Lexicon in Focus [Reprint 2015 ed.] 9783050073712, 9783050031156

Citation preview

Stiebeis, Wunderlich (eds.) Lexicon in Focus

studia grammatica Herausgegeben von Manfred Bierwisch unter Mitwirkung von Hubert Haider, Stuttgart Paul Kiparsky, Stanford Angelika Kratzer, Amherst Jürgen Kunze, Berlin David Pesetsky, Cambridge (Massachusetts) Dieter Wunderlich, Düsseldorf

studia grammatica 45

Barbara Stiebeis Dieter Wunderlich (eds.)

Lexicon in Focus

Akademie Verlag

Die Deutsche Bibliothek - CIP-Einheitsaufnahme Lexicon in focus / Barbara Stiebels and Dieter Wunderlich (eds.) - Berlin : Akad. Verl., 2000 (Studia grammatica ; 45) ISBN 3-05-003115-8

ISSN 0081-6469 © Akademie Verlag GmbH, Berlin 2000 Das eingesetzte Papier ist alterungsbeständig nach DIN/ISO 9706. Alle Rechte, insbesondere die der Übersetzung in andere Sprachen, vorbehalten. Kein Teil dieses Buches darf ohne schriftliche Genehmigung des Verlages in irgendeiner Form - durch Photokopie, Mikroverfilmung oder irgendein anderes Verfahren - reproduziert oder in eine von Maschinen, insbesondere von Datenverarbeitungsmaschinen, verwendbare Sprache übertragen oder übersetzt werden. All rights reserved (including those of translation into other languages). No part of this book may be reproduced in any form - by photoprinting, microfilm, or any other means - nor transmitted or translated into a machine language without written permission from the publishers. Druck und Bindung: GAM Media GmbH, Berlin Printed in the Federal Republic of Germany

Contents

Barbara Stiebeis and Dieter Wunderlich Introduction

1

Sharon Inkelas Phonotactic blocking through structural immunity

7

Cemil Orhan Orgun and Ronald Sprouse Understanding ungrammatically

41

Janet Grijzenhout and Martin Krämer Final devoicing and voicing assimilation in Dutch derivation and cliticization .... 55 James P. Blevins Markedness and blocking in German declensional paradigms

83

Carsten Steins How to account for non-concatenative phenomena in a morphemebased theory

105

Stephen R. Anderson Some lexicalist remarks on incorporation phenomena

123

Elisabeth Löbel Case alternation in Finnish copular constructions

143

Gisbert Fanselow Optimal Exceptions

173

Barbara Stiebels Linker inventories, linking splits and lexical economy

211

Dieter Wunderlich Predicate composition and argument extension as general options - a study in the interface of semantic and conceptual structure

247

Addresses of contributors

271

Introduction

This volume presents a selection of papers read at the international conference 'Lexicon in Focus', held under the auspices of the Sonderforschungsbereich 'Theory of the Lexicon' (Düsseldorf-Wuppertal-Cologne) at Wuppertal in August 1998. This conference focused on the lexical perspective on questions that dominated the linguistic discussion of the nineties. It brought together experts from different fields, ranging from lexical phonology to lexical semantics. The participants were asked to discuss three issues which cross these different fields and which we considered matters of controversy. The three issues are briefly characterized in what follows. A. Lexical constraints and the generation of candidates. Optimality ideas are widely accepted in several fields of phonology, morphology, as well as of syntax, and even in semantics they play a role. The crucial idea is that among several candidates, the optimal one has to be found by an evaluation procedure defined in terms of an ordered set of general constraints. Many questions are still open to discussion: What is the division of labor between lexical information and constraint evaluation? Is there more than one step of evaluation, that is, is the set of candidates restricted by some previous evaluation? What is the nature of morphological constraints, compared with phonological and prosodie constraints? Is lexical information itself (for instance, the morphemes listed in the lexicon) subject to specific constraints? What kinds of constraints have to be assumed in particular fields of grammar, such as argument linking, cliticization, and the formation of complex predicates? B. Economy principles in the lexicon. Both representational and derivational economy play a major role in modern conceptions of the lexicon. Representational economy serves to minimize the information to be stored, and is also connected with the possibility of multiple modes of storage (regular vs. irregular inflection and derivation). Derivational economy in the lexicon relates to processes of complex word formation. What kinds of lexical structures must be assumed to ensure minimal expense? What are possible intra-categorial and cross-categorial minimization strategies? How regular are irregularities in the lexicon, such as allomorphy, suppletion, syncretism, polysémie variation, and lexical case-marking? C. Semantic composition within and outside of words. Several proposals have been made for the internal structure of lexical items and for structure building mechanisms that compose minimal predicates to larger complexes. Is there one general syntactic component that determines properties of phrases, of complex words as well as the internal structure of verbs? Or are there composition mechanisms that are specific for either the morphology or the lexicon?

2

Barbara Stiebeis and Dieter Wunderlich

The collection of papers in this volume reflects the highly stimulating discussions of the conference only fragmentarily, but we think that these contributions (which underwent a lengthy procedure of writing, reviewing and rewriting) are still worth to be brought to a broader public. In Saussure's conception, words, morphemes, and phrases pair phonological and semantic information via morpho-syntactic features. Accordingly, this volume is organized along the line 'from phonology via morphology/syntax to semantics'. The papers address many of the questions raised above, but certainly not all of them, and they also pose additional ones. One major common ground of the papers is the concern for conflicting constraints; another one is the theoretical status of morphology vis-à-vis phonology and syntax. In the following, we briefly introduce the papers included in this volume. The first section comprises three articles that are concerned with the phonology-morphology interface. Sharon Inkelas ('Phonotactic blocking through structural immunity') proposes an analysis of nonderived environment blocking which crucially distinguishes between lexically prespecified vs. underspecified segments/structures. Inkelas first shows the empirical problems of three OT approaches to nonderived environment blocking phenomena, namely Neighborhood Protection by Ito and Mester (1996), Sequence Protection by Burzio (1997) and Parasitic Alternations by Lubowicz (1998). In Inkelas' Structural Immunity approach, nonalternating segments are fully specified, whereas alternating segments are underspecified. Faithfulness constraints of the MAX family ensure that lexical prespecifications are not altered in morpho-phonological processes. Inkelas argues that prespecification can involve all types of phonological structure, including syllable structure. She provides further evidence for her account on the basis of lexical exceptions. Cases of false blocking are accounted for by 'cophonologies' associated with compounding and affixation. Cernii Orhan Orgun and Ronald Sprouse ('Understanding ungrammaticality') address the question of absolute ungrammaticality, which they exemplify with English stress-shifting nominalization: verbs that contain more than two syllables do not give rise to grammatical stress-shifted nomináis. Orgun and Sprouse show that the attempt of generating ungrammaticality for these nomináis by using the constraint MPARSE (Prince and Smolensky 1993) is not successful. Rather than including null parse among the set of candidates, they propose to move all inviolable constraints to a new constraint system called CONTROL. Grammatical output forms, then, must satisfy two conditions: first, they have to be the optimal output in the evaluation component, and, second, they must satisfy all constraints in CONTROL. Thus, CONTROL eliminates all winners that fail to satisfy any of the inviolable constraints. The article by Janet Grijzenhout and Martin Krämer ('Final devoicing and voicing assimilation in Dutch derivation and cliticization') offers an exhaustive analysis of the variation that stem-final obstruents in Dutch show in derivation, compounding, or cliticization. The authors first show that the assumption of different levels in the lexicon (Trommelen and Zonneveld 1989) fails to explain the whole set of data, and that also the laryngeal analysis by Lombardi (1999) only captures some of the data. Grijzenhout and Krämer propose a prosodie analysis according to which 'internal affixes' form a prosodie word together with a preceding stem, while 'semisuffixes' and the members

Introduction

3

of compounds form a prosodie word of their own; clitics, however, are integrated into a prosodie phrase without forming a prosodie word. The use of ALIGNMENT and IDENTITY constraints accounts for all instances of final devoicing and voicing assimilation without any level-ordering or Output-to-Output constraints. The next two articles address issues concerning the status of morphological theory. In his study on German nominal paradigms, James P. Blevins ('Markedness and blocking in German declensional paradigms') shows that morphologically underspecified lexical entries do not impose any problems of blocking by the presence of more specific entries. Blevins distinguishes the lexical entries that define a paradigm from the structures or realizations that the entries sanction in a syntagmatic context. According to his Paradigmatic Blocking Principle, a realisation that satisfies an entry is the least structure that conflicts with every more specific entry in the paradigm. Given that only positive values are specified in the entries, this means that only negative values are inferred in the mapping from entries to their syntagmatic realizations. Blevins concludes that such an account, being essentially monotonie, is superior to a readjustment or impoverishment account that neutralizes features in a syntactic context. Carsten Steins ('How to account for non-concatenative phenomena in a morphemebased theory') reconsiders two of the phenomena that were central for the claim that important parts of morphology are a-morphous (Anderson 1992), namely morphological operations that delete segmental material (truncation), and operations where a single morpheme marks plural in one class of nouns and singular in another class of nouns (flip-flop). In inspecting the truncations in the perfective stems of Tohono O'odham ('Papago') in more detail, Steins shows that truncation is not the morphological exponent of perfective in this language (because there is a perfect auxiliary in addition) but follows from the requirement that the number of syllables of auxiliary + verb stem is retained in the perfective. For one alleged example of flip-flop, the Kiowa number system, Steins shows that the single morpheme under question maps a set of underlying number values onto a set of derived number values in a way such that each instance of number is recoverable, and no value switch arises. He concludes that the two investigated phenomena do not constitute counter-evidence to a morpheme-based account. The article by Stephen R. Anderson ('Some lexicalist remarks on incorporation phenomena') reconsiders several arguments put forward in favor of either a syntactic or a lexicalist account of noun incorporation. Anderson arrives at the conclusion that the majority of phenomena connected with this construction can be successfully analyzed under either account. Moreover, he argues that the phenomena discussed by Baker (1995) for Mohawk (such as agreement with the position of the incorporated noun, and coreference between the matrix subject and the argument corresponding to the incorporated noun) are not as clear as they should be in order to suggest a syntactic explanation as the preferred alternative; on the contrary, as Anderson concludes, the lexicalist account seems entirely viable, even for the phenomena that Baker treats as syntactic. The next three articles are concerned with the relation of argument structure and morphological case. Elisabeth Löbel ('Case alternation in Finnish copular constructions') considers the conditions under which predicative nouns and adjectives receive

4

Barbara Stiebeis and Dieter Wunderlich

case in Finnish. Löbel shows that with an impersonal copula, the case on the predicative can be the same as the case on the subject (case concord), or it can be nominative (by default), or partitive. Löbel argues that the selection of partitive vs. nominative does not only reflect the divisibility of the argument which is predicated of (mass or indefinite reading), but also the fact that this argument may only partially be predicated of, even if it is indivisible and definite. Therefore, she introduces two morphological features in order to account for the partitive-nominative alternation in Finnish. The other two articles devoted to case discuss the possible distributions of case for nonpredicative arguments, albeit from different perspectives. Gisbert Fanselow (Optimal exceptions') deals with exceptional case-marking in the framework of the Minimalist Program. In his study of German case patterns, Fanselow arrives at the generalization that exceptional case is realized only on the lowest available argument position, a result that he traces back to the assumption that all lexical exceptions must be specified at the root of a lexical verb rather than by functional categories. This conforms to the Parallel Movement Constraint (Müller 2000), which preserves the given syntactic configuration in the checking of case features. In the attempt to extend this account to Icelandic, Fanselow briefly considers the even more exceptional case options, and he suggests that Icelandic has a different constraint ranking from German. The article by Barbara Stiebels ('Linker inventories, linking splits and lexical economy'), which is coached in the framework of Lexical Decomposition Grammar (LDG, Wunderlich 1997), is typologically oriented. Stiebels considers morphological case as only one instance of argument linker (next to pronominal affixes and agreement). All linkers are specified in terms of features that also specify the underlying argument roles. In her typology of argument linking, the several inventories of linkers are evaluated under the question of how faithful they are with respect to the universal underlying feature distribution. Stiebels also shows that a linking split can be derived by Harmonic Alignment (Prince and Smolensky 1993) of the scale of case features with the Silverstein (1976) hierarchy. The results she arrives at differ both conceptually and empirically from those of Aissen (1999) and Kiparsky (1999). The final article by Dieter Wunderlich ('Predicate composition and argument extension as general options - a study in the interface of semantic and conceptual structure') discusses the minimal requirements for an account of resultative and 'possessor raising' constructions in the lexicalist framework of LDG (which adopts the two-level semantics proposed by Bierwisch 1983). Since these constructions extend the argument structure of a verb, Wunderlich argues that the verb as the lexical head licenses these constructions only if it is (covertly) combined with a further predicate in the Semantic Form (SF). The interface constraints of LDG, then, predict how the extended argument structure is realized in morphology and syntax. It is shown how the full meaning of the construction is derived in the Conceptual Structure (CS), in particular with respect to the possible identification of implicit and realized arguments. Finally, it is suggested that also SF formation can be subject to conflicting constraints.

Introduction

5

Regrettably, a good deal of papers presented at the conference are not included in this volume, for various reasons. Some of these papers have been published at other places. For reasons of fairness and completeness, we list those papers that were read at the conference but are not included here in alphabetic order of the authors: Emmon Bach (Amherst): Semantic relations in and outside of complex words. Laura Benua (University of Maryland): Transderivational correspondence and the Strong Domain Hypothesis. Manfred Bierwisch (Berlin): Modes of modification. Harald Clahsen (Essex), Sonja Eisenbeiß, Ingrid Sonnenstuhl and Martina Penke (Düsseldorf): Morphological paradigms in language processing and in aphasia. Cleo Condoravdi and Paul Kiparsky (Stanford) On the ambiguity of 'wieder ' and related adverbs. David Dowty (Ohio State University): On the origin of thematic role types. Jane Grimshaw (Rutgers): Optimization and variation in the functional lexicon. Pauline Jacobson (Brown University): Projection from the lexicon? The case of relative pronouns. Hans Kamp (Stuttgart): Lexical entries of speech act verbs. Ralf Naumann (Düsseldorf): A dynamic logic of events and states for the interaction between plural quantification and verb aspect in natural language. [Logic Journal of the IGPL 7 (1999), 591-627.] Martin Neef (Cologne): A non-universal analysis of the English nominal plural. Albert Ortmann (Düsseldorf): Where plural refuses to agree: feature specification and economy. [To appear in Acta Linguistica Hungarica.] Beatrice Primus (Stuttgart): Constraints on case selection. Irene Rapp and Arnim von Stechow (Tübingen): Fast 'almost' and functional heads. [Journal of Semantics 16 (1999), 149-204.] Cheiyl Zoll (MIT): Lexical underspecification and tone melodies: a different view.

Acknowledgments: Last but not least, we would like to thank the reviewers of the papers: Manfred Bierwisch, Ray Fabri, Gisbert Fanselow, Janet Grijzenhout, Martin Haspelmath, Ursula Kleinhenz, James Kilbury, Gereon Müller, Anne Vainikka and Richard Wiese. Our thanks also go to those people who helped us in the preparation of this volume: Thomas Gamerschlag, Gabi Krüger, Anja Latrouite, Albert Ortmann, Carsten Steins and Yi-Chun Yang.

Barbara Stiebels and Dieter Wunderlich, Düsseldorf, July 2000

6

Barbara Stiebeis and Dieter

Wunderlich

References Aissen, Judith (1999). Markedness and subject choice in Optimality Theory. Natural Language and Linguistic Theory 17, 673-711. Anderson, Stephen R. (1992). A-Morphous morphology. Cambridge University Press. Baker, Mark (1995). The polysynthesis parameter. Oxford University Press. Bierwisch, Manfred (1983). Semantische und konzeptuelle Repräsentation lexikalischer Einheiten. In Untersuchungen zur Semantik, Rudolf Rüzicka and Wolfgang Mötsch (eds.), 61-99. Berlin: Akademie Verlag. Burzio, Luigi (1997). Cycles, non-derived environment blocking, and correspondence. To appear in The Pointing Finger: Conceptual Studies in Optimality Theory, Joost Dekkers, Frank van der Leeuw and Jeroen van de Weijer (eds.). Oxford University Press. Itô, Junko and Armin Mester (1996). Structural economy and OCP interactions in local domains. Proceedings of the Western Conference on Linguistics (WECOL). University of California, Santa Cruz. Kiparsky, Paul (1999). Analogy and OT: Morphological change as emergence of the unmarked. Paper presented at the 21st annual meeting of the German Linguistics Society (DGfS), Constance, February 1999. Lombardi, Linda (1999). Positional faithfulness and voicing assimilation in Optimality Theory. Natural Language and Linguistic Theory 17, 267-302. Müller, Gereon (2000). Order preservation, parallel movement, and the emergence of the unmarked. To appear in Optimality Theoretic Syntax, Geraldine Legendre, Jane Grimshaw and Sten Vikner (eds.). Cambridge, Mass.: MIT Press. Lubowicz, Anna (1998). Derived environment effects in OT. Unpublished manuscript, University of Massachusetts, Amherst. Available as ROA-239-0198. Prince, Alan and Paul Smolensky (1993). Optimality Theory: constraint interaction in generative grammar. Ms., Rutgers University, New Brunswick and University of Colorado, Boulder. Silverstein, Michael (1976). Hierarchy of features and ergativity. In Grammatical categories in Australian languages, R.M.W. Dixon (ed.), 112-171. Canberra: Australian Institute of Aboriginal studies. Trommelen, Mieke and Wim Zonneveld (1989). Klemtoon en metrische fonologie. Muiderberg: Coutinho. Wunderlich, Dieter (1997). Cause and the structure of verbs. Linguistic Inquiry 28, 27-68.

Phonotactic blocking through structural immunity Sharon Inkelas

1. Introduction 'Phonotactic blocking' is the failure of phonotactic conditions to apply even though the structural description is apparently present. Three general types of phonotactic blocking have been the focus of work in the past two decades: (1)

a. Strict cycle effects (nonderived environment blocking) b. Lexical exceptionality c. Geminate inalterability These are all phenomena in which phonological generalizations are defied. Quite different approaches have been taken to them in the past. For strict cycle effects, conditions such as the Strict Cycle Condition (Mascaré 1976, Kiparsky 1982) have been proposed; lexical exceptionality has been handled by morphemic exception features (e.g. Chomsky & Halle 1968); geminate inalterability has inspired the Linking Condition (Hayes 1986) and Uniform Applicability Condition (Schein & Steriade 1986). Each of these various theoretical devices is targeted at specific types of phonotactic immunity. None extends to either of the other two types. By contrast, Structural Immunity is a general approach that has been applied to all three of these phenomena. Structural Immunity is the method of prespecifying (either lexically, or via the grammar) strings with structure that blocks the application of phonological alternations or phonotactic conditions. The idea of handling nonderived environment blocking using structural immunity is due to Kiparsky (1993). It has since been applied to geminate inalterability (Inkelas & Cho 1993) and lexical exceptionality (Inkelas, Orgun & Zoll 1997). The goal of this paper is to show that Structural Immunity, a time-honored technique, fares better than recent proposals couched within Optimality Theory (OT; Prince & Smolensky 1993) in handling nonderived environment blocking and in connecting that phenomenon to other types of blocking. In addition, we shall see that the structural approach leads to a new perspective on the 'Richness of the Base' hypothesis in OT.

8

Sharon Inkelas

2. Nonderived environment blocking: a parade example In this section we introduce the first of several examples which will inform the subsequent discussion of strict cycle, or derived environment, effects. The term 'strict cycle' comes from rule-based cyclic analyses of these and other similar examples, in which it appears that phonological alternations apply only when their environments are morphologically derived on the same cycle of rule application (see e.g. Mascará 1976, Kiparsky 1982).' The term 'derived environment effects', or 'nonderived environment blocking' (NDEB; Kiparsky 1993), captures in more general terms the generalization that these effects occur only across morpheme boundaries. In Turkish, intervocalic velars delete (Zimmer & Abbott 1978, Sezer 1981, Inkelas & Orgun 1995, Inkelas, Orgun & Zoll 1997), due to a well-known sound change that has given rise to productive synchronic alternations (2a).2 Analyzed in OT, the alternation could be attributed to a constraint against intervocalic velars, as in (2b); ranked above M A X - C but below M A X - V , *VGV would have the desired effect. However, Velar Deletion is subject to nonderived environment blocking. Velars internal to morphemes do not delete, as shown in (2c):3 (2)

Turkish Velar Deletion a. bebek 'baby' katalog bebe-i 'baby-ACC' katalo-u bebe-e 'baby-DAT' katalo-a b. *VGV ('no intervocalic velars') c. sokak 'street' *soak soka-i 'street-ACC' *soai

'catalog' 'catalog-ACC' 'catalog-DAT'

A few of the many other parallel examples from the literature include Polish First Velar Palatalization, Coronal Palatalization, as discussed by Rubach (1984) and Lubowicz (1998); Hausa palatalization (Inkelas & Cho 1993 and references therein); Korean affrication (Kiparsky 1993 and references therein), and Finnish assibilation (e.g. Kiparsky 1982, 1993). Nonderived environment blocking effects inspired a number of different approaches in the 1970's and 1980's. These fall into two camps:

1 I will ignore for purposes of this paper those strict cycle effects which are dependent on a phonologically derived environment; see e.g. Mascara 1976 for discussion. 2 Deletion does not apply when the preceding vowel is long, e.g. mera:k-i 'curious-ACC' (Sezer 1981). 3 Turkish data are given throughout the paper in standard Turkish phonemic transcription, 'i' = IPA [ui]; 'Ü' = IPA [y]; 'ö' = IPA [0]; 'ç' = IPA [tfl; 'c' = IPA [d3].

Phonotactic blocking through structural immunity

9

(3)

1970's and 1980's a. STRICT CYCLE CONDITIONS. Originating with the (Revised) Alternation Condition of Kiparsky (1973), various extrinsic conditions on rule application have been proposed to prevent alternations from occurring in nonderived environments. Mascaró's well known (1976) modification to Kean's (1974) Strict Cycle Condition stipulated that neutralizing cyclic rules may apply only in derived environments. b. ELSEWHERE CONDITION. Kiparsky (1982) proposed to make nonderived environment blocking an intrinsic, rather than extrinsic, property of rule application by reducing it to the Elsewhere Condition. Kiparsky's idea was that lexical identity rules (which stipulate the content of roots and of stems that have been created earlier in the derivation) would intrinsically block more general rules imposing phonological alternations. Only stems derived on the current cycle, which would not yet have corresponding lexical identity rules, would fail to block, and thus be subject to, structure-changing cyclic rules. Both of these approaches were rejected by Kiparsky (1993) on empirical grounds. Kiparsky observed that it is neither the case that all rules which are blocked in nonderived environments are cyclic, nor that all such rules are structure-changing. This finding contraindicates approaches like those in (3), which are inherently limited to neutralizing cyclic lexical phonology. Kiparsky's (1993) suggestion was to replace the conditions in (3) with a Structural Immunity approach to nonderived environment blocking: (4)

The early 1990's: STRUCTURAL IMMUNITY (Kiparsky 1993). Prespecified structure blocks structure-filling phonological rules. Kiparsky's inspiration was to use prespecified structure, including structure assigned by rule, which blocks structure-filling rules.4 Potentially alternating portions of roots would be underspecified, with their surface values filled in by contextual rules; only those underspecified segments which find themselves in a variety of surface contexts could potentially alternate. Those which always occur in the same context (i.e. morpheme-internal ones) would never alternate. This is very much the approach defended in the present paper, although Kiparsky (1993) assumes a more restrictive theory of underspecification than the one adopted here. Kiparsky achieved impressive results in his paper, gracefully handling many of the best-known examples of nonderived environment blocking. However, his approach faced objections within the Optimality Theory world from authors defending the 4 It is interesting to note that while Kiparsky (1982) had to assume that only structure-changing rules are blocked in nonderived environments, Kiparsky (1993) assumes that only structure-filling rules are blocked. Of course, the senses of 'structure' differ somewhat across the two terms. Structure-changing rules are defined by result; they are neutralizing. But the operation used to accomplish the neutralization is free to vary. By contrast, structure-filling rules are defined by operation, not result. Structure-filling rules are those which simply add structure and do not delete it. They may or may not be neutralizing.

10

Sharon Inkelas

'Richness of the Base' hypothesis of Smolensky (1996). This hypothesis, according to which grammars must be able to assign grammatical outputs to any conceivable lexical representation, has been popularly understood to be inconsistent with the use of underspecification to get analyses to work (see e.g. Burzio 1997). Thus many working in the Optimality Theory framework found themselves bereft of a general account of nonderived environment blocking. Into this vacuum several new approaches have been thrust. I will briefly review them here, comparing each to Optimality Theory instantiations of a modified version of Kiparsky's Structural Immunity account.

3. New approaches in Optimality Theory The first approach to be considered is that of Itô & Mester (1996), which relies on the notion that adjacency relations on the input must be preserved; segments in nonderived environments derive special protection from neighborhood preservation insofar as they have more input neighbors than segments in derived environments do.

3.1

Neighborhood Protection

Neighborhood, or adjacency, constraints in Optimality Theory are based on McCarthy & Prince's (1994:340) Linearity constraint, which is stated over two strings (their statement refers to R(eduplicant) and B(ase); here we will generalize to SI and S2): (5)

LINEARITY: S1 reflects the precedence structure of S2, and vice versa For r„ η G Dom(/), r, < η iff/fa) < / η )

In the above statement, / is the correspondence function relating strings S1 (the domain, or 'Dom' or f ) and S2 (the range o f f ) . Itô & Mester (1996:8) define a variant of the LINEARITY constraint, which they term NEIGHBORHOOD. (NEIGHBORHOOD differs slightly from LINEARITY in that it is unidirectional.) (6)

Neighborhood: • The neighborhood of a segment must be preserved • If α precedes/follows β, then the correspondent of α precedes/follows β.

In order to make use of these constraints, several clarifications/modifications are necessary: (7) Modifications to Linearity/Neighborhood a. Immediate precedence. I will assume that the notion of precedence being appealed to is immediate precedence. This appears to be the practice of Itô & Mester (1996) as well (on p. 9 of their paper, Itô & Mester say that "segments occupying stem edges ... are input-adjacent to only a single neighbor").

Phonotactic blocking through structural immunity

11

b. Directionality. I will assume that the constraints are evaluated only with respect to the one-way mapping from input to output. Doing otherwise would multiply the number of violations observed without, I believe, affecting outcomes in tableaux (for example, every affixed candidate would gratuitously violate the constraints by virtue of being one string in output but two nonlinearized strings in input). c. Mode of violation. I will assume that the constraints are evaluated for each pair of adjacent input elements (whether segments, or features on a tier), and that violations are cumulative. In summary, then, our working version of LINEARITY/NEIGHBORHOOD penalize corresponding strings in which a pair of adjacent input segments does not correspond to an adjacent pair in output. Relativized to a specific phonological tier, the constraints also penalize corresponding strings in which a pair of adjacent features does not correspond to an adjacent pair in output. To exemplify, consider a Turkish morpheme (sokak 'street') with intervocalic and root-final velars: a. b. c. d.

sokiak2-a sokiak2-a sokia-a soak2-a soa-a

NEIGHBORHOOD

*(ak2) **(oki, kia) ***(oki, kja, ak2)

Deletion of the root-final velar (8a,b,d) violates NEIGHBORHOOD once, since the velar is a member of only one pair of adjacent input segments. (Crucially, the suffix is not considered adjacent to the root in input, as its linear position is determined only in output, by Alignment constraints.) But deletion of the internal velar (8c,d) contributes two NEIGHBORHOOD violations, since it is a member of two adjacent input pairs. This observation is key to Itô & Mester's insight that two NEIGHBORHOOD violations are categorically worse than one. Internal segments, with larger neighborhoods to be loyal to, are costlier to disturb than peripheral segments, with smaller neighborhoods. To implement this insight, Itô & Mester appeal to the mechanism of selfconjunction (Smolensky 1997). Conjoining NEIGHBORHOOD with itself yields NEIGHBORHOOD2, a constraint which penalizes a string once for each pair of input segments that have two neighborhood violations. To illustrate, let us apply the Neighborhood approach to the Turkish velar deletion example already discussed. Given the constraint ranking in (9), (9) Neighborhood2 » *VGV » Neighborhood the correct prediction is that a tautomorphemic VGV sequence will be preserved, while a heteromorphemic one will be deleted:

12

Sharon Inkelas (10) Deletion of stem-final /k/ violates NEIGHBORHOOD only once, and therefore does not violate NEIGHBORHOOD2: a. b. c. d.

sokiak2-a sokiak2-a τ sokia-a soak2-a soa-a

NEIGHBORHOOD 2

*VGV

NEIGHBORHOOD

**|

*

*!(oki, kia) *!(ok,,k,a)

*(ak2)

*

***( oki, kia, ak2)

Although it appears to work well for this commonly attested type of nonderived environment blocking effect, the Neighborhood account suffers from a number of serious difficulties. These are summarized below: (11) Problems for Neighborhood account of nonderived environment blocking: a. Predicts equal alterability of stem-final and suffix-initial velars in Turkish b. Predicts equal alterability of peripheral sequences generally, not just under affixation c. Can't handle NDEB effects involving nonadjacent trigger and target d. Can't handle NDEB effects in which there is no phonological alternation 3.1.1 Problem 1 : disparate behavior of junctural segments To illustrate the first point, consider the following data from Turkish (which also, by the way, present a problem for Mascaró-style strict cycle accounts). Both the polygon name-former -gen and the relativizer -ki begin with velars and can combine with vowel-final stems, yielding the environment for velar deletion. Yet affix-initial velars never delete:5 (12) a. -gen iiç-gen '3-gon = triangle' dört-gen '4-gon = quadrilateral' '7-gon = septagon' yedi-gen *yedien b. -ki Orhan-in-ki 'Orhan-GEN-REL = the one belonging to Orhan' ora-da-ki *oradai 'there-LOC-REL = the one over there' This is contrary to the predictions of the neighborhood approach, which simply counts pairs of adjacent input elements. A CV suffix initial velar belongs to only one such pair; its deletion violates NEIGHBORHOOD but not NEIGHBORHOOD2, and should thus be permitted: (13) a. (•»-) b. 4 s

yedi-gen yedi-gen yedi-en

NEIGHBORHOOD 2

*VGV NEIGHBORHOOD *!

*(ge)

5 Harrison (2000) documents an interesting case from Tuvan, another Turkic language, in which suffix-initial velars do delete intervocalically, subject to some conditions on morpheme recoverability.

