarchy retain their influence over Faith(nasal place), every unit that clears the perceptual dimension runs into a stone wall in the articula- tory dimension. Three other rankings can be assembled into another group because they also have a common property: all members of the Perceptual Nasal-Stop Place Hierarchy outrank the faithfulness constraint. This becomes apparent in (15), 9. To facilitate their comparison, related rankings are presented in this alternative format. Dominance relations among members of the place hierarchies are indicated through the symbol >> and the rank of Faith(nasal place) with respect to both place hierarchies is signalled by the dashed arrows. A solid line separates different rankings. 84 CatJL 15, 2016 Carlos-Eduardo Piñeros where the structural constraints representing the perceptual dimension are now located on the top row, while those representing the articulatory dimension occu- py the bottom row. (15) Three additional rankings where the perceptual-cost threshold is nil *Dor-in-N >> *Cor-in-N >> *Lab-in-N Faith(nasal place) *Dor-in-C >> *Lab-in-C >> *Cor-in-C *Dor-in-N >> *Cor-in-N >> *Lab-in-N Faith(nasal place) *Dor-in-C >> *Lab-in-C >> *Cor-in-C *Dor-in-N >> *Cor-in-N >> *Lab-in-N Faith(nasal place) *Dor-in-C >> *Lab-in-C >> *Cor-in-C The number of members of the Articulatory Consonant Place Hierarchy over- taken by Faith(nasal place) is what makes the difference in this case. In an ini- tial stage, only the lowest member of that hierarchy falls below the faithfulness constraint, but in subsequent stages additional members lose their ground as well. Clearly, a unilateral ascent is unfolding here too but, unlike that in (14), it is at the expense of articulation. Because the perceptual dimension remains fully dominant, no unit can succeed. The general point that the rankings in (14) and (15) demonstrate is that, as long as one of the place hierarchies retains full power, the advances made by the faithfulness constraint are inconsequential. This reflects the autonomy of the grammatical-cost dimensions. Each can reject any place category independently of the other. The last of the rankings yielding an empty set is that in (16). It corresponds to the minimal ascent of Faith(nasal place) with respect to its two rivals. What makes Exceptional nasal-stop inventories CatJL 15, 2016 85 this move unfruitful is that the place hierarchies cancel each other out because they disagree as to which place category is the least costly. Thus, although the ranking Faith(nasal place) >> *Cor-in-C favors the selection of /n/, this unit can- not be recruited because *Cor-in-N disallows it. Similarly, although the ranking Faith(nasal place) >> *Lab-in-N is conducive to the selection of /m/, this unit remains unaffordable because *Lab-in-C precludes it. Such interactions corroborate that this is a three-way conflict: two families of structural constraints compete with each other, despite their alliance against Faithfulness. (16) Minimal ascent of the faithfulness constraint with respect to both place hier- archies *Dor-in-C *Dor-in-N *Lab-in-C *Cor-in-N Faith(nasal place) *Cor-in-C *Lab-in-N In conclusion, we now have a sensible understanding of why nasal stops are not universal phonemes. Some grammars deem that their articulatory cost, their perceptual cost, or both are too steep and choose not to invest in them. 4.2. Nasal solos After conquering the lowest member of each place hierarchy, the next step in the ascent of Faith(nasal place) is to overtake a second principle in either the articu- latory or the perceptual dimension. The first of these scenarios is illustrated in (17), where Faith(nasal place) has risen above the articulatory and the perceptual constraints against labial: *Lab-in-C and *Lab-in-N. A concomitant effect is that, since the place hierarchies are universal, *Cor-in-C is also surpassed. Given that this brings the articulatory and the perceptual cost of one place category below the critical point, the grammar is now able to make a minimal investment in nasal stops. That is how the solo /m/ is generated. The configuration in (17) makes it clear that this grammar is characterized by a state of imbalance between articulation and perception. Because a greater number of articulatory than perceptual constraints have lost their ground to Faith(nasal place), it is appropriate to say that this system sacrifices articulatory efficiency for the sake of perceptual ease. 86 CatJL 15, 2016 Carlos-Eduardo Piñeros (17) Grammar of languages with the solo /m/ *Dor-in-N *Dor-in-C *Cor-in-N Faith(nasal place) *Lab-in-C *Lab-in-N *Cor-in-C There is another arrangement which produces the same result. The configura- tion in (18) helps to see that, although Faith(nasal place) can make further progress by overtaking an additional member of the Articulatory Consonant Place Hierarchy, this is unavailing when no parallel progress is made on the perceptual dimension. (18) Another ranking leading to the solo /m/ *Dor-in-N >> *Cor-in-N >> *Lab-in-N Faith(nasal place) *Dor-in-C >> *Lab-in-C >> *Cor-in-C The other logical way to begin populating a nasal-stop inventory is to shift the pressure to the perceptual dimension. This development is portrayed in (19). The cost of [coronal] becomes fully affordable when Faith(nasal place) surpasses *Cor-in-C and *Cor-in-N. It should be noted here again that, because the place hierarchies are universal, this entails that *Lab-in-N is also surpassed. An interest- ing interpretation emerges from this arrangement. The solo /n/ is the product of prioritizing articulation over perception in a system which is willing to make only the minimal investment in nasal stops. This establishes a link between the attested nasal solos. The fact that (19) is the mirror image of (17) demonstrates that both stem from skewed grammars. As expected, further advances of the faithfulness constraint in the perceptual dimension are unrewarding when no parallel progress is made in the articula- tory dimension. That happens again in (20), a failed attempt to secure an additional unit. We have thus seen that four of the sixteen possible rankings lead to nasal solos. Exceptional nasal-stop inventories CatJL 15, 2016 87 (19) Grammar of languages with the solo /n/ *Dor-in-C *Lab-in-C *Dor-in-N Faith(nasal place) *Cor-in-C *Cor-in-N *Lab-in-N (20) Another ranking leading to the solo /n/ *Dor-in-C >> *Lab-in-C >> *Cor-in-C Faith(nasal place) *Dor-in-N >> *Cor-in-N >> *Lab-in-N A remarkable achievement of this constraint scheme is that it disallows the generation of nasal solos other than /m/ and /n/. This is proof that the crosslinguistic preference for anterior nasal stops has been successfully modeled. The position of *Cor-in-C at the bottom of the Articulatory Consonant Place Hierarchy entails that /n/ is the most affordable unit from an articulatory point of view and, similarly, the position of *Lab-in-N at the bottom of the Perceptual Nasal-Stop Place Hierarchy guarantees that /m/ will be the most affordable unit from a perceptual point of view. Given that all posterior units are articulatorily and perceptually more costly, selecting an anterior unit is the most sensible way to begin the construction of any nasal-stop inventory. This explains why /m/ and /n/ are quasi-universal. It is worth underscoring that the notion of grammatical cost is not one-dimen- sional. The proposed analysis demonstrates that both facets of speech have a say in the design of sound inventories. Thus, parallel to the articulatory dimension, there is a perceptual dimension, which can be equally influential. It is this duality that explains the variation exhibited by nasal solos. We have just seen that, when perception comes first, /m/ is the best choice; but when articulation takes the lead, /n/ is preferable. 88 CatJL 15, 2016 Carlos-Eduardo Piñeros 4.3. The predominant nasal duo The chances that the theory might overgenerate are substantially lower now. That is because twelve of the sixteen possible rankings have already been exam- ined, yet only three inventory types ―all of which have been attested― have emerged: the absence of nasal stops and the solos /m/ and /n/. This section adds to the positive results by showing that three of the remaining rankings converge on the duo /m n/. The introduction of a second unit becomes possible when Faith(nasal place) gains control over two members of each place hierarchy (21). This arrangement has the merit of resolving the disagreement between the grammatical-cost dimensions. In effect, while acting autonomously, the place hierarchies concur that [labial] and [coronal] are their top two choices; hence, the duo /m n/. (21) Grammar of languages with the duo /m n/ *Dor-in-C *Dor-in-N Faith(nasal place) *Lab-in-C *Cor-in-N *Cor-in-C *Lab-in-N Harmony is the hallmark of this grammar. Note that the demands of both place hierarchies are being equally considered, while the faithfulness constraint has gained just enough power to ensure a reasonable investment in nasal stops. As a result, none of the conflicting forces is overwhelmed by the others. The stability ensuing from such a harmonic system suggests that it should be a common stage in language evolution, a hypothesis validated by the fact that the duo /m n/ is the most common nasal-stop inventory of all. In ANSID, it is represented by 109 languages (24.12%). The fact that most languages have at least two nasal-stop phonemes can now be addressed. Why are there so few languages with nasal solos? A comparison among (17), (19) and (21) reveals the answer. Observe that, in the first two sce- narios, obtaining a single unit requires paying half the cost of a second one. This is evinced by the fact that, in order to obtain either /m/ or /n/, at least two constraints need to be dominated on one of the place hierarchies, while the domination of a single constraint is enough on the other place hierarchy (see 17 and 19). Having already paid for one and a half units, the grammar might as well try to get two. Reaching that target requires making a further push, but the compensation is worth it. It ensures that no part of the investment goes to waste (21). Exceptional nasal-stop inventories CatJL 15, 2016 89 Two additional rankings lead to the duo /m n/. As the reader can verify in (22) and (23), attempts to recruit a third unit by making further advances in the articulatory or the perceptual dimension alone fail. This corroborates that the pro- posed model is a system of checks and balances which keeps the grammar from overgenerating. (22) Attempt to enlarge of the duo /m n/ through a further trespass on articulation *Dor-in-N >> *Cor-in-N >> *Lab-in-N Faith(nasal place) *Dor-in-C >> *Lab-in-C >> *Cor-in-C (23) Attempt to enlarge of the duo /m n/ through a further trespass on perception *Dor-in-C >> *Lab-in-C >> *Cor-in-C Faith(nasal place) *Dor-in-N >> *Cor-in-N >> *Lab-in-N 4.4. The predominant nasal trio The last of the sixteen rankings corresponds to the ascent of Faith(nasal place) above the relevant fragments of both place hierarchies (24). This is the source of the trio /m n ŋ/, which, despite being more inclusive, perpetuates the bias towards anteriority. (24) Grammar of languages with the trio /m n ŋ/ Faith(nasal place) *Dor-in-C *Dor-in-N *Lab-in-C *Cor-in-N *Cor-in-C *Lab-in-N 90 CatJL 15, 2016 Carlos-Eduardo Piñeros 4.5. Ranking typology Figure 4 provides a global view of the analysis. This is the ranking typology ensu- ing from the proposed constraint system. Each position that the faithfulness con- straint can occupy with respect to the Articulatory Consonant Place Hierarchy has been identified with a letter, while the positions that the faithfulness constraint can occupy with respect to the Perceptual Nasal-Stop Place Hierarchy have been identified with a number. (The lowest and highest values have been assigned to the lowest and highest ranks, respectively.) This notation allows us to refer to each ranking as a function between a letter and a number: f(Letter, Number). The most restrictive ranking is (A, 1), the epitome of systems deprived of nasal stops. Six other rankings yield the same result, (B, 1), (C, 1), (D, 1), (A, 2), (A, 3), (A, 4), because each grammatical cost dimension is autonomous; hence, capable of counteracting the other one. Additionally, the ranking (B, 2) also precludes all nasal stops because the place hierarchies disagree as to which is the least costly place category. The first nasal-stop phoneme is invariably an anterior unit. Its recruitment becomes possible when the lowest member of one place hierarchy and at least two members of the other one succumb. Two rankings, (C, 2) and (D, 2), generate the solo /m/, while another two, (B, 3) and (B, 4), produce the solo /n/. Three rankings converge on the predominant nasal duo: /m n/. The epitome of such systems is (C, 3), a harmonic grammatical state given that none of the conflict- ing forces is oppressed by the others. Two other rankings yield the same result, (D, 3) and (C, 4), because all unilateral ascents are unavailing due to the autonomy of the place hierarchies. Figure 4. Ranking typology. Exceptional nasal-stop inventories CatJL 15, 2016 91 The last possibility gives rise to the predominant nasal trio: /m n ŋ/. At this stage, both fragments of the place hierarchies are conquered; hence, unable to prevent the introduction of posterior units. 5. Assessing the quality of phonological contrasts While perception has been recognized as one of the forces shaping nasal-stop inventories, it needs to be added that shouldering one of the grammatical cost dimensions is not the only way in which it intervenes. Ensuring that the quality of phonological contrasts maintains a certain standard is another area where per- ceptual factors take center stage. In this section, I draw on this additional function to account for the most challenging of the exceptions to normal nasality: the duo /m ŋ/ (see Table 2). The goal is to demonstrate that the unexpected structure of this inventory is justified by the need to meet the heightened demands of a third evaluation metric: a perceptual-distance hierarchy. That phonological contrasts differ in quality is one of the premises of Dispersion Theory (Flemming 1995, 1996, 2002, 2004a, b, 2006). This model holds that it is advantageous for languages to contrast sounds which are dispersed in the acoustic space because their separation makes them less confusable. It is then expected that languages seeking to reduce confusability will avert poorly dispersed contrasts. Signs of dispersion are rare in nasal-stop inventories, but they do exist. It will be shown that, despite being articulatorily and perceptually affordable, a nasal stop may still be rejected if its presence in the inventory is an obstacle to obtaining a more dispersed contrast. The method used to calculate dispersion involves the construction of a percep- tual scale where the relevant units are located according to the values of their acous- tic properties. Recall from Section 3.2 that the properties serving to distinguish one nasal murmur from another are the NZ and N-cluster. The scale in Figure 5 draws on these cues to estimate the perceptual distance between /m/, /n/ and /ŋ/. (Every integer indicates a span of 1000 Hz.) Because the NZ and N-cluster of /m/ are around 800 Hz, this unit has been placed low. By contrast, /n/ and /ŋ/ have been granted increasingly higher positions because their values for the same cues are around 1780 Hz and 3700 Hz, respectively. The acoustic proximity that exists between /m/ and /n/ yields a low dispersion value when these units are contrasted: 1. On the other hand, a contrast between a posterior unit such as /ŋ/ and an anterior one such as /m/ translates into a higher dis- persion value: 3. One can then foresee that a language aiming to obtain the best pos- sible contrast will have to abstain from using the acoustically intermediate unit: /n/. Figure 5. Nasal perceptual space. 92 CatJL 15, 2016 Carlos-Eduardo Piñeros A family of perceptual-distance constraints is introduced to formalize this pro- posal. Its general definition is given in (25). Its members, which are organized into the universal hierarchy in (26), require phonemic nasal stops to keep a certain distance between them. Note that members referring to low values are ranked higher than those referring to higher values. This is to promote dispersion. The further apart two phonemes stand, the less costly their contrast will be because fewer P-Distance constraints will object to it. (25) P-Distance(N:x):10 A perceptual distance equal or greater than x is required between contrasting nasal stops. (26) Universal ranking within the P-Distance family P-Dist(N:1) >> P-Dist(N:2) >> P-Dist(N:3) A feud between P-Distance and Faith(nasal place) is inevitable. These principles clash because, while discarding intermediate units is a logical way to improve the quality of contrasts, any simplification is detrimental to faithful- ness. As with any other constraint conflict, the resolution will depend on which principle is in power. To begin with the most common scenario, let us assume that P-Dist(N:1) is the only member of its family that dominates Faith(nasal place). A perceptual-distance requirement of 1 is thereby established, and since that happens to be the distance between /m/ and /n/, which are quasi-universal, it is possible to construe this ranking as a default setting, which few languages bother to reset. The fact that dispersion is a relational concept entails that it becomes relevant when several members of the same segmental class are affordable. A case in point is the trio /m n ŋ/. Returning for a moment to (24), it is easy to understand why it was possible to derive that inventory without considering that two of its units are acoustically adjacent. Since a perceptual-distance requirement of 1 sanctions adjacency, P-Distance is powerless in languages embracing the default dispersion setting. It has no choice but to accept the selections agreed on by Faith(nasal place) and the place hierarchies. This can be verified in (27) where the default dispersion setting has been added. A special situation arises when Faith(nasal place) falls below additional members of the P-Distance family. In (28), for instance, it has lost its ground to P-Dist(N:2), which now has the power to influence the selections. The new per- ceptual-distance requirement that is established thereby compromises the integrity of the nasal trio because, while the contrast between /m/ and /ŋ/ manages to meet it, the contrast between /m/ and /n/ falls short. As the only escape from this impasse is to discard one of the anterior units, the grammar must weigh the consequences to make the best choice. If /m/ is forsaken, the contrast between /n/ and /ŋ/ will be sanctioned by P-Dist(N:1) and P-Dist(N:2), but it will be censured by P-Dist(N:3). 10. To keep the name of perceptual-distance constraints short, the pertinent acoustic cues—the NZ and N cluster—are abbreviated as N. Exceptional nasal-stop inventories CatJL 15, 2016 93 On the other hand, if /n/ is forsaken, the contrast between /m/ and /ŋ/ will be sanc- tioned by all three principles. The crux of the matter is that, because the violation of Faith(nasal place) is inevitable, the decision falls on P-Distance, whose members collude to favor of the most dispersed contrast. (27) Grammar of languages with the trio /m n ŋ/ P-Dist (N:1) Faith(nasal place) *Dor-in-C P-Dist *Dor-in-N (N:2) *Lab-in-C *Cor-in-N P-Dist *Cor-in-C (N:3) *Lab-in-N (28) Grammar of languages with the duo /m ŋ/ P-Dist (N:1) P-Dist (N:2) Faith(nasal place) *Dor-in-C P-Dist *Dor-in-N (N:3) *Lab-in-C *Cor-in-N *Cor-in-C *Lab-in-N The picture emerging from the above is that the trio /m n ŋ/ and duo /m ŋ/ are akin. The latter is the product of contracting the former so that the best possible contrast can be obtained. Both diachrony and synchrony support this view as there are at least two languages where the transition from the larger to the smaller inven- tory occurred in the past (i.e. Palauan and Mekeo) and one where it is currently 94 CatJL 15, 2016 Carlos-Eduardo Piñeros active (i.e. Samoan).11 In its evolution from Proto-Austronesian, Palauan simplified the trio /m n ŋ/ by changing the manner of articulation of the intermediate unit; it rendered it lateral (Foley 1975). Mekeo arrived at the same result by shifting the place of articulation of the intermediate unit backwards; it converted it into velar (Blevins 2009). The latter development also occurs in modern Samoan, where /n/ takes over the space of /ŋ/ when speakers switch from tautala lelei ‘good speech’ to tautala leanga ‘bad speech’ (Duranti 1994, Hyman 2008). To summarize this section, the conundrum posed by the duo /m ŋ/ has been solved. The ousting of an anterior unit by a posterior one is induced by dispersion. One of the merits of this account is that it corrects a theoretical blunder. Reversing the dominance relations on which the place hierarchies are based has been avoided. We have learned that what appears to be a change in grammatical-cost values is a mirage created by another contending force: the drive to enhance phonological con- trasts. It is in response to this pressure that /n/ has to be left out of the exceptional nasal duo. The retention of /m/ is absolutely necessary because it plays a key role in achieving maximal dispersion within this consonant class. Another revelation this analysis makes is that, although posteriority is normally detrimental to the per- ception of nasal stops, it has the potential to become an advantage. That is because its combination with anteriority affords a better contrast than the combination of two anterior units. As a final remark, it should be noted that the introduction of a third evaluation measure does not make the theory excessively permissive. That is thanks to the con- dition that universal rankings must be strictly observed. If we go back to (28), for instance, there is only one alternative arrangement the system allows. Assume that, instead of 2, the perceptual-distance requirement is increased to 3; that is to say that Faith(nasal place) falls below P-Dist(N:3). It is easy to see that the outcome will still be the duo /m ŋ/ because no other pair can meet such a high standard. 6. Integration The analysis presented above differs from previous functional approaches in several important ways. Unlike Functional Phonology (Boersma 1997, 1998), the roles of speaker and listener have not been segregated; in other words, there is not a production and a perception grammar working independently of each other. The strategy I have adopted has been to integrate the structural constraints grounded on articulatory factors with those grounded on perceptual factors into a single gram- mar, where they are free to interact with one another as well as with the relevant faithfulness constraint. This integration has been the key to explaining the modicum of variation that exists in nasal-stop inventories. A notable difference with respect to Dispersion Theory (Flemming 1995, 1996, 2002, 2004a, b, 2006) is that, instead of introducing a constraint family dedicated to the promotion of contrasts (i.e. Maintain Contrasts or Maximize Contrasts), 11. The fact that these are all Austronesian languages suggests that the tendency to heighten the per- ceptual-distance requirement is a rare areal trait. Exceptional nasal-stop inventories CatJL 15, 2016 95 an independently needed constraint family has been charged with that task (i.e. Faithfulness). This has two positive effects. On the one hand, it avoids constraint duplication, and, on the other hand, it makes it possible to integrate sound inven- tories with the rest of the grammar because the principles that govern their struc- ture turn out to be the same ones that govern phonological patterns: *Structure, Faithfulness, and P-Distance.12 With regard to the latter principle, it should be added that it is superior to Flemming’s MinDist because, as shown above, the required minimal distance is not a primitive notion. It can be derived via interac- tion with Faithfulness. An additional difference with respect to both Functional Phonology and Dispersion Theory is that the present proposal recognizes that, parallel to the articu- latory effort dimension, there is a perceptual effort dimension. Accordingly, every nasal stop has two different grammatical costs: one articulatory and one perceptual. This understanding is missing from previous functional proposals, which, despite arguing that perception plays a central role in the phonology, fail to recognize a perceptual effort dimension. The view defended here is that perception plays dual roles. In addition to being the driver of dispersion phenomena (26), it is the founda- tion of one of the grammatical cost dimensions (11). In sum, full integration of articulation and perception into the phonology is the defining trait of the above analysis. 7. Conclusion This article has explored the topic of exceptionality in phonology focusing on a subset of nasal-stop inventories: those falling below the normal use of nasality (Table 2). The typology shows that this consonant class has a proclivity towards anteriority, or seen from the opposite angle, an aversion towards posteriority. This is evinced by the absence of nasal solos consisting of a posterior unit and the attestation of only one and extremely infrequent nasal duo where a posterior unit is present (i.e. /m ŋ/). One cannot get past these findings without asking the following question: Why is there not a balanced mix of places of articulation in these invento- ries? A sensible hypothesis is that anterior nasal stops make better phonemes than their posterior congeners, but this is not easy to defend because the patterns are not consistent enough to postulate an invariable order of precedence. The same difficulty arises in trying to rank the anterior units. Nasal solos, which should reveal which is the best member of the class, fail to do so because some of them pick /n/, while others opt for /m/. An easy way to deal with this variation is to grant the anterior place categories equal phonological status, but this is unsatisfac- tory because the structure of nasal solos becomes a matter of chance. Another intriguing fact is that, when languages decide to use nasal-stop pho- nemes, they are not satisfied with only one. The vast majority recruits at least two 12. Enhancement phenomena (e.g. prenasalization of voiced stops, aspiration of voiceless stops, round- ing of postalveolar fricatives, etc.) are proof that P-Distance constraints are not confined to sound inventories. They are operative throughout the phonology. 96 CatJL 15, 2016 Carlos-Eduardo Piñeros (92.70%). Logic tells us that, if languages want to be economical, they should limit the number of such phonemes to one. Why should they insist on a minimum of two? In order to cope with this range of facts, it is imperative that the theory be enriched. One of the conclusions emerging from this study is that grammars assess the cost of linguistic structures on multiple dimensions. The view that there is a single universal place hierarchy has therefore been rejected. It has been demonstrated that both the articulatory and the perceptual facets of speech influence the design of nasal-stop inventories and that the cost of articu- lating or perceiving place features varies across segmental classes. Differences in articulator velocity underpin the Articulatory Consonant Place Hierarchy (9), the assessment that, when it comes to forming a consonantal constriction, the corona is most adept, the lips are less so, and the dorsum is even less. Likewise, differences in acoustic salience buttress the Perceptual Nasal-Stop Place Hierarchy (11), the assessment that, when enveloped in the nasal murmur, dorsality, coronality, and labiality are gradually less taxing to the human ear. The point needs to be underscored that, despite being universal, these place hierarchies do not apply across the board. Their reference to co-occurring properties of the segment (e.g. [consonantal], [nasal], [occlusive]) makes them irrelevant to some segmental classes without compromising their universality. The application of different evaluation measures to the same segmental class has a sound justifica- tion as well. Although both hierarchies are concerned with the cost of nasal stops, there is no duplication of labor because one assesses them articulatorily while the other one does so perceptually. A third point to emphasize is that, since each place hierarchy is autonomous, variation is expected to arise. Languages will make dif- ferent selections depending on whether they set articulatory or perceptual cost as their leading priority. The grammar thus conceived is rich in interactions (Figure 4). The disagree- ment between the place hierarchies and their alliance against faithfulness gives rise to a three-way conflict, which individual languages arbitrate at their will. The one condition that applies is, of course, that universal rankings be respected. All possi- bilities that the constraint system allows have been checked in order to verify that the use of multiple evaluation metrics does not cause overgeneration. The results are highly positive. The proposed model is able to rule out unattested patterns, generate those that have been attested, and explain why, among the latter, there are some which are highly frequent and others which are extremely rare. The existence of two different nasal solos is no longer a mystery. As [coronal] is the best place feature a nasal stop can have from an articulatory point of view, languages seeking articulatory savings will pick /n/ for their nasal solos. On the other hand, since [labial] is the best place feature a nasal stop can have from a perceptual point of view, languages seeking perceptual savings will choose /m/ instead. Additionally, because there are no evaluation measures according to which a posterior nasal stop is the least costly member of its class, nasal solos such as */ŋ/, */ɲ/, or */ɳ/ are impossible to generate. The detriment to posterior units is perpetuated by nasal duos. It has been determined on empirical grounds that the next rank on each place hierarchy is Exceptional nasal-stop inventories CatJL 15, 2016 97 held by another anterior category. In the articulatory dimension, [labial] comes immediately after [coronal], while in the perceptual dimension, [coronal] comes immediately after [labial]. Since this means that the first two positions that become available will be seized by /m/ and /n/, there cannot be any nasal duos deprived of anterior units (e.g. */ŋ ɲ/, */ŋ ɳ/, and */ɲ ɳ/). There will be, by contrast, one nasal duo deprived of posterior units (i.e. /m n/). The normal minimum of two units now makes sense. The fact that the category sitting at the bottom of one place hierarchy occupies the immediately higher rank on the other place hierarchy entails that, in order to recruit its first nasal stop, a lan- guage must pay half the cost of a second one (17 and 19). As paying the remaining half facilitates the recovery of the extra payment, the avoidance of nasal solos is to be expected. Languages will seek a minimum of two units so that they can get a full return for their investment (21). With this additional revelation, the reasons behind the probabilistic universal in (2) are fully understood. Despite these gains, the analysis would not have been entirely satisfactory if a compelling explanation for the existence of the duo /m ŋ/ had not been found. This exception seems to contravene the universality of the place hierarchies because it escapes their prediction that the introduction of the first posterior unit should follow that of both anterior units. Fortunately, the problem is only ostensible. The duo /m ŋ/ is actually a contraction of the trio /m n ŋ/. This discovery was made by recognizing that there is a third evaluation measure at work: a perceptual-distance hierarchy (26). The drive to enhance phonological contrasts may cause languages to discard an acoustically intermediate unit while retaining the most peripheral ones. In this way, the fact that /n/, rather than /m/, is the anterior unit missing from the exceptional nasal duo has been explained and it has been further revealed that, although posteriority is normally disadvantageous for nasal stops, there are cases where it proves to be an advantage. We now have sensible answers to some of the key questions about the design of nasal-stop inventories. Rather than grammatical malfunctions, the exceptions were found to be the product of an intricate network of interactions, which lead to variation because individual grammars may settle conflicts in different ways. The diversity promoted thereby is highly limited, nonetheless, because there are substantive factors which impose a universal order on certain linguistic principles. The balance between restrictiveness and freedom that was needed to cope with the facts has been achieved. Universal rankings make the theory highly restrictive, while multiple interactions afford a modicum of variation. References Basbøll, Hans. 1985. Review of Patterns of sounds by Ian Maddieson. Phonology Yearbook 2: 343-353. Boersma, Paul. 1997. Inventories in Functional Phonology. Proceedings of the Institute of Phonetic Sciences of the University of Amsterdam 21: 59-90. [Rutgers Optimality Archive 232, Boersma, Paul. 1998. Functional Phonology. Ph.D. dissertation, University of Amsterdam. The Hague: Holland Academic Press. Borzone de Manrique, Ana María and Ángela Signorini. 1983. Segmental duration and rhythm in Spanish. Journal of Phonetics 11: 117-128. Blevins, Juliette. 2004. Evolutionary phonology: The emergence of sound patterns. Cambridge: Cambridge University Press. Blevins, Juliette. 2007. A Long Lost Sister of Proto-Austronesian?: Proto-Ongan, Mother of Jarawa and Onge of the Andaman Islands. Oceanic Linguistics 46.1: 154-98. Blevins, Juliette. 2009. Another universal bites the dust: Northwest Mekeo lacks coronal phonemes. Oceanic Linguistics 48: 264-273. Byrd, Dani. 1992. Perception of Assimilation in Consonant Clusters: A Gestural Model. Phonetica 49: 1-24. Clements, George N. 2009. The role of features in phonological inventories. In E. Raimy, and C. Cairns (eds.). Contemporary Views on Architecture and Representations in Phonology, 19-68. Cambridge, MA: MIT Press. Cooper, Franklin S., Pierre C. Delattre, Alvin M. Liberman, John M. Borst, and Louis J. Gerstman. 1952. Some experiments on the perception of synthetic speech sounds. Journal of the Acoustic Society of America 24: 597-606. de Lacy, Paul. 2002. The formal expression of scales. Ph.D. Dissertation, University of Massachusetts, Amherst. de Lacy, Paul. 2006. Markedness. Reduction and Preservation in Phonology. Cambridge: Cambridge University Press. Delattre, Pierre C. 1958. Acoustic cues in speech: First report. Phonetica 2: 108-118. Dukiewicz, Leokadia. 1967. Polskie gloski nosowe. Warsaw: Polska Akademia Nauk. Duranti, Alessandro. 1994. From Grammar to Politics: Linguistic Anthropology in a Western Samoan Village. Berkley: University of California Press. Fant, Gunnar. 1960. Acoustic Theory of Speech Production. The Hague: Mouton. Ferguson, Charles A. 1974. Universals of nasality. Working Papers on Language Universals 14: 1-16. Ferguson, Charles A. 1975. Universal tendencies and ‘normal’ nasility. In C. Ferguson, L. Hyman, and J. Ohala (eds.). Nasalfest: Papers from a Symposium on Nasals and Nasalization, 175-196. Stanford, California: Language Universals Project, Department of Linguistics, Stanford University. Flemming, Edward. 1995. Audio representations in phonology. Ph.D. Dissertation, University of California, Los Angeles. Flemming, Edward. 1996. Evidence for constraints on contrast: The dispersion theory of contrast. UCLA Working Papers in Phonology 1: 86-106. Flemming, Edward. 2002. Auditory Representations in Phonology. New York: Routledge. Flemming, Edward. 2004a. Contrast and perceptual distinctiveness. In B. Hayes, R. M. Kirchner, and D. Steriade (eds.). Phonetically Bases of Markedness, 232-276. Cambridge: Cambridge University Press. Flemming, Edward. 2004b. Speech Perception and Phonological Contrast. In D. B. Pisoni and R. E. Remez (eds.). The Handbook of Speech Perception, 156-182. Malden, MA: Blackwell Pub. Flemming, Edward. 2006. The role of distinctiveness constraints in phonology. Unpublished manuscript, MIT. Exceptional nasal-stop inventories CatJL 15, 2016 99 Foley, William A. 1975. Some rules involving nasals and their implications. In C. Ferguson, L. Hyman, and J. Ohala (eds.). Nasalfest: Papers from a Symposium on Nasals and Nasalization, 213-230. Stanford, California: Language Universals Project, Department of Linguistics, Stanford University. Fujimura, Osamu. 1962a. Analysis of nasal consonants. The Journal of the Acoustical Society of America 34: 1865-75. Fujimura, Osamu. 1962b. Formant-Antiformant structure of Nasal Murmurs. In Proceedings of the Speech Communication Seminar, Vol 1, Kungliga Tekniska Högskolan,1-9. Stockholm: Royal Institute of Technology, Speech Transmission Laboratory. Gubrynowicz, R., L. Le Guennec, and G. Mercier. 1985. Detection and recognition of nasal consonants in continuous speech: Preliminary results. In R. De Mori and C. Suen (eds.). New Systems and Architectures for Automatic Speech Recognition, 613-28. Berlin: Springer-Verlag. House, Arthur S. 1957. Analog studies of nasal consonants. The Journal of Speech and Hearing Disorders 22: 190-204. Hudgins, C. V., and R. H. Stetson. 1937. Relative speed of articulatory movements. Archives néerlandaises de phonétique expérimentale 13: 85-94. Hyman, Larry M. 2008. Universals in phonology. The Linguistic Review 25: 83-137. Klatt, Dennis H. 1976. Linguistic uses of segmental duration in English: Acoustic and perceptual evidence. Journal of the Acoustic Society of America 59: 1208-1221. Kuehn, David P., and Kenneth L. Moll. 1976. A cineradiographic study of VC and CV articulatory velocities. Journal of Phonetics 4: 303-320. Ladefoged, Peter and Ian Maddieson. 1996. The sounds of the world’s languages. Oxford, UK: Blackwell Publishers. Liberman, Alvin M., Pierre C. Delattre, Franklin S. Cooper, and Louis J. Gerstman. 1954. The role of consonant-vowel transitions in the perception of the stop and nasal consonants. Psychological Monographs 68. Maddieson, Ian. 1984. Patterns of sounds. Cambridge: Cambridge University Press. Maddieson, Ian and Kristin Precoda. 1990. Updating UPSID. UCLA Working Papers in Phonetics 74: 104-111. Malécot, André. 1956. Acoustic cues for nasal consonants: An experimental study involving a tape-splicing technique. Language 32: 274-284 Narayan, Chandan Raghava. 2006. Acoustic-perceptual salience and developmental speech perception. Doctoral dissertation, University of Michigan. Ohala, John J. 1975. Phonetic explanations for nasal sound patterns. In C. Ferguson, L. Hyman, and J. Ohala (eds.). Nasalfest: Papers from a Symposium on Nasals and Nasalization, 289-316. Stanford, California: Language Universals Project, Department of Linguistics, Stanford University. Ohala, John J and Manjari Ohala. 1993. The phonetics of nasal phonology: theorems and data. In M. Huffman and R. Krakow (eds.). Phonetics and Phonology: Vol. 5, nasals, nasalization, and the velum, 225-249. San Diego, California: Academic Press, Inc. Pagliuca, William and Revere D. Perkins. 1986. Review of Patterns of sounds. Lingua 70: 365-378. Penny, Ralph. 2002. A history of the Spanish language. Cambridge: Cambridge University Press. 100 CatJL 15, 2016 Carlos-Eduardo Piñeros Prince, Alan, and Paul Smolensky. 1993.[2004] Optimality Theory: Constraint inter- action in generative grammar. Colorado: Reuters University and University of Boulder. Revised version published by Blackwell, 2004. Recasens, Daniel. 1983. Place cues for nasal consonants with special reference to Catalan. The Journal of the Acoustical Society of America 73: 1346-1353. Recasens, Daniel. 1992. Phonology, and speech production and perception: The case of nasal stops. In W. Dressler, H. Luschützky, O. Pfeiffer, and J. Rennison (eds.). Phonologica 1988, 229-238. Cambridge: Cambridge University Press. Rice, Keren and Peter Avery. 1993. Segmental complexities and the Structure of Inventories. Toronto Working Papers in Linguistics 12: 135-153. Roon, Kevin, Adamantios I. Gafos, Philip Hoole, and Chakir Zeroual. 2007. Influence of Articulator and Manner Stiffness. In J. Trouvain and W. Barry (eds.). Proceedings of the 16th International Congress of Phonetic Sciences, 409-412. Saarbruecken, Germany. Simpson, Adrian P. 1999. Fundamental problems in comparative phonetics and pho- nology: does UPSID help to solve them? Proceedings of the 14th International Congress of Phonetic Sciences, 1: 349-352. Vaux, Bert. 2009. The role of features in a symbolic theory of phonology. In Eric Raimy, and Charles E. Cairns (eds.). Contemporary Views on Architecture and Representations in Phonology, 75-97. Cambridge, MA: MIT Press. ISSN 1695-6885 (in press); 2014-9719 (online) Catalan Journal of Linguistics 15, 2016 101-119 Diminutives: Exceptions to Harmonic Uniformity Péter Rebrus Hungarian Academy of Sciences [email protected] Péter Szigetvári Eötvös Loránd University [email protected] Received: February 25, 2016 Accepted: June 22, 2016 Abstract Front-back harmony in Hungarian is governed by the root of words: it disregards suffixes in most cases. That is, a back-vowelled root (B) followed by any number of neutral-vowelled suf- fixes (N) will take a back-vowelled suffix (B+N+N+B), but a root with a back vowel followed by several neutral vowels is possibly followed by a front-vowelled suffix (F): BNN+F/B. We call this Harmonic Uniformity. This is respected even in truncated stems: NB→N+N+B, although NN stems practically never take a back suffix (NN+F). Diminutive forms are the only exceptions to this pattern. We claim that this is so, because diminutive forms are much more loosely related to their “base” than is the case with any other types of suffixation. Keywords: morphology; vowel harmony; Hungarian; diminutives; truncation Resum. Diminutius: excepcions a la uniformitat harmònica L’harmonia anterior-posterior en hongarès és regida per l’arrel dels mots: en la majoria de casos no té en compte els sufixos. És a dir, una arrel amb una vocal posterior (P) seguida per qualsevol nom- bre de sufixos amb una vocal neutral (N) prendrà un sufix amb una vocal posterior (P+N+N+P), però una arrel que contingui una vocal posterior seguida de diverses vocals neutrals pot anar seguida d’un sufix amb una vocal anterior (A): PNN+A/P. Anomenem aquest efecte Uniformitat Harmònica. Aquesta es respecta fins i tot en radicals truncats: NP→N +N+P, encara que els radicals NN pràcticament mai no prenen un sufix posterior (NN+A). Les formes diminutives són l’única excepció a aquest patró. Proposem que la raó és que les formes diminutives estan relacio- nades a la seva «base» d’una manera molt més laxa que en qualsevol altre tipus d’afixació. Paraules clau: morfologia; harmonia vocàlica; hongarès; diminutius; truncament Table of Contents 1. Vowels and harmony in Hungarian 5. Diminutives and their roots 2. Harmonic uniformity 6. Diminutives and harmony 3. Truncation 7. Conclusions 4. Diminutives References 102 CatJL 15, 2016 Péter Rebrus; Péter Szigetvári In a nutshell the problem we are concerned with is the following. The back-har- monic root neːmɑ ‘mute’ loses its second vowel when suffixed: neːmiːt ‘make mute’. Yet, this stem remains back-harmonic: neːmiːtɑnɑk ‘make mute-3PL’. The back-harmonic root eːvɑ, a name, also loses its second vowel in its diminutive form: eːvi. This stem, however, is front-harmonic: eːvinɛk ‘Eva-DIM-DAT’. Why should the two stems be different? Diminutive forms are unique in the morphology of Hungarian. Although trunca- tion is not only applied in the case of diminutive formation, but also in the case of various other morphological operations, it is only in the case of diminutives that the harmonic properties of words may change. We submit that this is due to the facts that (i) diminutive formation is not entirely productive, (ii) truncation is potentially more radical in diminutive formation than in any other case, (iii) the semantic bond between a word and its diminutive form is weaker than in other comparable word pairs, therefore (iv) diminutive forms are not “derived” (i.e. they do not have the morphological structure X+DIM), but they can be treated as (monomorphemic) roots on their own right. Since the harmonic properties of words are influenced by their morphological complexity, diminutives behave exceptionally in appear- ing to be morphologically complex, but behaving—at least with respect to vowel harmony—as morphologically simplex forms. We first introduce the vowel inventory of Hungarian1 and the basics of front/ back harmony, with special emphasis on the front unrounded (i.e. the neutral) vow- els. We then show what harmonic uniformity means: in any noncompound word, the root morpheme’s harmonic property is inherited by the whole word. The next section looks at morphological operations that involve the truncation of the last vowel of the word, and the influence of truncation on vowel harmony. The most extensive part of the paper discusses the many ways diminutives are formed in Hungarian. We argue that a templatic diminutive word is a root: in many respects— most notably, with regard to vowel harmony—it behaves as morphologically sim- plex.2 Our main arguments for this claim are given in the last but one section. 1. Vowels and harmony in Hungarian Hungarian has root-governed front/back harmony, i.e. a root3 with a front vowel is typically followed by suffixes containing front vowels and a root with a back vowel is typically followed by suffixes containing back vowels (for a recent general survey, as well as references to earlier accounts see Törkenczy 2011). The vowel inventory is presented in (1). 1. In this paper “Hungarian” means standard Hungarian. In order to indicate morphemes clearly, we give fairly broad transcriptions (e.g. we ignore voicing assimilation and productive hiatus filling). 2. Diminutives are simplex as far as vowel harmony is concerned. However, such words do sometimes contain consonant clusters that are otherwise unprecedented within a morpheme: e.g. køsʧi ‘thanks- DIM’ (< køsønøm). Yet, as will be shown below, other diminutives even simplify well-formed morpheme-internal consonant clusters. 3. In this paper we use the term root for ‘monomorphemic stem’, i.e. in a form made up of X followed by two suffixes Y and Z, X is the root of XYZ (and of XY), and XY is the stem of XYZ. Diminutives: Exceptions to Harmonic Uniformity CatJL 15, 2016 103 (1) The vowel inventory of Hungarian front unrounded, front rounded back aka neutral i iː y yː u uː high — eː ø øː o oː mid ɛ — — — ɑ aː low Short and long vowels can be paired. The pairs are both phonetically similar (almost identical in the case of high vowels, rather different in the case of the low ɛ/mid eː and the low ɑ/aː pairs—as the symbols used here show) and mor- phologically related. All the short/long pairs participate in various morphological alternations: e.g. hiːd ‘bridge’ ~ hid-ɑk ‘bridges’, yːr ‘space’ ~ yr-yl ‘become empty’, huːs ‘twenty’ ~ hus-ɑd ‘one twentieth’, keːz ‘hand’ ~ kɛz-ɛl ‘handle’, tɛvɛ ‘camel’ ~ tɛveː-t ‘camel-ACC’, nøː ‘grow’ ~ nøv-ɛl ‘increase’, loː ‘horse’ ~ lov-ɑ ‘horse-3SGPOSS’, ɲaːr ‘summer’ ~ ɲɑr-ɑl ‘go on holiday’, pɑtɑ ‘hoof’ ~ pɑtaː-ʃ ‘hoofed’. This justifies coupling ɛ and eː despite their height difference. The height difference of the two vowels, however, cannot be ignored: it plays a role in the so-called height effect observed in neutral vowels, which we will briefly discuss below. We present vanilla cases of front/back harmony in (2) using the sublative suf- fix -rɑ/ɛ. (2) Front/back harmony a. viːz-rɛ ‘water-SUBL’, tyːz-rɛ ‘fire-SUBL’, uːt-rɑ ‘road-SUBL’ b. keːz-rɛ ‘hand-SUBL’, ʃør-rɛ ‘beer-SUBL’, bor-rɑ ‘wine-SUBL’ c. tɛj-rɛ ‘milk-SUBL’, vɑj-rɑ ‘butter-SUBL’, haːj-rɑ ‘fat-SUBL’ Besides front/back harmony, we can also observe rounding harmony in Hungarian. This is more limited than front/back harmony and it affects only a small number of suffixes (e.g. viːz-hɛz ‘water-ALLAT’, tyːz-høz ‘fire-ALLAT’, uːt-hoz ‘road-ALLAT’). In this paper we ignore rounding harmony. 1.1. Neutral vowels The front unrounded vowels in (1) are labelled neutral. This is justified by the behaviour of these vowels. Various types of disharmony characterize neutral vowels.4 Invariant suffixes — suffixes with a vowel that does not alternate due 4. The examples here and in the subsequent part contain the “totally” neutral vowels i, iː and eː, The “partially” neutral behaviour of the low vowel ɛ will be mentioned in §1.2. 104 CatJL 15, 2016 Péter Rebrus; Péter Szigetvári to vowel harmony — typically contain a neutral vowel. Some examples are given below. The adjectival -i is shown in (3a), the possessive -eː in (3b), the infinitival -ni in (3c), and the agentive -iːt in (3d). (3) Invariant suffixes containing a neutral vowel a. vi(ː)z-i ‘watery’, ty(ː)z-i ‘fiery’, u(ː)t-i ‘road-ADJ’ b. viːz-eː ‘water-POSS’, tyːz-eː ‘fire-POSS’, uːt-eː ‘road-POSS’ c. vin-ni ‘take-INF’, tyːz-ni ‘staple-INF’, un-ni ‘be bored-INF’ d. diːs-iːt ‘decorate’, yr-iːt ‘make empty’, uːj-iːt ‘make new’ Another type of disharmony that involves neutral vowels is standardly called transparency. Transparent vowels are invisible to vowel harmony, which accord- ingly works as if these vowels were not present. We give examples for back+neutral roots in (4a): these take back suffixes, and for front+neutral roots in (4b): these take front suffixes. (4) Transparency of neutral vowels a. fɑkiːr-rɑ ‘fakir-SUBL’, kɑʃteːj-rɑ ‘castle-SUBL, ɑleːl-vɑ ‘faint-PART’ b. kɛfiːr-rɛ ‘kefir-SUBL’, ʃøteːt-rɛ ‘dark-SUBL’, tɛkint-vɛ ‘look-PART’ The most spectacular type of disharmony is called antiharmony. Antiharmony is restricted to monosyllabic roots containing a front unrounded, i.e. neutral vowel. Examples for such roots are provided in (5a). The roots in (5b) are “normal” in that they take front suffixes as expected of a front-vowelled root. They are given for comparison. (5) Antiharmonic roots a. ʃiːr-rɑ ‘grave-SUBL’, heːj-rɑ ‘peel-SUBL’, irt-vɑ ‘destroy-PART’ b. hiːr-rɛ ‘news-SUBL’, eːj-rɛ ‘night-SUBL’, int-vɛ ‘wave-PART’ There is a very limited set of bisyllabic roots that are variably antiharmonic: feːrfi-rɑ/ɛ5 ‘man-SUBL’ and dɛreːk-rɑ/ɛ ‘waist-SUBL’, and fully so with vowel- initial suffixes: feːrfi-ɑk ‘man-PL’, dɛrɛk-ɑm ‘waist-1SGPOSS’. The fact that (almost) only monosyllabic roots may be unhesitatingly antiharmonic is named the polysyllabic split by Rebrus & Törkenczy (2015).6 5. Cf. the archaic variant feːrfiu(ː) ‘man’. 