Phonotactic blocking through structural immunity

13

There are various stipulative solutions available. One possibility would be to rank faithfulness to affix consonants above faithfulness to root consonants, thus running afoul of the universal tendency (elevated to a universal constraint ranking in Optimality Theory; see e.g. McCarthy & Prince 1995) that root faithfulness outranks affix faithfulness. Alternatively, one could stipulate that NEIGHBORHOOD2 must crucially include a violation of adjacency on the left, so that suffix-initial segments would never qualify. A third option, suggested by a reviewer, would be to stipulate that suffix-initial segments must be syllable-initial in output. (The relevant constraint must, of course, rank low enough to still permit intervocalic stem-final nondeleting consonants to syllabify as onsets.) 3.1.2 Problem 2: disparate behavior of stem-peripheral segments The second problem for the neighborhood approach, stated in (lib), is that it predicts that a stem-initial segment will be just as vulnerable to alternation conditioned by the following segment as a stem-final segment, which is clearly untrue. To see why, consider the following well-known example from Finnish (e.g. Kiparsky 1973, 1993). In Finnish, /t/ assibilates to /s/ before HI (14a). This alternation can be attributed to a constraint against /ti/ sequences (e.g. Burzio 1997:24), as in (14b). However, the alternation is limited to derived environments, such that when the underlying form /tilat/ is suffixed with HI, only the second /ti/ sequence — the one which spans the stem-suffix boundary — undergoes assibilation (14c): (14) Finnish assibilation a. Assibiliation turns /ti/ into /si/: halut-a 'to want' halus-i 'wanted' b. *ti c. Assibilation is subject to NDEB: tilat-a tilas-i

'to order' 'order-3SG.PRET'

*silat-a *silas-i

Although the initial Iti of /tilat/ never assibilates, the neighborhood approach predicts that it should, as shown by the following table. Here, NEIGHBORHOOD and NEIGHBORHOOD2 should be interpreted as applying to the [continuant] tier; changing Ν to Isl changes its value for [cont] and disrupts the underlying adjacency between the [-cont] specification of Iti and the [cont] specifications of the adjacent segments. (15) Problem: initial Iti of/tilas/ has only one neighbor, and is not protected by NEIGHBORHOOD2

a. b. o c. é*

/tilat, -i/ tilat-i tilas-i silas-i

NEIGHBORHOOD2

*TI *

NEIGHBORHOOD * **!

Sharon Inkelas

14

Candidate (15c), silas-i, does not violate NEIGHBORHOOD2 because no single pair of adjacent input segments suffers two NEIGHBORHOOD violations. Therefore, since silasi violates *Tl the least, it is (incorrectly) predicted to win. A possible solution would be to make NEIGHBORHOOD (and NEIGHBORHOOD2) sensitive to boundary symbols (á la Chomsky & Halle 1968), so that the initial word boundary would be counted as one of the neighbors of the word-initial segment. This would correctly protect the /ti/ of /tilat/: (16) Possible solution: m a k e NEIGHBORHOOD sensitive to b o u n d a r y symbols:

/#t,ilat2#, #a#/ #tiilat2#a# #s,ilat2#a#

a. b.

NEIGHBORHOOD2

*TI NEIGHBORHOOD * **

*!(#ti,t,i)

However, it would incorrectly block assibilation where we do want it to apply: (17) Problem: derived environments are protected too: a. («-) b.

/#t,ilat2#, #i#/ #t,ilas2#i# #s,ilas2#i#

NEIGHBORHOOD2

*TI NEIGHBORHOOD *

**!(#ti, 11i; at2, t2#)

**

In summary, protection by adjacency works only when the target of the alternation is completely morpheme-internal — and not all nonalternating segments meet this condi-

3.1.3 Problem 3 : stem-internal targets The third problem for the Neighborhood account is illustrated by the following example from Uighur, in which the suffixal trigger of a raising alternation is not string-adjacent to its stem-internal target. As reported by Orgun (1994, 1996), the rightmost vowel of an Uighur stem raises when in a nonfinal open syllable (18a). (Note that in the last form in (ii), deletion, which has a similar environment, has applied to the root-final vowel.) The raising process can be attributed to a constraint against a [-high] vowel immediately followed by a syllable, as in (18b). However, this constraint does not have any effect on high vowels in which the following syllable belongs to the same morpheme — i.e. to vowels in nonderived environments (18c): (18) a. i. qazan 'pot'

ii.

iii. b.

qazan-ni

'pot-ACC'

qazin-i

'pot-3.POSS'

bala bali-lar bal-lir-i

'child' 'child-PL' 'child-PL-3.POSS'

ameriqa ameriqi-liq *[-high]a

'America'

15

Phonotactic blocking through structural immunity c.

*qizan

'pot'

*bili-lar

'child-PL'

*imiriqa

'America'

A Neighborhood account of the immunity of vowels in non-stem-final syllables to Raising would protect two-member neighborhoods of high vowels by self-conjunction (Smolensky 1997), as in (19): (19)

NEIGHBORHOOD[HIGH] 2

The problem for this account is that the trigger and target of Uighur Raising are not string-adjacent. Thus, while NEIGHB0RH00D[HIGH] 2 correctly protects the vowel in stem-initial or stem-medial syllables from raising, it also incorrectly protects even those vowels in stem-final syllables, as long as a consonant follows in the stem. The reason is that any vowel surrounded by other segments in the stem is automatically protected by NEIGHBORHOOD 2 , as shown below: qazan, -i

NEIGHBORHOOD

*[-high]a

[HIGH} 2

a. b. («-) c.

qazani qazini qizini

NEIGHBORHOOD [HIGH]

**

*!(za, an) *!*(qa, az; za, an)

*

** ****

One might try to find a way around the problem of stem-internal targets by appealing to a more sensitive definition of adjacency. Perhaps only vowels project the [high] feature and comprise the neighborhoods in which faithfulness is computed (abbreviated below as V-NEIGHBORHOOD[HIGH]). But this does not work either. As (21) shows, in a stem with two input vowels, both have only one vowel neighbor. Thus neither is protected by V - N E I G H B 0 R H 0 0 D [ H I G H ] 2 , which is irrelevant to the outcome. As a result, the candidate which best satisfies *[-high]a incorrectly wins (21c). qazan, -i

V-NEIGHBORHOOD

*[-high]a

[HIGH] 2

a. b. (·-) c. Φ '

qazani qazini qizini

V-NEIGHBORHOOD [HIGH]

*|*

*!

*(aa) *(aa)

In summary, stem-internal undergoers of derived environment rules pose a challenge for the Neighborhood account because such undergoers have the same number of neighbors as the nonundergoers. 3.1.4 Problem 4: phonotactic requirements which induce no alternation A different kind of problem for the Neighborhood account is posed by phonotactic requirements that happen not to result in alternations. Such a case is found in Turkish and has been the focus of much work by Inkelas & Orgun (1995), Orgun (1996) and Orgun & Sprouse (1996, in press, this volume). As first documented by Itô & Hankamer (1989), some speakers of Turkish impose a disyllabic minimal size requirement

16

Sharon Inkelas

on derived words. (22a) shows that these speakers judge suffixed monosyllabic words as ungrammatical. This is modeled by a constraint requiring words to be minimally disyllabic (22b): (22) Turkish Disyllabic Minimality 'musical note /a-lSG.POSS = my fa' (cf. sol-üm 'my note soV) a. *fa-m *be-n 'letter ¿-2SG.POSS = your b' (cf. abece-n 'your abc') *ye-n 'eat-PASS = be eaten!' (cf. yut-ul 'be swallowed!') b. [σσ] ω 'musical note fa' c. fa 'letter è' be ye 'eat!' yen 'alight!' However, as (22c) shows, monosyllabicity is tolerated and even common among nonderived words. Thus, although the Disyllabic Minimality condition causes no alternations, it nonetheless is subject to nonderived environment blocking. A neighborhood account is entirely inappropriate here. When the phonotactic condition [σσ] ω is imposed on informed derived inputs, there is no output, hence no NEIGHBORHOOD 2 (or even NEIGHBORHOOD) violation. 3.1.5 Summary To summarize, the Neighborhood approach works only for stem-peripheral segments which are the target of an alternation conditioned by a trigger outside the stem. This is true of some cases of nonderived environment blocking, but it is by no means true of all.

3.2

Sequence Protection

A second type of approach to nonderived environment blocking is taken by Burzio (1997), who proposes faithfulness constraints on specified phonological sequences. The reason this approach can achieve nonderived environment blocking is that bases and affixes are not linearized in the input. Only morphologically derived tokens of such sequences escape sequence faithfulness and are permitted to alternate. Burzio (1997) specifically discusses the case of Finnish assibilation, proposing that input ti sequences are the target of special faithfulness constraints (23a). The same approach extended to Turkish velar deletion would involve faithfulness to the sequence VGV (23b). (23) a. b.

Finnish: Faithti » *ti Turkish: FaithVGv » *VGV

(Burzio 1997) (author's application of Burzio's approach)

These sequence faithfulness constraints would outrank the constraints responsible for the alternation (assibilation or velar deletion). The following table, generating the dative form of sokak 'street', illustrates how the approach would be applied to Turkish:

Phonotactic blocking through structural immunity (24) a. b. c.

sokak, -a sokaka sokaa soaa

FAITHVGV

17

*VGV ** ι *

*!

A characteristic of the Sequence Protection approach is its redundancy; the phonotactic constraint causing the alternation from which nonderived sequences must be protected is repeated verbatim in the protection constraint. There is clearly no generalization being captured here. Burzio's insight that those sequences which are immune are the sequences which are present in input is potentially valuable. But the analysis must stipulate this insight afresh for each invididual alternation. 3.2.1 Problem: 'overlapping' protected sequences The Sequence Protection approach fares better than the Neighborhood approach in handling the examples we have introduced thus far. However, at least one type of phenomenon does pose an empirical challenge. This is the Uighur raising example. Recall that in Uighur, syllable-final low vowels in stem-final syllables raise under affixation. Vowels followed by another vowel in the same morpheme do not raise. Under the Sequence Protection approach, this nonderived environment blocking effect would be attributed to a faithfulness constraint protecting a vowel-syllable (Va) sequence (abbreviated below as FAITHVO). However, as shown in the following table, the approach breaks down whenever the syllable in the relevant V a sequence itself contains a vowel that does undergo raising:

a. è b. (•»•) c.

/baila2, -lar/ FAITHVCT *[-high]a ** balalar * balilar *!(a,a2) bililar *!(a,a2)

No matter whether candidate (25c) is considered to violate the sequence faithfulness constraint once (as in the above table) or twice (since two segments in the same initial / a i W sequence sequence are altered), the fact remains that both candidates (25b) and (25c) violate the constraint while (25a) does not. Thus (25a), the candidate in which no raising occurs at all incorrectly wins. Note that not only does the intended winning candidate (25b) not win; it actually does the worst of the three candidates in (25)! Why does Sequence Protection fail to correctly identify candidate (25b) as the winner? Because the second /a/ in candidate (25b) is both the raising trigger for the first /a/ - which should not alternate - and the target of raising by the third (suffixal) /a/. Once the second /a/ raises, the sequence containing it and the first vowel is no longer intact. Once the sequence is altered, there is nothing to be gained by keeping the first vowel intact. As a result, candidate (25c), which better satisfies *[-high]a, outdoes (25b). Candidate (25a) of course does even better by keeping the sequence so intact that not even the second vowel raises.

18

Sharon Inkelas

In sum, Sequence Protection works only if no other alternation (or even the same one, conditioned by a different trigger) affects any part of the sequence. This is unrealistic, as the Uighur example illustrates. It is conceivable that the Uighur account could be rescued by constraint self-conjunction (Smolensky 1997). Depending on exactly how violations of sequence faithfulness are computed, the ranking FAITHvo2» *[-high]a » FaitHvo could correctly select candidate (25b), balitar, as the winner in (25). But this modification would still merely count violations without identifying them, and thus would still fail to capture the generalization that the second IdJ in a tautomorphemic V a sequence is alterable, while the first is not. Worse, the self-conjunction modification would no longer handle the Finnish example discussed by Burzio. A variety of rankings and degrees of sequence faithfulness would therefore have to be utilized, reducing the insight and generality of the analysis.

3.3

Parasitic alternations, or local conjunction

The Parasitic Alternations approach of Lubowicz (1998) can be seen as a more sophisticated version of Sequence Protection. Lubowicz proposes that local conjunction of faithfulness and markedness constraints (Smolensky 1997) can generate both phonological and morphological derived environment effects. The essential idea is that one alternation may be parasitic on another. The bulk of Lubowicz's discussion is devoted to phonologically derived environments, not a focus of the present paper, but she addresses morphologically derived environments as well. Noting that many such cases involve stem-final segments under affixation, Lubowicz's insight is that resyllabification across a morpheme boundary 'licenses' an alternation at the morphological juncture which otherwise could not occur: (26) "By means of local conjunction, violation of stem:syllable anchoring activates a markedness constraint, causing a phonological process. In case when there is no violation of anchoring the phonological process is blocked." [Lubowicz 1998:26-27] In other words, only if the syllable structure of a stem-final segment is being disturbed anyway is it all right to alter the segment in other ways. Lubowicz implements the stem-syllable anchoring insight with the following constraint schema: (27) [Markednessj X Anchor-R(stem, a)]Dom » Faithfulness » Markedness, Anchor-R(Stem, σ): "the rightmost segment of a stem in the input has a correspondent at the right edge of a syllable in the output" [Lubowicz 1998:25] Dom(ain): "the smallest domain within which both of the locallyconjoined constraints can be evaluated" [Lubowicz 1998:33]

Phonotactic blocking through structural immunity

19

Markednessi is the constraint responsible for the alternation in question. In its single state, it is ranked below Faithfulness, so that no alternations are expected to occur. However, Markedness, is also locally conjoined with an Anchoring constraint. The conjunction is ranked above Faithfulness. By convention (Smolensky 1997), conjoined constraints are violated only when both conjuncts are violated (crucially by the same portion of the string). Therefore the conjunction in (27) is violated only by those candidates in which a stem-final consonant surfaces as an onset (violating Anchor-R) and fails to conform to Markedness¡. Otherwise, the conjunction is satisfied. Therefore, the conjunction is satisfied by all candidates which satisfy Markednessi as well as by all candidates in which stem-final consonants surface as codas. On the assumption that a consonant must surface as an onset when followed by a vowel-initial suffix, the relevant candidates are only those whose stem-final consonants surface as onsets and which satisfy Markednessj. Only those consonants which resyllabify may alternate. Lubowicz uses Polish First Velar Palatatization to illustrate the analysis; here, we apply it instructively to using Turkish velar deletion data. In both Polish and Turkish, a stem-final consonant alternates before a vowel-initial suffix. In the Turkish example, Markednessj is instantiated by *VGV. Recall that a conjoined constraint is violated only if both components are violated within a given domain. Light shading is provided for columns which clarify how violations are computed for the conjoined constraint as a reminder that only the 'Product' column is relevant to determining overall constraint satisfaction/violation. The (subscripted) domain for the conjoined constraint is assumed to be an input VGV sequence, the minimal domain within which violations of *VGV are computed. /sok!ak2, -a/

a.

so.k1a.k2-a

b.

so.kia.-a

c.

so.a.-a

r*VGV X Anchor-Rlvov Relevant *VGV Anchor domains * /okja/ * * /ak2-a/ * /okia/ * /ak2-a/ /okja/ * /ak2-a/

Faith

*VGV

Product *okiai *ak2a *okia

*!

*k2 *k, *!k2

Lubowicz's Parasitic Alternations account is built on a valuable insight, namely that simply stipulates this (with Anchor-R), rather than deriving the special behavior of the stem-final segment in any general way. If a way could be found of linking alternations to resyllabification in general (as opposed to violations of Anchor-R specifically), the account would better capture Lubowicz's insight and bring her analyses of morphologically derived environments even more closely in line with her analyses of phonologically derived environments.

20

Sharon Inkelas

3.3.1 Problem 1: Préfixai targets An obvious sort of problem for the stem-anchoring account of nonderived environment blocking characterized by the quote in (26) is presented by targets of alternation which are not part of a stem. In Chumash, for example, an alternation which applies only in derived environments targets segments in a variety of prefixes. As documented by Applegate 1972, a rule of Pre-Coronal Laminalization turns [s] into [Π before another coronal (see Poser 1982, 1993; Kiparsky 1993 for further analysis and discussion):6 (29) Pre-Coronal Laminalization (PCL): s —» J" / _ [cor] s-lok'in J-lok'in 'he cuts it' P· 117 J-tepu? s-tepu? 'he gambles' P· 117 pil=c=nunux picnunux - > 'to fall and blunt tip' P· 118 kajtepet ka-s-tepet -» 'name of a hill, a long hog P· 203 back; it rolls' ma-l-is-tik-Vn - » malij'tikhm 'the first one; that which goes P· 206 first, which goes in' s-liyo? 'pool, pond' (cf. liyon 'to be -» Jliyo? P· 218 deep') ijtik h in is-tik-Vn -» 'to precede, go first' P· 243 Following Poser and Kiparsky, we assume that the feature in question is [-dist]; the constraint ranking in (30) represents the OT implementation of the alternation. — » — >

(30) * [-dist] [cor] » Faith[dist] Like other alternations we have seen thus far, PCL is subject to nonderived environment blocking, as illustrated below (data from Applegate, via Poser 1993:391):: (31) stumukun wastu? slow?

'mistletoe' 'pleat' 'eagle'

*jtumukun *wajtu? *J1ow?

PCL applies only when the intended target is morpheme-final. In this respect Chumash is just like Finnish Assibilation and Turkish Velar Deletion. The importance difference is that the morpheme-final consonant in Chumash belongs to a prefix, not a stem. (32) shows the result of plugging the relevant markedness constraint, *[-dist][cor], into the conjoined constraint schema: (32) * [-dist] [cor] X AlignR(Stem, σ) No tableau should be necessary to reach the conclusion that AlignR(Stem, σ) will always be vacuously satisfied by a segment which is not stem-final in the input. Thus, no prefix-final /s/ can ever violate AlignR(stem, σ) and, therefore, no prefix-final /s/ can ever violate the conjunction in (32).

6 Page numbers refer to Applegate 1972. Applegate does not provide morpheme glosses, and I have not ventured to create them myself. The s-final morphemes observed in these examples mark such functions as third person subject, possessive, and nominalization.

Phonotactic blocking through structural

immunity

21

It might be possible to avoid the Chumash problem altogether by stipulating that PCL affects only affixes (ranking Root faithfulness sufficiently high would accomplish this). Counterevidence would be a prefix containing an /sn/, /si/ or /st/ sequence (but even in such a case, one could point to the prefix as a lexical exception). 3.3.2 Problem 2: Alternation under perfect alignment A related challenge for the theory that 'derived environment' alternations apply exactly when anchoring is violated is presented by Basque. As documented by Hualde (1989) (see Orgun 1996 for discussion), final vowel raising applies only in suffixes - and only when those suffixes are word-final. In the data shown below, the underlying /a/ quality of the suffix vowel surfaces only when the vowel is 'protected' from the end of the word by another suffix (a-c). Crucially, final vowel raising does not apply to stems, as shown in (d): (33) Challenge: Basque a. muti-a-k muti-e b. ondaru-ra-ko ondaru-re c. ari-ka-ra ari-ke d. fabrika

final vowel raising (Hualde 1989) 'boy-DEF-ERG' 'boy-DEF' Raising applies 'bound for Ondarroa' 'to Ondarroa' Raising applies 'throwing of a stone' 'throwing a stone' Raising applies 'factory'

In Basque, the vowels that undergo the alternation are the ones which are word-final (and, of course, syllable-final). They are perfectly right-anchored in every way. The Basque example, like Chumash, might be handled in an entirely different manner. Rather than trying to make Raising parasitic on an Anchoring violation, it might be possible to characterize Raising as a condition on word-final vowels which is ranked below Root-faithfulness: (34) Faith-Root » *[-high]]„ This would account for the data, but weakens the theory of Parasitic Alternations. Based on the Chumash and Basque examples, nonderived environment blocking effects in affixes would have to be handled differently from those in stems. Affixes would fall outside the generalization that derived environment alternations occur under anchoring violations. 3.3.3 Problem 3: Stem targets which are not resyllabified Even when affixes are set aside, however, the generalization behind Parasitic Alternation theory does not extend to all nonderived environment blocking effects within stems. In particular, tying the application of alternations to violations of stem anchoring fails to account for nonderived environment blocking effects which involve no concomitant resyllabification or other phonological alternation. A striking example is presented by Uighur Raising, in which the intended target is a vowel. It never

22

Sharon Inkelas

resyllabifíes under suffixation. Although a following consonant may resyllabify with the suffix, resyllabification is coincidental and not necessary. Recall that Uighur Raising applies to the rightmost vowel in a stem when in an open syllable that is not word-final (i.e. which is followed by a suffix). As seen in the data below, repeated from (18), the event of a stem-final consonant resyllabifying with the suffix is irrelevant to the application of raising in the stem. The second example in (35) has no stem-final consonant: (35)

a.

qazan + i

- » qazini

Raising of lai in stem-final syllable; resyllabification of stem-final C b. bala + lar —> balilar Raising of /a/ in stem-final syllable; no resyllabification of stem-final C No tableau is needed to make the point that there is no consistently concomitant phonological alternation on which to make Raising parasitic. The only event that accompanies Raising in (35) is suffixation itself. The fact that Raising applies only in suffixed forms might suggest the possibility of reparameterizing Anchor to be sensitive to the prosodie word: (36) [*[-high]a X Anchor-R(stem, ω ^ . ^ σ » FAlTH[-high] » *[-high]a This constraint would be violated by any α σ sequence containing a stem-final segment which is nonfinal in a word. By definition (outside of metathesis and suffix deletion), suffixation renders a stem-final segment nonfinal in the word. The problem here is that in words like bala-lar, whose stem contains two lai vowels, both vowels violate the conjoined constraint in (36). The sequence ...la-lar, whose first vowel should raise, violates [*[-high]a X Anchor-R(stem, œ)][_high]CT by virtue of containing (double-underlined) stem-final lai. But [*[-high]a x Anchor-R(stem, co)][.hlgh]o is also violated by the sequence ...bala, whose first vowel should not raise. The following tableau illustrates the problem just described: 7

7 These analyses are highly dependent on the correct choice of domain in which to evaluate the conjoined constraint. Lubowicz (1998:31) explains that "the domain for LC is the smallest domain within which both of the locally-conjoined constraints can be evaluated. ... In cases of morphologically-derived environments ... the Anchoring constraint is still evaluated within a segment, but the markedness constraint requires two adjacent segments to be evaluated, and so the domain for LC is the window of two adjacent segments." In the example of morphological nonderived environment blocking that Lubowicz analyzes, from Polish, the two-segment window appears to come from the statement of the palatalization constraint, which mentions adjacent segments: •[dorsal][coronal,dorsal] [fn.17]. In Uighur I am therefore assuming that because the markedness constraint mentions a vowel and following syllable, that entire string is included in the domain in which the local conjunction constraint is evaluated. If both components of the conjunction are violated anywhere within that domain, I am considering the constraint to be violated. I trust that this is a correct extrapolation of Lubowicz's practice.

Phonotactic blocking through structural immunity

23

*[-high]a X Anchor-R(stem, )[_highio Faith *[-high]a Domain * r - h i g h l a Anchor-R Product * * ** a. ba.la.-lar /a.la/ *! * * * /a.-lar/ * * * * b. ( ['don.zax] 'down' -»'fluffy' Stem-final obstruents before suffixes with a full vowel a. /xoud/N xout + -achtigA - » ['xout^ax.tsx] 'gold' -»'gold-like' b. /di:v/N -» di:f + -achtigA - » ['di:f.,ax.t3x] 'thief -»'thievish' Stem-final obstruents before vowel-initial stems in compounds a. /xoud/N -» xout + aderN -» ['xout.,a:.d3r] 'gold' -»'gold-vein' b. /le:v/ v -» le:f + eenheidN - » ['le:f.,e:n.heit] 'live' -»'family unit'

(5)

Stem-final obstruents before clitics a. /ba:d/y -» ba:t + 'r -» ['ba:tar] 'bathe' -»'bathe her' b. /xe:v/ v -» xe:f + ' m —» [*xe:f3m] 'give' —»'give him' As is clear from examples (2a-b) and (3a-b), we can distinguish two types of suffixes, viz., (i) those in which stem-final obstruents may be resyllabified as onsets and (ii) those which trigger devoicing of a stem-final obstruent. We can also distinguish different types of suffixes on the basis of stress behaviour. The problem is that the suffixes that pattern together with respect to resyllabification are not necessarily part of the same class when we consider stress assignment. In Dutch underived words, main stress is assigned to one of the three last syllables in the word. Stress is assigned to the antepenultimate syllable if the penultimate syllable is open and the ultimate one is closed (see 6a).4 If the final syllable is 'superheavy', i.e., if it ends in a long vowel plus a consonant (-VVC) or a short vowel plus two consonants (-VCC), main stress is assigned to the ultimate syllable (see 6b,c). (6)

Stress assignment with a penultimate open syllable and an ultimate closed syllable a. lexikon ['lek.si:.kon] 'lexicon' b. abrikoos [ a:.bri:.'ko:s] 'apricot' c. president [pre:.si:.'dent] 'president'

In all other cases, main stress is usually assigned to the penultimate syllable: (7)

Stress assignment in other cases a. pyjama [pi:.'ja:.ma:] 'pyjamas' b. Alaska [a:.'las.ka:] 'Alaska' c. elektron [e:.'lek.tron] 'electron'

3 The subscripts Ά ' , 'N', 'V' indicate the morphosyntactic category of the steins and suffixes in question. Syllable boundaries are indicated by a dot. We will discuss syllabification of forms with clitics (as in 5a-b) in more detail in section 6 below. 4 In addition to the regular (predictable) pattern, Dutch allows irregular stress assignment in some words (e.g. peloton, celébes, amérika, olifant). For a more detailed analysis of Dutch regular and irregular stress patterns, we refer to Booij (1995) and Trommelen & Zonneveld (1989).

58

Janet Grijzenhouî and Martin Krämer

With respect to stress assignment, Trommelen & Zonneveld (1989) distinguish three types of suffixes in Dutch: one type changes the stress pattern of the base, another type requires stress to be assigned to the immediately preceding syllable, and the third type does not cause a stress shift in the base. Most suffixes belong to the first type. They conform to regular patterns of stress assignment. In the output of (8a), for instance, main stress is assigned to the antepenultimate syllable, because the final syllable is closed and the penultimate syllable is open (cf. 6a). In accordance with the rules of regular stress assignment, main stress is assigned to the final syllables in (8b) and (8c), because they are super heavy (cf. 6b,c): (8)

Class I derivation a. a:.no:.'ni:m + usN

—» [a:.'no:.ni:.mYs]

b. c.

—> [pro:.dYk.'ti:f] 'productive' [pro:.dYk.ti:.vi:.'teit] 'productivity'

pro:.'dYkt + ie/v/ A pro:.dYk.'ti:f + iteitN

'anonymous writer'

Some adjectival suffixes require that the main stress of the word be located on the syllable preceding the suffix. Examples are -(e)lijk [olok] and -ig [ax]: (9)

Class II derivation a. 'her.tox + -elijk b. do:t +-elijk c. sxon.'da:l + -ig d. leif + -ig

-> -> —» -»

[her.'to:.xo.lok] ['do:.do.lok] [sxan.'da:.lox] [Ίεί.νοχ]

'ducal' 'deadly' 'scandalous' 'corpulent'

Suffixes of type three do not affect the position of the main stress of the word they are attached to. Examples are the nominalising suffixes -aard, -dorn, -heid, -ling and -schap and the adjectival suffixes -achtig, -baar, and -loos. Main stress is on the first syllable in the words in (lOa-c) below, irrespective of the form of the penultimate or ultimate syllable: (10) Class III derivation a. 'her.tox + -dom b. 'tuei.fol +-achtig c. 'sxa:.dyu + -loos

-> ->

['her.toy.^om] ['ίυεϊ.ίοΐ.,αχ.ίοχ] [ , sxa:.dyu. l lo:s]

'duchy' 'doubtful' 'without a shadow'

Trommelen & Zonneveld (1989) assume that inflectional endings and the diminutive suffix belong to the third type, because they do not affect the stress-pattern of the words they are associated to (see 11a) and they may follow derivational suffixes (see llb,c): (11) Diminutive suffix and inflectional endings a. 'her.tox + -/jo/ - » ['her.tox.jo] b.

ro:d

+ -/axtox/ + -/or/

c.

veif

+ -/Inj/

—» ['ro:t.,ax.to.xor]

+ -/on/ - » ['νείί.,ΐιη.οη]

'little duke' 'more reddish'

'quintuplets (pi.)'