6. In some dialects there are further examples for polysyllabic antiharmonicity, e.g. piʃil-ok ‘pee-1SG’, fiktiːv-ɑn ‘fictively’, see Blaho & Szeredi 2013. Diminutives: Exceptions to Harmonic Uniformity CatJL 15, 2016 105 1.2. Gradient neutrality As shown in (1), there are four neutral vowels in Hungarian, i, iː, eː, and ɛ. It has been observed that these four vowels are not equally neutral (e.g. Kontra & Ringen 1986, Siptár & Törkenczy 2000: 70ff). Bisyllabic roots with a back vowel followed by i or iː (Bi:, we use B for any back vowel) practically always take the back variant of suffixes (e.g. fɑkiːr-rɑ ‘fakir-SUBL’; ɑbsint-rɑ/ɛ ‘absinthe-SUBL’ is a rare coun- terexample). Beː roots are not uniform, most are common with a back variant (e.g. kɑʃteːj-ra ‘castle-SUBL’), but some are variable (e.g. sɑteːn-rɑ/ɛ ‘satin-SUBL’). Finally, Bɛ roots are usually variable (e.g. fotɛl-rɛ/ɑ ‘armchair-SUBL’) and some take only the front variant of suffixes (e.g. haːrɛm-rɛ ‘harem-SUBL’). As can be verified, the higher a neutral vowel, the “more transparent” it is. This phenomenon is called the height effect by Hayes & Cziráky Londe (2006). The number of neutral (i.e. transparent) vowels in the string across which harmo- ny spreads also affects the degree of transparency. As we have seen Bi roots almost exclusively take the back variant of suffixes, but Bii roots are usually variable and perhaps tilt towards taking the front variant (e.g. ɑlibi-rɛ/ɑ ‘alibi-SUBL’, kolibri-rɛ/ɑ ‘hummingbird-SUBL’, klɑrineːt-rɛ/ɑ ‘clarinet-SUBL’; but again note the exclusively back harmonic aːpriliʃ-rɑ ‘April-SUBL’). Various combinations of neutral vowels yield different results (e.g. bɑkɛlit-rɛ/ɑ ‘bakelite-SUBL’ vs kɑbinɛt-rɛ/*ɑ ‘cabinet- SUBL’), but we need not go into further detail here (see Forró 2012 for more data). This phenomenon is called the count effect by Hayes & Cziráky Londe (2006). The four neutral vowels are also different in their distribution in suffixes. The two highest ones, i and iː, do not alternate in suffixes with a back counterpart: suf- fixes containing these vowels are invariant.7 The mid eː occurs in both invariant suffixes (e.g. -eː ‘POSS’, -eːk ‘FAM.PL’, -keːnt ‘ESSIV/FORMAL’, -neːk ‘1SG. COND’,) and harmonizing ones (e.g. -neːl/-naːl ‘ADESS’, -veː/-vaː ‘TRANSLAT’, -ʃeːg/-ʃaːg ‘-ness’). As expected, the low ɛ only occurs in harmonizing suffixes (e.g. -rɛ/-rɑ ‘SUBL’, -bɛn/-bɑn ‘INESS’). (For a list of suffixes, see Rebrus 2000: 775, 777ff, for a more detailed discussion of the graduality of vowel neutrality see Törkenczy & al 2013.) 2. Harmonic uniformity We have seen that the harmonic class of a monosyllabic root containing a neutral vowel is practically unpredictable. In the case of roots containing i or iː there is no statistical majority: about half of these roots take the back, the other half the front variant of suffixes. The other two neutral vowels, eː and ɛ, are usually front harmonic with a few cases of back harmony. Thus, at least in the case of monosyl- labic neutral-vowel roots we have to assume that these are lexically assigned to a harmonic class (front or back). 7. There is one suppletive suffix in which i alternates, the verbal 3SG.DEF -i/-jɑ: e.g. keːr-i ‘s/he asks for it’ vs vaːr-jɑ ‘s/he waits for it’. Note that even here i does not alternate with its high counterparts, y and u, despite the fact that these vowels are available in the inventory. So the ɛ/ø/o alternation is not parallelled by any i/y/u alternation. 106 CatJL 15, 2016 Péter Rebrus; Péter Szigetvári The harmonic class of a word is identical to the harmonic class of its root, be that root bound or free. In most cases this makes no difference, since suffixes fol- lowing the root agree with it in harmony. But the fact that harmony is determined by the root is visible if a root of the back-harmonic class is suffixed with invariant neutral-vowel suffixes. Consider the following examples. (6) The harmonic class of suffixed words a. hiːd-rɑ ‘bridge-SUBL’, hiːd-eː-rɑ ‘bridge-POSS-SUBL’ b. viːz-rɛ ‘water-SUBL’, viːz-eː-rɛ ‘water-POSS-SUBL’ c. in-nɑ ‘drink-3SG.COND’, in-ni-jɑ ‘drink-INF-3SG’ d. vin-nɛ ‘take-3SG.COND’, vin-ni-jɛ ‘take-INF-3SG’ e. paːriʒ-rɑ ‘Paris-SUBL’, paːriʒ-i-rɑ ‘Paris-ADJ-SUBL’ f. hɑmiʃ-ɑk ‘fake-PL’, hɑmiʃ-iːt-ok ‘fake-VERBAL-1SG’ g. kolibri-rɑ/rɛ ‘hummingbird-SUBL’, kolibri-eː-rɑ/rɛ ‘hummingbird-POSS-SUBL’ h. hotɛl-rɑ/rɛ ‘hotel-SUBL’, hotɛl-eː-rɑ/rɛ ‘hotel-POSS-SUBL’ i. haːrɛm-rɛ ‘harem-SUBL’, haːrɛm-eː-rɛ ‘harem-POSS-SUBL’ In (6a) we see that the bisyllabic stem hiːdeː is antiharmonic. We have claimed above that apart from two bisyllabic roots, all antiharmonic roots consist of a sin- gle syllable, this was referred to as the polysyllabic split. The examples above do not violate the polysyllabic split, since hiːdeː is not a single morpheme. The stem hiːdeː takes the back variant of suffixes because its root, hiːd, also takes the back variant of suffixes. This situation is perfectly regular in Hungarian for both nouns, like hiːd, and verbs, like iC-8 ‘drink’ as shown in (6c). If a root selects the front variant of suffixes, like viːz in (6b) and viC- ‘take’ in (6d), then the suffixed forms of these roots also select the front variant of suffixes. The bisyllabic roots paːriʒ, in (6e), and hɑmiʃ, in (6f), take the back variant, as almost all Bi roots do. When suffixed by the invariant neutral-vowelled adjectival suffix -i or verbal suffix -iːt, the harmonic class remains the same: paːriʒi ‘Parisian’ and hɑmiʃiːt ‘forge’ are not variable in their harmonic choice, like other monomorphemic Bii roots would be. As (6g) shows the vacillating Bii root, kolibiri remains vacillating after the addition of a neutral suffix, unlike a Biii or Biieː root, which is much more likely to attract a front suffix. Similarly, the vacillating root of (6h) remains vacillating with the invariant neutral-vowelled possessive suffix, while the nonvacillating root of (6i) remains nonvacillating. We call this phenomenon harmonic uniformity (Rebrus & Törkenczy 2015). 8. The stem of ‘drink’ and ‘take’ contains a variable consonant: is-ok ‘drink-1SG’, in-neːk ‘drink- 1SG.COND’, it-tɑm ‘drink-1SG.PAST’, iɟ-ɑk ‘drink-1SG.IMP’, iv-oː ‘drink-PART’. We mark this consonant by C. There are also forms in which this consonant is missing, e.g. i-hɑt ‘may drink’. Diminutives: Exceptions to Harmonic Uniformity CatJL 15, 2016 107 Harmonic uniformity is very pervasive in Hungarian vowel harmony. The bound root ind- ‘start’ is back harmonic (cf. ind-ul ‘start-MEDIAL’) and preserves this property irrespective of how many neutral (i.e. front-vowelled) suffixes are added: ind-iːt-vɑ ‘start-ACTIVE-PART’, ind-iːt-eːk-rɑ ‘motive-SUBL’, ind-iːt-eːk- eː-rɑ ‘motive-POSS-SUBL’ and ind-iːt-eːk-eː-i-rɑ ‘motive-POSS-PL-SUBL’. We do not observe any variation at all in these cases. The word indiːteːkeːi contains five phonetically front vowels and yet is invariably back harmonic. This is because its root, ind-, is antiharmonic, hence it will take the back variant of any variable suffix that follows it. Compare this with the front harmonic diːs ‘ornament’, where after the same neutral suffixes, we find the front variant of variable suffixes: diːs-iːt-eːk- eː-i-rɛ ‘decoration-POSS-PL-SUBL’. 3. Truncation Hungarian morphology is mostly concatenative (for its limits, the flectional phe- nomena in the verbal paradigm, see Rebrus 2005). The alternations across mor- pheme boundaries involve voicing and other types of assimilations, as well as trun- cation9 of the stem, or in some cases the suffix. Stem truncation usually involves the loss of the last vowel of the stem, which may be word-final or not. Representative examples are given in (7). (7) Truncation of the last vowel of the stem a. bɑrnɑ ‘brown’, bɑrn-ul ‘become brown’, bɑrn-iːt ‘make brown’ b. fɛrdɛ ‘oblique’, fɛrd-yl ‘become oblique’, fɛrd-iːt ‘make oblique’ c. domboru(ː) ‘convex’, dombor-od-ik ‘become convex-3SG’, dombor-iːt ‘make convex’ d. kɛʃɛry(ː) ‘bitter’, kɛʃɛr-ɛd-ik ‘become bitter-3SG’, kɛʃɛr-iːt ‘make bitter’ e. fɑkoː ‘pale’, fɑk-ul ‘become pale’, fɑk-iːt ‘make pale’ f. torok ‘throat’, tork-ɑ ‘throat-POSS3SG’, tork-oʃ ‘gluttonous’ g. ʃøpør ‘sweep’, ʃøpr-ynk ‘sweep-1PL’, ʃøpr-ød ‘sweep-2SG.DEF’ The words in (7a-e) illustrate the loss of word-final vowels. A truncated word- final vowel is typically ɑ or ɛ, as in (7a, b), but the truncation of u(ː), y(ː), and oː also occurs as shown in (7c, d, e). Truncated stems must be at least two syllables 9. Anonymous reviewers suggest that the deletion of a single vowel should not be referred to as truncation, this term should be reserved for the deletion of longer portions of the base. As we will see, diminutive truncation also often involves the deletion of a single vowel. Using different terms for the two types only because they have a different effect on the harmonic properties of the stems seems to be begging the question. So following Alber & Arndt-Lappe (2012) we use the term truncation for both types. Diminutive truncation is templatic, all other kinds of truncation in Hungarian are subtractive. 108 CatJL 15, 2016 Péter Rebrus; Péter Szigetvári long (but they might be longer, (7c, d)), the only vowel of a word is never lost. Word-internal truncation is also only possible in words that are at least two syl- lables long and end in -VCVC. The vowel that is lost in word-internal truncation is always short and predominantly ɛ, ø, or o, although any of the other short vowels also occurs in one or two examples each (Siptár & Törkenczy 2000: 215). Truncated roots provide further evidence for harmonic uniformity. Bisyllabic roots containing a neutral vowel followed by a back vowel are back harmonic, (8a, c), those containing a neutral vowel followed by a front vowel are front harmonic, (8b, d, e), as expected. (8) Neutral+back and neutral+front stems a. beːnɑ ‘lame’, beːnɑ-ʃaːg ‘lame-ness’ b. beːkɛ ‘peace’, beːkɛ-ʃːeːg ‘tranquility’ c. pisok ‘dirt’, pisok-rɑ ‘dirt-SUBL’ d. ikɛr ‘twin’, ikɛr-rɛ ‘twin-SUBL’ e. eːbɛr ‘alert’, eːbɛr-rɛ ‘alert-SUBL’ In the roots in (8) and their ilk it is the final vowel that explicitly shows the harmonic class they belong to. If this vowel is truncated there remains no phonetic representative of the harmonic class, i.e. the harmonic class of beːn- and beːk- cannot be read off their phonetic shape. Harmonic uniformity predicts, however, that their harmonic class does not change, and this is indeed so, as the examples in (9) show. (9) Truncated neutral+back stems a. beːn-iːt-unk ‘paralyse-1PL’ b. beːk-iːt-ynk ‘make peace-1PL’ c. pisk-iːt-unk ‘make dirty-1PL’ d. ikr-ɛk ‘twin-PL’ e. eːbr-ɛn ‘awake’ Roots whose back vowel is truncated leaving only a neutral vowel behind look and behave like monosyllabic antiharmonic roots, as shown in (10). (10) Comparison of truncated and antiharmonic roots a. ʃimɑ ‘smooth’, ʃim-ul ‘become smooth’, ʃim-iːt-unk ‘make smooth-1PL’ b. hiːd ‘bridge’, hid-ɑk ‘bridge-Pl’, hiːd-eː-rɑ ‘bridge-POSS-SUBL’ After this brief introduction of those properties of front/back harmony that are relevant to our discussion, we proceed to a survey of types of diminutive formation in Hungarian. Diminutives: Exceptions to Harmonic Uniformity CatJL 15, 2016 109 4. Diminutives Like in many other languages, there are several ways of forming diminutives10 in Hungarian and many types of diminutive forms are templatic (cf. Alber & Arndt- Lappe 2012). The diminutive suffix -(V)ʧkɛ/-(V)ʧkɑ is productive and has no tem- platic requirements, it can be added at the end of nominals of any length, the vowel before the suffix appears after consonant-final stems:11 e.g. gnuː ‘gnu’ ~ gnuː-ʧkɑ, tɛvɛ ‘camel’ ~ tɛveː-ʧkɛ, kɛnguru ‘kangaroo’ ~ kɛnguru-ʧkɑ, ɛlɛfaːnt ‘elephant’ ~ ɛlɛfaːnt-oʧkɑ, ʒiraːf ‘giraffe’ ~ ʒiraːf-oʧkɑ, mɑjom ‘monkey’ ~ mɑjm-oʧkɑ. This diminutive suffix follows patterns that many other suffixes do in Hungarian, including the lengthening of the stem-final ɛ (and ɑ) or the truncation of the last vowel. Compare the same stems combined with the plural suffix: gnuː-k, tɛveː-k, kɛnguru-k, ɛlɛfaːnt-ok, ʒiraːf-ok, mɑjm-ok. Another diminutive suffix, -kɛ/-kɑ is less free in its distribution: it cannot be productively added to monosyllables or to stems ending in ɑ or ɛ/eː. So we can have e.g. ɛmbɛr ‘man’ ~ ɛmbɛr-kɛ, dɑrɑb ‘piece’ ~ dɑrɑb-kɑ (in fact, ɛmbɛr-ɛʧkɛ, dɑrɑb-oʧkɑ are also possible, if rarer diminutive forms), but not ɑlmɑ ‘apple’ ~ *ɑlmaːkɑ (only ɑlmaːʧkɑ) or keːz ‘hand’ ~ *keːz-kɛ (only kɛzɛʧkɛ). Monosyllables must either take the unrestricted -(V)ʧkɛ/-(V)ʧkɑ suffix or augment the stem: søːr ‘(body) hair’ ~ søːr-øʧkɛ (or søːri-kɛ, with an augment to be long enough), rɑb ‘prisoner’ ~ rɑb-oʧkɑ, paːl ~ paːl-oʧkɑ (or pɑli-kɑ).12 4.1. Monosyllabic diminutives Other diminutive forms are more restrictively templatic: they set the exact size of the output. The most common diminutive template is bisyllabic, but there seems to exist a monosyllabic and a trisyllabic template too. Benua (1995) shows sev- eral cases of truncated diminutives (and other categories) that violate phonotactic constraints regulating other forms of the given language. Truncated diminutives of Hungarian do not show many signs of such misbehaviour, in fact they exhibit the harmonic behaviour of “normal” monomorphemic stems, unlike other morphologi- cally complex stems, which inherit the harmonic properties of their root. 10. In this paper we do not distinguish between diminutive and hypocoristic forms, since with a few exceptional cases, the two types of forms can both be produced applying the mechanisms described in this section. 11. We are not going to give a gloss for diminutive forms. Any other unglossed form is either a name, or has been glossed earlier. 12. There are archaic diminutives that do not obey this requirement, e.g. bøːr ‘skin’, bøːr-kɛ ‘bacon skin’, lɑp ‘sheet’ ~ lɑp-kɑ ‘flan/chip’, saːl ‘thread’ ~ saːl-kɑ ‘splinter’, huːr ‘string, intestine’ ~ hur-kɑ ‘black pudding’. This pattern is not productive, and the meaning relationship between the stem and its “diminutive” form is not clear. The regular, semantically transparent diminutive forms are bøːr ~ bøːr-øʧkɛ ‘skin-DIM’, lɑp ~ lɑp-oʧkɑ ‘sheet-DIM’, saːl ~ saːl-ɑʧkɑ ‘thread-DIM’, huːr ~ hu:r-oʧkɑ ‘string-DIM’. Another set of archaic diminutives truncate the stem to a bisyllabic tem- plate: bori ~ bor-kɑ, pɑnːi ~ pɑn-kɑ, ɟuri ~ ɟur-kɑ, judit ~ jud-kɑ, dorocːɑ ~ dor-kɑ. This truncation is not obligatory: bori-kɑ, pɑnːi-kɑ, ɟuri-kɑ, juʦi-kɑ, doːri-kɑ. Truncation is also possible at the end of longer forms: ilonɑ ~ ilon-kɑ, vɛronikɑ ~ vɛron-kɑ. (See Kiefer & Ladányi 2000 for a somewhat different account of this suffix.) 110 CatJL 15, 2016 Péter Rebrus; Péter Szigetvári There aren’t very many monosyllabic diminutives and they are mostly phatic ele- ments: e.g. køsønøm ‘(I) thank you’ ~ køs(ː) (also bisyllabic køsi), boʧaːnɑt ‘sorry’ ~ boʧ (also bisyllabic boʧi, boʧɛs), ɛgeːʃːeːgɛdrɛ reaction if someone sneezes ~ ɛgʃ, okʃi ‘okay-DIM’ ~ okʃ, ʧa(ː)oː greeting ~ ʧaː. There are also nonphatic words with a monosyllabic diminutive: e.g. profɛsːor ‘professor’ ~ prof(ː), ʒuʒɑ/ʒuʒɑnːɑ ~ ʒu. Nevertheless, this type of diminutive formation seems rather marginal in Hungarian, compared to, for example, English (cf. Sue, Joe, Vic, Bill, Kate, doc, vet, lab, bro). 4.2. Bisyllabic diminutives While monosyllabic diminutives have no fixed segments, the bisyllabic template ends in a fixed, albeit very variable set of endings: -u, -uʃ, -oː, -koː, -ʦoː, -os, -ɑ, -ʦɑ, -ʧɑ, -i, -iʃ, -si, -ʃi, -ʦi, -ʧi, -ɛk, -ɛs, -ːɛr13 (the last ending involves the gemination of the preceding consonant). Because of the size of the template, any word which is longer than a syllable must be truncated to give way for the ending of the template, which always includes a vowel. Using our native speaker’s competence, we have collected representative examples in (11), arranged by the vowel they contain. (11) Some examples of the bisyllabic diminutive template i. u-diminutives a. geːzɑ ~ geːz-u, ɑpɑ ‘father’ ~ ɑp-u, fizɛteːʃ ‘salary’ ~ fiz-u b. tɛreːziɑ ~ tɛr-uʃ, pɛlɛnkɑ ‘diaper’ ~ pɛl-uʃ, ɟɛngɛ ‘weak’ ~ ɟɛng-uʃ ii. o-diminutives c. kɑtɑlin ~ kɑt-oː, tɛʃtveːr ‘brother/sister’ ~ tɛʃ-oː, sɛndviʧ ‘sandwich’ ~ sɛnd-oː d. jaːnoʃ ~ jɑn-koː, tɛtovaːlaːʃ ‘tattoo’ ~ tɛt-koː, fɛsyltʃeːg ‘tension’ ~ fɛs-koː e. fɛrɛnʦ ~ fɛ-ʦoː, judit ~ ju-ʦoː, keːgli ‘flat’ ~ kɛ-ʦoː f. bioloːgiɑ ‘biology’ ~ bi-os, mɑʧkɑ ‘cat’ ~ mɑʧk-os, filoloːguʃ ‘philologist’ or filozoːfuʃ ‘philosopher’ ~ fil-os iii. ɑ-diminutives g. ɛtɛlkɑ ~ ɛt-ɑ, dolgozɑt ‘in-class test’ ~ dog-ɑ, fɛkɛtɛ ‘black’ ~ fɛk-ɑ ‘African American’ h. ilonɑ ~ i-ʦɑ, eːvi ~ vi-ʦɑ, laːsloː ~ lɑ-ʦɑ i. maːria ~ mɑr-ʧɑ, borbaːla ~ bor-ʧɑ, ɑnːɑ ~ ɑn-ʧɑ 13. Some of these diminutive endings are not productive, one example is -øː, which can only be added to front stems: bɛnɛdɛk ~ bɛnøː, peːtɛr ~ pɛtøː, gɛrgɛj ~ gɛr(g)øː. Historically it must have been the front variant of the still productive diminutive ending -oː. Similarly, we find dømøtør ~ dømɛ, domonkoʃ ~ domɑ, i.e. harmonizing ɛ/ɑ, but of this pair, it is again only the back variant that is still productive. Diminutives: Exceptions to Harmonic Uniformity CatJL 15, 2016 111 iv. i-diminutives j. zoltaːn ~ zol-i, tørteːnɛlɛm ‘history’ ~ tør-i, huːʃ ‘meat’ ~ huʃ-i, undoriːtoː ‘disgusting’ ~ und-i k. juːliɑ ~ jul-iʃ, borbaːlɑ ~ bor-iʃ, ɑndraːʃ ~ ɑndr-iʃ, ɑtilːɑ ~ ɑt-iʃ, aːbrɑhaːm ~ aːbr-iʃ l. ɟulɑ ~ ɟu-si, ɲuːl ‘rabbit’ ~ ɲu-si, ʧɑj ‘gal’ ~ ʧɑj-si m. mihaːj ~ mi-ʃi, ɑk(ː)um(u)laːtor ‘battery’ ~ ɑk-ʃi, okeː ‘okay’ ~ ok-ʃi n. jɛnøː ~ jɛn-ʦi, fudbɑl (earlier fodbɑl) ‘football’ ~ fo-ʦi, nøː ‘woman’ ~ nøː-ʦi o. lɑjos ~ lɑj-ʧi, puloːvɛr ‘pullover’ ~ pul-ʧi, finom ‘delicious’ ~ fin-ʧi, kiraːj ‘cool’ ~ kir-ʧi, unɑlmɑʃ ‘boring’ ~ un-ʧi v. ɛ-diminutives p. mɑtɛmɑtikɑ ‘mathematics’ ~ mɑt-ɛk, ʃɑpkɑ ‘cap’ ~ ʃɑp-ɛk, hɑpʃi ‘chap’ ~ hɑp-ɛk q. ʧɑbɑ ~ ʧɑb-ɛs, ɑlkoholiʃtɑ ‘alcoholic’ ~ ɑlk-ɛs, kol(ː)eːgium ‘dorm’ ~ kol-ɛs r. kɑlɑuz ~ kɑl-ːɛr ‘conductor’, kom(ː)uniʃtɑ ‘communist’ ~ kom-ːɛr, ɲugdiːjɑʃ ‘pensioner’ ~ ɲug-ːɛr We can see that there are many diminutive endings. Some stems may take more than one of these endings (e.g. tɛreːziɑ may be tɛruʃ, tɛri, tɛʦɑ, or tɛrʧi, fɛrɛnʦ may be fɛroː, fɛrkoː, fɛʦoː, fɛri, or fɛrʧi), but many stems only take one of them and in most cases it is unpredictable which one. In some cases we may detect a tendency to avoid homonymy, for example, rɛprɛzɛntaːʦioːʃ ɑjaːndeːk ‘hospitality gift’ ~ rɛpi ɑjaːndeːk vs rɛpyløː ‘airplane’ ~ rɛpʧi, tɛʃtveːr ‘brother/sister’ ~ tɛʃoː vs tɛʃtnɛvɛleːʃ ‘PE’ ~ tɛʃi. Moreover, there are cases where the stems are homonymous and their diminutive forms are different: lokomotiːv ~ loki name of a football team (DVSC) or lokʃi name of a rock band (Locomotiv GT). In other cases the diminutive forms are homonymous, e.g. sɛɲɑ can be a diminutive of sɛndviʧ ‘sandwich’ or sɛmeːt ‘bastard’ or unʧi of unɑlmɑʃ ‘boring’ or undori(ː)toː ‘disgusting’, piɲoː of pimpong ‘pingpong’ or pinʦɛ ‘cellar’, gɑbi of both gaːbor and gɑbrielːɑ, bɛni may be a diminutive of bɛnʦɛ, bɛnɛdɛk, bɛnjaːmin, or bɛndɛguːz. We also cannot fully predict the extent of truncation in these templatic dimin- utive forms. The endings in (11) all involve a vowel, so in order to fit the bisyl- labic template the second vowel of the word and everything that follows it has to be deleted. However, how much of the consonantal interlude between the first and the second vowel remains is only loosely governed by rules. Consonant plus liquid clusters are often simplified before the most common ending, -i: pɑtriːʦiɑ ~ pɑti, hɛnriɛtːɑ ~ hɛni, mikloːʃ ~ miki, kuplɛraːj ‘brothel’ ~ kupi, imrɛ ~ imi, 112 CatJL 15, 2016 Péter Rebrus; Péter Szigetvári but imruʃ, ɑmbruʃ ~ ɑmbi, but ɑmbroː, ɑlɛksɑndrɑ ~ sɑndi, but sɑndrɑ.14 There is variation in the case of other clusters: e.g. bɑrbɑrɑ ~ bɑrbi vs borbaːlɑ ~ bori (not *borbi), dolgozɑt ‘in-class test’ ~ dogɑ or doli (not *dolga or *dolgi and not *dolɑ or *dogi) vs olgɑ ~ olgi, borju(ː) ‘calf’ ~ boʦi (not *borʦi) vs maːrton ~ mɑrʦi, zoltaːn is usually zoli (rarely zolti), but ʒolt is usually ʒolti and only rarely ʒoli. The single vowel endings of the bisyllabic template are often preceded by a palatal consonant: piʃti ~ picu, peːtɛr ~ pɛcɑ, zoltaːn ~ zocɑ, paːl ~ pɑjɑ, sɛndviʧ ‘sandwich’ or sɛmeːt ‘bastard’ ~ sɛɲɑ, pinʦɛ ‘cellar’ ~ piɲoː, vinʧɛstɛr ‘hard disk’ ~ viɲoː, bɛnzin ‘gasoline’ ~ bɛɲɑ, bɛnɛdɛk ~ bɛɲuʃ, mɛrʦeːdɛs car type ~ mɛrʤoː. If the stem had a palatal consonant, it remains before -i (e.g. maːcaːʃ ~ mɑci, løcː ‘tasteless drink’ ~ løci, nɑɟmɑmɑ ‘grandmother’ ~ nɑɟi, boɲolult ‘complicated’ ~ boɲi), but unlike before the other single vowel endings, replacing the consonant by a palatal is not common before -i, although some examples occur: e.g. ʃaːndor ~ ʃɑɲi,15 litɛr ‘litre’ ~ lici. The length of the root-initial vowel is also rather unpredictable. We find many examples where this vowel shortens: paːl ~ pɑli or pɑlkoː, peːtɛr ~ pɛti or pɛcɑ, juːliɑ ~ juli or juloː, liːdiɑ ~ lidi, tiːmɛɑ ~ timi, loːvɛrʃɛɲ ‘horse race’ ~ lovi, kaːroj ~ kɑrɛs. In other cases we do not observe shortening: joːʒɛf ~ joːʒi (but joʦoː or joʒoː), loːraːnd ~ loːri, klaːrɑ ~ klaːri, beːlɑ ~ beːʦi, eːvɑ ~ eːvi, paːlinkɑ ‘brandy’ ~ paːlɛs. In fact, there also are a few cases where the vowel seems to lengthen: hɛdvig ~ heːdi, ʃpɛʦiaːliʃ ‘special’ ~ ʃpeːʦi. Following van de Weijer (1989) we suspect that in these cases the diminutive forms are loans from German. Another apparent case of lengthening is ʃɑroltɑ ~ ʃaːri, but this form is also the diminutive of ʃaːrɑ, which contains a long vowel. To fit the bisyllabic template some words are truncated at their beginning, too. This is more common for, but not limited to vowel-initial words. Examples are given in (12). (12) Initial truncation a. ɑlbɛrt ~ bɛrʦi, ɑlfreːd ~ freːdi, ɑlɛksɑndrɑ ~ sɑndi or lɛksi,16 ɑnːuʃ ~ nuʃi, ɑliːz ~ lizi, ɑmbruʃ ~ bruʃi b. ɛmɛʃɛ ~ mɛʃi, ɛdinɑ ~ dinɑ, ɛmaːnuɛl ~ mɑnoː, ɛlɛonoːrɑ ~ noːrɑ, eːvi ~ viʦɑ c. ignaːʦ ~ naːʦi, ilonɑ ~ lonʦi, imolɑ ~ moluʃ d. ødøn ~ dønʦi e. ʦɛʦiːliɑ ~ ʦili, fɛrdinaːnd ~ naːndi, bɛrnɑdɛtː ~ dɛtːi, ʃtɛfaːniɑ ~ faːni, mɑriʃkɑ ~ riʃkɑ, mɑrgit ~ gitːɑ, brigitːɑ ~ gitːɑ 14. Van de Weijer (1989) contends that this simplification always occurs when the cluster is not a possible “coda”. This is not the case, there are several counterexamples: e.g. ɑdriɛn ~ ɑdri, bodroʃ ‘frilly’ ~ bodri dog’s name, ugroː ‘jumping’ ~ ugri rabbit’s attribute. 15. It may be the case that ʃɑɲi is from ʃɑɲɑ, which itself is a diminutive of ʃaːndor and in which the palatal before -ɑ fits an attested pattern. 16. The male name ʃaːndor itself is a clipped form of ɑlɛksɑndɛr. Diminutives: Exceptions to Harmonic Uniformity CatJL 15, 2016 113 Besides initial truncation, the adding of phonetic material also occurs word initially. This is limited to a few vowel initial names, some of which are shown in (13). (13) Labial prefixation a. ɑnːɑ ~ pɑnːɑ, iʃti ~ piʃti b. ɑndraːʃ/ɑndor ~ bɑndi, ɛrʒeːbɛt ~ bøʒi In fact, this type reflects a common reduplication pattern in which a word is repeated with the insertion of a labial consonant or the replacement of its first consonant by a labial: e.g. eːdi (< eːdɛʃ ‘cute’) ~ eːdibeːdi, ʧigɑ ‘snail’ ~ ʧigɑbigɑ, ʦiʦɑ ‘cat’ ~ ʦiʦɑmiʦɑ. For labial-initial stems the form is repeated at the begin- ning without the initial labial: piʦi ‘small’ ~ iʦipiʦi, finʧi (< finom ‘delicious’) ~ inʧifinʧi, pirul ‘blush’ ~ irulpirul. In some cases this is also coupled by vowel change, front vowels in the first half, back vowels in the second half: e.g. mozog ‘move’ ~ izegmozog, zɛnɛ ‘music’ ~ zɛnɛbonɑ17 (cf. Sóskuthy 2012 and Patay 2015 for detailed surveys and analyses). In each case the second half begins with a labial, irrespective of whether it is the base or the reduplicant. Diminutives are not restricted to nouns and adjectives. Verbs, even clauses may have diminutive forms. Examples are given in (14). (14) Verbal diminutives a. mutɑʃ-d ‘show-2SG.DEF.IMP’ ~ muti/mutɑ, fiɟɛl-j ‘listen’ ~ fiɟi/fiɟu b. leːɟ sivɛʃ ‘be (so) kind (=please)’ ~ leːɟsi, ɑd(ː) idɛ ‘give-2SG.DEF.IMP here’ ~ ɑdi Yet another type of diminutive formation occurs only for given names, exception- ally for family names. This involves the reduplication of the first CV of the name, as the examples in (15a) show. The pairs in (15b) are phonologically less transparent diminutive forms, while (15c) shows family names exhibiting this process. (15) Reduplicated diminutives a. zoltaːn ~ zozoː, moːnikɑ ~ momoː, peːtɛr ~ pɛpɛ, vɛronikɑ ~ vɛvɛ, lɑurɑ/ lɑjoʃ ~ lɑlɑ, silaːrd/silviɑ ~ sisi, lujzɑ ~ lulu, zitɑ ~ zizi, aːgi ~ gigi b. ʒoːfiɑ ~ fifi, joːʒɛf ~ dodoː c. kovaːʧ ~ kokoː, sɑboː ~ sɑsɑ Some cases are ambiguous in their choice of the diminutive formation process, i.e. we cannot decide whether they exemplify truncation plus the ending i or redu- 17. Finally, in some cases only the vowels differ: mondɑ ‘legend’ ~ mɛndɛmondɑ, gɑz ‘weed’ ~ gizgɑz, gɑzoʃ ‘weedy’ ~ gizɛʃgɑzoʃ. 114 CatJL 15, 2016 Péter Rebrus; Péter Szigetvári plication, but we see no reason why the choice should be made anyway: eg lilːɑ ~ lili, viviɛn ~ vivi. We end this survey of bisyllabic diminutive forms by listing further examples that are totally idiosyncratic, (16a), involve hapax endings, (16b), metathesis,18 (16c), loan stems, (16d), or diminutives that do not have a phonetically similar nondiminutive form, (16e). (16) Idiosyncratic diminutives a. ɟørɟ ~ ɟuri, mɑrgit ~ mɑnʦi, løːrinʦ ~ loːʦi b. lɑbdɑ ‘ball’ ~ lɑsti,19 taːʃkɑ ‘bag’ ~ tɑcoː, diskoː ‘discotheque’ ~ diʒi, biʦikli ‘bicycle’ ~ biʦɑj, rɛpyløː ‘airplane’ ~ rɛpzɑj, klɑsː ‘cool’ ~ klɑfɑ or kɑfɑ c. forint ‘HUF’ ~ fronʧi, bɑlaːʒ ~ bɑʒi d. iʃkolɑ ‘school’ ~ ʃuli (< German ʃuːlə), fɛrɛnʦvaːroʃ a sports club ~ frɑdi (< German franʦʃtat Ferencváros, district of Budapest), baːc ‘elder brother’ ~ brɑcoː (< Slovak brat), mɛdvɛ ‘bear’ ~ mɑʦkoː (< Slovak maʦko, dimin- utive of matej), orː ‘nose’ ~ noːzi (< Yiddish noz ‘nose’) e. piʦi ‘small’, rɛcoː ‘loo’, duci ‘jail’, sycøː ‘pouch’, ʧɑʧi ‘donkey’, goːɟi ‘brains’, hɑjʧi ‘sleep’, buli ‘party’, pɑʧi ‘high five’, ʦunʦi ‘cunt’, bukʃi ‘head’, muki or mukʃoː ‘pal’. 4.3. Trisyllabic diminutives Most templatic diminutives are bisyllabic, a very little set is monosyllabic. There also is a group of diminutive forms that are three syllables long. These we list in (17). (17) The trisyllabic diminutive template a. kucɑ ‘dog’ ~ kuculi, fɑrok ‘tail’ ~ fɑrkinʦɑ b. piʦi ‘small’ ~ piʦuri, ? ~ pinduri ‘small’ c. ɑpɑ ‘father’ ~ ɑpuʦi, ɑɲɑ ‘mother’ ~ ɑɲuʦi, bɑbɑ ‘baby’ ~ bɑbuʦi d. aːɟ ‘bed’ ~ aːɟikoː, haːz ‘house’ ~ haːzikoː, laːdɑ ‘box’ ~ laːdikoː, laːb ‘leg’ ~ laːbikoː, hɑʃ ‘belly’ ~ hɑʃikoː, ɑnːɑ(?) ~ ɑnikoː For the patterns in (17a) we have found single examples. The word pɑʧuli ‘patchouli’, which is not a diminutive etymologically, also has sniffy connota- tions, probably due to its sound shape. Similarly, the tiny muʃliʦɑ ‘fruit fly’ is 18. In fact, bɑʒi may also be seen as an example of contiguity violation, the deletion of -laː- in the middle of the string. 19. Apparently this is a diminutive of ɛlɑstik, a brand name. Diminutives: Exceptions to Harmonic Uniformity CatJL 15, 2016 115 often muʃlinʦɑ. The first two examples in (17c) could be analysed as ɑpɑ > ɑpu > ɑpuʦi, but the ending -ʦi enforces a bisyllabic template in all other cases (cf. jɛnø ~ jɛnʦi, this is a possibility even for ɑpɑ ~ ɑpʦi). Finally, the ending -ikoː in (17d) apparently can only be added to stems containing ɑ or aː. If this stem is longer than one syllable, it is truncated, hence -ikoː too enforces a trisyllabic template. 5. Diminutives and their roots If we compare the phonetic form of a diminutive and its root we can observe that on the whole they are much more different from each other than other types of words and forms derived from them. In some cases diminutive forms happen to match the template by the simple addition of an ending (e.g. ʃyn ‘hedgehog’ ~ ʃyni, haːz ‘house’ ~ haːzikoː, ʧɑj ‘gal’ ~ ʧɑjsi, jaːnoʃ ~ jaːnoʃkɑ, fyløp ~ fyløpkɛ). In the majority of the cases, however, diminutive formation involves the loss of a smaller or larger portion of the base word, and it may also involve metathesis and other highly idiosyncratic phenomena, as discussed in the preceding section. Thus, as a category, diminutives are the least similar to their base in the whole system of Hungarian morphology. If we look at the semantic relationship between a word and its diminutive form, we also find that this relationship is much looser than in the case of most other suffixations. Diminutives often have strictly defined usages that cannot be derived from their form. The diminutive ɛgʃ for ɛgeːʃːeːgɛdrɛ can only be used as a reaction to someone sneezing, although the “full form” is also used before drinking, like ‘cheers’ (the compositional meaning of the word is ‘to your health’). The diminu- tives tøri ‘history.DIM’, føʦi ‘geography.DIM’, or bios ‘biology.DIM’ can only be used as school subjects, and we have seen the split in the meanings of loki and lokʃi above. Coupled with the fact that the selection of the dozens of different diminutive endings, the shortening or absence thereof of the first vowel of the stem, the simpli- fication or not of the consonantal interlude between the two vowels, the palatality of the consonant before the diminutive ending are largely unpredictable, we must conclude that diminutive forms are lexical items that are not phonologically related to their base word, i.e. they are suppletive forms. To sum up, we collect the differences of “regular” suffixation and templatic diminutive suffixation in a table in (18).20 20. The concatenative diminutive suffixes -kɛ/-kɑ and -(V)ʧkɛ/-(V)ʧkɑ are counted as regular as dis- cussed above. 116 CatJL 15, 2016 Péter Rebrus; Péter Szigetvári (18) Differences of diminutive and non-diminutive suffixation regular suffixation diminutive suffixation suffixation mainly agglutinative mostly two syllable template root alternations one vowel may be deleted radical: longer sequences can be deleted productivity mainly productive semi-productive: gaps cannot be predicted shape of the suffixed determined not determined which dim. forms suffix will be applied & random palatal/labial C insertion V harmony in suffix harmonizing not harmonized at all (weak tendency for counter-harmony) link between root semantically transparent weak: special pragmatics and suffixed form & referential differences for names Our claim that the connection between a diminutive form and its “base” is looser than in the case of other types of suffixation is further corroborated by the way diminutives trigger front/back harmony. This is what we discuss now. 6. Diminutives and harmony The endings of diminutive forms are unlike suffixes in that they do not have front/ back variants. This is not surprising in the case of endings containing i (and there are no diminutive endings containing eː), but it is unusual for those with a back vowel (u, oː, ɑ) or ɛ. These vowels do not occur in any other invariant suffix. Examples are given in (19). (19) Diminutive endings do not harmonize a. ilonɑ ~ iʦuʃ, tyndɛ ~ tynduʃ, ɛtɛlkɑ ~ ɛtuʃ, ɑnːɑ ~ ɑnːuʃ b. biʦikli ‘bicycle’ ~ biʦoː, tɛʃtveːr ‘brother/sister’ ~ tɛʃoː c. fɛkɛtɛ ‘black’ ~ fɛkɑ, kɑtɑlin ~ kɑtɑ d. kol(ː)eːgium ‘dorm’ ~ kolɛs, kaːroj ~ kɑrɛs, kɑlɑuz ‘conductor’ ~ kɑlːɛr e. kɑtɑlin ~ kɑtoː, tɛlɛfon ‘telephone’ ~ tɛloː Diminutives: Exceptions to Harmonic Uniformity CatJL 15, 2016 117 Although there exist the archaic diminutive endings -øː beside -oː and -ɛ beside -ɑ (cf. footnote 13), they are not used productively: tɛlɛfon ‘phone’ gives tɛloː (not *tɛløː, cf. sɛl ‘cut’ ~ sɛl-øː ‘cut-PART’ vs fɑl ‘devour’ ~ fɑl-oː ‘devour-PART’) and ɛtɛlkɑ ~ ɛtɑ (not *ɛtɛ). It seems that truncating diminutive endings are more independent of vowel harmony than other suffixes. Although neither -oː, nor -ɑ is added to a stem containing a front rounded vowel (which are among the less frequent members of the vowel inventory anyway), almost any front vowel may exhibit antiharmonic behaviour with diminutive endings. The fact that diminutive endings do not harmonize is relevant, because oth- erwise suffixes in Hungarian either alternate according to front/back harmony or contain a neutral vowel, i, iː, or eː—as has been mentioned above, the low neutral vowel, ɛ, does not occur in invariant suffixes. There do exist some endings that are invariant and contain nonneutral vowels, but these can be shown not to be suf- fixes, because they can be added to coordinate structures: e.g. -feːlɛ ‘kind of’ (fɑ vɑɟ bokorfeːlɛ ‘kind of tree or shrub’), -sɛry(ː) ‘type’ (fɑ eːʃ bokorsɛry(ː) ‘tree and shrub-like’), -kor (%neːɟ vɑɟ øtkor ‘at four or five’).21 Since many of the diminutive endings contain nonneutral vowels and yet are invariant, they clearly stick out of the system of suffixes. The two suffixes that show synchronic harmonic alterna- tion, -kɛ/kɑ and -(V)ʧkɛ/(V)ʧkɑ, are exactly the ones that do not squeeze their output into a template, hence do not truncate it, and are concatenated just like any “normal” suffix of Hungarian. Another indication that diminutive forms do not contain suffixes is the har- monic properties of these items. Harmonic uniformity does not apply to diminutive forms, as the examples in (20) show. (20) Diminutives defy harmonic uniformity a. ʃimaː-rɑ ‘smooth-SUBL’ ~ ʃim-iːt-vɑ ‘make smooth-PART’ b. ʃimon-rɑ ‘Simon-SUBL’ ~ ʃimi-rɛ ‘SimonDIM-SUBL’ c. joːʒɛf-rɛ ‘Joseph-SUBL’ ~ joːʒi-rɑ ‘JosephDIM-SUBL’22 d. kaːroj-rɑ ‘Charles-SUBL’ ~ kɑrɛs-rɑ/rɛ ‘CharlesDIM-SUBL’ e. tɛstveːr-rɛ ‘brother/sister-SUBL’ ~ tɛʃoː-rɑ ‘brother/sisterDIM-SUBL’ The words in (20a) are a reminder, they show the effect of harmonic uniform- ity: the participial suffix takes its back variant, -vɑ, because the root, ʃimɑ, which turns into ʃim- after truncation, is back harmonic. After any number of neutral suf- fixes following this root a variable suffix attached to a word containing this root will be back. The name ʃimon in (20b) is also back harmonic, as the untruncated form shows. After diminutive truncation, however, ʃimi becomes front harmonic. In (20c) we see a root that is front-harmonic because of the height effect: the back 21. Some speakers will not accept this form, only neːɟkor vɑɟ øtkor. For them -kor is an exceptional non-diminutive suffix with an invariant back vowel. 22. This example was pointed out by Catherine Ringen. Thanks! 118 CatJL 15, 2016 Péter Rebrus; Péter Szigetvári vowel of the first syllable is followed by the low neutral vowel, ɛ. In the diminutive form, this vowel is truncated and replaced by the high neutral vowel, i, the resulting diminutive name is back-harmonic, like a morphologically simplex Bi root would be. (20d) exemplifies the contrary situation: the back harmonicness of the stem is lost once its final vowel is replaced by ɛ: the result is variable. Finally, in (20e) the diminutive of tɛʃtveːr, which is front harmonic, turns out to be back harmonic, as it ends in the back vowel provided by the diminutive template, again defying harmonic uniformity. In these examples we omit the hyphen before the diminutive ending to visualize our claim: truncating diminutive forms are morphologically simplex. If a word like ʃimi, joːʒi, or tɛʃoː is morphologically simplex, then we predict that their harmonic class is not inherited from their root, since they do not have a root that is different from them. That is, the root of ʃimi is ʃimi, and therefore this word belongs to the front harmonic class like any other bisyllabic root with two neutral vowels, as the polysyllabic split predicts. Diminutive forms are also different from other suffixed words in that it is pos- sible to apply different diminutive formation processes to diminutives over and over again. This is not normally the case with other suffixes:23 a plural, a case, or a person-marking suffix cannot be added to a word that already has another instance of the semantically same suffix. With diminutives, on the other hand, this is very common: e.g. maːriɑ > mɑri > mɑrikɑ > mɑrikaːʧkɑ, ɑnːɑ > ɑnːuʃ > nuʃi > nuʃikɑ, eːvɑ > eːvi > viʦɑ > viʦaːʧkɑ, laːsloː > lɑʦi/lɑʦɑ > lɑʦkoː > lɑʦkoːʧkɑ. 7. Conclusions Most words of Hungarian are subject to harmonic uniformity, that is, whether a word takes the front or back variant of a suffix does not depend on the vowels of the word only the harmonic class of the root, that is, the first morph in the word. Needless to say, the harmonic class of the root is in most cases predictable from its vowel(s), although there is a sizable set of antiharmonic roots. We may take these to be exceptions to harmony. This exceptionality is inherited in the whole paradigm. Templatic diminutive forms in Hungarian, however, appear not to be subject to harmonic uniformity. Thus, a diminutive form does not “inherit” the harmonic class of the word it is the diminutive of. We submit that this is because templatic diminutive forms behave like morphologically simplex items of the lexicon. Thus these diminutive forms are exceptions to the overall pattern of harmonic uniformity, which normally preserves the exceptional antiharmonic property of a root. 23. A reviewer points out that the adjectival suffix -i, may be followed by another adjectival suffix, -ɑ/ɛʃ: e.g. hɑrʦ ‘fight’ > hɑrʦi ‘related to a fight’ > hɑrʦiɑʃ ‘militant’. These, however, are two different suffixes semantically: although both could be glossed as ‘ADJ’, hɑrʦi and hɑrʦoʃ only share their word category, adjective. Diminutives: Exceptions to Harmonic Uniformity CatJL 15, 2016 119 References Alber, Birgit & Arndt-Lappe, Sabine. 2012. Templatic and subtractive truncation. In Jochen Trommer (ed.). The Morphology and Phonology of Exponence, 289-325. Oxford: Oxford University Press. Benua, Laura. 1995. Identity effects in morphological truncation. In Beckman, Jill, Urbanczyk, Susan & Dickey, Laura Walsh (eds.). Papers in Optimality Theory, 77-136. University of Massachusetts Occasional Papers in Linguistics 18. Blaho, Sylvia & Szeredi, Dániel. 2013. Hungarian neutral vowels: A microcomparison. Nordlyd 40(1): 20-40. ISSN 1695-6885 (in press); 2014-9719 (online) Catalan Journal of Linguistics 15, 2016 121-143 Polish yers revisited* Amanda Rysling University of Massachusetts Amherst [email protected] Received: February 25, 2016 Accepted: July 8, 2016 Abstract It is common in linguistic research to attempt a unified analysis for similar patterns in related languages. In this paper, I argue that to do so for Polish and Russian vowel alternations would be a mistake. Although they share some notable phonological properties, they differ in their prevalence and their extensibility. I present an account of Polish under which vowel alternations are unexceptional, and exceptional blocking of alternation is achieved with lexically indexed constraints. This is the complement of Gouskova’s (2012) account of Russian, which I argue to be desirable on the basis of novel corpus statistics from the Polish lexicon and their divergences from the trends for analogous words in Russian. Keywords: Polish; Slavic; yers; vowel alternations; lexical indexation Resum. Un altre cop d’ull a les iers del polonès En la investigació lingüística és corrent buscar una anàlisi unificada per a fenòmes similars de llengües relacionades. En aquest treball, s’argumenta que fer-ho per a les alternaces vocàliques del polonès i del rus seria un error. Encara que comparteixen propietats fonològiques notables, difereixen respecte a la seva prevalència i la seva extensibilitat. Es presenta una anàlisi del polonès en la qual les alternances vocàliques no són excepcionals i el blocatge excepcional de les alter- nances s’assoleix amb restriccions indexades lèxicament. Aquesta és una anàlisi complementària de la de Gouskova (2012) per al rus, que s’argumenta que és desitjable sobre la base d’una nova exploració estadística d’un corpus del lexicó polonès i les seves divergències respecte a les tendències per a mots anàlegs en rus. Paraules clau: polonès; eslau; iers; alternances vocàliques; indexació lèxica Table of Contents 1. Introduction 4. The present account versus 2. Polish regular alternation alternatives with exceptional blocking 5. Conclusion 3. Lexical statistics of Polish References versus Russian * I would like to thank Maria Gouskova, Bruce Hayes, Gaja Jarosz, John Kingston, Claire Moore- Cantwell, Joe Pater, Jeremy Pasquereau, Robert Staubs, Shayne Sloggett, Kie Zuraw, the UMass Amherst Sound Workshop, and audiences at the Old World Conference in Phonology 12 Workshop on Exceptionality and Formal Approaches to Slavic Linguistics 24 for helpful discussion and comments. All errors are my own. 122 CatJL 15, 2016 Amanda Rysling 1. Introduction 1.1. Vowel alternations All modern Slavic languages display stem-internal vowel-zero alternations in the paradigms of some of their nouns. These alternating vowels are traditionally called ‘yers,’ and they derive from the same historical source: high front and back vowels that are no longer present in the modern languages. In Polish, there are near-mini- mal pairs of consonantal contexts that do and do not host vowel alternation, making straightforward phonological accounts of the phenomenon impossible. Examples of alternating vowels in Polish are given in (1), while non-alternating vowels in the same consonantal contexts are given in (2). Comparing (1a) with (2a), (1b) with (2b), and (1c) with (2c), it is evident that there is no way to describe the consonantal contexts of vowel alternation to the exclusion of the contexts of non-alternation.1 (1) Alternating vowels a. sfeter ~ sfetra ‘sweater’ nom. sg./gen.sg. b. idiotek ~ idiotka ‘female idiot’ gen. pl./nom. sg. c. kalek ~ kalka ‘carbon paper’ gen. pl./nom. sg. (2) Non-alternating vowels a. seter ~ setera ‘setter’ (dog) nom. sg./gen. sg. b. dɨskotek ~ dɨskotek ‘discotech’ gen. pl./nom. sg. c. kalek ~ kaleka ‘cripple’ gen. pl./nom. sg. Prominent analyses of yer vowel alternations have relied on abstract underlying representations of alternating vowels (Lightner 1972, Rubach 1986, Kenstowicz and Rubach 1987, Czaykowska-Higgins 1988, Szpyra 1992, Yearley 1995, Zoll 1996, Hermans 2002, Matushansky 2002, Steriopolo 2007). Under these approaches, all Slavic languages have been analyzed in a unified way: a rule that repairs an underlyingly defective yer vowel acts on that yer, and so allows it to surface, only when it is followed later in the UR string by another yer. These analyses thus assume that every output yer vowel is followed by an unrealized underlyingly present yer vowel at the end of the UR, i.e. [sfeter]~[sfetra] is underlyingly /sfetYr-Y/~/sfetYr-a/, where the (bolded and underlined) last yer of the UR /sfetYr-Y/ that conditions output [sfeter] is never seen overtly in the language. For accounts cast in Government Phonology (GP; Kaye 1990), the utility of these abstract URs has been used to argue for the correctness of GP assumptions about syllable structure (Rowicka 1999, Scheer to appear, inter alia), and the impossibility of casting these analyses in Optimality Theory (Prince and Smolensky 1993/2004) has been used to argue against OT as a whole (Scheer 2010, inter alia). For accounts that have not employed GP assumptions about syllable structure (Kenstowicz and Rubach 1987, Rubach 1987, inter alia), the 1. Data are taken from Bethin (1992), Gussmann (1990), and my own work with native speakers of Polish. Polish yers revisited CatJL 15, 2016 123 generalization about the locus of vowel alternation in the final syllable of a stem, and the syllable well-formedness consequences of this position of alternation, have been treated as accidental, rather than central to the explanation of the yer alternation (with the notable exception of Jarosz 2006, who uses syllable structure constraints to compel yer realization; this analysis is discussed in §4.1). Abstract UR accounts of Slavic have argued that there is no alternative by which to analyze Polish. In this paper, I advance three novel points. Firstly, my analysis diverges from previous accounts of Polish in that it treats non-alternation as the exceptional case in Polish; I argue that the lexical statistics of the language support this. Secondly, this account motivates Polish vowel alternation from syllable well-formedness pres- sures, and does not rely on underlyingly marked vowels. It thus incorporates gen- eralizations that most previous analyses have not captured about the phonological shape of the stems which host alternations, avoids the need to posit phonologically arbitrary rules, and provides a proof of concept for a whole-morpheme lexical indexation approach to exceptionality for Polish (something argued to be infeasi- ble by Rubach 2013). Thirdly, contra most of the earlier literature on Polish and Russian vowel alternations, I argue that Slavic languages should not all be analyzed in the same way, and present qualitative and quantitative comparisons of Polish and Russian to support this assertion. 1.2. Structure of the paper I begin by presenting the generalizations about words that do and do not host vowel alternations in Polish in §2. Next, in §2.1, I present my formal OT analysis of Polish, and compare it with Gouskova’s (2012) analysis of Russian. The statistics of the Polish lexicon are then given in §3, where the statistics of Russian found by Gouskova and Becker (2013) are also given for comparison. I argue in §3 that Polish and Russian diverge significantly in their observable attributes, such that the separate analyses for each language recommended here are preferable to a unified account of the languages’ vowel alternations. In §4, I discuss other analyses of Polish, and compare them with the one I advance here. §5 concludes. 2. Polish regular alternation with exceptional blocking In Polish, the vowel [e] alternates with zero in the stem-final syllables of some words but not others. Even though there are near-minimal pairs for these alternations, there are generalizations about morphemes that alternate, and contexts in which alternating vowels always or never occur. Words with alternating vowels like [sfeter]~[sfetr-ɨ] are illustrated in (3b). As noted by Jarosz (2006), there are no cases in which a vowel alternates before a consonant cluster, i.e. *[CeCC#]~[CCC-ɨ#]. The vowel [e] always appears in the case form of a noun’s paradigm that lacks an overt affix2 before the final consonant of that noun’s stem. 2. This is the nominative or accusative singular for masculine words, and the genitive plural for feminine or neuter words. 124 CatJL 15, 2016 Amanda Rysling Words like [seter]~[seter-ɨ] in (3a) have vowels that do not alternate, but which occur with the same front mid vowel quality within the same C_C consonantal contexts as the alternating vowels of the words in (3b). This can be seen by comparing no vowel alternation in [seter] with vowel alternation in [sfeter], even though both [e] vowels occur within the consonantal context [t_r]. Other nouns are vowelless throughout their declensions, even in what would be the expected cases to host a vowel. These words like [vʲatr]~[vʲatr-ɨ] in (3c) contain stem-final consonant clusters that are not broken apart by a vowel in their unaffixed cases, as opposed to (3a) and (3b). Thus, in the example of the [t_r] consonantal cluster, a mid front vowel does not alternate in [seter], does alternate in [sfeter], and does not even occur in the relevant form of [vʲatr]. As Hayes (2009: ch. 12) points out, whether the Polish vowel alternation is treated as deletion (Gussmann 1980, Bethin 1992, Jarosz 2008, Rubach 1986, 2013) or epenthesis (Czaykowska-Higgins 1988), there must be lexical exceptions: no phonological rule or constraint applying without restriction can pick out the words in (3b) to undergo alternation without also accidentally encompassing either those in (3a) or those in (3c). (3) Non-alternating, epenthetic, and absent vowels in Polish UR Unaffixed Case Suffix Diminutive Gloss a. Non-alternating /seter/ seter seter-ɨ seter-ek ‘setter’ /kalek/ kalek kalek-i kaletʃ-ek ‘cripple’ b. Epenthesis >1σ /sfetr/ sfeter sfetr-ɨ sfeter-ek ‘sweater’ /lalk/ lalek lalk-i laletʃ-ek ‘doll’ c. Blocking I /vʲatr/ vʲatr vʲatr-ɨ vʲater-ek ‘wind’ /katedr/ katedr katedr-ɨ kateder-ek ‘cathedral’ The generalization seen in the ‘Diminutive’ column in (3) is reliable: if the last consonant of a stem is a sonorant, a vowel appears between the last two consonants of that stem in the diminutive, even if insertion does not apply elsewhere3. Thus, the [t_r] cluster that remains intact in [vʲatr] is separated by an [e] vowel in the noun’s diminutive forms [vʲaterek]~[vʲaterk-a] (nom.~gen. sg.). In contrast to clusters that end with a sonorant, others that do not host vowel alternation with regular case morphology remain unbroken in the diminutive cases of the nouns in which they occur. This is shown in (4), in which the potential context [s_t] lacks a vowel across all of unaffixed [most], case suffixed [most-ɨ], and diminutive [mostek]. 3. Other examples of a vowel breaking up a word-finally preserved obstruent-sonorant cluster under diminutive suffixing include [blizn]~[blizna] but [blizenka] ‘scar,’ [bubr]~[bobra] but [boberek]‚ ‘beaver,’ [tɕfikwa]~[tɕfikw] but [tɕfikʲelka] ‘beetroot,’ [flandr]~[flandra] but [flanderka] ‘flounder,’ [jawmuʒn]~[jawmuʒna] but [jawmuʒenka] ‘alms,’ [kilometr]~[kilometra] but [kilometerek] ‘kilometer.’ Polish yers revisited CatJL 15, 2016 125 (4) No alternation in obstruent-obstruent clusters UR Unaffixed Case Suffix Diminutive Gloss Blocking II /most/ most most-ɨ most-ek ‘bridge’ /swuʒb/ swuʃp swuʒb-ɨ swuʒb-ek ‘service’ There are also words in which the occurrence of vowel alternation appears to be variable. It is unclear whether this variation is across or within speakers of Polish, but I observe that words with such optional alternations usually contain the same Slavic suffixes (-(e)v, -(e)n), as in (5). My analysis does not seek to provide an account for the alternations and variation in these cases. (5) Contexts of variation in Polish alternations UR Unaffixed Case Suffix Diminutive Gloss Variation /bit-v/ bitf, bitef bitv-ɨ bitev-ek ‘battle’ /vew-n/ vewn, vewen vewn-ɨ vewen-ek ‘wool’ The contents of examples (3), (4), and (5) are summarized in (6), with a novel row, (6b), that illustrates alternations in monosyllabic nominal stems. (6) Six types of patterns in Polish UR Unaffixed Case Suffix Diminutive Gloss a. Non-alternating /seter/ seter seter-ɨ seter-ek ‘setter’ /kalek/ kalek kalek-i kaleʧ-ek ‘cripple’ b. Epenthesis 1σ /dɲ/ dʑeɲ dɲ-i dʑon-ek ‘day’ /mɡw/ mɡʲew mɡw-ɨ mɡʲew-ek ‘fog’ c. Epenthesis >1σ /sfetr/ sfeter sfetr-ɨ sfeter-ek ‘sweater’ /lalk/ lalek lalk-i laleʧ-ek ‘doll’ d. Variation /bit-v/ bit(e)f bitv-ɨ bitev-ek ‘battle’ /vew-n/ vew(e)n vewn-ɨ vewen-ek ‘wool e. Blocking I /vʲatr/ vʲatr vʲatr-ɨ vʲater-ek ‘wind’ /katedr/ katedr katedr-ɨ kateder-ek ‘cathedral’ f. Blocking II /most/ most most-ɨ most-ek ‘bridge’ /swuʒb/ swuʃp swuʒb-ɨ swuʒb-ek ‘service’ 126 CatJL 15, 2016 Amanda Rysling 2.1. A whole-morpheme analysis of Polish In this section, I argue that Polish vowel-zero alternations should be analyzed as regular epenthesis with exceptional blocking captured by using lexically indexed constraints (Pater 2010; inter alia). This account of Polish alternations is the com- plement of Gouskova’s (2012) analysis of Russian, discussed in §2.2, in which the regular pattern of the language is to maintain mid vowels in the stem-final syllables of the null affix cases of nouns, and vowel alternation is the result of exceptional deletion. In Polish words that contain non-alternating vowels such as [seter]~[seter-ɨ] in (6a), the ever-present vowel is a part of the underlying representation of a noun’s stem. In alternating words such as [mɡʲew]~[mɡw-a] in (6b) and [sfeter]~[sfetr-ɨ] in (6c), the constraint that triggers epenthesis depends on the phonological size of the noun’s stem, for reasons discussed after the relevant constraint rankings are laid out. In polysyllabic stems that host alternating vowels, the vowels break up a word final consonant cluster (CC#). Two constraints that cannot be ranked with respect to each other serve to motivate the position of vowel epenthesis inside the diconso- nantal cluster; these are *ComplexCoda and *FinalAppendix, defined in (7) and (8). Together, these constraints are referred to as *CC# (9), and must outrank Dep-V (McCarthy and Prince 1995), the constraint against vowel epenthesis defined in (10). This is illustrated in (11a), where the insertion of a vowel is compelled by the ranking *CC# Dep-V in order to avoid a word-final consonant cluster. Epenthesis into non-coda stem-final consonant clusters is prevented in grammatical cases with overt suffixes ≫by Dep-V, as is shown in (11b). (7) *ComplexCoda - ‘assign a violation for every instance of a branching coda’ (8) *FinalAppendix - ‘assign a violation for every instance of an unsyllabified word-final consonant’ (9) *CC# - a cover constraint for the combined action of (7) and (8) (10) Dep-V - ‘assign a violation for every vowel present in the output that is not present in the input’ (11) Polysyllabic words with alternations a. /sfetr/ ‘sweater’ *CC# Dep-V → sfeter * sfetr *W L b. /sfetr-ɨ/ ‘sweaters’ *CC# Dep-V → sfetrɨ sfeterɨ *W Polish yers revisited CatJL 15, 2016 127 In monosyllabic alternating words like [mɡʲew]~[mɡw-a], Headedness, the pressure to have a syllable nucleus, drives insertion when the word would otherwise lack a vowel (Szpyra 1992, Hayes 2009). Headedness is undominated in the lan- guage, so some insertion is obligatory, but *CC# determines the site of epenthesis in /CCC/ words with underlying triconsonantal clusters: insertion must place a vowel before the last consonant of the stem, breaking up the potential final cluster. Insertion in monosyllables is demonstrated in (13)4. (12) Headedness - ‘assign a violation for every word that lacks a full syllable’ (13) Monosyllabic words with alternations / mɡw/ ‘mist’ Headedness *CC# Dep-V → mɡʲew * mɡw *W *W L meɡw *W L The instances in which there are no alternations between unaffixed and case forms such as [vʲatr]~[vʲatr-ɨ] and [most]~[most-ɨ] are specified as exceptions to epenthesis via lexical indexation. Lexically indexed ContiguityEx (McCarthy and Prince 1995), defined in (14), is ranked above *CC# and so prevents vowel inser- tion in unaffixed cases, as illustrated for sonorant-final words in (15a) and obstru- 5 ent-obstruent-final words in (15b) . ContiguityEx protects all stem-final sonority profiles in the words to which it is indexed. (14) ContiguityEx (ContigEx) - ‘assign one violation for every instance of two segments that are contiguous in the input but not contiguous in the output’ (15) Words without alternations in null affix cases a. /vʲatrEx/ ‘wind’ ContiguityEx *CC# Dep-V → vʲatr * vʲater *W L *W b. /mostEx/ ‘bridge’ ContiguityEx *CC# Dep-V → most * moset *W L *W 4. /mɡw/ historically contained a yer vowel between the first two consonants, such that a genitive plural would have been cluster-final [mьɡl]. Modern-day Polish has thus changed the word /mɡw-a/ to undergo the present vowel alternation in a different locus in the word’s stem from its historical source vowel. 