On the basis of the observations concerning stress assignment, Trommelen & Zonneveld (1989: 192, 200) propose to organise the lexicon in at least three different levels. This model of the lexicon poses a problem for devoicing phenomena. It presupposes that syllabification only applies at early levels, but not at the level where class III affixation, inflection, and compounding take place. Resyllabification of a stem-final

Final devoicing and voicing assimilation in Dutch derivation and cliticization

59

obstruent is indeed attested in words derived by suffixes of the first two types, i.e. suffixes that conform to rules of regular stress assignment (see 8) and suffixes that require main stress before the suffix in question (see 9). Obstruents before schwa-initial suffixes of type three and before inflectional endings are also resyllabified and surface as onsets (see 12). Before suffixes of the third type that have a full vowel, however, stemfinal obstruents are not resyllabified and they are voiceless (see 13). (12) Inflection a. /ro:dJ b. /le:v/

+ -er + -en

-»

fro:.dar] ['le:, van]

'redder' '(we) live'

(13) Class III derivation + -achtig -> a. /ro:d/ *[ro:.dax.tsx] b. /ro:d/ /ro:t/ + -achtig - » [ro:t.ax.t3x] 'reddish' + -ling *[vei.vliq] c. /veiv/ ->· d. /veiv/ - » /νεif/ + -ling [veif.lnj] 'quintuplets' Examples (12a,b) illustrate that resyllabification takes place in the case of inflectional endings, so that the stem-final consonant surfaces as an onset. Examples (13b,d) and (4a,b), respectively, illustrate that resyllabification does not take place in the case of other class III affixes and in the case of compound formation. It is thus questionable that inflection belongs to the same level as other class III suffixation and compounding. Alternatively, it could be argued that inflectional endings belong to a level on which the phonology is organised in such a way that resyllabification is ordered before final devoicing. Other class III suffixes and compounds may be said to belong to a later level on which resyllabification no longer applies, so that final devoicing may apply to stem-final obstruents. However, in this way, we predict that inflectional endings are attached prior to class III affixation and we incorrectly exclude the possibility that class III affixes may precede inflectional endings, as is the case in examples (1 lb,c) above. So far, we have demonstrated that level ordering does not help to explain the different behaviour of suffixes with respect to obstruent devoicing. In our view, the solution to this problem lies in the different prosodie structure of suffixes which trigger resyllabification and suffixes before which we find final devoicing. This will be the topic of section 3. Before we discuss our solution to this problem, we will first consider a related phenomenon in Dutch which poses an interesting problem for rule ordering, viz. voicing assimilation. Final devoicing is often obscured by voicing assimilation and this has been explained in rule-based analyses by a specific ordering of rules for final devoicing and rules for voicing assimilation. Consider in this respect that some class III suffixes which begin in a voiced obstruent trigger voicing assimilation. In most cases, voicing assimilation is regressive and as a result, a stem-final obstruent is voiced before these suffixes (see 14a,b):5

5 A rule of regressive voicing assimilation is formulated in Berendsen (1983) as follows: [-son] -> [avoice] / — (#) [-son, avoice]

Janet Grijzenhout and Martin Krämer

60

(14) Regressive voicing assimilation a. /e:t/ +-baar -» ['e:d.ba:r] 'edible' b. /uas/ + -baar —» ['uaz.ba:r] 'washable' Progressive voicing assimilation occurs when the right member of a cluster is a fricative, i.e. the fricative assimilates to the obstruent on its left:6 (15) Progressive voicing assimilation /uerk/

+ - z a a m —>

['uerk.sa:m]

'active; e f f e c t i v e '

In clusters with a fricative as the second member, voicing assimilation interacts with final devoicing, i.e. devoicing takes place before progressive voicing assimilation: (16) Final devoicing and progressive voicing assimilation a. /vri:nd/ —> /vri:nt/ + -schap —» ['vri:nt.sxap] 'friendship' b. /ra:d/ - » /ra:t/ + -zaam -> ['ra:t.sa:m] 'advisable' The data in (14) illustrate that devoicing may not take place after regressive voicing assimilation and the data in (16) illustrate that devoicing must take place before progressive assimilation. The order of rules must thus be that final devoicing applies first and that progressive assimilation precedes regressive assimilation. Voicing assimilation is also attested in compounds and across (grammatical) word boundaries. In cases where the second obstruent in a cluster is a stop, regressive voicing assimilation takes place: (17) Regressive voicing assimilation a. /klap/ +/d0r/ -» ['klab.dor] b. /stof/ + /duk/ -> ['stav.duk] c. /hand/ + /palm/ -» ['hant.palm] d.

/sto:v/

+ /pe:r/

-»

e. /klap do:r da mo: lan/ - »

['sto:f.pe:r]

'swing-door' 'duster' 'palm' 'stewing-pear'

[klab do:r da mo:1a] 'a slap by the windmill'

If the second obstruent in a cluster is a fricative, we find progressive voicing assimilation: (18) Progressive voicing assimilation a. /sla:p/ +/zak/ - » ['sla:p.sak] 'sleeping-bag' b. /dak/ + /venstar/ -» ['dak.fen.star] 'dormer' c. /as/ + /vat/ ['as.fat] 'dustbin' d. /klap van da mo:lan/ —» [klap fan da mo:la] 'a slap of the windmill' Final devoicing has to precede progressive voicing assimilation: (19) Final devoicing and progressive voicing assimilation a. /rond/ -»/rant/ + /va:rd/ —> ['rant.fa:rt] 'cruise' b.

/ x r a : v / - » / x r a : f / + /zYxt/ - » ['xra:f.SYXt]

'digging urge'

The order of rule application in the case of compounds and at the postlexical level is thus the same as the order of rule application for class III derivation. A problem arises when we consider voicing assimilation and the past tense affix -de. This suffix has an 6 A rule of progressive voicing assimilation is characterised in Berendsen (1983) as follows: [-son, +cont] -> [-voice] / [-voice] (#) —

Final devoicing and voicing assimilation in Dutch derivation and cliticization

61

initial voiced stop (see 20a) which, according to the rule of regressive voicing assimilation, should trigger voicing of stem-final obstruents. This prediction is not borne out by the facts. Voiceless stem-final obstruents are not voiced before the inflectional ending -de. Instead, after voiceless stem-final obstruents, the morpheme -de is pronounced with voiceless [t]:7 (20) Progressive a. zwaai b. stop c. maf d. tob e. stoov

assimilation with inflectional -de + -de—»zwaaide [zua:j.da] + -de—»stopte [stop.ta] /*[stob.da] + -de—»mafte [maf.t3]/*[mav.d9] + -de—»tobde [tob.da] + -de-»stoofde [sto:v.da]

'wave', 'waved' 'stop', 'stopped' 'snooze', 'snoozed' 'drudge', 'drudged' 'stew', 'stewed'

Another problem concerns devoicing of stem-final obstruents before clitics. In (21a) we find an instance of a clitic with an initial voiced plosive. According to the ruleordering established above, one might expect that a preceding obstruent is voiced due to voicing assimilation (see the last two columns in 22). This is not the case for all speakers, however:8 (21) No regressive assimilation with clitics a. kan + die—»kan-die [kandi:] 'can he' b. vond + die—»vontie [vanti] / *[vonddi] 'found he' c. of + die—>of-tie [ofti] / *[ovdi] 'whether he' Below, we illustrate that the rule-order suggested in the literature (e.g., Berendsen 1983, Zonneveld 1983) makes the wrong predictions for the past-tense suffix -de and the clitic die: stop-de vond-die of-die Final devoicing t u progressive ass. •U· regressive ass. *b d *d d *v d In summary, we found that the level ordering that can be established on the basis of stress assignment does not help to explain the different behaviour of inflectional suf-

7 Zonneveld (1983) proposes that the past tense suffix actually begins in a voiced fricative /δ/ (which is not a phoneme in Dutch). After the rules of final devoicing, progressive assimilation, and regressive assimilation have applied, the resulting voiced or voiceless fricative is transformed into a plosive. Booij (1995:61-64) proposes that the underlying initial stop of the past tense suffix is unspecified for [voice]. He furthermore assumes a phonological rule that spreads the Laryngeal node from a preceding vowel or consonant to this stop. In section 5 below, we present an alternative analysis that assumes neither an underlying voiced fricative nor underspecification. 8 The data below are taken from the literature (Berendsen 1983, Booij 1995), but there is a great degree of variety among speakers of Dutch with respect to these forms. We assume with Berendsen (1983:29) coherency for the data in the sense that they may occur in one speaker/ hearer.

62

Janet Grijzenhout and Martin Krämer

fixes and other class III suffixes with respect to final devoicing. Furthermore, we showed that the order of rules that is needed to explain examples with some class III suffixes, compounds, and the behaviour of consonant clusters across word boundaries does not explain forms which are built by the suffix -de and forms which involve clitics.

3.

The prosodie structure of affixes

To explain the different behaviour of suffixes with respect to final devoicing, voicing assimilation, and stress in Dutch and to explain phonological phenomena which may or may not apply in different morphosyntactic environments in other languages, Booij (1995:111 ff.) and Selkirk (1995) make the following suggestion. In their view, there are different categories of suffixes which may be prosodified differently. The first category consists of suffixes that do not form a Prosodie Word of their own. Following Selkirk (1995), we will refer to these suffixes as 'internal suffixes', because they become part of a larger prosodie unit, i.e., the Prosodie Word. The second category consists of suffixes that form a Prosodie Word of their own and that have an internal stress pattern. In this respect, they resemble morphological stems. From now on, these suffixes will be referred to as 'semisuffixes'. The prosodie structure of semisuffixes ensures that stem-final obstruents are not resyllabified as onsets of the suffix-initial syllable (23b).9 Selkirk (1995) proposes that the morphosyntactic word structure for all suffixes is the same (23a), but the prosodie word structure may be different (23b).10 We illustrate this idea by representing the morphosyntactic word structure and the prosodie word structure for the examples roder 'redder' (see 12a) and roodachtig 'reddish' (13b), respectively: (23) a.

Morphosyntactic word structure Lex Stem

I

/ro:d/

Affix

I

-hr/

b. Prosodie word structure 'internal affix' ( ro:.dar )ω

9

'semisuffix' (ro:t)m ( αχί3χ) ω

In the remainder of this paper, we will use round brackets as boundary markers for the prosodie categories 'Prosodie Word' (PWd) and 'Phonological Phrase' (PPh). Also, the symbol 'ω' denotes 'Prosodie Word'. 10 Following Selkirk (1995:440), 'Lex' (lexical word) designates a morphosyntactic word belonging to a lexical category, i.e., N°, Vo, or A0.

Final devoicing and voicing assimilation in Dutch derivation and cliticization

63

We assume in the remainder of this paper that semisuffixes have the same status as stems underlyingly. Under this assumption, prosodie structure will be assigned by the grammar (i.e., stems and semisuffixes will both get their prosodie word structure by the grammar).11 In (24b) it is shown that an underlyingly voiced obstruent {Ivi) is realised as such when it surfaces in onset position. In (24c), the fricative is also in front of a vowelinitial suffix (i.e. a semisuffix), but in this case, it does not resyllabify as an onset. Due to the fact that it finds itself in a Prosodie Word-final position, it is voiceless: (24) a. die/v/ ->(άί:ί) ω 'thief b. die/v/ +-en —»(Ήί:.ν3η)ω 'thieves' c. die/v/ + -achtig ->(Ήί:ί) ω (,αχΛ3χ)ω 'thievish' The challenge is to account for the fact that a stem-final obstruent may resyllabify before some vowel-initial suffixes, but not before other ones. In section 2, we argued that a derivational account for final devoicing has serious shortcomings. In the remainder of this paper, we will first discuss and criticise a previous constraint-based nonderivational account and, subsequently, we will present our own proposal.

4.

Lombardi's OT-account of final devoicing and voicing assimilation

A recent explanation for part of the Dutch data is found in Lombardi (1995, 1996, 1999). She defends the position that there is no feature specification [-voice]. In her view, voicing in obstruents is expressed by the presence of a Laryngeal node. To account for devoicing, Lombardi (1996) proposes a constraint which prohibits segments with features for voicing (i.e., *LAR). This markedness constraint interacts with faithfulness constraints which say that underlying featural specifications should remain the same. One group of familiar faithfulness constraints are the so-called "Identity constraints" introduced by McCarthy & Prince (1995) which say that a segment in the output should have the same value for a particular feature as the corresponding segment in the input: (25) IDENT(F) (McCarthy & Prince 1995:264):12 Let α be a segment in Si and β be any correspondent of α in S2. If α is [yF] then β is [yF]. (Correspondent segments are identical in feature F.) 11

The fact that semisuffixes never occur in isolation, even though they are stems, according to our view, can be explained by their semantically 'defective' content. They bear grammatical information, but no conceptual information. The linear order of 'real' stems and semisuffixes (e.g., a noun precedes a semisuffix and not the other way around) can be derived from the same principle that governs the order within real compounds ('right-hand head rule'). Since semisuffixes are derivational affixes, they determine the categorial status of the construction they belong to and this makes them syntactic/semantic heads. 12 A list of the constraints used in this paper is given in the appendix.

64

Janet Grijzenhout and Martin Krämer

Lombardi (1996, 1999) points out that it is usually more important to be faithful to onset laryngeal specifications than to coda - or elsewhere - specifications. She proposes to limit the scope of Identity constraints by so-called 'positional requirements' to ensure that obstruents in onset positions do not devoice. For example, the identity constraint in (26) requires that the underlying value for the feature [voice] of an obstruent surfaces as such if it is realised in onset position: (26) IDENTONSET(Laryngeal) (IDOnsLar): Onsets should be faithful to underlying laryngeal specification. According to Lombardi (1996), Dutch only allows a laryngeal specification (i.e., voicing in obstruents) in onsets. This is guaranteed in her analysis by the ranking IDONSLAR »

*LAR »

IDENTLAR:

(27) 'deed' Input: /da:d/ [da:d] Fta:t]

a. b. c.

IDONSLAR

*LAR IDENTLAR **! **

*! *

RDA:TL

*

Lombardi suggests that assimilation can be attributed to a constraint which she calls that requires adjacent obstruents to have the same value for voicing:

AGREE

(28) AGREE:

Obstruent clusters should agree in voicing.

In a cluster of two obstruents where one of the obstruents appears in onset position, the obstruent in onset position will be faithful to its underlying voicing specification at the cost of the preservation of the underlying voicing specification of its lefthand neighbour. This is ensured in Lombardi's analysis by ranking IDONSLAR above the markedness constraint on voiced obstruents (*LAR): (29) /kas/ + /buk/ 'cash-book' according to Lombardi (1996, 1999) a. b. c.

riP

Input: /kas/ + /buk/ [kaz.buk] [kas.puk] [kas.buk]

AGREE

IDONSLAR

*LAR

IDENTLAR

**

* *

*! *!

*

However, examples with fricatives as second members of consonant sequences (see 15, 16, 18, and 19 above) show that voicing is not always preserved in onsets. In Lombardi's analysis, the postobstruent fricative voicing constraint given in (30) is assumed to capture the fact that the identity constraint for onsets (IDONSLAR) may be overruled in the case of fricatives: (30) Postobstruent Fricative Voicing constraint (FRICVOICE):

* [-son] [-son] [+cont] [+voice] Lombardi acknowledges that this constraint is not very elegant and in section 5 we will propose an analysis which does not use this constraint.

65

Final devoicing and voicing assimilation in Dutch derivation and cliticization

In the following tableau, Lombardi's analysis is exemplified for a case where it works properly. Ranking FRICVOICE and AGREE above I D O N S L A R will give the correct output for a case where the input consists of a voiced plosive followed by a voiced fricative: (31) /ra:d/ + /za:m/ 'advisable' according to Lombardi (1996, 1999) a. b. c. d. • r

Input: /ra:d/ + /za:m/ [ra:d.za:m] [ra:t.za:ml [ra:d.sa:m] [ra:t.sa:ml

FRICVOICE

AGREE I D O N S L A R *LAR IDENTLAR **

*! *!

*

*

*

*!

*

* * **

*

Lombardi (1996, 1999) is not concerned with inflection and clitics and her proposal does not exhaustively explain the Dutch facts. This is demonstrated in the tableaux below. The little skull and crossbones {%) marks the winning candidates according to Lombardi's grammar and the sad smiley (Θ) marks the actual outputs of the Dutch grammar. (32) 'clapped', following Lombardi's analysis a. % b. © c. d.

Input: /klap/ +/ da/ [klabda]

FRICVOICE

AGREE I D O N S L A R *LAR IDENTLAR **

[klapta] rklabtal [klapda]

*

*! *

*!

*

*

**

*

*!

(33) 'snoozed', following Lombardi's analysis Input: /maf/ + lési a. £ [mavda] [mafts] b.® c. [mavtel d. [mafda]

FRICVOICE

AGREE I D O N S L A R *LAR IDENTLAR **

*! *

*!

* *

*

**

*

*!

(34) 'gave he', following Lombardi's analysis Input: /χαν/ + /di/ [xavdi] a.£ b.® rxaftil c. [xavtfl d. Txafdil

FRICVOICE

AGREE

I D O N S L A R *LAR IDENTLAR ** **

*! *!

*

*!

These data will be accounted for by us in the following sections.

*

*

*

*

66

Janet Grijzenhout and Martin Krämer

5.

Prosodie structure and voicing assimilation

Combinations of a stem plus an internal suffix form one domain for stress assignment and, hence, one Prosodie Word. We attribute this to the requirement that a Prosodie Word must correspond to a lexical word (stem + suffix). Selkirk ( 1 9 9 5 : 4 4 5 ) formulates such a type of constraint in terms of Generalized Alignment (McCarthy & Prince 1993):

Align the right edge of every Prosodie Word with the right edge of some lexical word (Ν, V, or A).

(35) ALIGNR P W D :

The prosodie structure in (36) for a form like schandalig [sxan.'da:.bx] 'scandalous' illustrates that the Prosodie Word is right aligned with a lexical word (here an adjective) and that internal suffixes are part of a Prosodie Word: (36)

PWd Foot

Foot

I

σ

Δ sxan

σ

Δda:

σ

Δ lax) Τ internal affix

In this paper, we defend the view that final devoicing may be formulated as a locally restricted markedness constraint against voicing in obstruents. In particular, we will argue in favour of a context-free constraint against voicing in obstruents (37a) and in favour of a specific constraint against voicing in obstruents at the end of a Prosodie Word (31b). (37) a. Markedness: *[+voice] i.e., obstruents are voiceless, b. Positional markedness: *[+voice] )ω i.e., Prosodie Word-final obstruents are voiceless. We will show shortly, that the fact that devoicing of obstruents is obviously ignored within the Prosodie Word (see 38a), but not across a Prosodie Word boundary (see 38b) can be attributed to the different ranking of (37a) and (37b) with respect to Identity constraints.

13

Note that it is not necessary to assume a positional markedness constraint on final devoicing which makes reference to the domain of the syllable, i.e., *[+voice])„, or a local conjunction like {NOCODA & *[+voice]}. (Such constraints are proposed by, for instance, Ito & Mester 1998 and Féry 1998 for German, and by Krämer, to appear, for Breton).

Final devoicing and voicing assimilation in Dutch derivation and cliticization

67

(38) PPh I PWd I F / \ σ σ I I a. / t o b / + / d a / A f f i x ( t o b . d a )pwd 'drudge' +past-> 'drudged'

PPh

b. /xav/+/di/Ciitic 'gave' + 'he'

I I ->((xaf) PW d ti )PPh —> 'gave he'

In (38a), the constraint *[+νοϊοε])ω is vacuous and the underlying voicing specification of both the stem-final obstruent and the affix-initial obstruent is unaltered. In (38b), *[+voice])m plays an important role, since the underlying voicing specifications of both obstruents are lost in order to comply with this constraint. The following tableau illustrates that the three constraints ONSET (i.e. syllables should have an onset), ALIGNR PWD, and *[+νοίΰβ])ω give the correct output for the verbal stem /xe:v/ 'give' and the nominalizing affix -/ar/: (39) /xe:v/ + -/ar/ 'someone who gives' a. b. c. d. e.

Input: xe:vStem + arAffix f(xe:v) .ar] f(xe:f) .ari [(xe:).var 1 [(xe: .var)l [(xe:v. ar)]

ALIGNR PWD *! *!

* [+voice] )ω ONSET * * *

*!

*I

( ) = Prosodie Word boundaries; [] = lexical boundaries Candidate (39a) violates *[+voice])M, because the Prosodie Word-final obstruent [v] is not voiceless. In candidates (39a,b,c), the right edge of the Prosodie Word is not the right edge of a lexical word in violation of ALIGNR PWD. Candidates (39a,b,e) violate ONSET. Candidate (39d) does not violate any of the three constraints mentioned above and it is thus the winner or the 'optimal' candidate. Below we show a tableau for a lexical word which consists of a stem followed by two internal affixes. The constraint requiring that a lexical word must correspond to one Prosodie Word (ALIGNR PWD) is violated in (40a,b) because the only Prosodie Word is not at the right edge of a lexical word. The winning candidate (40d) has an onset and, hence, one violation less than (40c):

68

Janet Grijzenhout and Martin Krämer ( 4 0 ) /prordYkt/ + -/i:v/ + -/itsit/ 'productivity' Input: pro:dYkt St em + i : v A f f i x + iteit A ffix

a. b.

Γ (pro:dYkt).i:.viteit ]

c.

[ (pro:dYk.ti:f.i.teit) 1

d.

f (pro:dYk.ti:.vi.teit)]

Γ (pro:dyk ti:v).i.teit ]

ALIGNR

*[+νοίΰε3) ω

ONSET

PWD *

*! *

*!

*

*!

We still have to account for the fact that stem-final obstruents are voiceless before the so-called 'semisuffixes' and in compounds. Stem-final obstruents cannot be onsets of semisuffixes due to the fact that the latter form a Prosodie Word of their own. Recall that, in section 3, we assumed that these suffixes behave like stems. The constraint that ensures that morphological stems (and semisuffixes) form an independent Prosodie Word domain is a morphology/prosody Alignment constraint which says that the left edge of a stem (or a semisuffix) coincides with the left edge of a Prosodie Word: Align the left edge of every stem with the left edge of some Prosodie Word. In tableau ( 4 2 ) , A L I G N L STEM is violated in candidates a-d, because the left edge of the semisuffix -/axtax/ does not coincide with the left edge of a Prosodie Word in these candidates. The constraint A L I G N L STEM must be ranked higher than A L I G N R P W D , because it incurs a fatal violation in candidate (d) below. Candidates (e) and (f) violate A L I G N R PWD. Candidate (e) violates *[+νοίΰβ])ω as well and candidate (F) wins, even though it violates O N S E T (which must, hence, be ranked below *[+voice])ω ). ( 4 1 ) ALIGNL STEM:

(42) /di:ν/ + -/dxtax/ 'thievish'

a. b.

c. d. e. f.

Input: diVstem + aXtaXsemisuffix f(di:v). ax.tax ] [(di:f). ax.tax ] Ï(di:).(vax.t3x)] [(di:.vax.t9x)] f(di:v).(ax.t3x)] f(di:f) .(ax.tex)l

ALIGNL

ALIGNR

STEM

PWD

*!

*

*!

*

*!

*

*[+voice])ra

ONSET

*

* *

*! * *

*!

* *

Now consider the situation where both an internal affix and a semisuffix are involved. The constraints proposed in this section account for final devoicing before semisuffixes as well as for the fact that final obstruents of semisuffixes may be onsets of internal affixes. In candidate (43d) below, the violation of A L I G N L STEM is fatal, and candidate (43 F) wins, even though it violates A L I G N R P W D once.

69

Final devoicing and voicing assimilation in Dutch derivation and cliticization (43) /ro:d/ + -/axtax/ + -/ar/ 'more reddish' Input: ro:dStem ALIGNL ALIGNR *[+νοίοε])ω STEM PWD + axtaxsemisuffix + arAfr,x * * a. [(ro:d). ax.tax .ar] *! * b. [(ro:t). ax.tax .ar] *! c. i(ro:t). (ax.tax).ar] d. [(ro:. dax. ta. xar)l *! **! e. [(ro:t). (ax.ta).xar] * [(ro:t). (ax.ta.xar)] ALIGNL STEM

ONSET ** ** **

*

•

also ensures that members of a compound form a Prosodie Word of

their own: (44) /xoud/ + /a:dar/ 'gold-vein'

a. b. c. d. e.

Input: xoudstem + a:darStem f(xoud).a:.dar 1 [(xout). a:.dar] [(xou).(da:.dar)l [(xou.da:.dar)] [(xoud).(a:.dar)] [(xout) ,(a:.dar)l

ALIGNL

ALIGNR

STEM

PWD

*!

*

*!

*

*!

*

•f+voice]^

ONSET

*

* *

*! * *

*!

* *

This is not the end of the story, however. There is another element with the same surface form as the suffix -/ar/. This element, a reduced form of the pronoun haar 'her', is a clitic and it behaves differently. In particular, stem-final obstruents may be voiced before the suffix -/ar/ (see 36d), whereas stem-final obstruents are always voiceless before vowel-initial clitics. We here suggest that this may be attributed to the different prosodie requirements of stems, internal affixes, and clitics. Schwa-initial affixes have to be incorporated into a prosodie structure, whereas the weak forms of personal pronouns are clitics which are not part of a Prosodie Word. Rather, clitics are directly linked to a Phonological Phrase and do not belong to the same Prosodie Word as their hosts:14

14

Selkirk (1995) refers to elements which do not belong to a Prosodie Word as 'free' elements. We will discuss the prosodie structure of clitics in more detail in section 6.

70

Janet Grijzenhout and Martin Krämer (45)

PPh

( (χε:ί) ω

3r)pph

'give her'

If we assume that the host of a pronominal clitic is a lexical word, the constraint ranking established thus far correctly predicts final devoicing before clitics. The tableau in (46) differs from the one in (39) above in that there is no internal affix involved and a Prosodie Word boundary is not required to the right of the clitic -ar. Tableau (46) also illustrates that *[+voice])O, and ALIGNR P W D may not be violated in the optimal form, whereas ONSET may be violated. ONSET should therefore be ranked below *[+νοίοβ])ω a n d ALIGNR P W D .

(46) /xe:v/ + /ar/ c. /vond/ -»

Lexical Level (devoicing) /hep/ +/ik/ /xe:f/ + /at/ /vont/ + hkJ

Postlexical Level (resyllabification) -> [he.pik] 'have I' —» [xe:.fat] 'give it' - » [von.tik] 'found I'

In contrast Booij, we note that there is no reliable means to determine the syllabic position of medial obstruents in words like hepik, xe:fot, or vontik and we assume that they should be treated as ambisyllabic obstruents rather than obstruents in onset position. This is most obvious in the case of hepik, since most linguists (including Booij) assume that there are no syllables in Dutch that end in a short lax vowel. Hence, the syllable [.he.] is impossible and the syllabification given in (70a) is therefore questionable.

78

Janet Grijzenhout and Martin Krämer

In the account given in this paper, intervocalic obstruents in cliticized forms belong simultaneously to the right edge of a Prosodie Word and the left edge of the next syllable. As such, they are subject to *[+νοίεε])ω which outranks general IDENTITY constraints. Furthermore, the underlying voicing specification of these obstruents is not "protected" by IDENT PWOS, because they are not in the onset of a prosodie word. Another issue which may cause potential problems for accounts of Dutch cliticization is that the pronunciation is subject to variation. Ernestus (1997) found that in both frequent and non-frequent instances of a stem plus a clitic, a stem-final stop may be pronounced voiced or voiceless when followed by a vowel-initial clitic. Her findings do not show a preference for either pronunciation. In this respect, we note that for postverbal pronouns three possibilities are attested. The pronoun may be realized as a full pronominal form in which case we find stem-final devoicing (71a), or it can be a clitic in which case we also find stem-final devoicing (71b), or it may form a stronger unit with its host in which case we do not find stem-final devoicing (71 c): (71) Different realisations of heb ik 'have I': a. hep Á (two words) b. hepik/hebak (Xo plus clitic) c. hebak (Xo plus affix) According to Selkirk (1995), function words can constitute independent Prosodie Words (as in 68a, 71a), they can be 'free' and adjoined to a higher prosodie category (see the clitic in 68b, 71b), or they can be part of a Prosodie Word (71c). If Ernestus is right, we have to assume that weak forms of personal pronouns are sometimes considered to be 'free clitics' (as in 71b) and sometimes they are considered affixes (as in 71c).

7. Conclusion Stem-final obstruents in Dutch behave differently in derivation, compounding, and cliticization. With respect to derivation, we have argued that one group of affixes are 'internal affixes' which do not form independent Prosodie Words, but form a Prosodie Word together with a preceding stem (e.g., -ar, -da, -ax, -dak, -(t)ja). Another group of affixes are 'semisuffixes' which, like the members of compounds, form a Prosodie Word on their own (e.g., -αχtax, -ba:r, -lo:s, -sxap, -za:m). We have also argued that weak forms of personal pronouns are clitics that are not incorporated in a Prosodie Word and do not form a Prosodie Word of their own (e.g., ar, at, za):

Final devoicing and voicing assimilation in Dutch derivation and cliticization (72) a.

PPh

b.

c.

PPh

PWd

PWd PWd

Ft

Ft

79

PPh

Ft

Λ

σ σ σ σ σ σ Stem AffiXj, Stem Semisuffix Stem Clitic •int (e:d). (ba:r) (ba:t). ar (ro:. dar) 'redder' 'edible' f (Num) * Num f (Num) = { q, p, r } Agreement between è- and -gb can easily be accounted for when agreement features are represented as indices on the lambda abstractors of verb and noun (see Wunderlich 1994 for a complete formalization). The 'switch' morpheme consists of the function f (Num), which applies to the inherent number specification of the noun ('Num', [+pl] in (28a)), deriving a distinct set of feature values ('f (Num)', f ([+pl]) in (28b)). Because in the Kiowa case this new set is distinct from the basic set (f (Num) ψ Num, cf. (27)), distinct agreement with both basic and derived forms is possible. (28) a.

b.

â: [+pi] pole λχ[+ρ1] POLE(x)

[class II] 'poles' (>1)

á : [+piydÒ fl;Num)

pole-GO λχ Γ (i+pl]> POLE(x)

'one pole' (singular)

In 'real' flip-flop phenomena, the basic and derived feature value sets are identical as shown in (29). Consequently, agreement with f (Num) would be impossible, because it would be indistinguishable from Num, only allowing one and the same verbal affix to agree with both basic and derived form.

15

Even though orthographically similar, the affix is phonologically distinct from the 'switch' affix è- by nazalization.

120 (29) Num

Carsten Steins =

{ χ, y } 4r 4 => f (Num) = Num f (Num) = { y, χ } The affixes of the nominal and verbal paradigm are shown in (30). In the nominal system only the 'switch' morpheme applies. The verbal system is somewhat richer: It has got two 'switch' morphemes: the default è- and the marked ά- for [+human] nouns. Additionally, the verbal system also has proper number prefixes: è- for [+pc,+pl] (dual) and gyà- for [+pl]. (30) a. Nominal suffixes -go f (Num): α -α b. Verbal prefixes èf (Num) àf (Num),+human è+pc,+pl gyà- +pl These affixes and specifications are sufficient to derive the correct combinations as in (31). The specification of the 'switch' morpheme is more specific than the proper number features [+pc,+pl] and [+pl], taking precedence over these affixes and blocking their affixation. (31) a. à:-dò è-cél pole[+pl]-f(Num) f(Num)-be.set Ά pole is set up (there). ' b. à: è-cél pole[+pl] [+pc,+pl]-be.set 'Two poles are set up (there).' c. à: gyà-cél pole[+pl] [+pl]-be.set 'Several poles are set up (there).' The agreement behaviour of class IV nouns still needs to be explained though. Class IVa is the agreement we would expect all of them to take. Recall that class IV nouns are unmarked for number in the lexicon, thus not taking the 'switch' morpheme -go. Consequently, in verbal agreement è- cannot apply since there is no f (Num) index it could agree with. In dual and in plural they agree with the standard dual and plural morpheme è- and gyà-. Assuming that class IVb nouns also lack inherent number marking, we have to explain why they fail to induce agreement with the plural morpheme gyà-. According to Watkins (1984), all nouns of this subclass are collective nouns. This would explain why they fail to trigger plural agreement, since they are already collective.16 In order to formalize this, the context feature [-collective] has to be added to the specification of

16 Note that there are also collective nouns in Class II. These, too, fail to receive the plural agreement marker gyà-.

How to account for non-concatenative phenomena in a morpheme-based theory

121

the plural agreement prefix gyà-, preventing it from agreeing with any collective nouns:17 (32) gyà- +pl / [-collective]

4. Conclusions The present analysis has dealt with some of the main data that have been presented in the literature as counter-evidence for morpheme-based theories. A close inspection of these data has shown that a morpheme-based theory is actually able to analyze this data. At no stage has the notion of 'morpheme' as a lexical entry consisting of a combination of exponent and semantic information been at stake. While this obviously is not proof that morpheme-based theories are superior to 'a-morphous' theories, it nevertheless shows that they certainly are capable of analyzing data which have previously been thought impossible for morpheme-based theories to handle.