5. An anonymous reviewer asks why Dep-VEx is not preferred over ContiguityEx here. This is because of the nature of lexical indexation; an indexed constraint can only apply to the segments that are underlyingly part of the stem to which it is indexed. An inserted vowel in the output is not part of that stem in the input, thus, it would not violate a prohibition against insertion that were only applied to the stem segments. 128 CatJL 15, 2016 Amanda Rysling But ContiguityEx must be dominated, since there is obligatory epenthesis for sono- rant-containing cluster-final stems in the context of diminutive suffixes. This requires the action of two constraints, *InternalAppendix and SonoritySequencing, defined in (16) and (17), respectively, which cannot be ranked with respect to each other. Together, *InternalAppendix and SonoritySequencing will be referred to as *CRC (18). The word-final [r] in (15a) is assumed to be a prosodic word appendix, and so it does not violate *CRC. (16) *InternalAppendix - ‘assign one violation for every instance of an unsyllabi- fied word-internal consonant’ (17) SonoritySequencing - ‘assign one violation for ever pair of tautosyllabic conso- nants the sonority of which does not rise in the direction of the syllable nucleus’ (18) *CRC - a cover constraint for the combined action of (16) and (17) Following Jarosz (2008), I argue that the diminutive suffix /-k/ introduces a pres- sure to optimize the paradigm in which it occurs by consistently placing stress on the same vowel throughout. Because Polish has regular penultimate stress in all but a small set of Latinate words, this amounts to a pressure to keep the same vowel in the penultimate syllable throughout a word’s declension. This is achieved by indexing the paradigmatic stress faithfulness constraint in (19) to the diminutive suffix /-kk/. (19) OPStressk (OPStrk)- ‘assign one violation for every difference between the main stress placements of any two output forms in the same inflectional paradigm’ Because OPStressk selects among paradigms, outputs from multiple underlying representations are considered at once in tableaux including it. Since OPStressk and *CRC do not conflict, they cannot be ranked with respect to each other. Comparing (15a) and (20a), we find evidence that *CRC dominates ContiguityEx: insertion is blocked when it would break up the word-final obstru- ent-sonorant cluster of the lexically indexed stem of [vʲatr] in (15a), but it occurs when an unbroken obstruent-sonorant-obstruent cluster would otherwise result in 6 *[vʲatrka] in (20aiii) . OPStressk rules out the candidate paradigm that shifts stress 6. A handful of words such as [pʲotrek]~[pʲotrka] ‘Peter,’ [kref]~[krvi] ‘blood,’ and [tʃosnek]~[tʃosnku] ‘garlic,’ which host trapped sonorants, are all indexed to another, higher-ranked ContiguityEx2 constraint, as shown in (i). In the case of [kref], Headedness compels insertion, as discussed for [mɡla] in (13) and (21). (i) Exceptional nominal stems with diminutive affixes /pʲotrEx2-k/, / pʲotrEx2-k-a/ ContigEx2 OPStr *CRC ContigEx1 *CC# Dep-V ‘Peter’ (dim.) → pʲótrek, pʲótrka * * pʲótrek, pʲotréka *W L **W pʲotérek, pʲotérka **W L ***W Polish yers revisited CatJL 15, 2016 129 between the stem vowel and the diminutive vowel in (20aii), and so the optimal paradigm in (20ai) merely incurs violations of ContiguityEx and Dep-V. In (20b), stress can be maintained on the stem vowel without violating the sonority sequenc- ing principle, so no vowel is ever inserted into the stem-final cluster. (20) Selection of nominal stems with diminutive affixes a. /vʲatrEx-kk/, / vʲatrEx-kk-a/ OPStrk *CRC ContigEx *CC# Dep-V ‘wind’ (dim.) → i. vʲatérek, vʲatérka ** *** ii. vʲátrek, vʲatréka *W L **L iii. vʲátrek, vʲátrka *W L *L b. /mostEx-kk/, /mostEx-kk-u/ OPStrk *CRC ContigEx *CC# Dep-V ‘bridge’ (dim.) → i. móstek, móstku * ii. mosétek, mosétku ***W The proposed ranking of Headedness >> OPStressk, *CRC >> ContiguityEx >> *CC# >> Dep-V accounts for the above generalizations about alternating and non-alternating environments in Polish. Words that contain vowels that do not alter- nate are faithful mappings from the UR of a stem to its output. Polysyllabic words that alternate with regular case morphology are the general pattern in the language, subject to the ranking of constraints against word-final consonant clusters above the constraint against epenthesis, *CC# >> Dep-V. In monosyllabic words that alternate, however, it is an undominated pressure that requires each word to have at least one well-formed syllable that accounts for vowel epenthesis, as captured by Headedness >>*CC# >> Dep-V; avoidance of final consonant clusters merely determines the location of insertion. Indexed ContiguityEx prevents epenthesis into the unaffixed case forms of some morphemes by dominating *CC#, as in Headedness >> ContiguityEx >> *CC# >> Dep-V, but it could never do so for the monosyllables in the language. Even if a learner were to mistakenly index the stem of monosyllabic [mɡwa] ‘mist’ to ContiguityEx, Headedness would still compel insertion, and *CC# would determine the position, because indexed constraints participate in a language’s general ranking, as in (21). (21) Headedness compels insertion where necessary / mɡwEx/ ‘mist’ Headedness ContiguityEx *CC# Dep-V → mɡʲew * * mɡw *W *W L meɡw * *W L 130 CatJL 15, 2016 Amanda Rysling For the diminutive forms in the language that could result in an unsyllabified sonorant, *CRC dominates normal ContiguityEx, and paradigm correspondence pressures determine the site of epenthesis throughout the diminutive declension. The analysis I advance here captures both what happens in the language and what does not. Vowel alternation occurs as a result of insertion in order to prevent stem-final consonant clusters, whether as a branching coda or as a singleton coda followed by a prosodic word appendix. This accounts for the positional general- izations of epenthesis: it always occurs in the stem-final syllable and before the stem-final consonant, i.e. *[CeCC#]~[CCC-ɨ#]. But regular epenthesis is blocked in exceptional cases, where final cluster faithfulness is maintained. Importantly, exceptional blocking is not undominated; it interacts with the ranking of the whole language and is prevented from over-extending. For this reason, all words obliga- torily contain at least one syllabic nucleus, even if a monosyllabic stem is somehow indexed, and unsyllabified word-internal sonorants are prevented in indexed stems. This indexation analysis does not suffer from over-generation, and it does not rely on arbitrary undominated exceptions, unlike the strawman indexation analysis of Rubach (2013). Analyses that attribute yer vowel realization to the surfacing of the rightmost of two abstract vowels in the UR cannot capture the syllable structural generalizations that my analysis does. For these sequential realization analyses, vowel alternation before a cluster *[CeCC#]~[CCC-ɨ#] or even in a non-final syllable of a stem *[CeCVC#]~[CCVC-ɨ#] should be well-formed if the underlying positions of the yer vowels happen to fall into this configuration. These sequential abstract analyses also have to posit that there are two underlyingly different overtly null suffixes. Stems that do not host a vowel in their regular case morphology, but do display a vowel in their diminutive case morphology, would be suffixed with a truly null morpheme in their overtly unaffixed regular forms, such as [vʲatr] from /vʲatYr-Ø/, but stems that do host alternation would be suffixed with an unrealized yer in their unaffixed forms, to yield [sfeter] from /sfetYr-Y/. One purported advantage that these sequential abstract UR analyses have over the analysis I have presented for Polish is that the abstract account extends equally well to other Slavic languages, while my account of Polish would not be applicable to, for example, Russian. The rest of this paper argues that such a unified analysis is in fact undesirable. Below in §2.2, I present Gouskova’s analysis of Russian that also uses lexical indexation, and captures the same syllable structure generaliza- tions that my analysis does, but does so with different constraints. I then turn to the lexical statistics of Polish as evidence for my assertion about the regularity of alternation in Polish, and compare these trends with those of Russian, which do the opposite. I return to comparison with other accounts of Slavic alternations in §4. 2.2. Comparison with Russian Gouskova (2012) argues that exceptionality is represented at the level of designat- ing whole morphemes that host exceptional behavior, and proposes that Russian vowel alternations are the result of exceptional deletion, while the words that retain Polish yers revisited CatJL 15, 2016 131 a non-alternating vowel in a stem-final syllable with a single consonantal coda are the regular pattern. A lexically indexed *Mid constraint, defined in (22), trig- gers deletion of a mid vowel as in (23a), when it would not cause a violation of *ComplexCoda or *FinalAppendix. But when such deletion would violate *CC, the candidate with a mid vowel in the last syllable is preferred, as in (23b). (22) *Mid - ‘assign a violation for every instance of a mid vowel’ (23) Russian words with alternations a. /moxEx/ ‘moss’ *CC# *MidEx → mxa moxa *W b. /xʲitʲerEx/ ‘clever’ *CC# *MidEx → xʲitʲor * xʲitʲr *W L For unindexed/unexceptional Russian words, normal *Mid is ranked below Max-V, the constraint against vowel deletion defined in (24). The normal marked- ness of mid vowels (that Gouskova argues for based on the realization of unstressed reduced vowels in Russian) cannot compel their deletion in regular words, even when the result could be parsed into well-formed syllables. We see this in (25b), in which [bʲi.lʲe.ta] is preferred to [bʲil.ta], even though all of the syllables of [bʲil. ta] are well-formed. Notably, although *CC is ranked above *MidEx, it cannot cause vowel insertion due to high-ranking Dep-V. For this reason, the fully faithful candidate [mʲet.r] is selected in (25c), even though it ends in a consonant cluster. (24) Max-V - ‘assign a violation for every instance of a vowel present in the input that is absent in the output’ (25) Russian words without alternations a. /lʲes-a/ ‘forest’ Dep-V *CC# *MidEx Max-V *Mid → lʲesa * lsa *W L b. /bʲilʲet-a/ ‘ticket’ Dep-V *CC# *MidEx Max-V *Mid → bʲilʲeta * bʲilta *W L c. /mʲetr/ ‘meter’ Dep-V *CC# *MidEx Max-V *Mid → mʲetr * mʲeter *W L **W 132 CatJL 15, 2016 Amanda Rysling The high ranking of Dep-V in Russian may be a part of the explanation for the persistence of paradigm gaps in the unaffixed cases of words like ‘mist,’ which in Russian lacks a genitive plural, but is a cognate of the Polish [mɡwa] in its nomi- native form [mɡla]. This word that has a gap in its null affix-, and so potentially vowel-containing, case did contain a yer vowel in an earlier stage of the historical development of the language, and so after the “fall” of the yers, it could have con- tained an alternating mid vowel at some point. But the likely rarity of the genitive plural form of a word like ‘mist’ could have prevented successive generations of learners from receiving positive evidence for a stem-internal vowel’s exist- ence, making it unclear whether the underlying form of the stem was vowelless / mɡl/ or vowelful indexed /mɡVlEx/ that always happened to appear with its vowel deleted. Faced with such uncertainty, Russian speakers might have accommodated a paradigm gap rather than risk insertion into an output for which the UR may not contain a vowel. 3. Lexical statistics of Polish versus Russian In this section, I present the results of an analysis of a corpus of Polish nouns, and compare these with Gouskova and Becker’s (2013) findings for a similar study of Russian nouns. On the basis of this, I contend that the lexical statistics of Polish support the proposal that vowel alternation is the general case, while a lack of an alternating vowel within stem-final consonant clusters is the exceptional case in the language. I further argue that the lexical statistics of the two languages together support the conclusion that they should be analyzed in complementarily different ways. The POLEX lexicon of Polish (Vetulani et al. 1998) contains 41,742 nouns. Of these, 15.8% exhibit stem-final vowel alternation, as reported in (26a). Another 6.3% of these nouns contain a non-alternating [e], as shown in (26b). Yet another 16.1% of the lexicon ends in an unbroken word-final consonant cluster in some grammatical case, as illustrated in (26c). Of these nouns that end in a consonant cluster, all of which could possibly be considered exceptions to alternation under my analysis, the majority end in the suffixes [-oɕtɕ], [-izm], [-ist], [-stv], [-ovɲ] and [-ɨtm]; these are counted in (26ci). These suffixes do not host alternating vowels in their regular case paradigms, and tend to be part of a more formal register in language use. Words that end in such consonant cluster-final suffixes represent 11.1% of the whole lexicon, leaving 5.0% of the lexicon ends in CC#, but does not contain these particular suf- fixes. These unsuffixed cluster-final words are reported in (26cii). Polish yers revisited CatJL 15, 2016 133 (26) Corpus statistics for Polish Count Of Lexicon Example a. Alternating [e] 6,581 15.8% sfeter ~ sfetrɨ b. Non-alternating [e] 2,624 6.3% seter ~ seterɨ c. Ends CC# cluster 6,729 16.1% i. Suffixed 4,630 11.1% markɕizm ~ markɕizmu ii. All unsuffixed 2,099 5.0% swuʃp ~ swuʒba d. CCV# or non-[e] CVC 25,808 61.8% azja ~ azji Total 41,742 100% If Polish speakers know that the particular Latinate and Slavic suffixes counted in (26ci) are unacceptable contexts for vowel insertion, and so rank morpheme- specific faithfulness to them above the pressures to break up a word final consonant cluster, then the CC# words in the lexicon that must be treated as idiosyncratic exceptions to epenthesis, and would be indexed to ContiguityEx under the analy- sis presented here, would be 5% of the nouns of Polish. This is comparable to the percentage of alternators in Russian, to which I now turn. Gouskova and Becker (2013) performed an analysis of the 20,563 masculine second declension nouns from Zaliznjak’s (1977) dictionary. They focused on masculine nouns, rather than including feminine ones, because feminine nouns are those more likely to have paradigm gaps in exactly the case expected to contain an alternating vowel. Their findings for forms comparable to those in (26) are given in (27)7, which lists the alternators that are exceptional in (27a), the unexceptional unbroken clusters in (27b), and all other words in (27c). (27) Corpus statistics for Russian Count Of corpus Example a. Alternating [e]/[o] 1,902 9.2% ʃatʲor ~ ʃatrof b. Ends CC# cluster 3,177 15.5% most ~ mastof c. Other VC# or CCV# 15,484 75.3% krot ~ kratof Total 20,563 100% Under an account of Polish and Russian yers that combines the present analy- sis with Gouskova’s (2012), the percentage of exceptional non-alternating words (e.g. [most]~[most-ɨ]) in Polish is smaller (5%) than the percentage of exceptional alternating words in Russian (e.g. [ʃatʲor]~[ʃatrof], 9%), while the percentages of words that follow the majority pattern are comparable across the two languages 7. The apparent [a]~[o] alternation in the example Russian words in (27) is the result of stress shift, with [o] under primary stress and [a] when pretonic. 134 CatJL 15, 2016 Amanda Rysling (Polish 15.8% alternating, e.g. [sfeter]~[sfetrɨ], to Russian 15.5% not hosting a vowel, e.g. [most]~[mastof]). On the basis of these data, I argue that while Polish and Russian arose from the same historical source, they have taken divergent paths in their evolution. The languages today differ in multiple ways that likely impact the appropriate analysis of their vowel alternations. In Russian, both [e] and [o] alternate, and so the back- ness distinction of the original yers is preserved, but the percentage of alternators in the lexicon is almost half of the percentage of non-alternators. This contrasts with Polish, in which only the [e] vowel alternates, but the percentage of alterna- tors in the lexicon is three times the percentage of unpredictable non-alternators. Thus, Polish and Russian are the opposites of each other on extensibility to loans, predictability of vowels, paradigm gaps, and the relative sizes of the proportions of the lexicon that undergo alternation versus retain word-final unbroken clusters in null affix cases. These differences are summarized in the table in (28). (28) Qualities of Polish versus Russian alternations Polish Russian Extended to loans? Yes No Vowel predictable? Yes No Paradigm gaps? No Yes Unbroken CC# 5.0% 15.5% Alternation 15.8% 9.2% The vowel alternations of Polish and Russian occur in phonologically similar contexts, but with different rates and limitations. While a unified analysis of the two languages’ vowel alternations would capture the phonological contextual parallels, it would ignore the particular quantitative attributes of the inputs to which learners of Polish and Russian have access. I discuss the impact of these differences in §4 as part of a greater comparison of the present account with others. 4. The present account versus alternatives In this section, I compare the account that I have advanced here with alternatives by focusing on three differences. The first difference between my and most other approaches that I discuss is that my approach to Polish vowel alternations builds the syllabic contexts of the alternations into the analysis, while, to my knowledge, all other approaches besides Jarosz (2006) treat the pre-final consonant positional restriction of the alternations as accidental. The second difference between my account and others’ discussed below is that I argue that Polish and Russian should be analyzed differently, while a major argument advanced in favor of adopting an abstract-UR approach to alternating vowels has been that an abstract-UR approach can give a unified analysis to all Slavic languages. As I explain above in §3, I argue for analyzing Polish and Russian alternations differently on the basis Polish yers revisited CatJL 15, 2016 135 of their speakers’ qualitatively divergent behaviors and their lexical statistics’ quantitatively divergent patterns. The other difference between the present and most alternative accounts that I focus on here is that my account treats Polish alternations as epenthetic, while most other accounts of Slavic alternating vowels have treated alternations as the result of deletion. If one assumes that the segmen- tal inventory of Polish is composed of only those segments that surface in the language, there is no phonological markedness constraint that could be used to trigger vowel deletion in a lexical indexation analysis (Rubach 2013). While the pioneering whole-morpheme analysis of Polish advanced by Jarosz (2008) makes use of vowel deletion, its central insights and triggering faithfulness constraint may be retained in an account that assumes an epenthetic alternation, as I show below. For these reasons, I argue that an epenthetic analysis of Polish is simpler to posit from the perspective of markedness constraints, and does not lose any of the surface-true generalizations that deletion-based analyses of the Polish alter- nation can cover. 4.1. Capturing syllable structure generalizations Most abstract-UR accounts of Slavic vowel alternations have posited the exis- tence of underlying ‘yer chains,’ and a special corresponding rule of yer realiza- tion. In this section, I will demonstrate that, because of the way that this rule is written, the yer chain approach does not capture the phonological generalizations about Polish vowel alternation laid out in §2. It would allow alternations in syl- lable structural and stem-internal positions in which vowel alternation is never observed. In yer chain analyses of Slavic, the process of yer vowel realization is the repair and subsequent surfacing of all the yer vowels but the very last one in an underlying string. This requires the assumption that, for every alternating yer vowel observed in the surface forms of the language, there is another yer vowel that occurs later in the underlying representation of a word that is never observed. The output forms of some words of the language, URs assumed by abstract yer chain accounts, and the URs that I propose for those same words are given for comparison in (29). For the unaffixed case form [sfeter] in (29a), a yer chain account would require an underlying yer vowel in the position that the alternating vowel (underlined) occurs, as well as one appended to the end of the stem as a case suffix s that i not overtly phonologically realized. In contrast with (29a), [vʲatr] in (29b) would not have an unrealized yer vowel suffix in the yer chain account. Instead, its suffix would be truly null, and this would prevent the stem-internal yer that surfaces between the stem-final consonants in [vʲaterek] of (29c) from being realized in the non-diminutive paradigm of the word. When a yer-containing stem is realized with an overt case suffix, the yer chain account again relies on its rule of only realizing a yer if another follows, yielding forms such as [mɡwa] in (29d). 