In the examples cited in Watkins (1984) there are also a few nouns in this class which arguably are not collectives (e.g. road, pond). I would have to assume these to be irregularly marked as [+collective], bearing the morphological feature without providing interpretative relevance.

122

Carsten Steins

References Hale, Kenneth (1965). Some preliminary remarks on Papago morphophonemics. Internationaljournal of American Linguistics 32, 295-305. Hale, Kenneth (1983). Papago (kjc. International Journal of American Linguistics 49, 299327. Hill, Jane; and Zepeda, Ofelia (1992). Derived words in Tohono O'odham. International Journal of American Linguistics 58, 355-404. Lombardi, Linda; and McCarthy, John (1991). Prosodie circumscription in Choctaw morphology. Phonology 8, 37-72. Martin, Jack (1988). Subtractive morphology as disassociation. West Coast Conference on Formal Linguistics 7, 229-240. Noyer, Rolf (1997). Features, positions, and affixes in autonomous morphological structure. New York: Garland. Stonham, John T. (1994). Combinatorial morphology. Amsterdam: Benjamins. Watkins, Laurel J. (1984). A grammar of Kiowa. Lincoln, NB: University of Nebraska Press. Wonderly, William; Gibson, Lorna F.; and Kirk, Paul L. (1954). Number in Kiowa: Nouns, demonstratives, and adjectives. International Journal of American Linguistics 20, 1-7. Wunderlich, Dieter (1994). Towards a lexicon-based theory of agreement. Theoretical Linguistics 20, 1-35. Wunderlich, Dieter; and Fabri, Ray (1995). Minimalist Morphology: An approach to inflection. Zeitschrift für Sprachwissenschaft 14, 236-294. Zepeda, Ofelia (1983). A Papago grammar. Tucson: The University of Arizona Press.

Some lexicalist remarks on incorporation phenomena Stephen R. Anderson

1. Introduction: Two approaches to Noun Incorporation* An important theme in the study of language during the twentieth century has been the balance between the grammar of words (morphology) and that of phrases and sentences (syntax). Some writers, such as Morris Swadesh (1939), have seemed to treat this as a merely terminological issue, referring to (much of) morphology simply as 'internal syntax'. More recently, some (e.g. Selkirk 1982, Williams 1989) have argued explicitly that the mechanisms necessary to describe the internal organization of words are a direct extension of syntactic principles. Others, in contrast (cf. among others Anderson 1992b, chap. 2; Zwicky 1992) have argued that there are substantive differences between morphological and syntactic principles which have important implications for the organization of linguistic knowledge. This discussion has largely centered on the analysis of a small number of constructions for which both syntactic and (more purely) morphological accounts seem to be available. If the syntactic account can be substantiated in these cases, the more general conclusion seems warranted that syntactic mechanisms are (or at least can be) responsible for combining pieces into complex words. If, on the other hand, there are reasons to prefer the autonomously morphological analysis in such circumstances (which are, after all, the best candidates for syntactic word formation prima facie), that conclusion casts doubt on the more general perspective from which structurally autonomous elements are manipulated syntactically to form unitary words, a notion that pervades much recent work. * The work represented here was supported in part by grants number SBR-9514682 and SBR9876456 from the US National Science Foundation to Yale University; an earlier version appeared as Anderson (1997). This paper has benefited from discussion in seminars at Yale during the past several years, and at the 'Lexicon in Focus' conference in Wuppertal on 19 August, 1998. I am especially grateful to the members of my seminar on Clitics at Yale during 1996-97 (David Harrison, Lizanne Kaiser, Matt Richardson and Jennifer Vanloon), some of whom also participated in a seminar devoted to a prepublication version of Baker (1995) that was kindly provided to us by Mark Baker. The comments and criticisms received on these occasions have been very helpful. I have also benefited from the insightful comments of two referees for the current volume.

124

Stephen R. Anderson

The topic of this paper is one widely discussed instance of such a morphological structure that looks as if it might be formed within the syntax: Noun Incorporation. This is a feature of the grammar of a wide range of languages, including (as a semirandom, quite small sample) Mohawk, Chuckchee, Southern Tiwa, Classical Nahuatl, and many others. Noun Incorporation has attracted attention for quite some time: grossly, it is a construction in which, on the surface, a verb stem and a noun stem constitute a single word, with the noun stem representing an argument of the verb. The most typical incorporated nouns correspond to the direct object of a transitive verb, as in the examples in (1) below. (1)

Southern Paiute (Sapir 1911, p. 263) a. qam'úyaai- num- puva' jackrabbit- hunt- USIT- REM.PAST 'He used to hunt jackrabbits' b. cû'q'uc'" qam'úv ax qa- q-a' one jackrabbit- killCOMP.SS 'Having killed one jackrabbit,...'

In some languages, the underlying object/derived subject of an unaccusative verb can incorporate; in other languages, unergative subjects can incorporate too. 1 These possibilities are illustrated in (2) below. (2)

Chukchee (Polinsky 1990) a. qeyk-ak ?9l-3-lg-a-g?i hill-LOC

b.

snow-EVID-melt-EVID-AOR.3SG

'On the hill, the snow melted' atlon qinqe-et-3-l?et-g?e he

child-lNTENS-EViD-come-AOR.3SG

'He got many children' In a few cases, the incorporated noun appears to be some sort of adjunct, like an instrumental or a locative. (3)

a.

b.

(Huauhtla) Nahuatl (Merlan 1976): ya? ki-kocillo-tete?ki panci 3SG 3SG.O-knife-cut bread 'He cut the bread with a knife' Chukchee (Skorik 1948, apud Spencer 1995): gatg-alqat-gîe walwagan lake-go-3SG.S

raven.ABS

'Raven went to the lake' Noun Incorporation has been an important topic for a long time. In the early years of this century, it was argued by some to constitute a particularly characteristic feature of North American languages. Kroeber (1909), in an attempt to rebut that notion, argued that in fact there was no such thing as Noun Incorporation, but in a classic paper Sapir 1 Polinsky (1990) provides arguments for the 'unaccusative' vs. 'unergative' contrast in the verbs of these examples.

Some lexicalist remarks on incorporation phenomena

125

(1911) provided the first reasoned analysis of Noun Incorporation in a variety of languages. There are two basic views of the nature of Noun Incorporation, and both have been around for a very long time. One of them, which we can refer to as the 'syntactic' story, treats Noun Incorporation as a syntactic process, by which an argument of the verb (or at least part of that argument) is actually moved from its syntactic A-position to adjoin to the verb (see e.g. Mardirussian 1975). In favor of this analysis a priori is the natural account it offers of how the incorporated stem comes to fill the semantic (Θ) role of a corresponding unincorporated noun. Proponents of this view in the recent literature include Jerry Sadock (1980, 1986), who argued, from the apparent need to form words in the syntax by moving a noun into the same word as a verb, that the lexicalist prohibition against syntactic manipulation of the internal form of words must be wrong. Actually, the construction discussed by Sadock in Eskimo is one (the formation of 'denominal verbs') that several other authors (Sapir 1911; Mithun 1984, 1986; Gerdts 1997) have wanted to distinguish from Noun Incorporation, but as he puts the issues they are quite similar, regardless of terminology.2 A conceptually similar (but technically very different) position is developed by Mark Baker, who has explored it at some length in one book (Baker 1987) and made considerable use of it in another (Baker 1995). Baker also maintains that nouns move in the syntax to take up their incorporated positions inside of verbs. But Baker's theory of the kind of movement involved goes much farther, and in fact is the basis of the rather more general notion of'head movement' in syntax. This, in turn, is at the technical heart of much contemporary theorizing about clause structure, because it is head movement that is central to the notion that the inflectional content of a clause is composed of a large number of functional heads that move around syntactically. If head movement were not an established notion, much of the 'split-INFL' account of clause structure would have rather less plausibility; and if Noun Incorporation were to turn out not to involve syntactic movement, the original (and empirically richest) support for head movement would disappear. The alternative to this syntactic movement analysis is the 'lexical' account of Sapir (1911), Mithun (1984), Anderson (1992b) and others, which says that the noun plus verb combinations we call 'Noun Incorporation' are built in the lexicon, not in the syntax. As with the syntactic analysis, the lexical one has some apparent prima facie advantages. One sort of fact that seems to point to a lexical analysis is that the shape of incorporated nominal elements often differs phonologically from a corresponding free form in unpredictable ways. In some instances, such variation is limited to a small number of suppletive forms: e.g., Mohawk "-nahskw- 'domestic animal' appears only incorporated, while a semantically equivalent stem -tshenv appears only as an independent N[oun]" (Mithun 1984, p. 876). In other languages, however, this phenomenon may be quite pervasive. Thus, in the Australian language Tiwi, "[incorporated forms and free forms are generally not cognate, [...] as the free forms which were cognate with the 2 For some discussion of Sadock's arguments in a context quite similar to that of the present paper, see Anderson (2000).

126

Stephen R. Anderson

existing incorporateci forms have long since disappeared, owing to the rapid rate of lexical replacement in Australian languages" (Osborne 1974, p. 48). Some representative pairs of incorporated and free forms are given in (4). Note that in some cases, some degree of phonological resemblance can be seen (especially in some body part names), but in others there is no apparent similarity. Also, in some cases a single free form corresponds to more than one incorporated form, while in others the reverse is true. (4)

Tiwi (Osborne 1974): incorporated wiuatiuqintamiawariawarimaratiqa giliwante(i]9) parplajnti-

free -wiuatiqa -jaraqintamiua -wuiara tamuripa teraka teraka -taraqa

gloss forehead nose belly navel (live) wallaby (dead) wallaby buffalo

A language may show semi-systematic relations between free and incorporated forms, where these do not in fact reflect phonological regularities found elsewhere. A particularly interesting instance of this is found in the Munda language Sora and some of its relatives, as documented and analyzed by Zide (1976). In this language, (the free forms of) a number of nouns can be regarded as derived from verbal or other roots by the addition of prefixes, suffixes or infixes: thus, db-ga 'feed' is related to dr-db-ga 'food (being fed to an infant)'; kot) 'shave' is related to k-sn-oy 'razor', etc. In some cases, the combining form of such a noun can be made by simply removing such added morphological material; thus, the combining form of k-an-oy 'razor' is - koy, and that of gatas i 'play' is - gasi (cf. the verb gasi- 'to play' from which the noun is derived by infixation). This relation has been massively over-generalized and inverted, however; in that a huge number of polysyllabic free form nouns have incorporated forms produced by removing non-existent prefixes, suffixes and infixed -VC- sequences so as to reduce the form to a canonical -CVC shape. For details, see Zide (1976); our point is simply that the elements appearing in incorporation structures are quite distinct from semantically similar free forms in ways that do not follow from the independent phonology of the language, but rather point to lexical processes. Just as the phonological form of incorporated and non-incorporated elements may require specifically lexical description, the semantics of an incorporated nominal construction may also differ, again unpredictably, from the interpretation of a corresponding construction in which the nominal is represented by a free NP. (5)

a. i.

Mohawk (Baker 1995): tu-s-a-yu-[a]t-háh-a-hkw-e' DUP-ITER-FACT-F.SG.S-SRFL-road-0-pick.Up-PUNC

'She started her journey (lit.: picked up the road)' ii. # tu-s-a-yú-([a]te)-hkw-e' ne DUP-ITER-FACT-F.SG.S-SRFL-pick.up-PUNC

'She picked up the road (literal reading only)'

oháha

PRT road

Some lexicalist remarks on incorporation b. i.

phenomena

127

(Huauhtla) Nahuatl (Merlan 1976) i. 0-nec-maka-0-k pa?tli 3 SG-1 SG-give-PAST-SG

medicine

'He gave me medicine' ii. 0-nec-pa?-maka-0-k 3sG-1 SG-medicine-gi ve-PAST-SG

ii. i.

'He doctored me' tesiwitl w e c i - 0 - 0 hail

fall-PRES-SG

'(What's falling? - answer:) Hail is falling' ii. tesiwi-weci-0-0 hail-fall-PRES-SG

'(What's the weather like? - answer:) It's hailing' Neither of these types of idiosyncrasy is to be expected if Noun Incorporation is simply a syntactic process that may (or may not) apply to a fully articulated underlying syntactic structure to move some of its lexical material from one position in the tree to another. Such arguments have not received much attention in the syntactic literature, but this does not mean they are negligible: syntacticians tend to assume that both the phonology and the lexical semantics will take care of themselves, and the possibility that this might not always be the case often carries little weight. These sorts of idiosyncrasy point away from the syntax in the analysis of incorporation structures, and toward the locus of item-specific information, the lexicon. We must ask, of course, whether such [ v [n stem] stem] combinations are in fact within the scope of the type of regularity found in the lexicon, but the answer is surely in the affirmative. We build such combinations in the lexicon anyway in the case of 'synthetic' compounds: duck hunting ([ v [n duck] [v hunting]]) is built from [n duck] and [v hunt] and has an interpretation in which the noun is taken as specifying one of the arguments of the associated verb. While some have argued for a syntactic incorporation analysis of such compounds, the appeal of such a view is limited, since compounds are evidently structurally parallel to non-compound members of basic lexical categories. The possibility of building all lexical compounds in the syntax was tried out in the early days of generative grammar (Lees 1960), but that was really before there was any theory of the lexicon. Nowadays most authors agree that the formation of compounds takes place in the lexicon, not the syntax. Of course, compounds like duck hunting involve not only a relation between the noun and the argument structure of the verb, but one which is like the one we find in Noun Incorporation constructions. Indeed, in both cases the noun most typically corresponds to an argument of the verb that would fill the syntactic direct object position and the θ-role of THEME. This is not always the case, however. In compounds like earthquake, sunrise, landslide, etc., the noun apparently represents the subject of an intransitive verb. In these examples, we might invoke the Unaccusative hypothesis and say that the argument in question represents an underlying direct object. Such an account is less plausible, though, for examples such as crybaby, flashlight, workman,

128

Stephen R. Anderson

playboy, hand laundry, cottage industry, etc. that seem to involve a noun specifying the agentive argument of an unergative verb. What does seem constant about all of these cases is the fact that the associated noun consistently corresponds to the θ-role of THEME. Other compounds, though, correspond to other non-thematic NI types: e.g. hand laundry, cottage industry. Once we admit the possibility that all of these verb-argument(/adjunct) relations can be established by a lexical rule, as they must be if we are to deal with true compounds, the initial motivation for a syntactic account of NI disappears. Sapir (1911) was probably the first to propose that Noun Incorporation constructions are actually instances of lexical compounding. Since the possibility of a 'syntactic' analysis in the modern sense was not really open to him, though, his arguments for this position came mostly from the formal consideration that incorporated structures involve a combination of two stems, like compounds, together with phonological and semantic idiosyncrasies of the sort adduced above.

2. Some (ultimately neutral) arguments Noun Incorporation, then, is a construction type with some overall properties (the combination in a single word of verbal and nominal stems, where the nominal generally supplies information about one of a small number of thematic types from the argument frame of the verb, and where the whole functions as a verb), and whose analysis can be approached from either of two perspectives (syntactic or lexical) each of which seems to have some natural conceptual affinities with the construction's basic character. The question to be addressed is which of these (if either; or perhaps some combination of the two) is correct. This may seem like an issue of limited import, but in fact its resolution has major implications for linguistic theory. Baker's work has focused for some time on an account of syntactic incorporation and its generalization from the core cases of Noun Incorporation to other head-movements. It is quite important for him, then, that syntactic Noun Incorporation be at least possible (even if other sorts of Noun Incorporation exist too). On the other hand, for those who maintain that "the syntax neither manipulates nor has access to the internal form of words" (Anderson 1992b, p. 84), as at least some versions of the Lexicalist Hypothesis require, it is quite important not to allow syntactic rules to put words together in this way. The choice of analyses for Noun Incorporation constructions thus has a good deal of importance for the theory of how morphology and syntax are related. And as noted above, Baker's theory of head-movement (in support of which Noun Incorporation is the most obvious empirical domain) is central to the articulatedINFL account of clause structure common to much current syntax, so any challenge to the syntactic nature of Noun Incorporation undercuts this view, at least indirectly. A choice between the two accounts does not seem to follow from the basic descriptive properties of Noun Incorporation. Both views seem to be able to describe the fact that the Noun stem supplies content associated with a thematic position in the argument

Some lexicalist remarks on incorporation phenomena

129

structure of the verb. As far as the limitations on which positions can be involved, the lexicalist view accommodates these thematic restrictions rather straightforwardly: lexical rules often refer to the relation of THEME, and compounding in particular does so. If Noun Incorporation is simply a form of noun-verb compounding, this is exactly what we would expect. As we have seen, the incorporated noun is not always a THEME: it is sometimes a locative or an instrumental, as in (3). But this is again quite parallel to the facts about compounds, and supports the view that there is a single regularity at work in the two cases. The syntactic view derives this result in a way that is, at a minimum, rather less direct: some might find the account in Baker (1995, sec. 7.3) of why only the direct object position is accessible somewhat tortuous, but for the sake of discussion let us assume that such an account is at least possible. Similarly, we will assume that the phonological and semantic idiosyncrasies referred to above can find a home within the syntactic account: idioms, for instance, provide a clear precedent for the assignment of non-compositional interpretations to syntactically complex structures, and it is at least possible that suppletion and other non-phonological variation occurs as a function of syntactically derived environments. If we are to choose between the two theories of Noun Incorporation under consideration, it will probably be on the basis of detailed study, not simply as a consequence of the basic descriptive regularities that define the construction. Baker proposes to do just this. His primary evidence that Noun Incorporation is syntactic comes from Mohawk, for which an analysis is developed at length in Baker (1995). He suggests that a lexical noun can be generated (as the exhaustive content of a NP) in argument position, and then moved to adjoin to the governing verb. Why does this movement take place? According to an argument that he supports in great detail, overt NP's are not licensed in A-positions in Mohawk, but only appear in adjunct positions, where they serve as appositive expressions (a position similar to, but not identical with, the position of Jelinek 1984). If such a [NP [noun]] were generated in an A-position, it would thus be ill-formed. In order to get θ-marked, such a NP has to be co-indexed with an element in the verb (a key subpart of what Baker calls the Morphological Visibility Condition, or MVC). If that element were agreement, the agreement would (on Baker's hypothesis) absorb the case assigning property of the verb, so the overt NP would still be ill formed. But if we move the noun to adjoin it to the verb, the MVC is satisfied, and the otherwise caseless NP no longer has phonetic content, and no well-formedness conditions are violated. The Morphological Visibility Condition is a parametrically determined characteristic of certain languages (called 'polysynthetic' by Baker, in a usage that deviates somewhat from the traditional sense of this term). We may ask, naturally, which languages have (syntactic) Noun Incorporation? Baker's answer is: those that have to. That is, in Mohawk, incorporation is forced, as above. In English, on the other hand, as in most languages, the MVC does not hold, and so movement is not forced. But on minimalist assumptions, if you do not have to do something, you have to not do it. As a result, Noun Incorporation is impossible in English, since it is not forced.

130

Stephen R. Anderson

Now in fact many languages have constructions that look like Noun Incorporation (i.e., cases where noun plus verb together seem to form a single word, and the noun is interpreted as an argument of the verb). They differ quite a bit from one another, though. For one thing, in most Noun Incorporation languages, the incorporated noun is always interpreted as indefinite and/or generic. This is comparable to the interpretation of nouns in (English) lexical compounds. (6)

a. b.

She's a truck-driver, which is why she has back problems. *She's a truck driver, which is why it's parked over there.

If you say She's a truck-driver, that means she drives trucks, not (just) some specific truck. The truck in this compound is not available for anaphoric reference, since it is necessarily generic (or at least referentially underdetermined). In contrast, in Mohawk, an incorporated noun can refer to something that is referentially specific or definite. (7)

(Mohawk, cf. Baker 1995, p. 288) a. ThetÁre' wa'-ke-nakt-a-hnínu-' yesterday FACT-1 SG.S-bed-0-buy-PUNC Ί bought a bed yesterday' b. I-k-ehr-e' Uwári Λ-ye-núhwe'-ne' 0-lSG.S-think-IMPF

Mary

FUT-F.SG.S-like-PUNC

Ί think Mary will like it (the bed)' This is just what would be expected if the incorporated noun comes from an NP in an A-position, since such an NP can perfectly well be specific or definite. If the lexical analysis is to remain viable, it must do something to accommodate this possibility. It looks like the right thing to do is to allow the verb plus noun structure that constitutes a lexical compound verb to take arguments in the thematic position that is (also) specified by the incorporated noun stem. The most common view to be found in the literature probably is based on the assumption that when a noun is compounded with a verb, the noun satisfies (or 'saturates') the corresponding argument in the verb's argument structure, with the noun also being interpreted generically or indefinitely. Assuming that nouns themselves have an external θ-role to discharge (the 'R-role', connected with the noun's possibility of referring), this generic interpretation corresponds to a certain sort of binding of the noun's own external θ-role, along with the corresponding variable in the interpretation of the verb. We might say that the relevant compounding operation, like many others, introduces an appropriate generic operator binding the logical variable corresponding to the noun's R-role and the verbal θ-role identified with it. To accommodate the facts of languages where an incorporated noun is potentially referential, we need to extend this account. Let us say that in some languages, at least, noun-verb compounding is an operation that 'unifies' the semantics of the noun with the argument position of the verb, but without saturating the argument itself. That is, in such a language, fish-catch is a transitive verb meaning 'X catches Y, Y a fish'. This is essentially the nature of what Rosen (1989) has called 'classifier incorporation', and which others have proposed for at least some incorporation structures (e.g. Di Sciullo

131

Some lexicalist remarks on incorporation phenomena

& Williams (1987), Spencer (1995); a similar view was presented orally in Anderson 1985). On this view, the verb in (8a) would have an associated semantics something like (8b). (8) a. [v [N fish] [v catch]] = 'X catches Y, Y a fish. ' b.

fish R FISH

+ CATCH

catch

Agent

Theme

=> CATCH

fish-catch

Agent

Theme(R) FISH

This verb is still syntactically transitive; and while its THEME argument is specified as having the properties of a FISH, it is not logically or referentially bound, and so the verb still takes an argument (perhaps just pro) which can bear independent reference. In this picture, Mohawk differs typologically from other languages, including those in which incorporated arguments are always necessarily generic, not in having syntactic (vs. lexical) Noun Incorporation, but rather in the fact that the lexical Noun Incorporation rule does not saturate (or bind, logically or referentially) the argument position. The semantic operations here are of a sort that we need independently, to deal with lexical operations of compounding and also with phenomena like the lexical suffixes of Wakashan and Salish languages (see Anderson 1992b, Bach 1998), and the denominal verbs of e.g. Eskimo (constructions often conflated with Noun Incorporation: see Gerdts 1997). There is another typological difference among languages with respect to Noun Incorporation. In some languages an incorporated NP can be doubled by an external NP whose content also specifies the corresponding A-position. There are potentially two cases here: first, where the external NP consists only of modifier material, as in I a new (one) bed-bought; and secondly, where there is a head N, too, as in My father eight bullheads fish-caught?

3 An anonymous reader points out that the lexical and syntactic accounts make different predictions here in one possible case. Consider a language like English, with a required overt head (e.g. one) in this construction. Baker's approach seems to predict that such a language could nonetheless have headless NP's just in case the head was incorporated, since in syntactic terms, the phrase would have a head as required. The lexical approach taken here, in contrast, predicts that an overt head will be required, if at all, even in incorporation structures. I know of no data that bears on this issue, but simply record the prediction here.

132

Stephen R. Anderson

[e]j

fishj-

bought

b. Tohka niyohserá:ke tsi nahé several so.it.year.numbers so it.goes c. sha'térku nikú:ti rabahbót wahutsyahní:nu ki rake'niha eight of.them bullhead he.fìsh.bought this my.father '[Several years ago,] my father bought eight bullheads' The first case is straightforward for Baker's original head-movement account: you just say that the head alone moves, leaving any modifiers, determiners, etc. in situ. The second case was much more problematic for the analysis of Baker (1987), of course, because how could the head have moved if it is still in its original position? His response was to say that in these sentences, the doubling NP is actually an adjunct, related to a separate (phonetically null) NP in argument position. And of course, on the analysis presented in Baker (1995), that is claimed to be the case for all overt NP's. (9) above provides a schematic analysis in these terms of the relevant parts of a Mohawk sentence taken from Mithun (1984, p. 870). In this structure dotted lines indicate anaphoric links, and the solid line indicates a movement relation. How does the lexicalist describe constructions in which an incorporated noun stem is doubled by all or part of an NP in argument position? On that view, when an incorporated Ν is doubled, we can say essentially the same thing as Baker: the A-position is filled by an empty category, and the overt (adjunct) NP forms a chain with that position in the same way as other overt argument expressions. The only structural difference is that if there is an empty category in the A-position, as Baker argues for Mohawk, it is an instance of pro rather than of trace. Notice that the presence of the empty category itself follows not from anything about Noun Incorporation, but rather from Baker's independent arguments about the position of overt NP's in Mohawk. Without those, we could also just say that the overt NP is in the expected argument position. Since the incorporated Ν did not originate there, there is no syntactic reason why some other expression might not fill this position. In the case of the apparently headless NP's, where it looks as if head movement has taken place, we can note that the expressions themselves are well-formed as NP's in

Some lexicalist remarks on incorporation phenomena

133

the relevant languages. In Mohawk, as in most languages, there is no overt correspondent of English one, so the object phrase in I want a [new] one consists of just the Adjective new. Such a noun phrase arises not by movement of its head out of the NP to adjoin to a governing verb, but rather by the selection of pro as head of the NP, with the semantics of the nominal being supplied from context (or verb-internally, in a case like (8b). On this view the headed and headless cases of doubling NP's fall together, at least in principle. With this apparatus in place, we can approach a typology of referentiality and 'doubling' within the class of Noun Incorporation languages. Three major classes of construction must be distinguished: (10) Types of 'doubling' of incorporated N's: a. No independent reference, no doubling b. 'Classifier' incorporation (doubling with or without independent reference) c. Free reference, no doubling The first of these is the case in which the noun is necessarily interpreted as non-specific and/or generic, and where no doubling expression can be present. In this situation, we assume that the compounding operation (a) identifies the semantic variables corresponding to the noun's external θ-role and the appropriate verbal argument; and (b) introduces an appropriate logical operator binding this combined variable, resulting in a semantics like (11). The syntactic subcategorization of the resulting verb is reduced by the elimination of the argument position corresponding to this variable. In some languages, the result is that transitive verbs become intransitive as a result of incorporation. In others, the morphological properties of the derived verb may remain transitive, although no argument corresponding to the 'incorporated' element is present. We assume that this is an instance of the formal dissociation of morphological and syntactic transitivity, as discussed in Anderson (1992a). X (11) GEN

A

(CATCH

Agent

Theme FISH

Free reference with the possibility of doubling ('classifier incorporation') is parallel, but lacks the verb-internal logical operator binding (and saturating) the relevant argument: in this case, two semantic variables are available for external specification by expressions in A-positions. Of course, since the semantics of the incorporated noun stem have been unified with one argument in this representation, the externally specified expression must be consistent with that: thus, the information provided by the object of (8b) must (at minimum) be consistent with the property that what it designates is FISH.

Free reference but no doubling (as in e.g. Nahuatl or Southern Tiwa) must be the case where the semantics of Ν unifies with that of a verbal argument position, but the resulting verb has some property that is incompatible with an overt NP. Phonologically empty pro can appear in the A-position, however, and supplies the possibility of independent reference. What could the relevant property excluding overt argument expres-

134

Stephen R. Anderson

sions be? Baker's own account of a language like Southern Tiwa, where doubling is excluded but non-generic reference in association with an incorporated noun is allowed, is that the condition licensing adjuncts in association with A-positions only allows them to be licensed by pro, and not by a trace (of head-movement); but even this stipulation will not suffice within a theory in which incorporated nouns are produced lexically (and thus correspond to pro in A-position, rather than to the trace of head movement). What account of the difference between doubling and non-doubling languages is available on lexicalist assumptions? Apparently, argument positions corresponding to an incorporated Ν have some property that is incompatible with the structure of a verb derived by Noun Incorporation. One possibility is the following: suppose that Noun Incorporation in a language like Mohawk (which permits external doubling NP's) not only unifies the semantics of the Ν with that of the corresponding argument within the semantics of the verb, but also eliminates the noun's external R θ-role (or perhaps merges it with the θ-role assigned to that argument). In Southern Tiwa, on the other hand, the incorporated N's R θ-role is identified (but not merged) with the θ-role of the argument. That means that if the A-position contains an expression headed by a noun that assigns its own R θ-role, there is a θ-criterion violation. Another way in which languages may differ requires further investigation, but the outlines of a solution are reasonably clear. Languages that permit doubling may differ in which of the following schematic sentence types they admit: (12) a. b. c.

John fish-caught three trout. John fish-caught a fish. (*) John trout-caught a fish.

As far as is known, all languages that permit doubling of an incorporated noun stem allow sentences such as (12a), where the doubling expression is more specific and detailed than the incorporated noun alone. Many (though apparently not all) allow sentences like (12b), where the two are essentially synonymous. None however, appear to allow sentences like (12c), where the doubling expression is actually less specific and detailed than the incorporated noun. These facts appear to result from a requirement (semantic or pragmatic, depending on one's view of where the line between these is to be drawn) that overt expressions be at least minimally informative with respect to the information already provided as part of the verb's semantics. (12c) violates this because the information provided by the overt NP is actually a proper subset of that already provided by the verb, while the overt NP in (12a) clearly supplies information above and beyond that present in the verb. Languages apparently differ on the basis of whether they consider an essentially equivalent expression 'informative' or not (perhaps by virtue of the possibility it introduces of independent referentiality). One further point should be noted about the properties of doubling NP's. In many cases, these provide information about the nature of the corresponding argument itself: in a Mohawk sentence like (13), for instance, the overt argument expression and the incorporated noun combine to specify a single referent 'a polka-dotted dress':

Some lexicalist remarks on incorporation

phenomena

135

(13) Kanekwarúnyu wa'-k-akya'tawi'tsher-ú:ni it.dotted.DIST PAST-I-dress-make Ί made a polka-dotted dress' In other cases, however, the external NP may supply information about a possessor of the argument in question, as in (14). (14) Rembarrnga (Dixon 1980) tiq? paqa-warnta-na-0 woman lSG/3PL-track-see-PAST Ί saw the women's track' Note that although it is a single track that is referred to, the agreement is plural, due to the fact that it is the track of more than one woman, even though the NP tirj? 'women' is not overtly marked as plural. The overt argument expression, thus, is not simply something with a phonologically null head (as in English I was looking for a track, and I saw the women's [0]). Note that the agreement is with a third person plural, showing that 'women' is the head of the object NP, not simply a modifier. In such cases (often referred to as 'possessor ascension,' especially in the literature of Relational Grammar; or more theory-neutrally as 'external possession' in Payne & Barshi 1 9 9 9 ) , the semantics must be capable of construing an external argument expression as specifying the possessor of an (already partly specified) argument, not the argument itself.