136 CatJL 15, 2016 Amanda Rysling (29) Comparison of URs Output Yer Chain UR Proposed UR a. sfeter /sfetYr-Y/ /sfetr-Ø/ b. vʲatr /vʲatYr-Ø/ /vʲatrEx/ c. vʲaterek /vʲatYrYkY/ /vʲatrEx-kk/ d. mɡwa /mɡYw-a/ / mɡw-a/ The rule that realizes all but the last yer in a chain is phonologically arbitrary; it is not integrated into the well-formedness pressures of the rest of the language. Because of its arbitrariness relative to syllable structure, this rule would allow vowels to alternate before a word-final coda cluster. Since pre-final-cluster alter- nation is never observed in Polish, it is an undesirable prediction of any yer chain account. If yer chains were indeed the way that speakers represented these words, it might have been expected that historical genitive plural of ‘mist’ [mьɡl] could have been maintained as [meɡw] from /mYɡw-Y/, instead of being regularized to [mɡʲew]~[mɡwa] from /mɡw/~/mɡw-a/. It is possible to test whether speakers of Polish are aware of the inappropri- ateness of alternation in a non-final syllable or before a word-final coda cluster. Gouskova and Becker (2013) did this in a nonsense word experiment for Russian, and found that speakers dispreferred forms with alternations in these unattested conditions relative to forms that alternated in the appropriate conditions. Given that the Russian alternation is not as common in that language as the Polish alternation, it is reasonable to expect that Polish speakers would similarly downgrade the well- formedness of pre-cluster alternating nonsense words. Yer vowels in chains are assumed to be somehow defective relative to the segmental inventories of the languages in which they occur, and so the process of yer vowel realization in these accounts is the repair and surfacing of all but the very last yer in an underlying string. Thus, an abstract yer chain analysis requires speakers to hypothesize the existence of vowels that are never seen overtly, and to endow them with defectiveness of some kind (this varies by account). In order for speakers to arrive at a phonologically arbitrary grammatical analysis that requires specific representational assumptions about vowels that are never overtly observed, their innate knowledge would have to encompass a preference for such an analysis. Scheer’s (2004) CVCV account of alternating vowels is similar to yer chain accounts in that it posits underlying vowels never seen on the surface that occur in a licensing relationship with alternating vowels. It also assumes that syllables are underlyingly all consonant-vowel, that is, for every apparent coda consonant, there is in fact an unrealized underlying vocalic nucleus for which it provides an onset. The difference between loci of alternation and unbroken clusters is then one of dif- ferent underlying specifications of alternating vowels versus never-seen ones. This difference is taken to be identical across Polish, Russian, and all other languages (e.g., French) in which some form of vowel alternation is seen. It is not the aim of this paper to argue against the CVCV approach to syllable structure, under which Polish yers revisited CatJL 15, 2016 137 Scheer’s statement of the conditions of vowel alternation is indeed motivated by that framework’s syllable structural well-formedness criteria. But insofar as my analysis does not require that alternating and non-alternating vowels be underlyingly specified as different, does not require positing vowels that are never seen on the surface, and begins from the assumption that analyzing related Polish and Russian differently is desirable on the basis of those languages’ lexica, all the arguments presented here to favor my analysis over other approaches apply to Scheer’s as well. One earlier analysis of Polish captures the syllable positional generalization of vowel alternation under structural assumptions like those I follow here, and employs URs with vowels that are not attested in the language. Jarosz’s (2006) account pos- its that alternating vowels are underlyingly /ɪ/ vowels, which do not surface in the language due to the markedness of simultaneous [+high] and [-tense] features. Due to their markedness, these vowels only surface when the alternative would be the creation of a complex coda, but even then they are not realized faithfully. Instead, the [+high] specification is changed to a [-high] one to yield a mid vowel. This is shown in (30a), in which [sfeter] is chosen over possible [sfetr] or [sfetɪr]. When a complex coda would not otherwise result, the underlying yer vowel is deleted as in (30b), while non-alternating vowels are present as mid vowels underlyingly. (30) Unattested surface vowel in appropriate position of the UR a. / sfetɪr-Ø/ *ɪ *ComplexCoda Id[HI] Max-V → sfeter * sfetr *W L *W sfetɪr *W L b. /sfetɪr-ɨ/ *ɪ *ComplexCoda Id[HI] Max-V → sfetrɨ * sfetɪrɨ *W L sfeterɨ *W L Jarosz thus posits that alternation stems from the underlying defectiveness of the vowels, which are dispreferred by universal markedness constraints. The major differences between Jarosz’s (2006) analysis and the one I present here are its reli- ance on vowels that are not attested in the surface forms of the language and its characterization of the alternation as deletion-based. 4.2. Related languages with different grammars In this section, I argue that a unified account of Polish and Russian, in which one exceptional rule were posited to account for the alternations of both languages, would require assumptions about learning and the use of extra-grammatical gen- eralization mechanisms that an account with separate analyses does not require. 138 CatJL 15, 2016 Amanda Rysling As I have shown above in §3, Polish and Russian are qualitatively and quantitatively different. Qualitatively, Polish extends vowel alternation to loan words, does not have paradigm gaps, and only alternates the front mid vowel, while Russian does not extend alternation to loans, retains paradigm gaps, and alternates both of the front and back mid vowels. Quantitatively, the percentage of the Polish lexicon that undergoes alternation (15.8%) is comparable to the percentage of the Russian lexicon that does not undergo alternation (15.5%), while the percentage of non-alternators that are exceptional under my analysis of Polish (5.0%) is comparable to the percentage of exceptional alternators under Gouskova’s (2012) account of Russian (9.2%). If Polish and Russian are subject to complementary grammatical analyses like those I advance here, then Polish speakers are extending a regular rule of their language to apply to loans and select output forms for even low-frequency grammatical cases, while Russian speakers are not extending an exceptional rule of their language to loan words or gaps in existing paradigms. A unified analysis of the two languages would require speakers of Polish and Russian to be sensitive enough to the lexical frequencies of alternation in order to reach divergent decisions about whether to extend the pattern, but not so sensitive that they would adopt divergent grammatical analyses for their lan- guages as wholes. Presumably, extension would have to be determined by a post-grammatical mechanism. Any unified grammatical analysis necessarily makes representational assumptions that, as discussed above, require it to rely on innate knowledge about the appropriate grammatical analysis for Slavic in order to ensure that learners reach the desired outcome in both languages. This is because a single-rule analysis must be neutral to the differences between the two languages such that it does not build those generalizations into an account by, for example, treating alternations as regular in one language and exceptional in another. The motivation for alternation does not come from any phonologi- cal pressure that the learner observes at work in her target language; instead, somehow, learners must be constrained to prefer phonological arbitrariness over all other phonologically-grounded possible explanations. Unified analyses thus require substantial assumptions about both the innate knowledge with which speakers begin learning and the relationship of extra-grammatical mechanisms of rule extension to grammatical analyses. An account in which Polish and Russian differ in the way I have argued here would require speakers to observe the prevalence of vowel alternations in their native languages, and use this for both grammatical learning and later possible pat- tern extension. No special preference for phonological arbitrariness would have to be encoded in learners’ innate knowledge, and no analytical insensitivity combined with behavioral sensitivity would have to be posited for speakers’ awareness of their native languages. 4.3. Insertion rather than deletion Under the account that I advance here, [vʲatr] and [most] are exceptions to epen- thesis in Polish. The alternative analysis that uses lexical indexation is that [seter] Polish yers revisited CatJL 15, 2016 139 resists deletion. In this section, I discuss the reasons for which an epenthetic analy- sis is to be preferred. 4.3.1. Epenthetic analyses are simpler The only whole-morpheme account of Polish yers in the literature to date is Jarosz’s (2008). Jarosz argues that underlying full vowels delete in some words of Polish and not others, because some words of Polish are subject to a constraint ranking in which the constraint requiring all of the forms of a noun’s paradigm to have stress on the same segment, OPStress, outranks the constraint against vowel deletion, Max-e. Importantly for our discussion of epenthesis versus deletion, Jarosz (2008) does not motivate her choice of process or underlying representation. She assumes underlyingly present vowels, and determines that deletion is necessary to make an underlying vowel not surface when it would either interrupt completely regular penultimate stress placement or force a case form to be stressed on a different syllable from the rest of its paradigm. In Jarosz’s analysis, the paradigm in (31a), in which underlying final-syllable [e] is deleted with vocalic suffixes, is preferred over the paradigm in which stress shifts one syllable over to remain penultimate in (31b) and the paradigm in which a word- final [e] is epenthesized so as to keep the [ter] sequence both intact and stressed in (31c). (31) OT with co-phonologies for deletion /sfeter-∅/, /sfeter-a/ Dep-e OPStress Max-e /sfeter/: OPStress >> Max-e → a. sféter, sfétra * b. sféter, sfetéra *W L c. sfetére, sfetéra *W L But it is possible to formulate an account that uses the same triggering constraint and theoretical approach as Jarosz’s if one assumes that underlying con- sonant clusters are broken only when it would result in stress placement on the appropriate vowel in the word’s stem. This is seen in comparison of the tableaux in (31) and (32), which respectively show Jarosz’s use of OT with co-phonologies to choose among possible paradigms for an underlying stem that contains an alternat- ing vowel and a hypothetical use of OT with co-phonologies to choose among pos- sible paradigms for an underlying stem that contains a consonant cluster that hosts an alternating vowel. In the alternative co-phonology analysis using OPStress and epenthesis shown in (32), epenthesis of one [e] vowel as in (32a) is preferred over allowing stress to shift as in (32b) or epenthesizing both into- and after the stem as in (32c). 140 CatJL 15, 2016 Amanda Rysling (32) OT with co-phonologies for epenthesis /sfetr-∅/, /sfetr-a/ Max-e OPStress *CC# Dep-e /sfetr/: OPStress >> Dep-e → a. sféter, sfétra * b. sféter, sfetéra *W **W c. sfetére, sfetéra **W d. sfétr, sfétra *W OPStress can thus be used to motivate epenthesis if assumptions about under- lying representations are changed. As Rubach (2013) notes, it is not possible to identify a markedness constraint that would require deletion of underlying full [e] vowels that could be used to motivate alternation in Polish. Unlike in Russian, a Polish learner would not encounter good evidence for the action of *Mid in the lan- guage, because unstressed vowels are not reduced such that their qualities change, and only [e] alternates. Furthermore, a Russian-style account a la Gouskova would wrongly predict that [o] can alternate in the last syllable of nouns’ stems in the language, and such a deletion account would still have to posit epenthesis in Polish diminutives (which are the same shape as the Russian). Jarosz’s (2008) account captures diminutive-conditioned alternations by allowing the demands of the [-ek] suffix to override the independent phonology of the stems of nouns like [vʲatr], and thereby compel epenthesis into the cluster only when the diminutive suffix is present. This is the co-phonology-based analysis off of which the proposal for diminutives in §2.1 is modeled. Because Jarosz’s (2008) account can so easily be re-cast using epenthesis, its success does not provide an argument in favor of using deletion. This epenthetic alternative using paradigm uniformity, however, must include the same kind of constraints against word-final clusters that my analysis does in order capture two key generalizations about syllable structure. Firstly, if it did not include syllable or prosodic word well-formedess constraints such as a prohibition against clusters or final prosodic appendices, and only mandated that stress remain on the penultimate syllable of all output forms, it would permit vowel alternation before a stem final cluster to preserve the uniform location of stress across a paradigm. Secondly, it would select a candidate like [sfétr]~[sfétra] in (32d) as the optimal paradigm, because this paradigm does not include epenthesis, but also maintains stress on the same vowel throughout. This inclusion of cluster-prohibiting constraints in an epenthetic co-phonology account renders it extensionally equivalent to the present proposal. 4.3.2. Epenthetic analyses are not new Epenthetic analyses of Polish have been advanced before. Czaykowska-Higgins (1988) proposed an epenthetic account of the Polish alternation in a rule-based framework, and noted that the existence of near-minimal pairs made it impossible Polish yers revisited CatJL 15, 2016 141 to determine by potential consonantal environment alone whether a rule should apply to a stem. Bethin (1992) proposed to analyze alternations in loanwords as the product of epenthesis, but alternations in all other words as the product of filling in a lacking representational component as in a traditional underspecification account. My account of Polish most closely resembles the epenthetic analysis sketched as a pedagogical exercise by Hayes (2009). Under this analysis, vowel insertion is productive only in clusters that would otherwise be sonority rises, e.g. [t_r], but not in plateaux, e.g. [t_s] or [l_r]. The limited number of words that contain an obstruent-obstruent cluster that hosts alternation are exceptional cases of the appli- cation of insertion, while the words that contain an obstruent-sonorant cluster that does not host alternation are exceptional cases of blocking. The major difference between my account and the one that Hayes suggests is thus in the statement of the process and the extent of its application. I treat alternation as regularly applying in clusters of all sonority profiles, and treat all the cases of unbroken stem-final clus- ters that are not within protected suffixes as exceptional, in contrast to motivating epenthesis by avoiding sonority rises, and requiring two sub-classes of exceptions. It would be very difficult, if not impossible, to derive different predictions from Hayes’ account as could be derived from mine8; whichever account is preferred, vowel alternation is unexceptional in sonority rises, and Polish and Russian are still analyzed differently. 5. Conclusion I have argued that while Russian and Polish vowel alternations are historically related, they diverged. In Russian, alternations result from exceptionally triggered deletion. But in Polish, they are the result of a productive epenthesis process that is subject to exceptions. In Russian, general faithfulness to underlying vowels, exceptional deletion of fully specified underlying vowels, and a prohibition against vowel insertion contribute to explaining a lack of extension of alternation to loan words, the semi-predictable quality of alternating vowels, and paradigm gaps. In Polish, general epenthesis of a default vowel, exceptional blocking of epenthesis, and a prohibition against deletion contribute to explaining ready extension to loan words, entirely predictable alternating vowel quality, and a lack of paradigm gaps. The analysis of Polish that I advance is capable of capturing syllable structural and stem positional generalizations that other accounts miss, while not relying on vowels that are unattested in the language, and lending itself to future experimental investigation. 8. As Zuraw (p.c.) points out, a difference in the productivity of vowel alternation across clusters with different sonority profiles does not have to be the result of exceptionally fixed occurrence of alter- nation in some plateaux versus general application in rises. Rather, minimally different Contiguity constraints that penalize the break-up of different types of sequences may be ranked with respect to each other and the rest of the ranking of the language to yield more apparent extensibility in rises than plateaux. All of these analyses would capture the syllable and stem positional generalizations of Polish, and treat alternation as the regular rule, not the exception. 142 CatJL 15, 2016 Amanda Rysling References Bethin, Christina. 1992. Polish syllables: The role of prosody in phonology and mor- phology. Columbus: Slavica Publishers. Czaykowska-Higgins, Ewa. 1988. Investigations into Polish morphology and phonol- ogy. Doctoral Dissertation, MIT, Cambridge, MA. Dresher, B. Elan. 1981. On the learnability of abstract phonology. In Baker, C. L. & McCarthy, John J. (eds.). The logical problem of language acquisition, 188-210. Cambridge, Mass.: MIT Press. Gouskova, Maria. 2012. Unexceptional segments: a non-representational theory of Russian yers. Natural Language and Linguistic Theory 30: 79-133. Gouskova, Maria & Becker, Michael. 2013. Nonce words show that Russian yer alter- nations are governed by the grammar. Natural Language and Linguistic Theory 31: 735-765. Gouskova, Maria, Kasyanenko, Sofya & Newlin-Łukowicz, Luiza. 2015. Selectional restrictions as phonotactics over sublexicons. Ms. New York University. Gussmann, Edmund. 1980. Studies in Abstract Phonology. Cambridge, Mass.: MIT Press. Hayes, Bruce. 2009. Introductory Phonology. Malden, Mass., and Oxford, UK: Wiley- Blackwell. Hermans, Benjamin. 2002. Overapplication of yer vocalization in Russian. Linguistics in the Netherlands 85-95. Jarosz, Gaja. 2006. Polish yers and the finer structure of output-output correspondence. In Proceedings of the 31st Annual Meeting of the Berkeley Linguistics Society. Jarosz, Gaja. 2008. Partial ranking and alternating vowels in Polish. In Proceedings of CLS 43, volume 41, 193-206. Chicago Linguistic Society. Kaye, Jonathan. 1990. ‘Coda’ licensing. Phonology 7: 301-330. Kenstowicz, Michael & Rubach, Jerzy. 1987. The phonology of syllabic nuclei in Slovak. Language 63: 463-497. Lightner, Theodore. 1972. Problems in the Theory of Phonology, Vol 1: Russian Phonology and Turkish Phonology. Edmonton: Linguistic Research, Inc. Matushansky, Ora. 2002. On formal identity of Russian prefixes and prepositions. In MIT Working Papers in Linguistics 42, 217-253. Cambridge, MA: MIT. McCarthy, John J. & Prince, Alan. 1995. Prosodic morphology. In Goldsmith, John A. (ed.). The Handbook of Phonological Theory, 318-366. Cambridge, Mass., and Oxford, UK: Blackwell. Pater, Joseph. 2010. Morpheme-Specific Phonology: Constraint Indexation and Inconsistency Resolution. In Parker, Steve (ed.). Phonological Argumentation: Essays on Evidence and Motivation. London: Equinox. Prince, Alan, & Smolensky, Paul. 1993/2004. Optimality Theory: Constraint interac- tion in generative grammar. Rowicka, Grazyna. 1999. Prosodic optimality and prefixation in Polish. In Kager, René, van der Hulst, Harry & Zonneveld, Wim (eds.). The Prosody- Morphology Interface, 367-389. Cambridge: Cambridge University Press. Rubach, Jerzy. 1986. Abstract vowels in three-dimensional phonology: The yers. The Linguistic Review 5: 247-280. Rubach, Jerzy. 2013. Exceptional segments in Polish. Natural Language and Linguistic Theory 31: 1139-1162. Polish yers revisited CatJL 15, 2016 143 Scheer, Tobias. 2004. What is CVCV and why should it be? Vol. 68. Walter de Gruyter. Scheer, Tobias. 2010. Why Russian vowel-zero alternations are not different, and why Lower is correct. Journal of Language and Verbal Behavior (St. Petersburg) 9: 77-112. Scheer, Tobias. 2012. Variation is in the lexicon: yer-based and epenthetic vowel-zero alternations in Polish, volume Sound, Structure, and Sense: Studies in Memory of Edmund Gussmann. Wydawnictwo KUL. Steriopolo, Olga. 2007. Jer vowels in Russian prepositions. In Arbor, Ann. Formal Approaches to Slavic Linguistics. The Toronto Meeting 2006, 365-385. MI: Michigan Slavic Publications. Szpyra, Jolanta. 1992. Ghost segments in nonlinear phonology: Polish yers. Language 68: 277-312. Yearley, Jennifer. 1995. Jer vowels in Russian. In Beckman, Jill, Dickey, Laura Walsh & Urbanczyk Suzanne (ed.). Papers in Optimality Theory II: University of Massachusetts Occasional Papers in linguistics, 533-571. Amherst, Mass.: GLSA Publications. Zaliznjak, Andrej. 1977. Grammatiˇceskij slovar’ russkogo jazyka. Moscow: Russkij Jazyk. Zoll, Cheryl. 1996. Parsing below the Segment in a Constraint-based Framework. Doctoral Dissertation, University of California, Berkeley. ROA 143. ISSN 1695-6885 (in press); 2014-9719 (online) Catalan Journal of Linguistics 15, 2016 145-171 Polarized Variation* Kie Zuraw University of California, Los Angeles [email protected] Received: February 25, 2016 Accepted: June 22, 2016 Abstract In cases of exceptionality, there are usually many words that behave regularly, a smaller number that behave irregularly (the exceptions), and perhaps an even smaller number whose behavior varies. This paper presents several examples of exceptionality and variation that are polarized in this way: most items exhibit one behavior or the other consistently, with only a minority of items showing variation. The result is a U-shaped histogram of behavior rates. In some cases, this requires listing of surprisingly long units. There are, however, some cases of bell-shaped histograms, where most items show variation, and only a minority are consistent. Some simple simulations are pre- sented to show how polarized variation can result when variation is between two categorical outcomes, and both types of variation can result when variation is along a phonetic continuum. Keywords: exceptions; variation; corpus; diachronic change Resum. Variació polaritzada Quan parlem d’excepcionalitat, generalment hi ha moltes paraules amb un comportament regular, un grup menor amb comportament irregular (les excepcions), i potser un grup encara menor que varien. Aquest article presenta diversos exemples d’excepcionalitat i de variació que son pola- ritzats així: la majoria dels ítems exhibeixen un comportament o l’altre quasi constantment, amb només una minoria d’ítems que varien. Resulta un histograma en forma d’U, si tracem proporci- ons de comportament. Hi ha alguns casos que exigeixen allistar elements de llargada sorprenent. Tanmateix, existeixen casos d’histogrames en forma de campana: la majoria dels ítems exhibeixen variació, amb només una minoria que siguin constants. Es presenten simulacions senzilles que mostren de quina manera la variació pot resultar polaritzada quan és entre categories, i de quina manera les dues menes de variació son possibles quan la variació es troba en un contínuum. Paraules clau: excepcions; variació; corpus; canvi diacrònic Table of Contents 1. Introduction: free 3. Modeling U-shaped and lexical variation (and non-U-shaped) distributions 2. Case studies 4. Summary and conclusion of U-shaped distributions References * Thanks to audiences at the UCLA phonology seminar, the 2015 Old World Conference on Phonology, and the University of Massachusetts for their comments and questions, especially Bruce Hayes, Brian Smith, and Joe Pater. Thanks particularly to the organizers of the workshop, Eulàlia Bonet, Joan Mascaró, Eduard Artés, and Teresa Cabré, for bringing this paper into existence, and to two anonymous reviewers for their careful and stimulating feedback. 146 CatJL 15, 2016 Kie Zuraw 1. Introduction: free and lexical variation We can define two extreme types of variation, and a continuum between them. At one extreme lies purely free variation, where each relevant word itself shows variation. For example, many American English speakers optionally reduce /nt/ to [ɾ]̃ in roughly the same contexts where tapping applies (between two vowels, the second of which is unstressed): spli[nt]er varies with spli[ɾ]er̃ , pe[nt]atonic with pe[ɾ]atonic̃ . If the variation applied indiscriminately across words, with no words having an idiosyncratic tendency to favor one variant over the other, this would be a case of pure free variation. (In reality, there probably are some words with especially high or low tendencies to undergo the reduction, though this has yet to be documented as far as I know.) At the other extreme lies purely lexical variation, where each word consistently shows one variant or the other. The English past tense is close to being of this type: for the most part, some verbs have a consistent regular past (reach, reached), and others a consistent irregular past (teach, taught). Here, variation is over types (dif- ferent verbs, reach vs. teach), not over tokens (occasions of uttering splinter). In an experimental context, Richtsmeier (2008) uses the terms token variation and type variation to distinguish the two extremes of our continuum. There do exist mixed cases that fall somewhere between free/token and lexi- cal/type variation. Some English verbs do actually show free variation. For many speakers, sneak has an informal irregular past snuck alongside the prescriptively correct regular sneaked, and dive has interchangeable plurals dove and dived for many North American speakers. The real state of English past tense is thus close to being lexical variation, yielding a U-shaped distribution whose two arms are of uneven height, as shown in (1). If there were more words like sneak or dive (vary- ing), at the expense of words like teach (exceptional) and reach (regular), the case would be closer to free variation. (1) Schematic U-shaped distribution of exceptionality Polarized Variation CatJL 15, 2016 147 This paper has two parts. First, it presents several cases of U-shaped distribu- tions, showing that they are common, that they involved both affixed words and longer phrases, and that they suggest listing of fairly large units is possible. In many cases it is not clear which variant should be labelled the exception, so the variants will be given more-neutral labels. The second part of the paper shows two ways that U-shaped distributions could arise: listing, and thresholding on a phonetic continuum. 