3. Some less neutral arguments By and large, up this point, the syntactic and lexical theories of Noun Incorporation are 'tied' in that each can be said to account for roughly the same range of phenomena the other can. Baker is quite fair: he considers the lexical account (though not with all of the details supplied here), and says essentially the same thing. In fact, he says not only that some languages have lexical/morphological Noun Incorporation rather than syntactic Noun Incorporation but that even Mohawk has this, in addition to syntactic incorporation. This makes the theory rather close to unprovable: any fact that appears to argue against the syntactic story is dealt with by saying that in such a case, the incorporation is lexical. But Baker also discusses some phenomena which he feels argue for the syntactic account over the lexical one. As the syntactic and lexical views have been elaborated above, they converge to a great extent as far as the representations they assume. But there is one difference: for Baker, the empty category present in A-position in association with a verb that has undergone syntactic Noun Incorporation is a trace, while in the lexical Noun Incorporation case it is a pro. As a result, any way these two possibilities could be teased apart could provide a way of discriminating between the theories. On this basis, Baker ( 1 9 9 5 , pp. 3 1 4 - 3 2 9 ) offers three arguments that at least some cases of syntactic Noun Incorporation exist. The first of these concerns agreement. He argues that in general, there is no agreement with the position corresponding to the source of Noun Incorporation. In this

136

Stephen R. Anderson

respect, he differs explicitly from Postal (1979), who claimed that there is agreement with an incorporated noun. For Baker, agreement occurs with a position containing pro, while syntactic incorporation ought to leave not a pro, but a trace. Absence of agreement with Noun Incorporation verbs, as opposed to its presence with simple pro, would then argue for the kind of syntactic difference he assumes. To evaluate this argument, we must consider some limitations on Mohawk Noun Incorporation. In fact, it is almost exclusively inanimates that are incorporated. This is particularly interesting in light of the fact that in Mohawk, agreement with an inanimate object is indistinguishable from no agreement at all. Incorporation of animates is generally disfavored, and regarded as pejorative (implying the treatment of a person as an object, for example). Why should this restriction obtain? Baker admits to having no explanation for why incorporation of animates ought to be avoided. One possibility, though, is that it is a consequence of the fact that it is precisely in this case that the speaker can 'fudge' the issue of whether agreement is or is not present, similar to the use of modal constructions in English in those cases where no particular agreement seems right. /1 -».τ ·,» , j??am/??is/??are coming! t t . . (15) Neither cFred nor TI < ... f to the party. will come * J Other examples of the same sort are provided by Pullum & Zwicky (1986). They note, for example, that in German, the conjunction of verbs taking dative objects with those taking accusative objects is only felicitous when the object NP is ambiguous between the two cases. (16) a.

Er findet und hilft Frauen he finds(+ACC) and helps(+DAT) women 'He finds and helps women' b. * Sie findet und hilft Männer(n) she

finds(+ACC) and

helps(+DAT)

men(DAT)

'She finds and helps men' The tension in Mohawk which is resolved by limiting incorporation to inanimates might be described as follows: where the verb contains an incorporated Ν referring to an overt argument NP, that seems like 'enough' to a speaker, and so the presence of a separate agreement element seems superfluous. Where the transitive object is inanimate, however, the agreement morphology is ambiguous as to whether it contains a marker referring to the object or not, so no surface discomfort results. The bottom line, in any case, is that in most cases of Noun Incorporation, there is no overt indication of whether agreement is present with the 'incorporated' position or not, because of the formal similarity between Sbj/InanimateObj markers and intransitive Sbj markers. In addition, as Baker notes, in cases where there is no external 'doubling' NP we can always say that the lexical operation of incorporation has constructed an intransitive verb (by saturating the argument), so the absence of agreement would follow on either the lexical or the syntactic view.

Some lexicalist remarks on incorporation

137

phenomena

(17) Tu-t-a-yako-kétoht-e' DUP-CIS-FACT -F.SG.O-appear-PUNC 'She appeared carrying a baby.'

ts-e-wir-Aháwi ITER-F.SG.S-baby-carry.STAT

In some cases, though, there is clearly an animate noun incorporated: either there is no agreement with this noun, or else it is (exceptionally) treated as if it were inanimate. (18) Ra-wir-a-núhwe'-s thíkA (owirá'a) M.SG.S-baby-0-like-HAB that (baby) 'He likes that baby' This is the sort of incorporation structure that is crucial to Baker's argument, since (on his analysis) the lack of agreement is forced by the fact that the relevant position is occupied, after movement, by a trace. The lexicalist, on the other hand, is led to say that in such a case, the NP referring to the baby is treated as inanimate. Since (animate) agreement is obligatory in such cases when incorporation has not taken place, Baker concludes that the argument position in a sentence like (18) must be filled with a trace rather than with (inanimate) pro. Baker (1995, p. 335) asserts that "if Mohawks can freely view babies as inanimate entities, then one would expect that they could trigger inanimate agreement even when not incorporated [as in (19b) below -sra], contrary to fact." This conclusion does not necessarily follow, however. It might well be the case that the semantics of the incorporating stem -wir- 'baby' differ subtly from those of the independent noun owirá'a, perhaps in being underspecified for animacy. (19) a.

b.

c.

shako-núhwe'-s

(ne

owirá'a)

M.SG.S/3PL.O-like-HAB

PRT

baby

'He likes them (babies)' * ra-nuhwe'-s (ne owirá'a) M.SG.S-like-HAB PRT baby 'He likes them (babies)' ra-wir-a-núhwe'-s M.SG.S-baby-0-like-HAB

'He likes babies' On the other hand, it seems that with animate incorporated objects, agreement is at least optional. (20) a.

b.

Uwári ye(-ruwa)-kstA-hser-Áhaw-e' ne Mary F.SG.S(/M.SG.O)-old.person-NOM-carry-IMPF PRT 'Mary is holding my father' Wa'-ke(-hi)-kstA-hser-áhset-e'

rake-'níha my-father

FACT-1 SG.S(/M.SG.O)-old.person-NOM-hide-PUNC Ί hid the old person (the old man)'

And in fact Baker (Baker 1995, p. 336) observes in a footnote that "when the doubling material makes explicit the gender of the argument in question, the Noun Incorporation plus agreement construction is preferred where possible." Since agreement ought to be impossible in any case of (syntactic) incorporation via movement, this would appear to be a strike against his analysis.

138

Stephen R. Anderson

So how does the lexicalist derive the result that agreement and Noun Incorporation do not generally co-occur? The facts are obviously rather complicated. There appears to be a preference for avoiding a situation in which both overt agreement material and an incorporated noun refer to the same participant. One way to resolve this tension is to treat the agreed-with position in a Noun Incorporation construction as if it were inanimate, in which case no overt marker appears: that provides the basis for the preference for inanimates as incorporated elements, and the sense that incorporation of animates is somehow pejorative (since the overt agreement pattern would imply treatment of the NP in question as inanimate). If this could be maintained consistently, we could say that Noun Incorporation constructions do indeed have morphosyntactic agreement, even though this has no overt phonological consequence. Actually, Baker notes that in some languages which are otherwise syntactically like Mohawk (Tanoan languages like Southern Tiwa and Gunwinjguan languages like Mayali), overt agreement does appear with positions that are also associated with an incorporated NP. In Ainu, which he puts in the class of polysynthetic languages (incorrectly, as shown in Kaiser 1997), some dialects have agreement with the NP position associated with Noun Incorporation and some do not. Wherever we find agreement with an incorporated position in a 'polysynthetic' language, that presents a problem for Baker's analysis. On the other hand, the lexicalist account derives these cases with little difficulty, since they represent plain transitive verbs. In fact, quite adequate responses can also be given, apparently, to Baker's other two arguments. One of these derives from the fact that in Mohawk the object position associated with a Noun Incorporation verb cannot be questioned with a general-purpose question word. For Baker, this follows on the syntactic analysis from the fact that the A-position cannot simultaneously contain a noun to be incorporated and a question word. There are two subcases to this argument. The first is that of who questions. (21) Úhka wa'who

ke

khe

.

ksá-ht-a-ya'k-e'?

*lSG.S lSG.S/F.SG.O

'Who (a child) did I slap?' In this case, we see that if agreement is overtly present, the sentence is acceptable; while the absence of agreement leads to ungrammaticality. As Baker points out, the variant with agreement validates the predictions of the lexicalist account; but he suggests that the impossibility of the alternative without agreement supports the syntactic account, since it would only follow (according to him) from the incompatibility of agreement and movement traces. But this does not in fact follow: the badness of the sentence without agreement (treated on the lexicalist view as involving inanimate agreement) can be due simply to an evident agreement conflict. That is, the question word 'who' makes it explicit that the object is animate, which conflicts with apparent inanimate agreement.

Some lexicalist remarks on incorporation phenomena

139

For wAa¿-questions the account is a bit different. In (22) there is no animacy conflict between the question word and the incorporated object 'meat' but the sentence is still ungrammatical. (22) * NahótA wa-ha-'wáhr-a-k-e'? what FACT-M.SG.S-meat-0-eat-PUNC 'What did he meat-eat?' Baker again concludes that the incompatibility of question words with incorporation in such sentences follows from his syntactic analysis, but there is an alternative that is readily available to the lexicalist: the badness of (22) would also follow from the requirement that an overt argument has to be informative. Since 'what' adds nothing to the semantics of 'meat-eat' that would further specify the properties of its object argument, it fails to meet that requirement. Again, the lexical account accommodates these facts without syntactic movement. Baker's final argument is based on Condition-C effects, and appears to be subject to some of the same objections that apply to his other points. He notes that in sentences like those in (23), coreference between the matrix subject and the argument corresponding to an incorporated noun is at best marginal. (23) a.

b.

(Sak) ra-tsháni-s tóka Λ-ke-ksá-ht-a-yak-e' Sak M.SG.S-fear-HAB maybe FUT-lSG.S-child-NOM-0-hit-PUNC 'He (Sak)i is afraid that maybe I will slap the childj/??¡' Sak í-hr-ehr-e' ks-kstA-hser-akéras Sak

c.

0 - M.SG.S -think-lMPF

N.SG.S-old.person-NOM-stink

'Sak¡ thinks the old personj /?? ¡ smells bad' Sak í-hr-ehr-e' a-ke-kstA-hser-áhset-e' Sak

0 - M.SG.S -think-lMPF

OPT-lSG.S -old.person-NOM-hide- PUNC

'Sakj wants me to hide the old personj/??¡' Baker interprets the verb forms in the embedded clauses in these examples as involving an incorporated object and intransitive agreement. He suggests that the difficulty of getting coreferential interpretations would follow if the position corresponding to the incorporated noun contained a trace, rather than pro, and that trace exhibited Condition-C effects (rather than the Condition-B effects we would expect of a pro). Another interpretation is quite possible, however. We might say that the agreement with the positions corresponding to incorporated nouns in (23) is actually neuter, rather than missing (as is indeed overt in the case of the basic intransitive (23b). In that case, the difficulty of getting a coreferential reading for these sentences would follow from an apparent gender conflict between the matrix subject (overtly animate and masculine in all cases) and the lower NP, apparently identified as inanimate. This interpretation is strengthened by Baker's example in (24), where appropriate transitive agreement with the 'incorporated' position appears, and where coreference is perfectly acceptable. (24) (Sak) ra-tsháni-s Sak

M.SG.S-fear-HAB

tóka

Λ-hi-ksá-ht-a-ya'k-e'

maybe

FUT-lSG.S/M.SG.O-child-NOM-0-hit-PUNC

'He (Sak)i is afraid that maybe I will slap the childj/j'

140

Stephen R. Anderson

Given the uncertain status of Condition-C effects in the current syntactic literature, it would not be wise to put too much weight on this argument in either direction, but it certainly appears that even assuming the disjoint reference principles Baker invokes, the conclusion does not necessarily follow that the position corresponding to an incorporated noun is occupied by a trace, rather than pro. Indeed, Baker explicitly suggests that the relevant NP in (24) is pro rather than trace. In summary, the lexicalist account of Noun Incorporation seems entirely viable, even for the phenomena Baker treats as most centrally syntactic.

4. Conclusion So where does this survey of incorporation phenomena leave us with respect to a choice between the two primary views of Noun Incorporation? It appears that even Baker agrees that much Noun Incorporation is in fact lexical, not syntactic, even in the language for which he feels the strongest syntactic case can be made (Mohawk). But in fact, the limited sets of facts for which the syntactic account is said to be necessary can also be accommodated within the lexical account, without invoking mechanisms that have no precedent elsewhere. And that means that a purely lexical account of Noun Incorporation, without syntactic movement, is almost certainly possible. But that, in turn, means that the best putative support for an operation of syntactic head-movement may be non-existent — a conclusion with extensive consequences for many areas of contemporary syntax, especially the split-INFL analysis with its proliferation of functional heads.

Appendix: List of abbreviations ABS ACC AOR CIS COMP DIST DUP EVID F FACT FUT GEN HAB IMPF

absolutive accusative aorist cislocative subordinating affix distributive duplicative evidential feminine factual future genitive habitual imperfective

INTENS ITER LOC

M Ν NOM

O OPT PAST PL PRES PRT PUNC REM.PAST

intensive iterative locative masculine neuter nominative object agreement optative past tense plural present particle punctual remote past

Some lexicalist remarks on incorporation S SG SS SRFL

subject agreement singular same subject marker semireflexive

141

phenomena STAT SUP USIT

stative superessive usitative

References Anderson, Stephen R. (1985). ICak1" ala morphology. Paper presented at Conference on Canadian Native Languages in Theoretical Perspective, SUNY Buffalo. Anderson, Stephen R. (1992a). Syntactically arbitrary inflectional morphology. In Yearbook of Morphology 1991, Geert Booij and Jaap van Marie (eds.), 5-19. Dordrecht: Kluwer. Anderson, Stephen R. (1992b). A-Morphous morphology. Cambridge: Cambridge University Press. Anderson, Stephen R. (1997). Remarks on agreement and incorporation phenomena. In Y.A.L.E. 1: Studies in the morphosyntax of clitics, L. Kaiser (ed.), 29—44. Dept. of Linguistics, Yale University. Anderson, Stephen R. (2000). Lexicalism, incorporated (or incorporation, lexicalized). To appear in Proceedings of the Chicago Linguistic Society. Bach, Emmon (1998). Semantic relations in and outside of complex words. Paper presented at the conference 'Lexicon in Focus', Wuppertal, 19 August, 1998. Baker, Mark (1987). Incorporation: A theory of grammatical function changing. Chicago: University of Chicago Press. Baker, Mark (1995). The polysynthesis parameter. New York: Oxford University Press. Di Sciullo, Anna-Maria & Edwin Williams. (1987). On the definition of word. Cambridge MA: M. I. T. Press. Dixon, Robert M. W. (1980). The languages of Australia. Cambridge: Cambridge University Press. Gerdts, Donna B. (1997). Incorporation. In Handbook of morphology, Andrew Spencer & Arnold Zwicky (eds.), 84-100. London: Basil Blackwell. Jelinek, Eloise (1984). Empty categories, case and conflgurationality. Natural Language and Linguistic Theory 3, 39-76. Kaiser, Lizanne (1997). The interaction of noun incorporation and applicative formation in Ainu. In Yearbook of Morphology 1998, Geert Booij and Jaap van Marie (eds.), 157178. Dordrecht: Kluwer. Kroeber, Alfred (1909). Noun incorporation in American languages. In Proceedings of the sixteenth international congress of Americanists. Lees, Robert B. (1960). The grammar of English nominalizations. The Hague: Mouton & Co.

142

Stephen R. Anderson

Mardirussian, Galust (1975). Noun incorporation in universal grammar. Proceedings of the Chicago Linguistic Society 11, 383-389. Merlan, Francesca (1976). Noun incorporation and discourse reference in modern Nahuatl. International Journal of American Linguistics 42, 177-191. Mithun, Marianne (1984). The evolution of noun incorporation. Language 60, 847-894. Mithun, Marianne (1986). On the nature of noun incorporation. Language 62, 32-37. Osborne, C. R. (1974). The Tiwi language. Canberra: Australian Institute of Aboriginal Studies. Payne, Doris L. & Immanuel Barshi (eds.) (1999). External possession. Amsterdam: Benjamins. Polinsky, Maria S. (1990). Subject incorporation: Evidence from Chukchee. In Grammatical relations: A crosstheoretical perspective, Katarzyna Dziwirek, Patrick Farrell and Errapel Mejfas-Bikandi (eds.), 349-364. Stanford: CSLI publications. Postal, Paul (1979). Some syntactic rules of Mohawk. New York: Garland. Pullum, Geoffrey K. and Arnold M. Zwicky (1986). Phonological resolution of syntactic feature conflict. Language 62, 751-773. Rosen, Sara Thomas (1989). Two types of noun incorporation: A lexical analysis. Language 65, 294-317. Sadock, Jerrold (1980). Noun incorporation in Greenlandic: A case of syntactic word formation. Language 56, 300-319. Sadock, Jerrold (1986). Some notes on noun incorporation. Language 62, 19-31. Sapir, Edward (1911). The problem of noun incorporation in American languages. American Anthropologist (n.s.) 13, 250-282. Selkirk, Elizabeth (1982). The syntax of words. Cambridge MA: M.I.T. Press. Skorik, Pjotr Ja (1948). Ocerk po sintaksisu cukotskogo jazyka: inkorporacija. Leningrad: Ucpedgiz. Spencer, Andrew (1995). Incorporation in Chukchi. Language 71, 439-489. Swadesh, Morris (1939). Nootka internal syntax. International Journal of American Linguistics 9. 77-102. Williams, Edwin S. (1989). Maximal projections in words and phrases. In Alternative conceptions of phrase structure, Mark R. Baltin and Anthony S. Kroch (eds.), 280-291. Chicago: University of Chicago Press. Zide, Arlene R. K. (1976). Nominal combining forms in Sora and Gorum. Oceanic linguistics special publication 13 (Austroasiatic Studies, Part II), 1259-1294. Zwicky, Arnold M. (1992). Some choices in the theory of morphology. In Formal grammar: Theory and implementation, Robert Levine (ed.), 327-371. Vancouver: Oxford University Press.

Case alternation in Finnish copular constructions Elisabeth Löbel

1. Introduction* In syntax, as a rule, copular verbs, especially the verb to be, which represents the paradigm case, are said to be 'defective' due to the fact that they are not capable of assigning a theta-role, i.e. they are assumed to have no theta-grid, no descriptive content. Their main function is supposed to be that of bearing tense, and, therefore, they exhibit some affinity to functional categories. In this paper, I will argue that copular verbs, in fact, do have a theta-grid. In elaboration of the distinction made by Rizzi (1990), I propose that they are part of the small subclass of verbs such as weigh or measure, which select an argument bearing a nonparticipant role (Measure, Name, etc.). The non-participant role selected by copular verbs is Property, i.e. property-denoting expressions are looked upon not as a predicate (in semantic terms), but as a complement. This line of reasoning implies that, as a rule, two types of semantic selection must be distinguished: (i) selection of a participant role for verbs such as hit or write, and (ii) selection of a non-participant role for verbs such as be, become, measure, weigh. Selected non-participant roles differ from participant roles by the following properties: (i) they are coindexed with their respective subjects, and (ii) they are not assigned case. Case-marking is either (i) triggered by case concord (syntactic condition), or it is (ii) free, i.e. triggered by semantic conditions or by default. The paper is organized as follows: Section 2 is a summary of Löbel (2000) where my own approach for copular verbs and their lexicon-theoretic aspects is presented. In section 3, this approach is applied to Finnish, a language with an elaborated case system and case alternation, which represents a challenge to the hypothesis that copular complements are not assigned case. Furthermore, this language is of special

* This paper has been presented on various occasions at the Universities of Cologne, Wuppertal, and Leipzig. For their useful comments and suggestions, I wish to thank Peter Gallmann, Joachim Jacobs, Thilo Tappe, an anonymous reviewer, and last not least, the editors. I would also like to thank Leea Wallraf for assistance on the Finnish examples. The results presented in this paper are part of the DFG project 'Predicative Noun Phrases', SFB 282 "Theory of the Lexicon", Diisseldorf/Cologne/Wuppertal.

144

Elisabeth Löbel

interest as not only nouns, but also adjectives are case-marked, independently of whether or not they are used as predicatives. Section 4 concludes this paper with a summary of the results.

2. Copular verbs and argument structure 2.1

The status of copular complements

There are, in principle, three alternatives to describe sentences with copular verbs: (i) predication is category-neutral, i.e. there is no difference with regard to whether or not the predicate consists of a verbal or a non-verbal category (Williams 1980 and subsequent work), (ii) the copula selects a constituent whose head is non-verbal (Déchaîne 1993), and (iii) the copula selects a small clause (SC), i.e., the subject of the sentence is part of the argument selected by the copula. In the following, I will confine myself to the latter approach. The following characterization of the SC analysis is based on Heggie (1988:117ff.). In contrast to thematic verbs, the copula be does not have a theta grid. The primary function of be is to indicate a predicate in its postverbal position, not to select arguments; it may only select one type of complement, a 'small clause'. Furthermore, any phrasal element may be a predicate under the copula: "Because be does not select for specific phrasal categories, the lexical entry does not constitute a θ-grid. Be never assigns θ-roles" (Heggie 1988:122). Based on these arguments, it is proposed that the copula is listed in the lexicon as a "verbal element which must coindex a constituent which becomes the predicate position of a small clause" (ibid.). Therefore, the lexical entry of be is indicated as follows: (1) be: I i.... XPi] Heggie proposes the D-structure in (2); the copula itself does not assign Case to the predicate NP, whereas the subject NP is raised to the specifier position of IP in order to get Case. Following Williams (1980), Heggie assumes that the index placed on the predicate spreads to the subject via predication: (2) a. John¡ is [NP t¡ [Np a man¡]] b. IP (Heggie 1988:122)

The predicative NP a man is base-generated as a predicate, this being justified by the assumption that a man, in fact, has "more meaning" than the copula, which is supposed

Case alternation in Finnish copular

constructions

145

to be rather meaningless. This kind of syntactic structure, however, does not take into account that the NP/DP a man is a property-denoting expression, and, therefore, a 'predicate', only in virtue of the copula, otherwise it would be an individual-denoting expression. In other words, the base-generation of a man as a predicate is stipulated and not derived by any other criteria. Equative sentences and, therefore, postcopular definite NPs/DPs constitute an apparent problem for the SC analysis. In Rothstein (1995), the difference between definite predicative noun phrases and indefinite ones is accounted for by different c-selection properties. Based on the assumption that predication is a relation between two syntactic constituents, one of which is an argument and the other an open predicate, equative constructions cannot involve a small clause. Thus, for equatives, Rothstein instead proposes to return to the idea developed in Stowell (1978): Instead of subcategorizing for a small clause, be can c-select an NP. Accordingly, the sentence John is Mr. Smith has the structure indicated in (3): (3)

[John [is [Mr. Smith] NP ]vp]ip

In this structure, be "turns out to be a raising verb very like consider, in that it selects either an NP or a small clause. When it selects a small clause, raising of the θ-marked small clause subject is obligatory, since be does not assign case. [...] Where be selects for an NP [= (3), E.L.], no raising has taken place; yet, if the verb is the same as raising be, it cannot have assigned case to Mr. Smith or an external θ-role to John, either" (Rothstein 1995:45). As case is only obligatory for arguments, where arguments are either θ-marked or subjects of predicates, and as the NP Mr. Smith does not constitute such a predicate, the complement of be does not need to be assigned case. Therefore, the question arises as to what the definite postcopular NP/DP Mr. Smith in (3) above constitutes, since, according to her analysis, it is neither an argument nor a predicate. Williams (1997) explicitly denies the existence of real equative sentences : "Often it is said that the subject is referential, and the predicate not; and in particular, when the predicate is a nominal, it is said that it must be a nonreferential nominal. I think that this misses the real difference. Even when both the subject and the predicate are referential, there is an asymmetry" (Williams 1997:324). Based on examples such as (4) for definite NPs and (5) for proper names and their potential inverse orders, he concludes that "(i) predication is an asymmetric relation always, (ii) the semantic content of the asymmetry is epistemic priority based on 'directness of acquaintance', and (iii) the copula permits the asymmetry to run in either direction, but right-to-left is the normal or basic order" (Williams 1997:330): (4)

a. b.

The mayor is the police commissioner, The police commissioner is the mayor.

(5)

a. b.

Clark Kent is Superman. Superman is Clark Kent. (Williams 1997:327f.)

According to Williams, equative sentences reveal that there is no absolute notion of referentiality: "It is sometimes thought that a predicate cannot be 'referential'. It seems to me though that the best we can say is that the predicate is less 'referential' than its subject, and that what we really mean is something having to do with directness of

146

Elisabeth Löbel

acquaintance" (Williams 1997:331). Intrinsically quantified NPs such as every person are excluded from predicate position, as a quantifier must bind a position to which a theta-role is assigned ( *John is every person in that room). In contrast, definites, like indefinites, are intrinsically nonquantificational; what the indefinite and the definite article contribute to meaning is "perhaps nothing more than that the signals 'you can tell which one I mean' and a/n signals 'you can't tell which one I mean'": (6) [John]Np¡ is [the/a mayor]NPi (Williams 1997:332). To summarize the assumptions found in the literature: (i) postcopular constituents are base-generated as predicates on the grounds that they have 'more meaning' than the copula, (ii) 'any phrasal category' may function as a predicate, a criterion which is taken as indicative for the lack of any theta grid for copular verbs, (iii) different cselectional properties (NP complements vs. SC complements) account for the 'referentiality' of definite predicative NPs/DPs in equative sentences, and (iv) the notions of argument and theta-role are implicitly understood as 'referential argument' and 'referential theta-role', respectively. The crucial test case for all these criteria are noun phrases, as they may be propertyor individual-denoting expressions. Their respective readings, however, are dependent on the selectional properties of the verb of which they are the complement. For this reason, in what follows, I take the copular verb itself as a starting-point and present criteria in favor of a refinement of the notions of 'argument' and 'theta-role'.

2.2

Participant vs. non-participant roles

Based on examples such as (7), it is often argued that it is not the type of theta-role which is relevant in syntax (in (7a) John is Agent, in (7b), Patient), but rather the number of arguments: (7)

a. John broke a vase, b. John broke a leg.

The following observation by Rizzi (1990), however, demonstrates that it is necessary to distinguish two kinds of theta-roles and arguments, namely 'referential' and 'nonreferential' arguments and theta-roles. This is illustrated by the contrastive use of ambiguous verbs like weigh. In (8a), the agentive weigh takes a direct object, in (8b), the stative weigh selects a measure phrase: (8) a. John weighed apples. b. John weighed 200 lbs. (Rizzi 1990:78) The two readings of weigh differ in their theta grid: (9) a. weigh (Agent, Patient) b. weigh (Theme, Measure) Patient as well as measure, according to Rizzi, are both theta-marked complements. As it is intuitively plausible to say that the direct object of weigh is referential whereas the measure phrase selected by stative weigh is not, he proposes that the notion of 'referen-

Case alternation in Finnish copular constructions

147

tiality' should be made precise in terms of thematic theory. All selected elements are theta-marked, but two kinds of theta-role should be distinguished: "Some selected elements refer to participants in the event described by the verb [...]; other selected elements do not refer to participants but rather qualify the event (compositionally (measure, manner, etc.) [...]. We can thus distinguish argumentai or referential thetaroles (agent, theme, patient, experiencer, goal, etc.) and quasi-argumental, non-referential theta-roles (manner, measure, [...])" (Rizzi 1990:85f.). This distinction is used to show that referential indices, as a rule, should be confined to referential theta-roles. In contrast, Frampton (1991), on the basis of sentences such as (10), shows that referential indices, hence coindexation, are independent of whether a noun phrase has reference or not: (10) [Not a single guest]j did I wonder why you invited tj. As measure phrases are intrinsically 'non-referential', Frampton proposes to derive sentences such as (1 la) from a small clause structure: (11) a. Jack weighed 100 pounds b. e weigh [Jack 100 pounds] (Frampton 1991:38) This structure implies that "the measure phrase is not an argument of any kind, but a predicate" (Frampton 1991:39). The analysis in (1 lb) suggests that there seems to be a common denominator for verbs such as weigh, which, in Rizzi's approach, select a 'non-referential theta-role' or, according to Frampton's proposal, a small clause, and copular sentences, for which a small clause analysis is commonly assumed. This approach can be justified on the grounds that the measure phrase 100pounds in Jack weighed 100 pounds, in fact, denotes a property which is predicated of John. To be more precise, it denotes a property in the dimension of weight which is predicated of the subject JohnFrom this point of view, the similarity to a predicate nominal such as a teacher in John is a teacher is rather obvious, as the predicate nominal a teacher, too, is a property-denoting expression. In the following, I will take the affinity illustrated in (11) between an NP as a 'nonreferential' argument on the one hand and a (base-generated) predicate on the other as the starting point for arguing exactly the opposite way: Copular verbs, too, select a non-participant role, which I call Property. In essence, I assume that what is a predicate in semantic terms (i.e. a property-denoting expression) does not necessarily imply that it must be one in syntactic terms; therefore, the postcopular property-denoting expression is not a (syntactic) predicate, but a (predicative) complement selected by the copular verb. This is the topic of the next subsection.

2.3

Predicatives as complements

Measure verbs such as weigh (two pounds), measure (three meters), last (two hours), but also verbs of naming such as German heißen 'to be called' as well as copular verbs

1 Thanks to Joachim Jacobs for drawing my attention to this important aspect.

148

Elisabeth Löbel

such as be, become, remain all have in common that they have an external, referential argument bearing the participant role of Theme, and an internal argument bearing a non-participant role, as outlined in (12): (12) external argument internal argument participant role non-participant role a. weigh, measure Theme Measure b. Germ, heißen 'to be called' Theme Name c. last Theme Duration d. be, become, remain Theme Property etc. The common semantic denominator of all these selected non-participant roles is 'Property'. Whereas, due to their specific meanings, the verbs mentioned under (a) to (c) select an internal argument which denotes a property specified for a certain dimension (weight, measure, duration, name),2 copular verbs in (d) have in common that the property which the internal argument denotes is not specified in such a way. What is required is semantic compatibility of the non-participant role with the referent of Theme: (13) a. The meeting was at 10 o'clock. b. *The woman was at 10 o'clock. Differing from Heggie (see (1)), I propose the following entry (14a) for copular verbs; for comparison, the lexical entry of a transitive verb such as hit is presented in (14b): (14) a. be [XP h XP2] [THEMEJ,

b. hit

[DP,,

PROPERTY; ] 3

DP 2 ]

[AGENT, PATIENT]

In contrast to the participant-role Agent which is canonically realized as a DP, the nonparticipant role Property is not tied to any specific lexical or syntactic category, although the adjective might be looked upon as the canonical lexical realization of Property. In other words, it is not contingent to conclude from this potential categorial variation that it is the 'predicate' which is category-neutral.4 Furthermore, as it is indicated by the indices, the lexical entry in (14) entails that Property is to be understood as 'Property of Theme', since Property is a notion which is relational in itself. In fact, all the well-known peculiarities which predicative NPs/ DPs exhibit in contrast to argumentai ones (different interpretation of the (in)definite article, lack of scope, restrictions with regard to long movement, cf. Williams 1994:

2

Cf. the characterization of this type of verbs in Levin (1993); measure verbs "can be used intransitively with a postverbal phrase expressing the measurement; this noun phrase does not show the properties of a direct object" (Levin 1993:272, emphasis E.L.). 3 The thematic grid applies analogously for other copular verbs such as become or remain. 4 Cf. Larson (1988), who claims that lexical properties of verbs should be expressed in terms of thematic grids instead of subcategorization frames.

Case alternation in Finnish copular constructions

149

40ff.), may be dealt with in terms of the non-participant role of Property, without having to assume that they are 'predicates' in syntactic terms. The basic structure in (15) illustrates that the main difference between copular verbs and other two-valued verbs does not lie in the copular verb itself, as it projects just like any other verb, but in the nature of the selected complement:5 (15) Basic structure:

VP subject}

V'

V predicative complementi I be As it is commonly assumed, the subject is base-generated in VP, i.e. it is not necessary to assume a small clause in order to license coindexation; the theta-role Theme of the subject is assigned compositionally. The arguments presented above imply that the relation between the copula and the predicative complement is not functional selection, which is defined in terms of syntactic categories (D selects NP, I selects VP, etc.); rather, it is semantic selection. Within semantic selection, therefore, two kinds of roles are distinguished: (i) selection of participant roles such as Agent, Patient, etc., and (ii) selection of non-participant roles such as Property, Measure, Name, etc. As a rule, selected arguments bearing a participant role are R-expressions, i.e. they are referentially independent and are assigned structural (or lexical) case by their respective verbs. In contrast, copular verbs select an argument bearing a non-participant role. These arguments are coindexed with their subjects, hence they do not exhibit referential independence. In this sense, they are [-R], This characterization, however, does not imply that they do not have any reference at all. More often than not, not only definite NPs/DPs, but even indefinite ones exhibit the same syntactic behavior as [+R] expressions with regard to accessibility, i.e. they can be referred to anaphorically:

5 With regard to coindexation, cf. Williams (1980:204); the full representation of (i), where the predicate rich does not c-command its subject John, is said to be the one in (ii), where the predicate does c-command the subject: (i) John, became rich¡. (ii) John¡ [became rich¡ ]¡. "Given such a predicate, we could then say that a predicate has to c-command its subject, except where a predicate is contained in and coindexed with another predicate" (ibid.). A closer look at this line of reasoning, however, reveals that the notion of predicate is used in two different ways: (a) as 'predicate-of in syntactic terms (i.e. in correlation to 'subject-of ), and (b) as 'predicate' in semantic terms. - Ultimately, it is the aim of this paper to disentangle these two notions and restrict the term 'predicate' to the use mentioned under (a).

150

Elisabeth Löbel

(16) Das

Veilchen

ist eine

the.NEUTR violet.NEUTR is

Blume.

a.FEM flower.FEM

'The violet is a flower.' Es/Sie blüht im Frühling. It.NEUTR/it.FEM flowers in spring 'It flowers in spring.' The example in (16) emphasizes the complement-like character of predicative noun phrases. On the other hand, they are also affined to adjective phrases, as it is amply illustrated in Leech/Li (1995). Being property-denoting expressions, they are capable of being coordinated with APs (17a); another correlate of their "quasi-adjectival" quality is gradability (17b,c): (17) a. b. c.

... whose house was still unoccupied and a ruin. He was too much of a dreamer to do any good. Last winter was more of a panic than anything else ... (Leech/Li 1995:186ff.)

I assume that copular verbs, which select a non-participant role, do not assign case. There are three alternatives by which their selected complements may be case-marked: (i) they agree in case with the subject DP (syntactic condition), (ii) case is triggered by semantic conditions, or (iii) by default (nominative). An obvious example for the first and the third type is shown by the alternation in (18). 6 (18) Ich lasse ihn I

let

him

einen guten

Arbeiter/ ein

a.ACC good.ACC worker/

guter

Arbeiter sein.

a.NOM good.NOM worker

be

Ί let him be a good worker.' Prepositions constitute further evidence for the hypothesis that the copula does not assign case. In Williams (1994:222), it is observed that a grammatical preposition does not project a PP that has a predicative use; predicative use of PPs is confined to semantic prepositions: (19) a. b.

The letter to John by himself, The letter was to John.

(20) a. b.

*The destruction of the city was by the enemy, The book was by Harry.

In (19a), to functions as a grammatical preposition, in (19b), as a semantic one: "There is really no contradiction here as long as we allow to to be ambiguously grammatical or semantic" (Williams 1994:222). Similar observations hold for (20). These examples may suffice to illustrate the ambivalent nature of predicative noun phrases between accessibility on the one hand, which emphasizes their nominal status, and gradability as well as coordination with predicative adjectives on the other hand, which emphasize their 'quasi-adjectival' status due to the fact that they are propertydenoting expressions.

6 I owe this example to Gisbert Fanselow (p.c.).

Case alternation in Finnish copular

constructions

151

To summarize: There is a small subclass of verbs which select a non-participant role (Measure, Name, etc.); copular verbs, which select the non-participant role of Property, are part of this subclass. All these verbs exhibit the following characteristics: (i) they do not assign case, i.e. case-marking is triggered by case concord (syntactic condition) or it is free (semantic condition or default); (ii) the selected complements are coindexed with their respective subjects; in this sense, they are 'non-referential', as they have no independent reference. Hence, they are [-R]-expressions in contrast to [+R]-expressions which are defined in terms of referential independence. Languages exhibiting case alternation are especially interesting for mechanisms of case assignment. At first sight, they represent a challenge to the hypothesis that predicative complements are not assigned case by copular verbs. From the typological point of view, Comrie (1997:45) suggests that, as far as case marking of predicatives in copular constructions is concerned, "general linguistic theory must allow at least two possibilities: case can be assigned by government, seen in differences between the case of the subject and that of the nominal predicate 7 [...]; case can be assigned by agreement, seen in the co-variation of the case of the nominal predicate with that of the subject". A well-known example of case alternation is Finnish, which has a very elaborate case system. The hypothesis that copular verbs do not assign case even in languages such as Finnish constitutes the topic of the next sections.

3. Predicatives and case alternation in Finnish 3.1

Case alternation and case concord

Finnish is an agglutinating language with a rich case system; there is no absolute agreement on the exact number of cases, but it is commonly agreed upon that nominative, accusative and partitive are to be considered as structural cases in contrast to the other cases, which are primarily semantic. More often than not, the latter ones correspond to prepositions in English. 8 In languages such as English and German, the distinction between structural cases in the sense of being assigned in a purely configurational way, such as nominative and accusative for subject and object, and inherent or lexical cases which are assigned by properties of lexical items is rather clear-cut. There are, however, languages in which cases are structurally assigned, but simultaneously obey semantic conditions, since they are "sensitive to the NP's definiteness, animacy, or quantificational properties, or to the aspectual character of the VP, or to some combination of these factors. The Finnish partitive is a particularly clear instance of this apparently hybrid category of semantically conditioned structural case" (Kiparsky 1998a:265). Kiparsky examines in detail the alternation of accusative 7 The term 'nominal' comprises both nouns resp. noun phrases and adjectives (Comrie 1997:39). 8 For an overview of the Finnish case system, cf. Sulkala & Karjalainen (1992) and Holmberg & Nikanne (1993:1-20). Case alternation in Finnish, especially with regard to nominative, accusative and partitive, is discussed by various contributors in Holmberg & Nikanne (eds.) 1993.

152

Elisabeth

Löbel

and partitive in Finnish. Whereas this kind of case alternation is constitutive of object NPs, it is the alternation of nominative and partitive which is relevant for case-marking of predicative adjectives and nouns or NPs. On the one hand, it is well-known that this alternation correlates with the potential divisibility of the referent of the NP, i.e. for NPs which are either plural or mass. Conceptual shifts similar to English ones in kakku 'cake' as an entity- or mass-denoting noun are reflected by different cases on the predicative adjective, with nominative for non-divisibility and partitive for divisibility (21): (21) a.

b.

Kakku on hyvää. (mass-denoting noun) Cake.SG.NOM be.3SG delicious.SG.PART 'Cake is delicious.' Kakku on hyvä. (entity-denoting noun) Cake.SG.NOM be.3SG delicious.SG.NOM 'The/this cake is delicious.'

As it is indicated by the translation, this difference in meaning can be reflected in English by means of the obligatory occurrence vs. non-occurrence of an article. 9 There are, however, other instances where the difference in meaning is not so obvious. In (22), the subject NP is divisible, in (23) it is non-divisible, but both instances allow for case alternation of the predicative complement with a slight difference in meaning. In (22), the nominative emphasizes the totality of the amount referred to by the subject NP; in (23), the partitive signals that the property expressed by the predicative NP is to be understood as a temporary characteristic of the subject referent: (22) a.

b.

(23) a.

b.

Paidat ovat uusia. Shirt.PL.NOM be.3PL new.PL.PART 'The shirts are new.' Paidat ovat uudet. Shirt.PL.NOM be.3PL new.PL.NOM 'All the shirts are new.' Hän on kova poika. He be.3SG hard.SG.NOM boy.SG.NOM 'He is a tough guy.' Hän on kovaa poikaa. He be.3SG hard.SG.PART boy.SG.PART 'He is somewhat/sort of a tough guy.' (Fromm 1982:175)

Example (23) illustrates furthermore that this type of case alternation is possible not only for predicative APs, but also for predicative NPs. In the following, I will show that the nominative, too, parallel to the partitive, can be subsumed under the category of semantically conditioned structural case. Whereas case alternation along the lines of (21) is conditioned by the lexicon (in these instances, the distinction of N[+divisible] vs. N[-divisible] is somewhat parallel to the distinction of N[+count] vs. N[-count] in English), this line of reasoning cannot be 9

Regarding the obligatory interpretation of the subject-NP in (21b), (22), and (23) as definite, cf. section 3.3 below.

Case alternation

in Finnish copular

constructions

153

applied to the examples mentioned under (22) and (23). Furthermore, there is another type of case alternation in copular sentences which, to my knowledge, has not been dealt with extensively in the literature, and which, at first sight, seems to be rather marginal, namely case alternation in subordinate, non-finite sentences. The following example quoted in Comrie (1997) illustrates the predicative use of adjectives: In finite sentences (24a), the subject as well as the predicative adjective shows nominative casemarking, whereas in the corresponding subordinate non-finite sentence the subject is invariably in the genitive, but case-marking of the predicative complement may vary between genitive (by case concord) and nominative: 10 (24) a.

Kirja on valkoinen. Book.SG.NOM be.3SG white.SG.NOM 'The book is white.' b. Tiedän kirjan olevan valkoisen/ valkoinen know.lSG book.SG.GEN being white.SG.GEN/ white.SG.NOM Ί know that the book is white.' (Comrie 1997:47) This kind of variation contrasts with the inessive, which indicates locality and, therefore, is a purely adverbial semantic case. Complements marked by the inessive do not exhibit the variation in (24); as a rule, they remain constant throughout: (25) a.

b.

Tiedän että sinä olet Helsingissä. know.lSG that you.NOM are Helsinki.INE Ί know that you are in Helsinki.' Tiedän sinun olevan Helsingissä. know.lSG you.GEN being Helsinki.INE Ί know that you are in Helsinki.' (Comrie 1997:45)

For the phenomenon illustrated in (24b), Comrie suggests that this "variation between government [= nominative, E.L.] and agreement [= genitive, E.L.] within individual languages [...] seems to be an independent 'parameter' on which languages can vary, independent in the sense that it does not correlate with other syntactic properties" (Comrie 1997:48). At first sight, this type of variation seems to undermine the hypothesis that the nominative, too, constitutes a semantically conditioned structural case, especially if we take into account that the phenomenon illustrated in (24b) is not confined to predicative adjectives, but also holds for predicative nouns: (26) a.

b.

Pallo on lelu. Ball.SG.NOM be.3SG toy.SG.NOM 'The ball is a toy.' Tiedän pailón olevan lelun / lelu. know.lSG ball.SG.GEN be.PRT.PRES toy.SG. GEN / toy.SG.NOM Ί know that the ball is a toy.'

10 Cf. Kiparsky (1998b) where an empty complementizer is assumed which assigns genitive case to the subject of non-finite sentences. The present participle olevan 'being' is an invariant form of the copula olla 'to be'.

154

Elisabeth

Löbel

This variation between case concord (genitive) and case-marking by default (nominative) is reminiscent of the case alternation found in the German example (18).There is, however, an important difference between Finnish and German: The following examples in the plural show that the case alternation exhibited in (24) and (26) is confined to the singular in Finnish. Neither predicative nouns (27) nor adjectives (28) in the plural allow for case concord: (27) a.

b.

Pallot

ovat

leluja.

ball.PL.NOM

be.3PL

toy.PL.PART

'Balls are toys.' Tiedän pallojen

olevan

leluja /*lelujen.

k n o w . l S G ball.PL.GEN be.PRT.PRES Ί k n o w that balls are toys.'

(28) a.

b.

toy.PL.PART/*toy.PL.GEN

Kirjat

ovat

valkoiset/valkoisia.

book.PL.NOM

be.3PL

white.PL.NOM /white.PL.PART

'The books are white.' Tiedän kiijojen

olevan

valkoiset /valkoisia/

know.lSG

be.PRT.PRES

white.PL.NOM/white.PL.PART/ *valkoisten. white.PL.GEN

book.PL.GEN

Ί know that the books are white.' In sum, these examples show that case concord with the genitive subject of non-finite subordinate sentences is confined to the singular. This means that, due to the interaction with number (singular vs. plural), it is only the nominative singular which could be looked upon as the default case with regard to the genitive singular. Parallel to the finite sentences, there is case alternation of partitive and nominative plural, the latter one emphasizing the totality of the amount referred to by the subject NP (cf. also (22)). For the following example (29a), it could possibly be maintained that there is agreement in the nominative between the subject and the predicative N P koira 'dog' in the nominative, as possessive constructions with the possessor in the genitive could be described as being elliptical: 11 (29) a.

Koira

on

pojan

(koira).

dog.SG.NOM

be.3SG

boy.GEN

(dog.SG.NOM)

'The dog is the boy's (dog).'

11

Cf. Vainikka (1993:129ff.) for evidence that the genitive case in Finnish is the structural default case for the specifier position. Her arguments also provide further evidence for our assumption that the copula is to be treated as a main verb: "In order to maintain a parallelism between subjects of sentences with predicative APs and all other subjects [...], I would have to treat olla 'be' in (12a)[= (i), E.L.] as a main verb, and Pekka would be base-generated in the Spec(VP) position of this V": (i) Pekka on Riitan pitu-inen. Pekka.NOM is Riitta.GEN Height-ADJ 'Pekka is (as) tall (as) Riitta.' (Vainikka 1993:135)

Case alternation in Finnish copular constructions b. Koira dog.SG.NOM

on

pojan

isällä..

be.3SG

boy.GEN

father.ADE

155

'The dog is with the boy's father.' (Toivainen 1993:112) This argument, however, cannot be extended to cases such as (29b) and (30) where the predicative complements are marked by cases which are unambiguously semantic in nature, with the inessive being a local case and the essive expressing a temporary property: (30) a.

Olin

sotilaana

Berliinissä.

be.IMPF.lSG

soldier.ESS

Berlin.INE

Ί was a soldier in Berlin.' b. Olin sotilas kolme vuotta. be.IMPF.lSG soldier.NOM three year.PART Ί was a soldier for three years.' (Sulkala/Karjalainen 1992:233) The diversity of case-marking for predicative complements can be taken as evidence that one cannot assume that the copula olla 'to be' assigns case by government; instead, the examples in (29b) and (29a) illustrate that predicative complements can be case-marked by adverbial, i.e. purely semantic cases. In what follows, however, I will concentrate on case-marking of predicative complements by nominative and partitive and the conditions which are involved when using these semantically conditioned structural cases. Before that, a short characterization of the accusative/partitive alternation, which is constitutive of complements of transitive verbs, is given in the following excursus. Excursus: Accusative/partitive alternation In combination with transitive verbs, the partitive has two functions: an aspectual VPrelated function with regard to irresultativity or unboundedness, and an NP-related function, as partitive is assigned to "quantitatively indeterminate NPs (including indefinite bare plurals and mass nouns)" (Kiparsky 1998a:266f.). In contrast to the VP-related function of the partitive, where it can be used in both the singular and plural, the NP-related function of the partitive is restricted to the singular (cf. (31)). For intrinsically bounded verbs such as saada 'get', the objects are "partitive when they are quantitatively indeterminate (in particular, when they are indefinite plurals or mass nouns), otherwise accusative; [...] Here, the partitive in effect marks the indefiniteness of bare plural or mass noun objects [...]" (Kiparsky 1998a: 267):12

12

In fact, karhua 'bear' in (32a) with partitive singular "is OK with 'bear' understood as a mass noun [...]" (ibid.).

156

Elisabeth

(31) a.

saa-n

*karhu-a/karhu-j-a

get. 1SG

*bear-PART/bear-PL-PART

Löbel

' I ' l l get the ( a ) bear / bears.' b.

saa-n

karhu-n / karhu-t

get. 1 SG

bear-ACC /bear-PL.ACC

'I'll get the (a) bear /the bears.' (Kiparsky 1998a:268) I assume that, for copular sentences, it is the NP-related function which is relevant for case-marking of predicative complements with either nominative or partitive. Therefore, the following questions arise: (1) Why is the case-marking o f singular count nouns restricted to the accusative? (2) How can the case alternation exhibited in (24) and (26) between nominative and genitive, which is also confined to singular count nouns, be explained? In the following sections, it is argued that these two seemingly unrelated aspects (only accusative and no partitive on the one hand, and case alternation (nominative vs. genitive) on the other) can be traced back to one and the same phenomenon.

3.2

Main characteristics of Finnish copular sentences

In the following, I will confine myself to the verb olla 'to be', as this is the verb most commonly used in copular sentences.13 As we are interested primarily in the question of how the case o f predicative NPs or APs is determined, I will concentrate on the core cases nominative and partitive. Due to its rich morphology, word order in Finnish is relatively free. For copular sentences, however, more often than not, word order is rather rigid and follows the general pattern subject - copula - predicative complement, as otherwise one and the same sentence may have different readings. This is illustrated in (32); a difference in word order correlates with a difference in meaning: (32) a.

b.

Pallo

on

lelu.

Ball.SG.NOM

be.3SG

toy.SG.NOM

'The ball is a toy.' Lelu on

pallo.

Toy.SG.NOM

ball.SG.NOM

be.3SG

'The/this toy is a ball.' In both sentences, the subject is in the nominative. Sentence (32a) is generic; the subject, too, has a generic reference, whereas (32b) presupposes that the subject lelu 'the toy' has been mentioned before, i.e. the referent of the subject is 'locatable'. The common denominator o f these two kinds of possible interpretation with regard to the subject marked by nominative case is inclusiveness (of reference). I will return to this aspect below (section 3.3).

13

The other copular verb, namely tulla 'to become', seems to be "very rare in a proper predicative clause" (Sulkala/Karjalainen 1992:217).

Case alternation in Finnish copular

157

constructions

For subjects whose referents are neither generic nor locatable, i.e. which are introduced as new referents, the copula olla 'to be' is used as an impersonal verb confined to the 3rd person singular (33): (33) a.

Kannussa on Jug.iNE

b.

maitoa.

be.3SG milk.SG.PART

'There is (some) milk in the jug.' Kadulla on autoja. Street.ADE be.3SG car.PL.PART

c.

'There are (some) cars in the street.' (Chesterman 1991:117) Kadulla on auto. Street.ADE be.3SG car.SG.NOM

'There is a car in the street.' These examples illustrate the relevance of the concept of divisibility for case marking: "All Finnish NPs are either divisible or non-divisible. A non-divisible noun is one whose referent is an individual unit which can be multiplied but not divided [i.e. singular count nouns like auto 'car' in (33c), E.L.].[...] A divisible NP has a conceptually divisible referent - a subpart of the referent may still be designated by the same noun [mass nouns like maito 'milk' in (33a) or plural count nouns like autoja 'cars' in (33b), E.L.]" (Chesterman 1991:93). The relevance of this distinction for case-marking is obvious when comparing the postcopular NPs in (33): It is well-known that the partitive constitutes the default case for divisible NPs (33a,b) and the nominative that for non-divisible NPs (33c). 14 This fundamental distinction is found even in interrogative pronouns: The use of an interrogative pronoun in the nominative implies that a discrete entity is being asked for (34), whereas the use of the partitive is indicative of the fact that the speaker is expecting an answer in terms of a non-discrete entity, i.e. a substance (cf. Schot-Saikku 1990:32ff.): (34) a.

b.

Mikä/Mitä

tuo

what.SG.NOM/ what.SG.PART

that.SG.NOM be.3SG

'What is that (there)?' Se on juusto/ it.SG.NOM be.3SG

on?

juustoa.

cheese.SG.NOM/cheese.SG.PART

'It's a cheese / it's cheese.' The following example (35), however, illustrates that a characterization in terms of default does not suffice in order to describe copular sentences and their postcopular complements. If the copula is used as a finite verb, it agrees in number with the subject

14

One could argue that it is not contingent to conclude that auto 'car' is truly nominative, as it could also be one of the variants of the accusative. When the relevant NP is questioned, however, the Wh-word is unambiguously nominative: (i) Kuka/*kenet on kadulla? Who.NOM/who.ACC be.3SG Street.ADE 'Who is in the street?' Thanks to the anonymous reviewer for pointing out this fact to me.

158

Elisabeth Löbel

which is invariably in the nominative (35) (emphasis in the translations by the author, E.L.): (35) a.

Maito

on

kannussa.

milk.SG.NOM

be.3SG

jug.lNE

'The milk is in the jug.' b. Autot ovat kadulla. car.PL.NOM

c.

be.3PL

street.ADE

'The cars are in the street.' (Chesterman 1991:117) Auto on kadulla. car.SG.NOM

be.3SG

street.ADE

''The car is in the street.' At first sight, the opposition illustrated in (33) between divisible NPs marked by partitive and non-divisible NPs marked by nominative seems to be neutralized in (35), as the subject obligatorily occurs in the nominative, irrespective of whether the NP is divisible (35a,b) or not (35c). A closer look at the core meaning of nominative and partitive, however, reveals that, in addition to the concept of divisibility, another concept must be taken into account in order to characterize the NP-related function of these cases, namely totality. This is the topic of the next section.

3.3

Feature specification of nominative and partitive

From the point of view of the lexicon, Finnish nouns can be characterized in terms of the feature [± divisible]: (36)

Singular N[+divisible]: maito 'milk' N[-divisible]: palio 'ball'

Plural pallot 'balls' sakset 'scissors' (plurale tantum) 15

In syntax, however, if these nouns are assigned case, it is necessary to assume a second feature in addition to divisibility in order to describe nominative and partitive as semantically conditioned structural cases. Let me start with the core meaning of the partitive, which, in its NP-related function, is characterized as being assigned to "quantitatively indeterminate NPs (including indefinite plurals and mass nouns)" (Kiparsky 1998a:266). In a similar way, according to Chesterman (1991:92), the partitive "marks the object for partial quantity, i.e. a quantity which also allows the possible existence of a (contextually) relevant additional quantity". In contrast, the nominative marks an NP as an "entity in its entirety or a total, determinate quantity" (Karlsson 1984:97, emphasis E.L.).16 15 Body-part terms such as kädet 'hands' constitute a special case: As relational nouns, they are conceptualized as non-divisible, parallel to sakset 'scissors'; in their non-relational use, they are like ordinary common nouns, parallel to pallot 'balls'. Cf. section 3.4.1 below. 16 "Der Partitiv drückt eine unbestimmte, nicht näher begrenzte Menge der teilbaren Wörter aus" (Karlsson 1984:77); "Der Nominativ drückt eine konkrete oder abstrakte Ganzheit oder eine bestimmte begrenzte, totale Menge aus" (Karlsson 1984:97).

Case alternation in Finnish copular

constructions

159

A closer look at these definitions reveals that they contain two different notions: divisibility and totality. The latter one is a concept which, in addition to locatability, is used for the description of determination in languages with articles, in particular for the characterization of lexically realized determiners.17 In the referential subject NP of non-finite copular sentences (see (33)), nominative correlates with [-divisible] and partitive with [+divisible], I will assume that predicative NPs, required as complements of the copula, have in addition the feature [total]. Nominative, then, realizes the feature value [+total], and partitive the value [-total]. However, the latter is still restricted to [+divisible], so that the feature combination *[-total,-divisible], which would be odd for semantic reasons, cannot be instantiated. On the basis of these assumptions, I propose that predicative NPs with nominative or partitive are given the following feature specifications: (37) a. b.

Predicative noun phrases in the nominative: NP [+total, ± divisible] Predicative noun phrases in the partitive: NP [-total,+ divisible]

The definite or indefinite interpretation of an NP such as auto 'a/the car' in (33), (35) is dependent on its position in the sentence: clause-initial (35c) as definite, and clausefinal (33c) as indefinite. This condition holds for all Finnish sentences in which the matrix verb is finite, and in which the subject is obligatorily nominative: (38) a.

Poika

rikkoi

ikkunan.

boy.SG.NOM

broke

window.SG.ACC

'The/*a boy broke a window.' b. Joku poika rikkoi some.SG.NOM

boy.SG.NOM

broke

ikkunan. window.SG.ACC

Ά boy broke a window.' (Chesterman 1991:106) In (38b), the indefinite reference must be realized lexically, i. e. the "indefinite function word [= joku 'some', E.L.] thus becomes necessary to preserve the indefinite reading" (Chesterman 1991:106).18 According to Chesterman (1991:137), "the subject is 'total' in the sense that the whole of the subject performs the action in question".

17 Cf. Hawkins (1978) and his notion of inclusiveness; his theory is discussed in Chesterman (1991), where the Finnish cases nominative, partitive, and accusative are defined in terms of the features [locatable], [all], and [one] and are compared with definiteness in English. This approach, however, does not take into account the use of the partitive with singular count nouns (cf. section 3.4.2 below). 18 That nominative subjects do not always have a definite interpretation can be illustrated by the following example: (i) Varkaat varastivat tavarani thieves.NOM stole things.ACC.my 'Some thieves stole my things.'

160

3.4

Elisabeth Löbel

Case-marking of predicative complements

It has been illustrated in the preceding section that subjects of finite sentences (i) obligatorily occur in clause-initial position and (ii) are obligatorily nominative (cf. examples (35) and (38)); hence, the definite interpretation of these subjects corresponds to the core meaning of this case. Accordingly, subjects of finite sentences can be specified by the feature combination [+total,¿divisible]. This feature combination also holds for subjects in clause-initial position (39a) of non-finite sentences with participles, which are case-marked by the genitive (emphasis by the author, E.L.): (39) a. Kalle uskoi karhun olevan pesässä. Kalle.NOM believed bear.GEN being den.lNE 'Kalle believed the bear was in the den.' b. Kalle uskoi pesässä olevan karhu. Kalle.NOM believed den.INE being bear.NOM 'Kalle believed there was a bear in the den.' (Chesterman 1991:141) It is the "association between genitive subject and clause-initial position that seems to account for the tendency to interpret such subjects as definite" (Chesterman 1991: 141).19 Based on these observations, I assume that both the subjects of finite and nonfinite sentences in clause-initial position (with the participle olevan 'being') are specified as [¿total]. This also holds for example (24), repeated here for convenience under (40) (emphasis E.L.): (40) a. Kirja on valkoinen. book.SG.NOM

be.3SG

'The book is white.' b. Tiedän kirjan know.lSG

white.SG.NOM

olevan

valkoisen /valkoinen

book.SG.GEN be.PRT.PRES

white.SG.GEN/white.SG.NOM

Ί know that the book is white.' (Comrie 1997:47) These observations can be summarized as follows: In analogy to the basic structure for copular sentences, given in (15), the following feature specifications can be assumed for copular sentences with olla 'to be' with regard to nominative and partitive as semantically conditioned structural cases: (41)

VP NPj [¿total] [¿divisible]

V' V [¿finite]

APi /NPj [¿total] [¿divisible]

olla 'be' 19 This statement is based on observations made by Hakulinen, Karlsson & Vilkuna (1980); using a corpus of over 10,000 clauses, they found that "genitive subjects occur almost without exception in clause-initial position (97 percent of all occurrences)" (Chesterman ibid.).