2. Case studies of U-shaped distributions 2.1. Tagalog nasal substitution Zuraw (2010) examines Tagalog nasal substitution, using data from English’s 1986 dictionary. Undergoers of nasal substitution fuse a prefix nasal with a stem-initial obstruent, as in /maN+bigáj/ → [ma+migáj] ‘distribute’. Non-undergoers simply assimilate the nasal in place: /paN+diníg/ → [pan+diníg] ‘hearing’. (The non-un- dergoers could be labeled exceptions, but then again the rule applies only at the prefix-stem boundary, so it would also be reasonable to label the undergoers as exceptions.) If we tally up the number of words, across all phonological and mor- phological sub-cases, that undergo the rule, fail to undergo, or are listed with both variants in the dictionary, we derive a real U-shaped distribution, similar to the hypothetical one from (1): (2) Tagalog nasal substitution data from Zuraw (2010), six most-common mor- phological constructions combined 2.2. Tagalog tapping at prefix-stem boundary Zuraw (2009) examines tapping in Tagalog, using data from the written corpus described in Zuraw 2006. Tapping occurs in a few environments, but here we will 148 CatJL 15, 2016 Kie Zuraw look at only the prefix-stem boundary. If a stem begins with /d/ and receives a vowel-final prefix, that /d/ is now in the environment V__V, and can undergo tap- ping: /ma+dumí/ → [ma+ɾumí] ‘dirty’. Failure to undergo tapping is also common, as in /ma+dáhon/ → [ma+dáhon] ‘leafy’. Because the data here are from a corpus, we can ask not just if a word undergoes tapping or not, but how often it undergoes tapping, and plot the number of words that show each rate of tapping. Very-low- frequency items must be avoided, because they will look artificially polarized. For example, if a word occurs only once, its observed rate of undergoing tapping must be 0% or 100%, no matter what its true underlying rate of undergoing might be. Because there are many low-frequency items (the lower the frequency, the more items there are with that frequency), this inflates the 0% and 100% counts. The data shown here are limited to items that occur at least 10 times. As the histogram in (3) shows, there are a large number of words that consist- ently fail to undergo tapping (on the far left bin, rates of 0% to 10% tapping), and a large number that consistently do undergo tapping (far right, 90% to 100% tap- ping rate) (3) Histogram of tapping rates for prefixed words with frequency ≥ 10 Using corpus data also allows us to see how a word’s frequency is related to its tapping rate. In (4), we see histograms for five groups of words: those with frequency 2 to 4 (Group A, not included in (3)); frequency 5 to 9 (Group B, also not included in (3)); frequency 10 to 99 (C); 100 to 999 (D); and 1000 and up (E). Because the groups contain different numbers of words, the vertical axis is the percentage of words in each group that fall into each tapping-rate bin, rather than the raw number. Polarized Variation CatJL 15, 2016 149 (4) Histograms of tapping rates, broken down by word frequency We can see that the U shape remains in each frequency bin, although the bal- ance of non-undergoers (left arm) and undergoers (right arm) changes from bin to bin, with more and more undergoers as frequency increases. Returning briefly to Tagalog nasal substitution, the same corpus data can be used to get nasal-substitution rates for individual words (see Zuraw 2009 for limits on using written data for nasal substitution). The result, shown in (5), is similar: a U-shaped distribution in every frequency tranche, but with more undergoers at higher frequencies. (5) Histograms of nasal substitution rates (all prefixes), broken down by word frequency 2.3. English aspiration In English, word-initial voiceless stops are aspirated, regardless of stress, as in [pʰ]ercéption. The best way to describe the aspiration environment is probably prosodic-word-initial, because aspiration also applies after a productive prefix, again even if the stem’s first syllable is unstressed: mìs-[pʰ]ercéption (Zuraw & Peperkamp 2015). If misperception were a single phonological word, then the stem-initial /p/ would not be eligible for aspiration, even if it is syllable-initial (mis.per.cep.tion), because the syllable it initiates is unstressed. But if the stem perception forms its own prosodic word, with the prefix mis- adjoined to form a larger prosodic word, then the /p/ remains eligible for aspiration. Some words, however, act as monomorphemic, even if they are arguably still prefixed, such as mistake, with an unaspirated [t] (Ogden et al. 2000, Baker, Smith & Hawkins 2007, Smith, Baker & Hawkins 2012, Clayards, Hawkins & Gareth 150 CatJL 15, 2016 Kie Zuraw 2012). Mistake does contain the meaning of mis-, although not really the meaning of take, but nonetheless it is pronounced as though monomorphemic, with the /t/ unaspirated because it is not syllable-initial (mi.stáke, or perhaps an ambisyllabic s). The studies just cited compare in detail a small number of words predicted to have a clear prefixed or a clear monomorphemic behavior. Zuraw & Peperkamp (2015) conducted a production study of a large number of words (110), looking for factors that affect the variation, and for items that vary. There were up to 16 tokens of each word (one for each participant, with data occasionally discarded for disfluency or spirantization), yielding a good picture of cross-speaker variation rates. As shown in (6), there were many words that were consistently unaspirated (far left: disposed, disparities, mistakes, disturbing, and 36 others), a smaller num- ber that were consistently aspirated (far right: miscalculations, misquote, mistrial, discontinuity, and 22 others); in the middle, there were 44 items with variation (displacement, discourteously, discolored, and 41 others).1 (6) Histogram of aspiration rates in English prefixed words 2.4. Baroni’s Northern Italian s-voicing Baroni (2001) examines the rule of Northern Italian intervocalic s-voicing. Within monomorphemes, the rule is straightforward: while casa ‘house’ would be [ˈkasa] in other dialects of Italian, for Northern speakers it is [ˈkaza]. The rule is typically blocked at a prefix-stem boundary, as in a[s]immetrico ‘asymmetric’—that is, just as in English, stem-initial consonants behave as word-initial. But, also as in English, some prefixed words behave as though monomorphemic, and do undergo voicing: pre[z]upposizione ‘presupposition’. Baroni tests many factors that could affect this variation, through a production study in which 10 Northern Italian speak- ers read aloud 102 real words in sentence context, 5 times each. Plotting the data from Baroni’s appendix, we see in (7) that the largest group of words consistently 1. Percent aspirated here excludes tokens coded as “unsure”, and thus are different from the percent- ages shown in the appendix of Zuraw & Peperkamp (2015). Polarized Variation CatJL 15, 2016 151 resist voicing (rightmost column), the next-largest consistently undergo it (leftmost column), and there are a substantial number of words—27—in between. (7) Histogram of [s] rate in Baroni’s data 2.5. Tagalog noun/adjective order The above cases have all involved prefixed words. Now, we move to larger, multi- word units. In Tagalog, a noun and its modifying adjective can occur in either order, with a “linker” morpheme in between (Schachter & Otanes 1972): maganda-ng babae (beautiful-linker woman) and babae-ng maganda (woman-linker beautiful) both mean ‘beautiful woman’. Shih & Zuraw (2014, submitted) examine phonological and other factors that influence the order variation, using data from the Zuraw (2006) corpus. The data are polarized: as shown in (8), at every frequency tranche most adjective/noun pairs consistently have the order adjective-linker-noun, with a smaller group of consistent noun-linker-adjective pairs (far left column in each histogram). (8) Histograms: rate of taking adjective-linker-noun order 2.6. French h aspiré In French, certain function words and adjectives undergo changes before a vow- el-initial word, as in /lə eʁitaʒ/ → [l eʁitaʒ] ‘the heritage’, where the determiner 152 CatJL 15, 2016 Kie Zuraw loses its final vowel. There are some vowel-initial words that exceptionally block these changes, as in /lə eʁisõ/ → [lə eʁisõ] ‘the hedgehog’; because most of these words are spelled with h, they are known as “h-aspiré” words. They have been subject to various analyses, but several phenomena can be covered at once if the words are treated as requiring a syllable boundary at their left edge (Frei 1929, Damourette & Pichon 1930, Malmberg 1943, Tranel 1981, Cornulier 1981, Tranel & Del Gobbo 2002, one possibility presented in Boersma 2007). Zuraw & Hayes (submitted) examine variation in h-aspiré, using data from the Google n-grams corpus (Michel et al. 2011, Lin et al. 2012), which compiles frequencies of word sequences from digitized books. Zuraw and Hayes look only at words that are likely to show h-aspiré behavior: words that begin with a glide sound (yacht ‘yacht’); words that begin with a vowel sound and are spelled with initial h (héritage or hérisson); and a few words that begin with a vowel sound and are spelled with a vowel letter, but are claimed in dictionaries and prescriptive sources to behave as h-aspiré (uhlan ‘uhlan [type of soldier]’). Even within this group, the distribution is highly polarized at every frequency tranche, as shown in (9). The left arm of each U is words that are consistently h-aspiré, and the right arm is words that are consistently regular. (9) Histogram of rate of regular (non-h-aspiré) behavior 2.7. French de omission in coordination Just as in English, French de ‘of’ can be omitted in coordination: morceaux de carottes et (de) tomates ‘pieces of carrots and (of) tomatoes’ (Tseng 2003, Abeillé et al. 2004, Abeillé et al. 2005, Abeillé et al. 2006). Unlike in English, however, it is more common to retain the second de: (10) Two options for French de more common: de is repeated a. morceaux [de carottes] et [de tomates ] b. pieces [of carrots ] and [of tomatoes] less common: first de takes wide scope / second de is omitted c. morceaux de [carottes et tomates ] d. pieces of [carrots and tomatoes] Polarized Variation CatJL 15, 2016 153 Zuraw (2015), again using data from Google n-grams, examines phonological factors that influence de omission, using simple cases of the form de X et (de) Y ‘of X and (of) Y’ and de X ou (de) Y ‘of X or (of) Y’, where X and Y are single words. The strongest factor is the initial sounds of X and Y: if either X or Y starts with a vowel (h-aspiré words are excluded here), the second de is more likely to appear. As the histograms in (11) show, most X-et-Y and X-ou-Y triples consist- ently retain the second de, and thus are counted in the rightmost column of their histogram. In the top two rows (VER = verbs, NOM = nouns), are no triples that consistently drop the second de. The effect of beginning with a consonant or vowel is seen mostly in the balance of variers vs. consistent retainers. If both X and Y begin with a consonant (C C), then there are more variers than if one begins with a vowel (C V or V C), or if both begin with vowels (V V). This is thus a one-sided polarization, with only one arm of the U present. But in the bottom row (ADJ = adjective), where rates of de retention are lower over all, instead of just seeing more and more variers, we begin to see a small group of triplets that consistently omit the second de (the small columns in the far left of each ADJ histogram), outnumbering triples that retain at low rates. (11) Histograms of rate of retaining second de 2.8. English binomials English coordinated pairs of words have become known as binomials. As Mollin (2014) discusses, some have a fixed order: law and order is common, and is also the name of a television program; ??order and law, even outside the context of television, sounds strange. But others vary freely, such as knowledge and skills/ skills and knowledge. Mollin points out that cases of total irreversibility are rare; in her corpus data, it was more common for a pair to have a certain order 90% to 99.99% of the time than to have a certain order 100% of the time: 154 CatJL 15, 2016 Kie Zuraw (12) Number of two-word pairs, adapted from Mollin’s Figure 3.5, p. 45 (In the cases examined until now, there were two well-defined behaviors to examine: aspirated vs. unaspirated; noun-linker-adjective order vs. adjective-linker- noun order, etc. Here, there are no such labels available: does law and order belong in the same category as knowledge and skills, or as skills and knowledge? There seems to be no reasonable answer to this question. Therefore, Mollin simply asks, for each pair, how frequent the more-frequent order is. If the answer is 50%, the pair exhibits maximal variation; if 100%, minimal variation. The most polarized possible pattern, then, is one in which there is a tall bar at the right of the plot, but, unlike in the plots we have seen so far, there is not a tall bar at the left of the plot.) The bar plot in (12) is not a typical histogram, however, because the horizontal axis is nonlinear: most of the bins cover 10 percentage points, but the rightmost covers only 0.01 percentage points. To make a histogram comparable to the ones used for the other cases here, we add the items on the far right to the adjacent bar: (13) Histogram version of (12) The result is strongly polarized: most of Mollin’s pairs strongly favor one order, even if they do not show that order 100% of the time. Polarized Variation CatJL 15, 2016 155 2.9. Other cases The cases examined here have been chosen opportunistically, mainly by going through my own previous and ongoing research, looking for cases where big enough numbers were available. I know of two other cases currently under investigation: Yanovich’s (2014) work on Old High German ð and d, and Smith’s (in preparation) reexamina- tion of Hilpert’s (2008) data on English comparatives. English adjectives can form their comparative as either more X or Xer (more proud vs. prouder, to use Hilpert’s example). In Hilpert’s data, lower-frequency adjectives (frequen- cy < 100) show rather free variation, but higher-frequency adjectives (≥100) show a strong tendency to take –er, with a smaller group consistently taking more—this group only barely outnumbers the adjacent group of variers, though (Brian Smith, p.c.). 2.10. Within-speaker variation? One worry for the data discussed here is whether we see any true free (token) vari- ation at all, or just a mix of lexical (type) variation. Consider this scenario: suppose that for most items, speakers agreed on whether each item should have one behavior or the other; these items appear in the outer arms of the U-shaped histogram. Then suppose that for a few items, some speakers have one variant listed in their mental dictionaries, and some speakers the other; these are the items with intermediate observed rates of behavior. In that case, the only thing we need to explain is why there are few items with middle rates—that is, why speech communities are gener- ally good at agreeing on one pronunciation. In most of the cases here, we lack the within-speaker data we would need to assess this scenario, either because the data come from a multi-writer corpus, with individual authors not distinguished (the Tagalog and French cases), or because we have individual data but with only one token of each item from each speaker (English aspiration). In the Northern Italian s-voicing case, Baroni elicited five tokens of each item from each speaker, but does not report data broken down by speaker in his appen- dix. He does, in discussing one item, coseno ‘cosine’, give a breakdown for that item though: 2 speakers consistently said co[z]eno, 4 consistently said co[s]eno, and 6 had variation.2 In the French case, although the corpus data offer no insight into individual variation, h-aspiré has attracted the attention of so many grammarians and linguists that the literature does have anecdotal examples of within-speaker or within-writer variation. Cohen (1963) mentions a flyer with de un mois and d’un mois ‘of one month’ occurring two paragraphs apart; de un treats un ‘one’ as aspiré, while d’un does not. Cohen also mentions a history manual that similarly varies between de 2. Baroni reports here the data for 12 speakers; the data shown in (7) are for just the 10 speakers whose results were strongly correlated with one another. The data remain similarly polarized if the results for all 12 speakers are plotted. 156 CatJL 15, 2016 Kie Zuraw Henri IV and d’Henri IV ‘of Henry IV’. Grevisse & Goosse (2011), a widely used prescriptive manual, mentions several examples of variation within French authors (pp. 57-63). Cornulier (1981), drawing on his own impressions, states that “a sen- tence containing both de Hugo and d’Hugo would not be remarkable in conversa- tion” (p. 203; translation mine). He also remarks that “many people supposed to speak well seem to flip a coin each time” when it comes to pronouncing a final [n] in the article un ‘a’ in un handicap ‘a handicap’, un HLM ‘a social housing project’, or un hameçon ‘a hook’, “but, when questioned, decide in a definitive and sincere fashion, like the grammarians” (p. 203). Pronouncing un with a final [n] (liaison) means treating the vowel-initial following word as regular, while pronouncing with without [n] means treating it as aspiré. Gabriel & Meisenburg (2009) find intra- speaker variation in a reading-aloud task. Although firm data are scant, then, it seems that within-speaker variation is real, at least in some cases. The next section of this paper asks how polarized (and non- polarized) patterns of variation can arise when there is within-speaker variation. 3. Modeling U-shaped (and non-U-shaped) distributions 3.1. Not all distributions are polarized: English t/d deletion, Tagalog o/u alternation There are cases of variation that are not polarized. Coetzee & Kawahara (2013) examine English /t/ and /d/ deletion in words with final consonant clusters, such as sexist and girlfriend. Coetzee and Kawahara use spoken data from the Buckeye corpus (Pitt et al. 2007); their online appendix makes it possible to compile the rate of deletion for each word. The histograms in (14) show that the distribution of deletion rate, in each frequency tranche, is actually rather unimodal (except for the lowest-frequency items, which can show spurious U-shapedness, as discussed in Section 2.2), with the greatest number of words showing intermediate rates of deletion, and smaller numbers showing extremely high or extremely low rates. (14) Histograms of English t/d deletion rates in Coetzee & Kawahara (2013) broken down by frequency Data from Zuraw (2009) on Tagalog o/u alternations show a mixed pattern. Tagalog tends to have [o] in final syllables and [u] in non-final syllables, though Polarized Variation CatJL 15, 2016 157 contrast in loans and partial productivity of the alternation impose many compli- cations. A simple example of an alternation would be [tákot] ‘fear’, [takút-in] ‘to frighten-patient focus’. When a two-syllable word such as [háloʔ] ‘mixture’ is reduplicated, the end of the first copy is pulled in two directions: on the one hand, it is not final in the syntactic word, and thus should have [u] rather than [o] ([hálu- háloʔ] ‘a dessert of mixed fruits, ice, and sweets’]); on the other hand, it is arguably final in a prosodic word, and also corresponds to a word-final syllable, and on both counts should have [o] ([hálo-háloʔ]). Both variants do tend to occur: [hálo-háloʔ] ~ [háluháloʔ]. If we compile the rate of having (written) [u] in the first copy for each reduplicated word in the corpus, the pattern is U-shaped for lower frequencies and flat for higher frequencies: (15) Histograms of Tagalog o-raising rates in reduplication, broken down by fre- quency Not all distributions are polarized then. This section considers two mechanisms by which polarization could arise, then how non-polarized distributions could arise under the same mechanisms. 3.2. Polarization through lexical diffusion Reali & Griffiths (2009) show how polarized distributions can be obtained if agents have a prior expectation of polarization. Can we obtain a U shape without such a prior bias? One possibility is to assume that what we observe in the polarized cases discussed above is a snapshot of the process of lexical diffusion (e.g., Wang 1969, Bybee 2001, Phillips 2006), a process by which words are moving, one by one, from one category to another, such as from exceptional to regular. 3.2.1. A simple model We will begin with a very simple model and gradually build in more realism, though the final model will still be quite simplified. Suppose that each word changes abruptly—that is, there is only a short transi- tion period for each word during which it displays variation (we will return within this subsection to why the transition period should be short). Suppose further that different words transition at different times. In our simple model there are 240 words, transitioning from a conservative variant to an innovative variant over the course of 24 time units, which could 158 CatJL 15, 2016 Kie Zuraw perhaps be thought of as decades for a slow change, or years for a fast change. Each word’s rate of showing the innovative variant begins near zero, then quick- ly increases and tops out near 1 (using a logistic function, rate(time, intercept) = 1 / (1+e-(k*time*intercept)) ). The words begin their transitions at different times though (each word has its own value for intercept). The plot in (16) shows the behavior of individual words over time, but only for every 10th word, so that the individual traces can be distinguished. (16) Increase in innovative behavior for each word over time (only every 10th word shown) At each of the 24 time-steps, we can make a histogram like the ones in the case studies above: that is, we can compile the number of words that show each rate of innovative behavior. The histograms in (17), which plot every second time-step, show that at the beginning, nearly all the words consistently show the conservative behavior; over the time, more and more words consistently show the innovative behavior, until nearly all do. The key is that each word spends only a short time in the middle bins, moving quickly from the leftmost bin to the rightmost bin; thus, the middle bins are never very populated, and the distribu- tion is always U-shaped, though the relative heights of the U’s two arms change over time. Polarized Variation CatJL 15, 2016 159 (17) Histograms across time of rate of innovative behavior In the model above, we simply stipulated that each item changes abruptly (and thus spends little time in the middle bins of the histogram). Can we derive this behavior instead of stipulating it? Starting with Bloomfield (1933), linguists have noted that we can derive “S-shaped” change—change starts slowly, then is fast, then slows again, as shown for each word in (16)—if we adopt an epidemiologi- cal model. A speaker is infected by an innovation when she or he hears it from another speaker. Change is slow at first, because there are few innovators who can infect others, and change is slow at the end because almost everyone has already been infected. A simple simulation illustrates this. There are 10,000 speakers. At first, just one is innovative. At each time-step, speakers interact in randomly chosen pairs. If both are innovators or both are conservative, no change occurs; but if they mismatch, the innovator infects the non-innovator, who is henceforth an innovator too. The number of innovators in the population grows over time as shown in (18). 160 CatJL 15, 2016 Kie Zuraw (18) S-shaped growth in number of innovating speakers for an item The above model is one where variation is only across speakers. To model the possibility of variation within each speaker, we allow each agent to have two listed forms (this is Baroni’s 2001 proposal for Northern Italian s-voicing). Production depends on the strength of each form. Agents again meet in pairs. If one produces the innovative variant, the other increments its probability of producing the inno- vative form by 10 percentage points (unless it is already at 100%). The simulation begins with one agent at 10% innovative, and the rest at 0% innovative. Although it takes longer for change to spread, the spread is again (asymmetrically) S-shaped: (19) S-shaped growth in overall population rate of producing innovative variant The final adjustment to the model here is to allow not just the innovative form to be incremented, but also the conservative form. Specifically, an agent increments its innovative-form strength by 0.1 when it hears the innovative form, and incre- ments the conservative form’s strength by 0.0015 when it hears the conservative Polarized Variation CatJL 15, 2016 161 form. It is crucial that the conservative form be less amenable to strengthening; if the conservative increment is too large, change never takes off: (20) Growth in overall population rate of producing innovative variant when both variants can be incremented The model presented here was very crude, for expository purposes. More- serious models of language change have incorporated elements such as social net- work structure, or a more-sophisticated decision rule for speakers (Niyogi 2009, Pierrehumbert, Stonedahl & Daland 2014, and many others). 3.2.2. Why should an item change? The model of change just developed stipulates an advantage for an innovative form: listeners are more likely to adopt, or increase their tendency to use, an innovative form they hear than they are to adopt a conservative form that they hear. Why should this be? Two answers have been proposed. Sometimes one form has an inherent advantage over the other. Mollin (2014) argues that this is often the case for English binomials. One form could have a phonological advantage over the other, such as a better rhythm: párt and párcel has an alternating strong-weak-strong-weak rhythm, while ??párcel and párt has a lapseful strong-weak-weak-strong rhythm (see also Shih 2014). Or one form could have a processing advantage, such as greater retrievability of the first coor- dinand because of frequency, saliency, or recent occurrence (Benor & Levy 2006). A linguistic or psycholinguistic advantage then helps a form become the frozen form, Mollin proposes. In terms of the model above, this could be implemented by having the probability of producing part and parcel vs. parcel and part be a function of both the strength of the listed entry for each order and the inherent probability of each order. It is also possible, though, that innovation can be preferred for its own sake. As part of a series of simulations, Pierrehumbert et al. (2014) allow some speak- ers to be biased to prefer novelty, and others to shun novelty. The balance of such 162 CatJL 15, 2016 Kie Zuraw agents in the population will of course affect the spread of an innovation, but so can their distribution. Pierrehumbert et al. find that innovations are especially likely to spread if the agents socially connected to an innovator themselves pre- fer innovation. As Pierrehumbert et al. discuss, this may be linked to individual personality/cognitive traits and to tendencies for individuals with similar traits to affiliate socially (pp. 18-19). (See also Yu 2010, Stevens & Harrington 2014 on experimental findings with implications for the role of individual differences in sound change.) 