Case alternation in Finnish copular constructions

161

3.4.1 Case-marking of predicative adjective phrases In contrast to nouns (cf. section 3.4.2), adjectives themselves are neither specified for totality nor for divisibility. Subjects of finite copular sentences are always case-marked by nominative, and, therefore, are specified as [+total] throughout. It is therefore maintained that, apart from number agreement (syntactic condition), case-marking of postcopular APs is dependent on the feature [divisible] of the corresponding subject (semantic condition). In the following, a survey of the possible feature combinations and their semantic interpretations is given. Plural subjects, which are [+divisible], allow for case alternation. The predicative adjective is case-marked either by nominative ([+total]), cf. (22b), repeated here as (42a), or by partitive ([-total]), cf. (22a), repeated here as (43a). Case-marking remains constant in the corresponding embedded non-finite sentences (cf. (42b) and (43b)): (42) Subject NP: [+divisible,+plural] Predicative AP: [+total,+plural] a. Paidat ovat uudet. shirt.PL.NOM

be.3PL

new.PL.NOM

'(All) the shirts are new.' b. Tiedän paitojen olevan know.lSG

shirt.PL.GEN

uudet. be.PRT.PRES new.PL.NOM

Ί know that (all) the shirts are new.' (43) Subject NP: Predicative AP: a. Paidat

[+divisible,+plural] [-total,+plural] ovat uusia.

shirt.PL.NOM

be.3PL

new.PL.PART

'The shirts are new.' b. Tiedän paitojen know.lSG

shirt.PL.GEN

olevan

uusia.

be.PRT.PRES new.PL.PART

Ί know that the shirts are new.' There is a subtle difference in meaning between the use of the partitive and the nominative, which cannot be reflected unambiguously in English. The nominative, due to the feature [+total], emphasizes the totality of the amount referred to. In contrast, the partitive, specified as [-total], "allows the possible existence of a (contextually) relevant additional quantity" (Chesterman 1991:92). This is quite obvious in (44a), whereas in (44b,c), the partitive can be interpreted as implying a contrastive reading: (44) a. Jussin paidat ovat uusia. John.GEN

shirt.PL.NOM

be.3PL

new.PL.PART

'(Some of) John's shirts are new.' b. Hänen paitansa ovat he.GEN

shirt.PL.NOM.POSS

be.3PL

uusinta

muotia.

newest.SG.PART

fashion.SG.PART

uusia,

mutta eivät

new.PL.PART

but

'His shirts are new, but not very fashionable.'

be.not.3PL

162

Elisabeth Löbel c. Nokian Nokia.GEN

saappaat

ovat

pitkiä.

boot.PL.NOM

be.3PL

high.PL.PART

'Nokia's boots are high (in contrast to other boots).' Plurale tantum nouns such as sakset 'scissors', verhot 'curtains', although morphologically plural, are conceptualized as non-divisible. As a rule, the adjective is case-marked by nominative plural. A comparison of (45a) with (45b), however, reveals that the use of the partitive is by no means ungrammatical. Sakset 'scissors' in (45b) is understood as referring to several pairs of scissors, i.e. the opposition of nominative vs. partitive on the predicative adjective is used to signal the opposition between [-divisible,+plural] (one pair of scissors) and [+divisible,+plural] (several pairs of scissors): (45) a.

Subject NP: Predicative AP: Sakset

[-divisible,+plural] [+total,+plural] ovat uudet.

Scissors.PL.NOM

'The scissors are b. Subject NP: Predicative AP: Sakset

be.3PL

new.PL.NOM

new.' [+divisible,+plural] [-total,+plural] ovat uusia.

Scissors.PL.NOM

be.3PL

new.PL.PART

'The (several pairs of) scissors are new.' Body-part terms such as kädet 'hands' are also conceptualized as being non-divisible. When they are used as relational nouns, they obligatorily occur with a possessive suffix and the predicative adjective is case-marked by nominative (46a,b). If they are used in a non-relational way, they do not occur with a possessive suffix and are casemarked like ordinary common nouns in the plural, i.e. the adjective occurs in the partitive (47): (46) a.

Hänen

silmänsä

ovat

siniset.

he.GEN

eye.PL.NOM.POSS

be.3PL

blue.PL.NOM

'His eyes are blue.' b. Tiedän hänen silmiensä olevan siniset. know. 1SG he.GEN eye.PL.GEN.POSS be.PRT.PRES blue.PL.NOM Ί know that his eyes are blue.' (47) Kaikki kädet ovat mustia, all.PL.NOM hand.PL.NOM be.3PL black.PL.PART ja päät verisiä. a n d head.PL.NOM bloody.PL.PART

'All hands are black, and the heads bloody.' For plural subjects, therefore, the opposition found in the case-marking of predicative adjectives can be summarized as follows: (i) ordinary common nouns may trigger either nominative or partitive case-marking with a slight difference in meaning, (ii) for plurale tantum nouns, the opposition is used to differentiate between one vs. more than one entity, (iii) for body-part terms, case opposition corresponds to their relational vs. non-relational use.

163

Case alternation in Finnish copular constructions

For singular subjects, case opposition of nominative vs. partitive is used to differentiate between a mass or a count noun reading of the respective subject. This is illustrated in (21), repeated here for convenience under (48a) and (49a). A comparison with their use in the corresponding subordinate, non-finite construction shows that it is only the 'count noun' reading which allows for case concord with the subject (genitive, (48b) vs. (49b)): (48) Subject NP: Predicative AP: a. Kakku

[+divisible,-plural] (mass-denoting noun) [-total,-plural] on hyvää.

cake.SG.NOM

be.3SG

delicious.SG.PART

'Cake is delicious.' b. Tiedän kakun know.lSG

cake.SG.GEN

olevan

hyvää.

be.PRT.PRES

delicious.SG.PART

Ί know that cake is delicious.' (49) Subject NP: Predicative AP: a. Kakku

[-divisible,-plural] (entity-denoting noun) [+total,-plural] on hyvä.

cake.SG.NOM

be.3SG

delicious.SG.NOM

'The/this cake is delicious.' b. Tiedän kakun olevan know.lSG

cake.SG.GEN

hyvä / hyvän.

be.PRT.PRES

delicious.SG.NOM/SG.GEN

Ί know that the/this cake is delicious.' Example (50) illustrates that conceptual shifts such as kakku 'cake' and 'a cake' are more common in Finnish than in English:20 (50) a.

Omena

on

makea

Apple.SG.NOM

be.3SG

sweet.SG.NOM

'The apple is sweet.' b. Omena on Apple.SG.NOM

be.3SG

makeaa. sweet.SG.PART

'An apple/apples is/are sweet.' (Fromm 1982:176) [lit.: 'Apple is sweet.', E.L.] Body-part terms in the singular behave analogously to singular count nouns, i.e. in the corresponding non-finite sentence, the predicative adjective may be case-marked either by genitive or nominative: (51) a. Hänen päänsä on iso. he.GEN

head.SG.NOM.POSS

be.3SG

big.SG.NOM

'His head is big.'

20 This kind of mass noun reading is not possible in English. One example corresponding to (51) would be (i) 'Banana is good for you' (in contrast to *'Apple is good for you'). According to Fromm (ibid.), the term 'undivided' (instead of 'non-divisible') is more suitecfto characterize the use of the partitive in these instances.

164

Elisabeth Löbel b.

Tiedän

hänen

päänsä

K n o w . l S G he.GEN head.SG.GEN.POSS Ί k n o w that his head is big.'

olevan

iso/ison.

be.PRT.PRES

big.SG.NOM/GEN

The issue of case concord between subjects and predicative adjectives in finite and corresponding non-finite sentences can be summarized as follows: Adjectives themselves do not contain any inherent specification for totality or divisibility. They agree with their subjects in number. Furthermore, subjects which are [+divisible] allow for case alternation of the corresponding postcopular AP, with the nominative for total, and the partitive for non-total, i.e. partial predication. Postcopular APs in the partitive (singular or plural) and nominative plural do not exhibit case variation between nominative/partitive and genitive in non-finite sentences, these cases remain constant throughout; in this respect, they behave like semantic cases. This aspect indicates that case is not assigned in a purely structural way, but is triggered by semantic conditions. There is one notable exception which seems to undermine these findings, namely the alternation between case concord with the subject in non-finite sentences (invariably genitive) and the default nominative. This kind of variation, which is confined to the singular, could be explained as follows: The subject NP in finite copular sentences is invariably in the nominative, and invariably in the genitive in the corresponding non-finite sentences. Case alternation of the predicative adjective in non-finite sentences in the singular is possible due to the fact that in the feature combination [+total, -divisible], the feature [-divisible] entails the feature [+total], that is, a single non-divisible entity cannot be non-totally predicated of. Therefore, the opposition [+total] vs. [-total], which is relevant in the plural, is neutralized in the singular. This means that there is free variation between genitive and nominative, where the formally unmarked nominative is preferred if the respective adjective is morphologically complex. 21 On the other hand, it is exactly this type of variation between case-marking triggered by concord (genitive for totality) and by semantic conditions (nominative for nondivisibility) which shows that, as a rule, both types of features must be assumed not only for the plural, but also for the singular, despite the fact that one is entailed by the other in the singular. 3.4.2 Predicative noun phrases Generally speaking, adjectives never contain an inherent specification for divisibility, i.e. the corresponding feature specification is dependent on the subject NP with which they agree in number. Nouns, however, are inherently specified for divisibility, and predicative NPs in addition for totality. This means that case-marking of predicative NPs is determined by the features of the predicative NPs themselves. Lexically marked instances (body-part terms and collectives) will be considered first. Similar to adjectives, body-part terms, if used as relational nouns, do not allow 21

Simultaneously, this observation explains the findings in Comrie (1997:48), where it is shown that an adjective like heikko 'weak' is most likely to allow the genitive, whereas for valkoinen 'white', the nominative is preferred. The latter is morphologically more complex than the former.

Case alternation

in Finnish copular

constructions

165

for case alternation, the predicative NP is invariably in the nominative. Comparable to predicative adjectives, there is agreement in terms of the feature [-divisible]: (52) Subject NP: [-divisible,+plural] Predicative NP: [+total,-divisible,+plural] a. Hänen silmänsä ovat vinosilmät. he.GEN eye.PL.NOM.POSS be.3PL slit.eye.PL.NOM 'His eyes are slit eyes.' b. Tiedän hänen silmiensä olevan vinosilmät. know.lSG he.GEN eye.PL.GEN.POSS be.PRT.PRES slit.eye.PL.NOM Ί know that his eyes are slit eyes.' Variation in terms of totality is most obvious in the following contrast for collective nouns such as sisarukset '(one pair of) brother and sister': 22 (53) Subject NP: [+divisible,+plural] a. Predicative NP: [+total,+divisible,+plural] He ovat sisarukset. They.NOM be.PL brother.and.sister.PL.NOM 'They are brother and sister (i.e. those two).' b. Predicative NP: [-total,+divisible,+plural] He ovat sisaruksia. They.NOM be.3PL brother.and.sister.PL.PART 'They are brother and sister (relationship).' (Schot-Saikku 1990:39) Apart from these two lexically marked instances (body-part terms and collectives), the following observation holds for predicative NPs with count nouns in the plural: The partitive, i.e. agreement in terms of the feature [+divisible], constitutes the default case. (54) Subject NP: [+divisible,+plural] Predicative NP: [-total,+divisible,+plural] a. (Esan) paidat ovat työvaatteita / *työvaatteet. Esa.GEN shirt.PL.NOM be.3PL work.cloth.PL.PART/ *w.cloth.PL.NOM '(Esa's) / The shirts are work clothes.' b. Nämä miehet ovat murhaajiaI *murhaajat. these.NOM man.PL.NOM be.3PL murderer.PL.PART/*PL.NOM 'These men are murderers.' Nominative case-marking of the predicative NP is possible only in those instances in which definiteness (and hence, the feature [+total]) is lexically realized on the predicative N P itself, either by a demonstrative (55a), a prenominai genitive (55b), or a restrictive adjective (55c). The predicative NP is interpreted as having an identificational reading:

22 The German equivalent, the collective noun Geschwister is more comparable to the Finnish equivalent than the English translation.

166

Elisabeth

Löbel

(55) Subject NP: [+divisible,+plural] Predicative NP: [+total,+divisible,+plural] a. Nämä miehet ovat ne murhaajat. DEM.PL.NOM man.PL.NOM be.3PL the.PL.NOM murderer.PL.NOM 'These men are the murderers.' b. Nämä paidat ovat Esan työvaatteet. DEM.PL.NOM shirt.PL.NOM be.3PL Esa.SG.GEN work.cloth.PL.NOM 'These shirts are Esa 's work clothes.' c. Varpuset ovat pihan ainoat linnut. sparrow.PL.NOM be.3PL yard.SG.GEN only.SG.NOM bird.SG.NOM 'The sparrows are the only birds in the yard.' (Schot-Saikku 1 9 9 0 : 3 4 ) However, example (56) illustrates that the prenominai genitive does not necessarily imply that the predicative NP is to be interpreted as [+total]. For all these examples, the case of the predicative NPs remains constant throughout in the corresponding nonfinite constructions. (56) Subject NP: [+divisible,+plural] Predicative NP: [-total,+divisible,+plural] Nämä paidat ovat Esan työvaatteita. DEM.PL.NOM shirt.PL.NOM be.3PL Esa.SG.GEN work.cloth.PL.PART 'These shirts are (some of) Esa's work clothes.' In the singular, the independence of the predicative NP with regard to case-marking is most obvious in the following example, where the use of the partitive vs. nominative correlates with the obligatory non-occurrence vs. occurrence of the definite article in English: (57) Subject NP: [-divisible-plural] Predicative NP: [-total,+divisible,-plural] a. Se on kahvivettä. this.SG.NOM be.3SG coffee.water.SG.PART 'This is coffee water.' Subject NP: [-divisible,-plural] Predicative NP: [+total,+divisible,-plural] b. Se on kahvivesi. this.SG.NOM be.3SG coffee.water.SG.NOM 'This is the/our coffee water.' (Schot-Saikku 1990:36) This type of opposition shows that the nominative cannot be regarded as being triggered by case concord; rather, it is semantically conditioned. Moreover, mass nouns used in the partitive generally do not agree with the subject in terms of number: (58) Subject NP: [+divisible,+plural] Predicative NP: [—total,-(-divisible, -plural] a. Nuo turkit ovat inkkiä. DEM.PL.NOM fur.coat.PL.NOM be.3PL mink.SG.PART 'Those fur coats are (made of) mink.' (Schot-Saikku 1990:35)

Case alternation b.

in Finnish copular

Nämä

constructions

paidat

DEM.PL.NOM shirt.PL.NOM

ovat

villaa.

be.3PL

wool.SG.PART

167

'These shirts are (made of) wool.' The semantic nature of the 'partitive' in the literal sense of the word is most obvious in examples such as (59). Recall that the feature combination [-total,-divisible] is not possible for semantic reasons; thus the partitive with the feature [-total] always implies [+divisible]. Since, however, proper nouns cannot possibly be considered as expressing a divisible concept, the partitive induces a conceptual shift: the predicative NP in the partitive must be interpreted as 'part of whole': (59) Subject NP: [-divisible-plural] Predicative NP: [-total,+divisible,-plural] Oulu on Pohjanmaata. Oulu be.3SG Ostrobothnia.SG.PART 'Oulu is a part of Ostrobothnia.' (Sulkala/Karjalainen 1992:62) There are examples for which one could argue that "partitive case can coerce a count noun into a mass noun" (Kiparsky 1998a:291 (emphasis in the translation, E.L.): (60) Sita käsikiijoitusta oli sängy-n alla-kin that.PART manuscript.PART be.PAST.3sg bed-GEN under-even '{Parts o f ) that manuscript were even under the bed.' (Kiparsky 1998a:291) However, examples such as the following ones reveal that it is not coercion into mass nouns, but rather the forbidden feature combination * [-total,-divisible] which is relevant for these instances (emphasis in the translation, E.L.): (61) Venettä näkyi niemen takaa. boat.SG.PART was.seen peninsula.SG.GEN behind 'A/The boat was seen behind the peninsula.' (literally: part of the boat, not the whole outline of one)(Chesterman 1991:164) (62) Peiton alta näkyi vauvan jalkaa. 2 3 blanket.GEN under show.PAST.3SG baby.GEN foot.PART 'Part of the baby's foot showed (from) under the blanket.' (Vainikka 1993:131)

A similar reinterpretation is found in (23b), repeated here as (63). (63) Subject NP: [-divisible-plural] Predicative NP: [-total,+divisible,-plural] Hän on kovaa poikaa. He

be.3SG

hard.SG.PART

boy.SG.PART

'He is somewhat/sort of a tough guy.' (Fromm 1982:175)

23

In this example, "the partitive can be analyzed as semantically based, 'part of something'" (Vainikka 1993:131); according to her, it could also be argued that the partitive involves an abstract quantifier such as osa 'part'. In the approach proposed here, however, this usage of the partitive corresponds to the feature combination [-total,±divisible].

168

Elisabeth Löbel

Parallel to the arguments with regard to the nominative/partitive opposition presented in connection with example (61), the contrast in (64a,b) constitutes another instance of the observation that, for predicative NPs, it is not contingent to argue that the nominative of the predicative complement 'agrees' with the nominative of the subject NP: (64) Subject NP: [-divisible-plural] a. Predicative NP: [+totaI,-divisible,-plural] Se on Liisa. This.SG.NOM be.3SG Lisa.SG.NOM 'This is Lisa.' b. Predicative NP: [-total,+divisible,-plural] Se on Liisa-a. This.SG.NOM be.3SG Lisa.SG.PART 'This is typical of Lisa.' (Schot-Saikku 1990:40) The nominative of the predicative NP in (64a) is independently motivated, and it is only coincidental that, in this example, both constituents (subject and predicative complement) exhibit nominative case marking. There are, however, other instances of predicative NPs marked by nominative for which, in contrast to (63), case alternation in the matrix sentence is not possible. Instead, case alternation is allowed in the corresponding non-finite subordinate sentences: either (i) case concord with the subject in the genitive, or (ii) nominative. Consider (26), repeated here as (65): (65) a.

b.

Pallo on lelu. ball.SG.NOM be.3SG toy.SG.NOM 'The ball is a toy.' Tiedän pailón olevan

lelu/

lelun.

know.lSG

toy.SG.NOM/

toy.SG.GEN

ball.SG.GEN be.PRT.PRES

Ί know that the ball is a toy.' It has been demonstrated in section 3.4.1 that this type of case alternation also holds for predicative adjectives in the singular (examples (49) and (51)). Adjectives themselves are not specified inherently for divisibility, but they agree in number with their subjects which are specified as [-divisible] in the singular. For adjectives, therefore, variation between genitive (case concord) and nominative is possible because the feature [-divisible] entails the feature [+total]. For this reason, also predicative NPs allow case concord with their corresponding subject. As for adjectives, the morphologically unmarked nominative is preferred if the noun itself is morphologically complex, or if it is preceded by an adjective, i.e. if the NP is syntactically complex: (66) a.

b.

Tiedän

pailón

know.lSG

ball.SG.GEN be.PRT.PRES

olevan

arvokas

lelu.

valuable.SG.NOM

toy.SG.NOM

Ί know that the ball is a valuable toy.' Tiedän pailón olevan arvokkaan

lelun.

know.lSG

toy.SG.GEN

ball.SG.GEN

be.PRT.PRES

valuable.SG.GEN

Parallel to the use of predicative adjectives in the nominative singular, examples such as (66) show that, despite the implicative nature of the feature combination [+total,

Case alternation in Finnish copular constructions

169

-divisible], both features have to be assumed not only for the plural, but also for the singular in order to motivate the case alternation illustrated in (65).

4. Conclusion In this paper, I have argued that, against common assumptions, copular verbs are twovalued verbs which, in contrast to transitive verbs, select two different kinds of arguments: An internal (non-referential) argument bearing the non-participant role of Property, and an external (referential) argument which is assigned the participant role of Theme. This approach implies that the constituent selected by a copular verb is not a predicate (in semantic terms), but a property-denoting complement. Whereas participant roles which are expressed by noun phrases constitute R-expressions since they exhibit referential independence, non-participant roles realized as noun phrases are 'nonreferential' since they have no independent reference, i.e. they are [-R]-expressions. Furthermore, this approach implies that two kinds of semantic selection must be distinguished: (i) selection of a participant role, and (ii) selection of a non-participant role. This fundamental distinction correlates with case-marking: Verbs which select a nonparticipant role do not assign case to their complement. Rather, case-marking is triggered either by case concord, or it is free, i.e. triggered by semantic conditions. At first sight, Finnish with its very elaborate case system constitutes a challenge to these assumptions. Apart from case-marking of predicative complements by purely adverbial cases which are undoubtedly semantically conditioned, there are two structural cases which are especially relevant for sentences with the copula olla 'to be': Nominative and partitive which, in Kiparsky's terms, belong to the "hybrid category of semantically conditioned structural case". A closer look at the core meaning of these cases reveals that, apart from divisibility (plural count and mass nouns), totality of predication is also decisive for case-marking. Whereas only nouns are inherently specified for the feature [divisible], the feature [total] must be specified for both predicative APs and NPs, with [+total] marked by nominative, and [-total] by partitive. This feature, therefore, is constitutive of the relation of predication in Finnish. These observations show that, in contrast to the claim made by Comrie (1997), it is not government by the copular verb which decides on the case-marking of predicative complements, but rather the interaction of syntactic conditions (case concord) with semantic conditions, as expressed by the features [± divisible] and [± total].

170

Elisabeth Löbel

Appendix: List of abbreviations ACC ADE ADJ ESS FEM GEN IMPF INE

accusative adessive adjectivizer essive feminine genitive imperfect inessive

NEUTR NOM PART PAST PL POSS PRES.PRT SG

neuter nominative partitive past tense plural possessive marker present participle singular

References Chesterman, Andrew (1991). On definiteness: A study with special reference to English and Finnish. Cambridge: CUP. Comrie, Bernard (1997). The Typology of Predicate Case Marking. In Essays on Language Function and Language Type. Dedicated to T. Givón, J. Bybee, J. Haiman and S. A. Thompson (eds.), 39-50. Amsterdam/Philadelphia: Benjamins. Déchaîne, Rose-Marie (1993). Predicates Across Categories. PhD Dissertation, Amherst. Frampton, J. (1991). Relativized Minimality: A Review. The Linguistic Review 8, 1-46. Fromm, Hans (1982). Finnische Grammatik. Heidelberg: Winter. Hakulinen, Α., F. Karlsson and M. Vilkuna (1980). Suomen tekstilauseiden piirteitä: kvantitatiivinen tutkimus. Helsinki: University of Helsinki, Department of General Linguistics Publications no. 6. Hawkins, J. (1978). Definiteness and indefìniteness. A study in reference and grammatically prediction. London: Croom Helm. Heggie, Lorie A. (1988). The Syntax of Copular Structures. PhD Diss., Univ. of Southern California. Holmberg, Anders and Urpo Nikanne (1993). Introduction. In A. Holmberg and U. Nikanne (eds.), 1-20. Holmberg, Anders and Urpo Nikanne (eds.) (1993). Case and Other Functional Categories in Finnish Syntax. Studies in Generative Grammar 39, Berlin/New York: Mouton de Gruyter. Karlsson, Fred (1984). Finnische Grammatik. Hamburg: Buske. Kiparsky, Paul (1998a). Partitive Case and Aspect. In The Projection of Arguments Lexical and Compositional Factors, Miriam Butt and Wilhelm Geuder (eds.), 265-307. Stanford: CSLI Publications.

Case alternation in Finnish copular

constructions

171

Kiparsky, Paul (1998b). Cases as Complementizers. Paper presented at the workshop 'Effects of Morphological Case', Utrecht, August 1998. Larson, R. K. (1988). On the double object construction. Linguistic Inquiry 19, 335-391. Leech, G. and L. Li (1995). Indeterminacy between Noun Phrases and Adjective Phrases as Complements of the English Verb. In The Verb in Contemporary English, Aarts and Meyer (eds.), 183-202. Cambridge University Press, Levin, B. (1993). English Verb Classes and Alternations. Chicago University Press. Löbel, E. (2000). Copular Verbs and Argument Structure: Participant vs. Non-Participant Roles. Working papers Theorie des Lexikons No. 115, 1-28; University of Cologne. To appear in Theoretical Linguistics. Rizzi, L. (1990). Relativized Minimality. Cambridge, Mass.: MIT Press. Rothstein, S. (1995). Small Clauses and Copular Constructions. In: Small Clauses, A. Cardinaletti and M. T. Guasti (eds.), 27-48. New York: Academic Press. Schot-Saikku, Päivi (1990). Der Partitiv und die Kasusalternation: zum Fall Partitiv in der finnischen Syntax. Hamburg: Buske. Stowell, T. (1978). What was there before there was there. Chicago Linguistic Society 14, 458-471. Sulkala, Helena and Merja Karjalainen (1992). Finnish. Series Descriptive Grammars, London/New York: Routledge. Toivainen, Jorma (1993). The Nature of the Accusative in Finnish. In A. Holmberg and U. Nikanne (eds.), 111-128. Vainikka, Anne (1993). The Three Structural Cases in Finnish. In A. Holmberg and U. Nikanne (eds.), 129-159. Williams, E. (1980). Predication. Linguistic Inquiry 11, 303-338. Williams, E. (1994). Thematic Structure in Syntax. Cambridge, Mass.: MIT Press. Williams, E. (1997). The Asymmetry of Predication. In Proceedings of the 1997 Texas Linguistics Society Conference, R. Blight and M. Moosally (eds.), Texas Linguistic Forum 38, 323-333.

Optimal Exceptions

Gisbert Fanselow

1.

Introduction

*

Syntactic studies of case have a bias towards the analysis of structural (regular) case, that is, they focus on cases that encode grammatical functions or hierarchies among the arguments (structural case), or that can be understood as semantic subclassifications of NPs in structural case positions (e.g. the accusative-partitive distinction in Slavic languages)1. In contrast, 'lexical' or 'inherent' case in the sense of a case linked in a (relatively) unpredictable way to a certain argument of a verb has only rarely come even remotely close to being the focus of syntactic approaches to case - presumably because it may seem obvious that one can learn only little about the regularities of a language from the marked periphery of lexicon-based exceptions. In this paper, I want to show that this prejudice against inherent case is not justified: exceptions, in particular in the domain of case, have interesting implications for the nature of lexical entries. As we will argue, the domain in which case exceptions occur is severely restricted. An interaction of grammatical principles implies that there is at most one case exception per verb, and this case exception must show up on the lowest argument position - unless other principles override the laws behind these theorems. We take these generalizations as evidence for the claim that a connection between thematic roles/arguments and individual cases cannot be established in the lexicon. The expressive power of the lexicon is too limited for that. If correct, our findings argue in favor of a syntactic rather than a lexicon-based view of case determination.

Parts of this paper have been presented at the 1998 conference 'The Lexicon in Focus' of the SFB 282. I would like to thank the organizers and participants for helpful discussion - a particular thank goes to Barbara Stiebels and Dieter Wunderlich (also for their patience). The paper owes much to inspiring and detailed discussions with Joanna Blaszczak, Karin Donhauser, and Susann Fischer. Gereon Müller's comments on a prefinal version made significant revisions possible, for which I am very grateful. The paper also owes a lot to discussions with Artemis Alexiadou, Caroline Féry, Hans-Martin Gärtner, Hubert Haider, Matthias Schlesewsky, and Peter Staudacher. Some of the observations in this paper had been made in Fanselow (1985), others in Fanselow (1991), but they never reached the light of a broader public. 1 This paper will ignore so-called adverbial cases.

174

Gisbert Fanselow

The next section discusses the case patterns of German. We establish a number of generalizations concerning exceptional case: nominative is never exceptional, agents are never marked exceptionally, there is at most one exceptional case per verb, and it always goes to the lowest position in the argument hierarchy. Section 3 derives these generalizations from a restricted view of lexical entries and a 'parallel movement' constraint borrowed from Müller (2000a). Section 4 extends this analysis to Icelandic. In section 5, we discuss case marking from a cross-linguistic point of view, with a particular emphasis on case realization. Section 6 is devoted to a brief discussion of alternative approaches - a different view on argument hierarchies (as proposed, e.g., by Hoberg 1981, Müller 2000b) will be considered as well as a different perspective on case linking proposed by Wunderlich (1997) and related work.

2.

Verbal case patterns in German

2.1

The generalizations

German syntax distinguishes four different cases: nominative, accusative, dative, and genitive. The three regular case patterns are listed in (1), and can be predicted from a small set of simple rules. (1)

one-place verbs: nominative two-place verbs: nominative accusative three-place verbs: nominative dative accusative In addition, case can be determined in an exceptional way, too, and an inspection of the range of case exceptions leads to the following observations: Gl: G2: G3:

The highest argument of causative/agentive verbs is never marked exceptionally. At most one argument of a verb can be marked exceptionally. Only the lowest (DP-) argument of V can be marked exceptionally.

German respects G2 in a very obvious sense; other languages may allow a few exceptions to G2 that can be explained away (see below). G2 has been noted a number of times in the literature (see Czepluch 1982, Fanselow 1985, Yip, Maling & Jackendoff 1987).2 G3 faces an apparent (but not real) problem with psychological and other 'inverted' predicate types. A descriptive list of the case patterns for (noncomplex) verbs in German can be found in table 1.

2 Primus (1994) and Lenz (1997) propose markedness rankings for case frames. The markedness of a frame is a function of the number of deviations from the regular pattern. This predicts that case frames with two exceptions should be rare. Therefore, their proposals are related to what the references cited above imply. Empirically, the markedness approach comes close to what our system implies, but our model tries to derive rather than state markedness scales.

175

Optimal Exceptions Argument 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14

Argument 2

NOM

acc dat nom

nom

NOM

ACC

nom nom dat ??acc nom nom nom nom

dat gen nom nom acc acc acc acc

NOM

DAT

Argument 3 Example kommen 'come' frieren 'be cold' grauen 'dread' sein 'be' sehen 'see' helfen 'help' gedenken 'commemorate' gefallen 'please' ärgern 'annoy' acc nennen 'call' acc lehren 'teach' aussetzen 'expose' dat gen anklagen 'accuse' geben 'give' ACC

Table 1: Case Patterns in German (1) The contents of table 1 will be revised below. The numbering of the arguments reflects their position in the argument hierarchy, with 1 being highest. Regular patterns appear in small capitals. G2 and G3 impose virtually no restrictions on exceptions for monoargumental constructions (patterns 1-3), but Gl correctly implies that patterns 2 and 3 are restricted to nonagentive verbs. Biargumental verbs should tolerate an exceptional case on the lower argument 2 only - if we can account for 8 and 9, this prediction is borne out. Expectations for triadic verbs depend on whether the exceptional case position is ignored by the structural case rules or not. The former seems to hold in German: We expect free case variation on the lowest argument only. The intermediate argument receives dative in a fully regular pattern, and structural accusative case in the irregular paradigms. The options listed above are therefore derived.

2.2

Details of the paradigms

2.2.1 One-place verbs Among one-place verbs, nothing spectacular can be discovered. In addition to verbs constructed with default nominative (2c), a handful of verbs selects accusative case3, and an even smaller set of verbs combines with the dative. The nonexistence of oneplace verbs governing genitive case is probably a historical accident. 3

The verbs selecting inherent accusative (frieren 'be cold', hungern 'be hungry') are compatible with a regular construction ich friere 'I.nom am cold' quite easily. This suggests that subject agreement is tied with the case system in German.