3.2.3. What is the starting point for change from one category to another? The starting point in the simulation above was that all items had the same behavior (the conservative variant); change then consisted in items’ moving to the other category. This subsection considers starting points that seem plausible for the case studies presented above. Tagalog tapping, English aspiration, and Northern Italian s-voicing (sec- tions 2.2, 2.3, 2.4) seem to be in line with an all-conservative starting point. They involve an alternation that is blocked at the beginning of a prosodic word, including at the prefix-stem boundary. This is shown as Stage 1 in (21), using the Tagalog example. Some words, especially frequent words, could get reanalyzed as monomorphemic, or at least treated as not quite productively prefixed; such a reanalysis would both promote and be promoted by semantic drift from the unprefixed form (see Hay 2003). (21) Genesis of exceptional tapping in Tagalog Stage 1 Stage 2 Unprefixed: tapping [áɾaw] ‘day’ same as before [áɾaw] Prefixed: no tapping [ma+dáhon] ‘leafy’ same as before [ma+dáhon] Prefixed: no tapping [ma+dumí] ‘dirty’ not quite prefixed: tapping [maɾumí] The starting point then is of most items’ being consistently conservative (tap- ping is blocked). Some items could even have already been reanalyzed as mono- morphemic before tapping even arose, so that they were exceptions from the begin- ning. Individual words’ analyses would change over time, crossing from prefixed to monomorphemic and perhaps the other direction too. These changes would have had to be abrupt, along the lines sketched in section 3.2.1, so that varying items are always rarer than consistent items. In French h-aspiré (section 2.6), simplifying and idealizing somewhat, the starting point was that all items were regular. Consonant-initial words behaved as consonant-initial, and vowel-initial words as vowel-initial. This is shown as Stage 1 in (22), where consonant-initial words take the full article [la], and vowel- initial words take the form [l] to avoid a vowel-vowel sequence. The consonant /h/ then deleted, stranding some previously consonant-initial words as exceptional (Stage 2). Polarized Variation CatJL 15, 2016 163 (22) Genesis of French h-aspiré (idealized), using the example of la ‘the-feminine’ Stage 1 Stage 2 C-initial: no deletion [la tabl] ‘the table’ same as before [la tabl] C-initial: no deletion [la haʃ] ‘the axe’ exception: no deletion [la aʃ] V-initial: deletion [l arbr] ‘the tree’ same as before [l arbr] French thus had a polarized starting state, with most vowel-initial words being consistently regular, and others (the h-aspiré words) being consistently excep- tional. Vowel-initial items could then change category through regularization and hypercorrection, and new words entering the language would have to be assigned to a category. English binomials present a third scenario: there is no clear starting point for an individual item, no reason why either books and papers or papers and books should have been the original order. The starting point here is probably free variation: any new coordination can occur in either order, though perhaps with a preference for one order if phonological, semantic, and processing factors favor it. Because these factors would be just as likely to conflict as to conspire to favor one variant strongly, it seems likely that most items would start out in the intermediate range. But, as a binomial begins to be lexicalized, it would move towards one arm of the U (this type of change was not implemented in section 3.2.1). To summarize section 3.2, we can derive polarized, U-shaped distributions by assuming that individual items change category abruptly. Each item spends little time in the intermediate range, so that there are always few variers. In order for this model to work, most of the items in question must be listed. This is not too contro- versial for affixed words, but is more surprising for the longer units: Tagalog noun/ linker/adjective vs. adjective/linker/noun (2.5), French coordinations with de (2.7), and English binomials (2.8), though the English binomial literature does typically assume that many items are listed (see Mollin 2014 for overview). 3.3. Polarization through thresholding The question of how to derive a non-polarized distribution still has not been addressed. This section discusses how both polarized and non-polarized distribu- tions could arise from variation along a phonetic continuum. Suppose that Tagalog nasal substitution began (in a proto-language) as variable deletion. For example, the proto-language equivalent of /maN+bilí/, originally pro- nounced with simple nasal place assimilation as [mambilí], began to be pronounced as [mamilí]. It seems plausible that variation along these lines represents variation in gestural timing: the duration of the nasal gesture remains constant, but that of the labial closure shortens, producing a phonetic continuum from [mb] to [m]. When the duration of the [b] portion becomes short enough, /b/ is perceived as deleted, as the word could subsequently be lexicalized as /mamilí/. (Other phonetic changes would be relevant too: Beddor & Onsuwan 2003 show that for Ikalanga speakers 164 CatJL 15, 2016 Kie Zuraw differentiating between sequences like [amba] and [ama], carryover nasalization is even more important as a cue than oral closure duration.) We can simulate this situation with 1,000 words, each having a baseline preferred duration of non-nasal closure (that is, the [b] portion of [mb], the [t] portion of [nt], etc.). These baselines could be related to word frequency, stress, place of articulation, perhaps even other nearby segments, and other factors. These baseline preferred durations are normally distributed, as shown in (23)a, centered on the value 2 in arbitrary time units, with a standard deviation of 1. Each item is uttered 200 times, adding a small amount of Gaussian noise to the duration each time (mean 0, s.d. 1, divided by 0.5). The value of the noise can be positive or negative, so the oral stop duration can be longer or shorter than is usual for that word. The resulting observed durations, shown in (23)b, remain strongly unimodal. (23) Simulation of variable stop deletion (a) Words’ baseline preferences (b) Tokens’ observed oral-stop for oral-stop duration durations To convert these observed durations into observations of deletion or non- deletion, we can impose a duration threshold under which the oral stop will be perceived as deleted, as shown in (24)a, which repeats (23)b, adding a line at 1.5 to mark the chosen threshold. The result in this case is that, although the tokens as a whole are distributed in a bell curve, most words’ tokens are mostly on just one side of the threshold. For example, if a word’s baseline preferred duration is 2.5, with the small amount of noise we added, only about 2% of its tokens will fall below the threshold of 1.5, and thus this word will contribute to the leftmost bin of the histogram in (24)b (consistent non-deletion). Polarized Variation CatJL 15, 2016 165 (24) Thresholded perception of stop deletion (a) (23)b, with threshold shown (b) Histogram of words’ deletion rates If listeners and learners had a threshold along these lines for deciding whether deletion had applied in words like [mambili] ~ [mamili], then each word would have a tendency to get lexicalized either with or without deletion, rather than as varying. The scenario proposed for nasal substitution seems very much like English t/d deletion, though, which plausibly results from overlap by neighboring consonant gestures, and/or shortening of the t/d gesture itself.3 Why is the distribution of t/d deletion in the Buckeye corpus unimodal then, instead of U-shaped? The U shape in (24) depended on having the noise added to durations in pro- duction be small relative to how spread out the baseline duration preferences were. If noise is bigger, or if the baseline durations are clustered closer to the threshold, then things change. The plots in (25) show what happens if we repeat the simulation but multiply the Gaussian noise by 2 instead of dividing it by 2. In (25)a we have the distribution of baseline preferred duration (as for nasal substitution, mean is 2, standard deviation is 1); (25)b shows the distribution of observed durations, with the threshold again marked at 1.5. The distribution is more spread out in (25)b than it was in (23)b/(24)a, because we have added more noise to each token. Finally, (25)c shows the distribution of perceived deletion rate per word. It is unimodal, like Coetzee and Kawahara’s (2013) t/d data, because most words have a good chance of having a healthy number of tokens on both sides of the threshold. To take again the example of a word with a baseline preferred duration of 2.5, now about 31% of its tokens will fall below the threshold. 3. Compare British English /t,d/, where Temple (2014) documents acoustic continua of release strength, constriction degree, and glottalization, among others, though duration was not examined; or Mitterer & Ernestus’s (2006) similar findings for Dutch /t/ (gradience in amplitude reduction as compared to surrounding sounds). 166 CatJL 15, 2016 Kie Zuraw (25) Larger noise creates a unimodal distribution (a) (b) (c) (There is at least one example of an English word whose t-deletion rate is so high that it has been lexicalized by many speakers without the t: [tɛks(t)] ‘text’, in the sense of ‘mobile-phone text message’ or ‘to send a text message’. For these speakers, we can tell that the underlying form is /tɛks/ rather than standard /tɛkst/, because the plural or third-person singular is [tɛksɨz] rather than standard [tɛks(t)s], and the past tense is [tɛks(t)] rather than [tɛkstɨd].) We can thus get either a unimodal or a bimodal distribution from the same basic scenario, by changing how much variation there is in production. We could get the very same result by varying how much noise there is in the perceiver or transcriber’s threshold. That is, instead of holding the threshold constant at 1.5, we could allow it to vary from token to token; the more it varies, the more chance a word’s tokens have of falling on the other side of the threshold, and the more unimodal the distribution will be. 3.4. Summary of models This section has presented two ways to achieve a bimodal (polarized, U-shaped) distribution of item behavior, and one way to achieve a unimodal (bell-shaped) dis- tribution. The decision tree in (26) schematizes the decision. The first branch of the tree asks what the nature of the variation is. In some cases it is clearly categorical, such as in English binomials: there is no phonetic continuum from fruits and vegetables to vegetables and fruits; similarly for Tagalog noun/adjective order and French de omission (though in that last case there could in principle be a continuum). In other cases, the phonetics are unclear or unknown. As far as I know, Tagalog nasal substitution is categorical—the oral stop is either present or not—but this has not been studied rigorously. Kevin Ryan examined tapping in a corpus of Tagalog (Guevara et al. 2002); he found that there were two sharp categories along the duration dimension: [ɾ] (short duration) and [d] (long duration), with very few tokens in between, so Tagalog tapping is probably on the categorical branch (Zuraw & Ryan 2007). Baroni (1998) finds that, on both acoustic and electroglottographic grounds, Northern Italian [s] and [z] form two Polarized Variation CatJL 15, 2016 167 separate categories. As mentioned above, English t/d deletion probably belongs on the continuous branch. Tagalog o/u variation may also belong on this branch; anecdotally, speakers and transcribers are often unsure whether a Tagalog vowel is [o] or [u]. Detailed articulatory study may even reveal a mixed pattern: Ellis & Hardcastle (2002) found two English speakers who produced an articulatory con- tinuum from no place assimilation (re[d] coat) to full assimilation (re[ɡ] coat), but also two who produced only “full alveolars or complete assimilations” (p. 373).4 (26) Summary of models variation type categorical phonetic continuum degree of noise in production or in perceptual threshold small large U shape (i.e., bell shape polarized, bimodal) (i.e., unimodal) 4. Summary and conclusion Section 2 presented several cases of U-shaped variation: a tendency for each item to show a consistent behavior rather than an intermediate rate of variation. That is, most of the variation here was type variation, though all cases did have at least some admixture of token variation. Section 3 showed that listing of two options produces a U-shaped distribution, if each item moves from one variant to the other abruptly. When the two variants are categorically different, we saw a plausible mechanism for abrupt change. One consequence of this model was that it required us to posit listing for fairly long units, such as French de X et/ou (de) Y sequences or Tagalog adjective/noun pairs (or English binomials, where listing is widely acknowledged). When variation is along a phonetic continuum, two outcomes were derived. If noise in producing tokens is small (compared to the variance of baselines across items), the result is U-shaped, but if token noise is large, the result is bell-shaped. 4. In Bybee’s (2012) classification of diachronic changes as sound change or phonological change, changes that are articulatorily gradual will eventually become lexically regular, but while the change is in progress, lower-frequency words will tend to resist it. Changes that are articulatorily abrupt can leave behind high-frequency lexical exceptions. It is hard to say whether the cases reviewed here accord with this typology, because they involve the complicating factor of whether complex structures have become listed as units, which is more likely to occur for more-frequent items. 168 CatJL 15, 2016 Kie Zuraw The general prediction that emerges is that when variation is categorical and is over listable units, U-shaped distribution will tend to result. When variation is continuous, either a U-shaped or a bell-shaped distribution is possible. References Abeillé, Anne, Bonami, Olivier, Godard, Danièle & Tseng, Jesse. 2004. The syntax of French de-N’ phrases. In Müller, Stefan (ed.). HPSG 2004: Proceedings of the 11th International Conference on Head-Driven Phrase Structure Grammar, 6-26. Stanford: CSLI Publications. Abeillé, Anne, Bonami, Olivier, Godard, Danièle & Tseng, Jesse. 2005. Les syntagmes nominaux français de la forme de-N’. Travaux de linguistique 50: 79-95. Cornulier, Benoît de. 1981. H-aspirée et la syllabation: Expressions disjonctives. In Goyvaerts, Didier L. (ed.). Phonology in the 1980s, 183-230. Ghent: E. Story- Scientia. Damourette, Jacques & Pichon, Édouard. 1930. Des mots à la pensée: essai de gram- maire de la langue française. Vol. 1. Paris: Éditions d’Artrey. Ellis, Lucy & William J. Hardcastle. 2002. Categorical and gradient properties of assim- ilation in alveolar to velar sequences: evidence from EPG and EMA data. Journal of Phonetics 30(3): 373-396. Ogden, Richard, Hawkins, Sarah, House, Jill, Huckvale, Mark, Local, John, Carter, Paul, Dankovičová, Jana & Heid, Sebastian. 2000. ProSynth: an integrated pro- sodic approach to device-independent, natural-sounding speech synthesis. Computer Speech & Language 14(3): 177-210. Wang, William S.-Y. 1969. Competing Changes as a Cause of Residue. Language 45(1): 9-25. Coberta CJL 15.pdf 2 24/10/16 12:21 Catalan Journal of Linguistics Editors Centre de Lingüística Teòrica de la Universitat Autònoma de Barcelona Institut Interuniversitari de Filologia Valenciana Catalan Journal of Linguistics style sheet Informació general i subscripcions General information and subscriptions • Document format. Manuscripts should be should be listed chronologically, with the CATALAN JOURNAL OF LINGUISTICS és una revista de lingüís- CATALAN JOURNAL OF LINGUISTICS is a journal of theoretical written in English. Font 12 point Times New oldest publication at the top and the newest at tica teòrica, publicada pel Centre de Lingüística Teòrica linguistics published by the Centre de Lingüística Roman, interline at 1.5, use margins 2,5 cm the bottom. Please, provide the DOI of articles de la Universitat Autònoma de Barcelona (format per Teòrica de la Universitat Autònoma de Barcelona (1 inch). Phonetic symbols must be set in if available. lingüistes dels departaments de Filologia Catalana i de (comprising linguists from the departments of Catalan Doulos SIL or Charis SIL. a) Book: Filologia Espanyola) i per l’Institut Interuniversitari Philology and Spanish Philology) and by the Institut • Heading. Title, with title capitalisation, in Borer, Hagit. 1983. Parametric Syntax. de Filologia Valenciana (que integra lingüistes de les Interuniversitari de Filologia Valenciana (which boldface. Name of the author(s), affiliation(s) Dordrecht: Foris. universitats de València, d’Alacant i Jaume I de Castelló). comprises linguists from the universities of València, and e-mail address(es). Acknowledgements, DOI: http://dx.doi.org/10.1515/97831 CATALAN JOURNAL OF LINGUISTICS publica articles d’inves- Alacant and Jaume I at Castelló). as a footnote, called by a flying asterisk. 10808506 tigació en lingüística teòrica que desenvolupin teories CATALAN JOURNAL OF LINGUISTICS publishes research papers sobre l’estructura de les llengües particulars dins la in theoretical linguistics developing theories on the • Abstracts. A 150 word abstract is required. b) Journal article: perspectiva d’una teoria general del llenguatge humà. structure of particular languages within the perspective A Catalan version of the abstract should be Kayne, Richard S. 1993. Toward a modu- Els volums seran sovint monogràfics, editats per un of a general theory of the human language. Volumes will included. Authors are also asked to provide lar theory of auxiliary selection. Studia especialista en el tema. often be monographic, edited by a specialist in the topic. the keywords of the article. Linguistica 47: 3-31. € € DOI: http://dx.doi.org/10.1111/j.1467- El preu de subscripció anual és de 18 . The yearly subscription price is 18 . • Sections and paragraphs. Headings of 9582.1993.tb00837.x main sections: in boldface. Headings of first Comitè assessor Redacció order paragraphs: in italics. Headings of c) Book article: Josep Maria Brucart Marraco Universitat Autònoma de Barcelona second order paragraphs or others: in italics Jaeggli, Osvaldo & Safir, Ken. 1989. The (Universitat Autònoma de Barcelona) Departament de Filologia Catalana with text beginning at the same line. Null Subject Parameter and Parametric Maria Teresa Espinal Farré 08193 Bellaterra (Barcelona). Spain Theory. In Jaeggli, Osvaldo & Safir, Phone: +34 93 581 23 71 (Universitat Autònoma de Barcelona) • Text. The beginning of a paragraph should Ken (eds.). The Null Subject Parameter, C +34 93 581 23 72 Anna Gavarró Algueró be left indented, except for the first paragraph 1-44. Dordrecht: Kluwer. (Universitat Autònoma de Barcelona) Fax: +34 93 581 27 82 of a section. Words, phrases or sentences in a DOI: http://dx.doi.org/10.1007/978-94- M Maria Lluïsa Hernanz Carbó E-mail: [email protected] language other than English appearing in the 009-2540-3_1 Y (Universitat Autònoma de Barcelona) Intercanvi body of the text should be followed by a Sabine Iatridou (Massachusetts Institute of Technology) Universitat Autònoma de Barcelona translation in simple quotation marks. d) Dissertation, manuscript, or work in CM Michael J. Kenstowicz Servei de Biblioteques References should follow this format: preparation: (Massachusetts Institute of Technology) Secció d’Intercanvi de Publicacions Pesetsky, David. 1982. Paths and MY Chomsky (1981), Kayne (1981, 1982), Belletti Joan Mascaró Altimiras Plaça Cívica (edifici N) & Rizzi (forthcoming), Jackendoff (p.c.). If Categories. PhD dissertation, MIT. (Universitat Autònoma de Barcelona) 08193 Bellaterra (Barcelona). Spain Chomsky, Noam. 1994. Categories and CY the page number or an example is expressed: Gemma Rigau Oliver Phone: +34 93 581 11 93 Rizzi (1990: 20), Rizzi (1990: ex. (24)). When Transformations. Unpublished manus- CMY (Universitat Autònoma de Barcelona) Fax: +34 93 581 32 19 the reference is included between parentheses, cript, MIT. Henriette de Swart (Universiteit Utrecht) E-mail: [email protected] K • Publication rights. CATALAN JOURNAL OF Rafaella Zanuttini (Yale University) drop the ones enclosing the year: (see Rizzi Subscripció, administració, edició i impressió 1990: 20). LINGUISTICS is published under the licence Consell editor Universitat Autònoma de Barcelona • Examples. Example numbers should go system Creative Commons, according to the Teresa Cabré Monné Servei de Publicacions modality «Attribution-Noncommercial (by-nc): (Universitat Autònoma de Barcelona) 08193 Bellaterra (Barcelona). Spain between parentheses, left aligned (Arabic digits for examples in the text; lower case derivative work is allowed under the condition Ángel Gallego Bartolomé Phone: +34 93 58110 22. of non making a commercial use. The original (Universitat Autònoma de Barcelona) Fax: +34 93 5813239 Roman for examples in footnotes). Example work cannot be used with commercial Jesús Jiménez Martínez (Universitat de València) E-mail: [email protected] texts should start at the first tab. If there is Sandra Montserrat Buendía (Universitat d’Alacant) http://publicacions.uab.cat more than one example within a number, purposes». Therefore, everyone who sends a Xavier Villalba Nicolás ISSN 1695-6885 (in press); ISSN 2014-9719 (online) lower case alphabet letters should be used manuscript is explicitly accepting this publica- (Universitat Autònoma de Barcelona) Dipòsit legal: B-4.112-2003 followed by a stop and a tab. Examples must tion and edition cession. In the same way, Imprès a Espanya. Printed in Spain be referred to as (1), (5a,b), (6b-d), (ii), (iii he/she is authorizing CATALAN JOURNAL OF Imprès en paper ecològic c-e), (iv a,b), etc. LINGUISTICS to include his/her work in a journal’s issue for its distribution and sale. The Bases de dades en què CATALAN JOURNAL OF LINGUISTICS està referenciada • Glosses. Examples from languages other cession allows CATALAN JOURNAL OF LINGUISTICS — Bibliographie Linguistique — Linguistic Bibliography than English should be accompanied by a to publish the work in a maximum period of two — CARHUS Plus+ — Linguistic Working Papers Directory gloss and, if necessary, by a translation into years. With the aim of favouring the diffusion — CiteFactor — Linguistics Abstracts English. Align glosses with examples word of knowledge, CATALAN JOURNAL OF LINGUISTICS — Dialnet — LLBA (Linguistic and Language Behavior Abstracts) by word. Translations should be in simple joins the Open Access journal movement — Francis — Scopus quotation marks. — Índice Español de Ciencias Sociales (DOAJ), and delivers all its contents to different y Humanidades (ISOC) • References. References should be according repositories under this protocol; therefore, to the following pattern, with full names of sending a manuscript to the journal also entails CATALAN JOURNAL OF LINGUISTICS es publica sota el sistema de llicències Creative Commons segons la modalitat: authors included. Reference entries for the explicit acceptation by its author/s of this Reconeixement - NoComercial (by-nc): Es permet la generació d’obres derivades multiple works by the same author(s)/editor(s) distribution method. sempre que no se’n faci un ús comercial. Tampoc es pot utilitzar l’obra original amb finalitats comercials. CMY CY MY CM K Y M C Centre deLingüística Teòrica delaUniversitat Autònoma deBarcelona 145 121 101 67 53 27 9 5 Index Harmonic Uniformity. Constraints. in Maximum Entropy Grammar with Lexically Specific in CentralCatalanandtheProblemofMissingBase. Zuraw,Polarized Variation. Kie. Rysling, Amanda. Rebrus, Péter;Szigetvári,Péter. Diminutives: Exceptions to Piñeros, Carlos-Eduardo. Moore-Cantwell, Claire;Pater,Joe.GradientExceptionality Joan. Mascaró, Baković, Eric Bonet, Eulàlia; Torres-Tamarit, Francesc.Introduction. Institut Interuniversitari deFilologia Valenciana ISSN 1695-6885(inpress);2014-9719(online) Catalan Journal ofLinguistics Catalan Journal . ExceptionalityinSpanishStress. http://revistes.uab.cat/catJL Exceptions inPhonology Morphological Exceptions to Vowel Reduction Reduction Vowel to Exceptions Morphological Polish yersrevisited. Vol. 15,2016 Exceptional nasal-stopinventories. Exceptions in Phonology CCATATALAN JJOURNAL OFL LINGUISTICS Vol. 15 Centre deLingüística Teòrica delaUniversitat Autònoma deBarcelona OF Institut Interuniversitari deFilologia Valenciana ISSN 1695-6885(inpress);2014-9719(online) Eulàlia Bonet&FrancescTorres-Tamarit C Exceptions inPhonology C J J L L OURNAL INGUISTICS AT ATAL http://revistes.uab.cat/catJL V olume Vol. 15,2016 Edited by 2016 15 AN