176 (2)

Gisbert Fanselow a.

mich

mich

I.ACC

dürstet thirsts Ί am thirsty' b. mir vor dir graut I. DAT dreads about you Ί fear you' c. ich komme I.NOM come

I.ACC

friert freezes

In contrast to the situation in Icelandic, the exceptionally marked arguments in (2a) or (2b) do not fulfill the criteria for subjecthood, the constructions are subjectless. None of the predicates with an exceptional case is agentive. Two classes of semantically monoargumental verbs with a lexical case specification do not follow the patterns in (2). One class requires the insertion of an overt expletive into the subject position (see 3a), for which one cannot make the Extended Projection Principle responsible because of the grammaticality of subjectless sentences such as (2a,b). We assume that an extra quasi-argument (corresponding to an unspecified Cause) is present in these structures. Similarly, so-called 'inherently reflexive' verbs as in (3b) combine with two cases although they possess one argument only. Again, sich might be treated as a quasi-argument in (3b), perhaps also with a causative interpretation, along the lines proposed by Pesetsky (1994). (3)

2.2.2

a.

dass that 'that b. dass that 'that

es mich hingehauen hat it I.ACC downknocked had I fell' der Zweig sich biegt the branch REFL bends the branch bends'

Two-place verbs

(4) illustrates case patterns 4-7 from table 1 for two-argument verbs: (4)

a. b. c. d.

ich ich ich ich

bin/werde ein Held gedenke des Helden helfe dem Helden sehe den Mann

I.NOM am/become a.NOM hero I.NOM commemorate the.GEN hero I.NOM help the.DAT hero I.NOM see the.ACC man

All four cases of German show up in the object position of two-place verbs. The absence of subjectless two-place verbs, that is, the fact that at least one DP argument bears nominative case, is striking and calls for an explanation. It supports G 2: a verb with, e.g., a dative-accusative case pattern would have to bear two exceptional lexical specifications (if, as seems likely, the assignment of structural accusative case presupposes the presence of structural nominative, see below). None of the verbs with an exceptional pattern assigns the exceptional case to an agent or a cause. Two-place verbs thus support G l as well. A closer look at (4) reveals two further aspects of case linking. First, it is not entirely clear whether the second nominative in (4a) is 'assigned' in the same way as other 4 The EPP states that every clause must have a subject.

Optimal Exceptions

177

object cases, or whether it simply instantiates the general rule that predicate nomináis agree in case with the noun phrase they are semantically linked to, a generalization that is motivated independently, as (5) illustrates: the expression 'one after the other' derives its case from the noun phrase which it modifies semantically. (5)

a.

dass die Männer ihn einer nach dem anderen anriefen that the.NOM men him one.NOM after the other called 'that the men called him up one after the other' b. dass er die Männer einen nach dem anderen anrief that he the.ACC men one.ACC after the other called 'that he called the men up one after the other' Given that double nominative constructions are restricted to verbs which combine with a predicate nominal (sein 'be', werden 'become', and bleiben 'remain'), the agreement facts in (5) make it likely that a second nominative case instantiates case agreement with the subject rather than an exceptional object nominative (see also Yip, Maling & Jackendoff 1987 for the assumption that there is no exceptional nominative). Bittner & Hale (1996) suggest that nominative is a null case, and we can derive the absence of exceptional nominative case from this: an exceptional case must be listed in the lexicon. Suppose that cases are defined by some (yet to be identified) features, and suppose that the 'null case' nature of the nominative is a reflection of the fact that nominative noun phrases have no case-related feature at all. An exceptional nominative is then a contradictio in adiecto: assigning an exceptional case is equivalent to lexically prespecifying the corresponding features. If the nominative corresponds to no (markedness) features, prespecifying nominative would mean to prespecify no feature at all but this is the representation for a fully regular pattern. (4b) and (4c) are entirely straightforward: they represent the systematic possibility of assigning exceptional cases in the lexicon. (4d), on the other hand, instantiates the regular option for structural accusative in German. Recall from section 2.2.1 that accusative can be an inherent case in German. One might therefore expect that the case pattern nom-acc comes in two varieties, one in which the accusative is a structural case, and one in which it is lexically prelinked. This expectation is not borne out, though, as can be seen from the following observation: inherent cases are retained in the German werJe/j-passive (6b,c), but there is no twoplace verb that keeps an accusative object in such a passive. This shows that accusative is not a legal lexical exception with two-place verbs. (6) a.

der Professor wurde im 'Tresor' gesehen the.NOM professor was in.the 'Tresor' seen 'one saw the professor in the "Tresor"' b. mir wird geholfen I.DAT is helped 'one helps me' c. seiner wird gedacht his.GEN was commemorated 'one commemorates him'

178

Gisbert Fanselow

Verbs like those in (2a) rule out the idea that exceptionality is incompatible with the accusative as such, and the next section will provide an argument for the claim that accusative can be exceptionally linked to the direct object position of a triadic verb, that is, exceptional accusative case assignment is ruled out for the direct object position of dyadic verbs only - for those constellations, in other words, in which a lexical accusative would be formally indistinguishable from a structural one. This suggests that the restriction can be derived from an obvious learnability consideration. The patterns 4-7 of two-place verbs instantiate G3, because nominative case shows up on the higher of the two argument positions. That the exceptional case always goes to the lower argument position, as G3 claims, seems to be refuted by classes 8 and 9 of table 1, however. Class 8 (the dative-nominative class) represents a variation among psychological and other nonagentive predicates that one finds in many languages: the 'regular' nomacc pattern alternates with a dat-nom constellation. (7)

a.

der Ritter mag die Prinzessin the.NOM knight likes the.ACC princess 'the knight likes the princess' b. dem Ritter gefällt die Prinzessin the.DAT knight pleases the.NOM princess 'the princess pleases the knight' A number of verbs in class 8 (like entgleiten 'slip off or auffallen 'strike') are unaccusatives, since they form their perfect with sein 'be' rather than haben 'have', see Grewendorf (1989) for a discussion. Essentially, there are two ways of treating class 8, and we will consider them both. If one accepts the standard view reflected in table 1 that the experiencer is the higher of the two arguments of gefallen 'please' or auffallen 'strike', the dative marks the higher argument, but this does not necessarily imply that G3 is refuted - it is not, in fact, if the dative we find there is a structural rather than an exceptional case (see e.g. Fanselow 1991, Baylin 1995). Notice in particular that one wants the dative of a three-place verb to be structural both in the active nom-acc-dat and in the passive nom-dat pattern. In the latter, the dative combines with a nominative argument only, and this nominative argument is the direct object, which is normally assumed to be lower in the thematic hierarchy. We have to conclude, then, that dative-nominative is a regular case pattern of German (in the passive). The deep parallel between passives and unaccusative constructions has often been noted, so that it would be no surprise if a regular dat-nom pattern shows up in both construction types. It is an empirical question, then, whether the dative of class 8-verbs is a lexical exception, or an effect of a structural case rule applying to unaccusatives. There is no straightforward evidence in German for or against the inherent vs. regular nature of the dative of psychological predicates. The fact that unaccusative predicates combine regularly with a dative DP (noted by Grewendorf 1989), quite in contrast to standard intransitives, suggests indeed that the dative is regular rather than exceptional in pattern 8. A further argument in favor of this view comes from the fact

179

Optimal Exceptions

that other languages like Russian (see, e.g., Baylin 1995) behave in the same way. This would not be expected if the dative is a lexical exception. Alternatively, one might question the claim that the dative corresponds to the higher argument position of psychological predicates. Depending on the kind of evidence one considers crucial, there may be reasons to believe that the theme argument is higher than the experiencer. The datives of class 8 would then show up on the lower argument slot. For the validity of our generalizations, it would be of little importance, then, if the dative is exceptional or regular. A discussion of class 8 can be found in the next section and in section 6. Pattern 9 (putative acc-nom verbs) has to be dealt with in the latter way. There is evidence against the assumption that the accusative argument occupies a higher position in the thematic grid than the nominative argument. For instance, Sternefeld (1985) and Grewendorf (1989) observe that the PRO argument of an infinitive may be left uncontrolled only if the infinitive represents the structurally highest argument. Note, then, that an arbitrary interpretation of a PRO-subject of an infinitive is possible with nearly all accusative-nominative verbs. (8) a. PROarb sich zurückzuziehen würde die Polizisten irritieren REFL to withdraw would the.ACC policemen irritate 'it would irritate the police if one retreated now' b. PROarb sich auf sie zuzubewegen würde die Polizisten REFL towards them to move

would the.ACC policemen

ärgern annoy '*it would annoy the policemen if one moved towards them' In other words, these predicates behave as expected under the approach proposed by Pesetsky (1994): the nominative argument is a Cause, the accusative argument an Experiencer. No exceptional case shows up on the higher Cause argument. Note that normal word order is a bad guide for identifying argument hierarchies in German in other areas as well, as Hoberg (1981) and others have observed: there is an orderingby-animacy principle for arguments in German that is tied in the OT-sense with ordering principles based on grammatical functions (see, e.g., Müller 2000b). There seems little reason, then, to believe that class 9 exists - it is a subcase of the regular pattern. But note that this insight contributes a further argument against the idea that class 8 might involve an exceptional dative on the highest argument position: it would then be quite unclear why it is just psychological verbs and just datives that manage to violate G 3. 2.2.3 Three-place verbs The generalizations formulated above imply far-reaching but correct predictions concerning the case patterns of three-place verbs in German. Because of G2/ G3, the legal patterns must be of the form (9), in which linear order corresponds to the hierarchy of the arguments of the predicate.

180 (9)

Gisbert

Fanselow

Regular: NOM

-

DAT -

ACC

With an exceptional case: NOM

-

STRUCTURAL CASE

-

{AGREEING, GEN, DAT, ACC}

The nature of the structural case in the exceptional pattern in (9) depends on whether the presence of the DP with an exceptional case is 'ignored' or not by the case assignment system of the language, as has been observed by Fanselow (1985) and Yip, Maling & Jackendoff (1987). In German, inherent case seems to make the lowest argument position 'invisible' for the case assignment rules, so that the structural case mentioned in (9) is accusative, see (10), similar to what we observe in (most) constructions that combine a DP with a PP object. (10) Regular: NOM

-

DAT -

ACC

With an exceptional case: NOM

-

ACC -

{AGREEING, GEN, DAT, ACC}

The 'agreeing' pattern of (10) is exemplified in (11a) (=class 10 of table 1): the two occurrences of accusative case on the objects of nennen 'call' arise by case agreement, a fact that becomes obvious when one considers the passive version ( l i b ) , in which both underlying objects switch to nominative. There are no three-place verbs in German in which the lowest argument is a predicate of the subject rather than of the second object (presumably a Minimal Link Condition effect), so that the pattern nom-accnom does not arise. (11) a. b.

ich nenne ihn einen Idioten I call he.ACC an.ACC idiot.ACC er wird ein Idiot genannt he.NOM is an.NOM idiot.NOM called 'he is called an idiot'

German possesses a handful of nominative-accusative-genitive verbs, as (12) illustrates. The genitive appears on the lowest argument, and the accusative borne by the intermediate argument is a structural one, as the passive test (12b) shows. (12) a.

b.

ich beschuldige ihn des Mordes I accuse he.ACC the.GEN murder.GEN Ί accuse him of murder' er wird des Mordes beschuldigt he is the.GEN murder.GEN accused 'one accuses him of murder'

The pattern (10) is compatible with the existence of verbs selecting an inherent accusative case and combining it with a structural accusative for the intermediate argument. Such nom-acc-acc verbs differ from the agreeing system exemplified in (11) in their behavior under passivization: the inherent accusative of the lowest argument is retained, while the intermediate argument appears in the nominative. The passivization test shows that only one of the two occurrences of accusative case is systematically construed as the structural one - it is the one that alternates (13b). Again, we see that

Optimal

181

Exceptions

the theme receives the lexical marking (13c) and not the higher goal/experiencer argument (13b). 5 (13) a. b.

c.

er

fragt

mich

etwas

Schweres

he

asks

I.ACC

something.ACC

difficult.ACC

ich werde etwas Schweres I.NOM am something.ACC difficult.ACC 'One asks me something difficult' *etwas Schweres wird mich gefragt

gefragt asked

If dative case is never exceptional with three place verbs, we could defend a generalization proposed in Fanselow (1985): regular patterns must never be formally identical with irregular ones. Following Haider (1993), it has become a common practice, however, to distinguish two/three classes of dative-accusative verbs: there are verbs that have an unmarked dative before accusative serialization, while others prefer accusative before dative, viz. the class 12 verbs of table 1 (and there is a third class, the largest one, for which no preferred serialization can be established). If one makes the further assumption that unmarked order reflects the hierarchy of arguments, then dative would be assigned to the lowest argument position of verbs like aussetzen 'expose' in (14a-c). The accusative object switches to nominative case in the passive (14d). If irregular, the case pattern of (14) must thus be due to a lexical dative appearing on the lowest argument,

5 Gereon Müller and Dieter Wunderlich (p.c.) observe that the acceptability of a passive involving lehren 'teach' decreases sharply when the nonstructural accusative is marked overtly, as in (ii), and is not homophonous with the nominative, as in (i). (i) die Kinder werden Mathematik gelehrt the children are mathematics.ACC taught 'one teaches the children mathematics' (ii) ?* die Kinder werden den Rhythmus gelehrt the children are the.ACC rhythm taught 'one teaches the children the rhythm' It is not clear what conclusions one should draw from this observation. Consider (iii): DPs in PPs headed by als derive their case from the DP the PP modifies semantically. The appearance of accusative case with grossen Blödsinn (grammatical for some but not all speakers) in (iii) thus shows that nichts does bear accusative case. (iii) der Mensch wird nichts. als grossen Blödsinn gelehrt the man is nothing.ACC but great.ACC nonsense taught 'one does not teach human beings anything but complete nonsense' In any event, passives formed with double accusative verbs are grammatical in Latin and in Ancient Greek, with the retention of the lexical accusative case, see e.g. Smith (1996:88). In some dialects, lehren can be constructed with a regular case pattern, too, in which case it passivizes as expected.

182

Gisbert Fanselow

quite in line with the three generalizations introduced above, but not in line with the conjecture of Fanselow (1985).6 (14) a. er setzt die Kinder der Kälte aus he

sets

the.ACC

children

the.DAT

coldness

out

b. ?er setzt der Kälte die Kinder aus c. er setzt der Kälte die KINDER aus 'he exposes the children to coldness' d. die Kinder werden der Kälte aus-gesetzt the.NOM children are the.DAT coldness out-set 'the children are exposed to the coldness' Recall that unmarked word order of German is also governed by an animacy hierarchy that can override serialization by grammatical function. Normal order facts therefore do not constitute compelling evidence for a separate case pattern 12. We may observe, however, that the standard correlation between case and thematic roles is 'inverted' with verbs like aussetzen 'expose', unterziehen 'subject somebody to', etc.: the accusative noun phrase realizes the intermediate Experiencer/Goal role, while the dative appears on the Theme. This is different from the standard scheme, in which the dative links to the Experiencer/Goal. To the extent that thematic roles determine base positions uniformly (Baker 1988), we can conclude that Class 12 exists. One would also like to see some evidence for the claim that the dative of Class 12 verbs is lexical. If lexical cases cannot alternate with structural ones (see Chomsky 1981 and subsequent work; see also below for some discussion), the well-formedness of the bekommen/kriegen-a\temation of German (see Wegener 1985, Fanselow 1985) might help in deciding about the lexical nature of some occurrences of the dative. (15) a. er stiehlt mir ein Buch he steals me.DAT a book 'he steals a book from me' b. ich bekomme ein Buch gestohlen I.NOM get a book stolen 'a book was stolen from me' (16) a.

ich mische Wasser der Säure bei I mix water the.DAT acid to Ί mix water to the acid' b. Die Säure bekommt Wasser bei-gemischt The.NOM acid gets water to-mixed 'water is mixed into the acid'

6 Meinunger (1995) observes that all irregular dative verbs are morphologically complex - they are a combination of Ρ + V. Typically, such combinations assign the case governed by the preposition, as nach-telefonieren 'after-phone', vor-singen 'before-sing' selecting dative, vs. durch-pflügen 'through-plough', an-fliegen 'at-fly' selecting accusative illustrate, see Fanselow (1985:579ff.). The exceptional dative of class 12 verbs may thus 'reduce' to case government by the preposition.

Optimal

Exceptions

(17) a.

183

ich setze die Kinder der Kälte aus I set the children the.DAT coldness out Ί expose the children to coldness' * die Kälte bekommt die Kinder aus-gesetzt the-NOM coldness gets the children out-set 'the children are exposed to coldness'

b.

(18) a.

ich unterziehe die Studenten einer schweren Prüfung I subject the students a.DAT difficult test Ί subject the students to a difficult test' * eine schwere Prüfung bekam die Studenten unterzogen a.NOM difficult test got the students subjected 'the students were subjected to a difficult test'

b.

The data in (15)-(18) suggest that verbs with a standard correlation of case and thematic roles are almost always fine in the bekommen-consimcXion (15/16), in which the dative of the active construction is replaced by the nominative. Class 12 dative-accusative verb such as aussetzen 'expose' however seem to resist entering the bekommenalternation, which is to be expected if the dative is lexical. 7 If this line of reasoning is correct, the dative may be assigned to the lowest argument of ditransitive verbs as a lexical case. The fact that the correlation between object cases and thematic hierarchies of class 12 is inverted relative to the standard pattern class 14 may be sufficient for the lexical nature of the dative to be learnable by children8.

7 Gereon Müller (p.c.) objects to this by pointing (a) to the existence of standard ditransitive verbs that fail to enter the bekommen-constmction, and (b) the alleged existence of class 12 verbs that allow the bekommen-alternation. I see no problem with (a), because passive in general fails to be applicable to some verbs. According to Müller, the paradigm in (i) is a case in point for his claim (b). I do not agree with the classification of entziehen as a class 12 verb, however, because the accusative noun phrase clearly is the theme of this verb, so that the case pattern is not inverted, as required for being an instance of class 12. Unmarked serialization as illustrated in (ii) is indeed deceptive in this case: the dative is preceded by the accusative because of animacy, and not because of thematic hierarchy. (i)

(ii)

a.

ich entziehe dem See Wasser I remove the.DAT lake water Ί remove water from the lake' b. der See bekommt Wasser entzogen the.NOM lake gets water removed 'one removes water from the lake' sie entzieht die Kinder seinem Einfluss she removes the children his.DAT influence 'she removes the children from his influence'

8 If class-12 does not exist, as Müller (p.c.) suggests, one only has to assume that the inverted nature of the argument-case correlation is not sufficient for the irregularity of a dative being learnable. Nothing in our approach hinges on this issue.

Gisbert Fanselow

184 2.2.4

Summary

In the light of this section, we can simplify table 1 by eliminating nonexisting or agreement-triggered case patterns. Regular patterns appear in small capitals. Argument 1 1 2 3 5 6 7 8 11 12 13 14

Argument 2

Argument 3

NOM

acc dat NOM

ACC

nom nom dat nom nom nom

dat gen nom acc acc acc

acc dat gen

NOM

DAT

ACC

Example kommen 'come' frieren 'be cold' grauen 'dread' sehen 'see' helfen 'help' gedenken 'commemorate' gefallen 'please' lehren 'teach' aussetzen 'expose' anklagen 'accuse' geben 'give'

Table 2: Case Patterns in German (2)

3.

Explaining the patterns - deriving the constraints

3.1

Deriving the restrictions of German case patterns

The facts discussed in the previous section have not gone unnoticed, and several ideas have been proposed as an account for them. In her contribution to a 1996 workshop on restrictions on the number of verbal arguments, Marga Reis suggested that an OCP-like statement (see also Woolford, to appear) rules out the nonexistent nom-dat-dat and nom-gen-gen patterns. According to this idea, a single case must not be used twice in the case grid of a verb. This is incompatible with the existence of double accusative verbs, but the OCP is likely to be a violable constraint. The OCP cannot explain, however, why there are no nominative-dative-genitive verbs in German, so that it has at least to be complemented by further principles. Czepluch (1982), Fanselow (1985), and Yip, Maling & Jackendoff (1987) propose that each lexical entry may specify one exceptional case only, i.e., they share our generalization G2, repeated here for convenience. G l : The highest argument of causative/agentive verbs is never marked exceptionally. G2: At most one argument of a verb can be marked exceptionally. G3: Only the lowest (DP)- argument of V can be marked exceptionally.

Optimal Exceptions

185

If the nominative cannot be exceptional, the case patterns are restricted to (19) by G2 (*means the pattern is ungrammatical; ! identifies 'inverted' psych verbs). (19) nom, *gen, dat, acc nom gen, nom dat, nom acc, *gen nom, *acc nom, !dat nom nom dat acc, nom acc acc, nom acc dat, nom acc gen, *nom gen acc, *acc nom acc, *dat nom acc, *gen nom acc G3 filters out all patterns in which the exceptional case is not the lowest in the hierarchy. This leaves us with the attested patterns only - and the possibly problematic case of psychological predicates. (19') nom, *gen, dat, acc nom gen, nom dat, nom acc nom dat acc, nom acc acc, nom acc dat, nom acc gen, !dat nom

3.2

Deriving Gl

We will now try to derive the three generalizations on case assignment. In line with Woolford (to appear), we assume that case determination is a two-step process: a set of general case assignment principles in the syntax identifies a number of potential cases for an NP (corresponding to the GEN part in an OT account), and an evaluation component that selects the best case from these candidates. Gl reduces to a simple and quite standard assumption concerning case assignment/ checking and the thematic structure of verbs. In line with much recent work, including Chomsky (1995), (20) may be taken to represent the basic structural grid for the predicate phrase: agents and causes are not introduced into the syntactic structure by the verbal head itself, but by a 'light verb' ν selecting the VP. The slot for agent and causer subjects, X, is thus not c-commanded by the lexical verb. (20) ...TNS [.. ..ASP... [VP Χ ν [ VP Y [V Ζ]]] ...]....] If case is assigned to DPs by heads in syntactic structures such as (20), and if case marking by a head Η has the m-command9 domain of H as its upper bound, then V cannot assign case to X. Exceptional cases must be specified in the lexical entry of the full verb, however, i.e., they must be assigned by the lexical verb. Therefore, the absence of exceptional cases on agents and causers is derived.10 (20) allows, however, for the assignment of ergative case to X by v, if ergative is a case that is assigned to specific thematic roles independent of agreement projections, as (more or less) argued for by Woolford (1997), among others.

9 10

A head H m-commands all XPs not excluded from its projection HP. Müller (p.c.) observes that this reasoning presupposes that V does not assign case after it has moved to Infi or higher. This is, however, a standard background assumption in, e.g., the Government-and-Binding model or the Minimalist Program.

186

Gisbert Fanselow

The (fairly standard) system envisaged here allows for the possibility that the case of X is determined by any of the functional heads c-commanding the agent. Thus, the finiteness of Infi (e.g., nominative vs. accusative case in Latin) may influence the case that X bears. A related example may be genitive assignment to structural subjects by a modal head in Finnish (if the phenomenon is to be analyzed as such, see Laitinen & Vilkuna 1993 for a discussion). Likewise, agents may receive dative case in Serbian (Wunderlich, p.c.) when they form part of a modal construction. This can be accounted for in terms of structural case assignment by a higher modal head - it is not an instance of an exceptional marking of the agent by the verb.

3.3

Deriving G3

What comes to mind as a possible source for the generalization G3 is the concept of structural case government, too. Recall that we presuppose a case theory roughly as proposed by Woolford (to appear), in which the GEN component of syntax generates structures in which cases are assigned or checked, and the EVAL component selects among the options so generated. We assume that structural nominative and structural accusative are assigned by functional heads c-commanding VP: nominative by TNS, and accusative by an aspectual or voice head, by the light verb v, or by AGR-O. On the other hand, by their very nature, exceptional cases must be assigned by the lexical verb V. One could assume that government by lexical items is restricted to strict c-command. The verb V in (20) could thus check/assign the case of Σ only if V c-commands Σ in the base structure (i.e., V can assign a case to Ζ only), and this is equivalent to saying that Σ is the first, the lowest argument of V. The c-command restriction on lexical case government would therefore not only imply G3 - if syntactic branchings are always binary, G2 would be derived as well (if we presuppose standard downwards restrictions on government). The idea of restricting verbal case government to the c-command domain of V is not, however, compatible with a number of facts, among them being the case situation of verbs selecting prepositional objects. Prepositional objects, so it seems, always occupy the lowest position in VP, at least this is suggested by binding and constituent order facts - the prepositional object immediately precedes the verb in (the OV-language) German, and can hardly leave this position by scrambling (although it can undergo topicalization). If prepositional objects occupy the lowest argument position of a verb, and if exceptional cases have to go to the lowest position in VP because of the c-command requirement on case government, we expect that prepositional objects are incompatible with the additional presence of a lexically governed case. In fact, most verbs that select prepositional objects either do not tolerate further object arguments, as denken 'think (of)' or rechnen 'reckon (with)', or they combine with a structural accusative case, as aufklären über 'inform sb. about', interessieren für 'interest sb. in', befreien von 'liberate sb. from', warnen vor 'warn about', erinnern an 'remind o f , do:

Optimal Exceptions

187

(21) a.

Ich erinnere ihn an sein Versprechen I remind he.ACC at his promise Ί remind him of his promise' b. Ich warne ihn vor seinem Hund I warn he.ACC before his dog Ί warn him of his dog'

But there is a handful of verbs like abraten von, dienen zu, which select a prepositional object and a lexical dative. The simple account of G3 in terms of a c-command restriction must therefore be false. (22) a. ich rate dir von der Fahrt ab I advise you.DAT from the trip off Ί advise you not to go for the trip' b. das dient mir zur Entspannung this serves I.DAT to.the (=for) recreation 'this serves me for recreation' Prepositional objects are exceptional in many cases (as in (22)), because the selected preposition cannot be predicted. For a verb like dienen, the lexical entry thus specifies two exceptional features for arguments - one encoding the preposition zu, and the other dative case, a situation which is not ruled out by any of the generalizations considered so far, but which sharpens the problem of deriving G 2: if there can be two exceptional linking features for the arguments of a verb, it is quite implausible that there is a direct ban against two exceptional case features. The exceptional linking features of dienen exclude each other, so they must be checked by two different XPs. And only one of them, viz. the PP, will occupy a position that is c-commanded by the lexical verb. Assume that prepositional object constructions always involve the incorporation of the preposition into the verb, as seems likely. Recall that head movement is restricted to complements (Baker 1988), which implies that the PP checking the zw-feature must occupy the complement position of dienen, that is, it must be the lowest argument. The configuration simply could not be different: Ρ cannot incorporate into V if PP is not the lowest complement c-commanded by V. The same line of reasoning applies, of course, to all prepositional object constructions. We derive, therefore, that the presence of a prepositional object shifts the realization of the lexically governed case necessarily to the second-lowest argument position available. The existence of data like (22) thus implies that two arguments can be introduced in the projection of the lexical root: Ζ of (20) can be a PP, and Y an exceptionally case marked element. Furthermore, V must be able to exceptionally case-mark Y in (20), i.e., the case marking domain of V is its whole maximal projection. (22) finally shows that the exceptional case goes to the lowest available argument position: if other principles like the need to incorporate the head of prepositional objects exclude the possibility that the exceptional case aligns with the lowest position, it appears on the immediately higher one. Therefore, G3 must be reformulated as follows: G3*

Exceptional case is realized on the lowest available position.

188

Gisbert Fanselow

If the case checking/assigning domain of the verb comprises the whole VP, and if VP can contain two arguments, why does the exceptional case go to the lower of the two arguments whenever that is not ruled out on other grounds (such as the additional presence of a prepositional object?) The facts seem related to a condition on case checking/assignment that can be expressed in terms of (23) (23) If α, β check the case of γ, δ, respectively, and α c-commands β, then γ c-commands δ as well. The lexical verb V is the lowest category in (20) that checks case, it is c-commanded by all heads such as TNS, ASP(ECT), V, etc., that check structural case. For obvious reasons, lexical exceptions must be specified with the lexical verb root. By (23), we derive that all DPs bearing structural case must c-command the DP with lexical case. Why should (23) hold? (23) implies that the associations between case assigners and recipients should involve crossing rather than nested dependencies, and one might simply state, following an insight of Williams (1999), that case belongs to the realm of argumentai dependencies, which are always crossing. (23) is also reminiscent of Müller's (2000a) Parallel Movement Constraint (PARMOV). Müller shows that a considerable array of data can be explained if (unless other principles interfere) movement must NOT change c-command relations. In fact, (23) reduces to PARMOV if (a) case determination comes about by the checking of features of noun phrases by case checking heads, and if (b) checking implies the attraction of the (formal features) of a DP by the case checking head (as in Chomsky 1995) PARMOV:

If α c-commands β at structure L, then α c-commands β at structure L'. Consider first (24a), an abstract representation of a structure 'preceding' overt movement serving case checking/assignment. Suppose that HI and H2 are the heads checking the case of α and β, respectively, and that the two heads check case in their specifier position (so that they trigger overt movement). Then PARMOV favors those movement configurations in which the c-command relations between α and β are preserved, i. e., the output (24b) is favored over (24c) because c-command relations between α and β have not changed. For case checking by overt movement, we have thus derived (23) from PARMOV.

(24) a. Hl H2 α...β b. α-Η1... β-Η2 c. β-Η1... a-H2 Consider now case checking/assignment that is not accompanied by visible overt movement. Several options exist. Assume that case checking arises by covert movement. If covert and overt movement have in common that a phrase Ρ is copied from position 1 to position 2, and differ only in terms of where the phonetic matrix of Ρ is spelt out, then case checking by covert movement does not differ syntactically from the situation we have just discussed (see e.g. Groat & O'Neil 1996 and Fanselow & Cavar (to appear) for such a view).

Optimal Exceptions

189

If, on the other hand, covert movement affects the set of formal features of a phrase only (see Chomsky 1995), (23) can still be derived. Consider (25) as a more explicit version of (24), where α, β represent the phonetic information of A and Β, a' and β' the formal features, and A" and Β" the semantic information. PARMOV favors the 'crossing' way of case assignment: now, just a' and β' are displaced, but the c-command relations between these feature complexes are preserved in (25b) only, and not in (25c).11 (25) a. Hl H2 b. α'-Η1... β'-Η2 c. ß'-Hl... a'-H2 ...

3.4

... direct object > verb (or its mirror image). The linking features [hr] and [lr] primarily play a role in those cases of positional linking in which the word order cannot be analyzed as default linking, (ii) MAX(+hr) or ΜΑΧ(+1Γ) violations can only be observed in positional linking if a structural argument is dislocated in order to fulfill a discourse constraint or some similar constraint. In principle, languages could exhibit an n-ary contrast in the subject of intransitive verbs, which would increase the number of possible linking systems drastically. However, languages with a more complex intransitive split (e.g. Kambera, Klamer 1998: 5-way split) show that the split is not thoroughly instantiated in the whole verbal paradigm and that it often fuses tense/aspect and modal information with linking information.

222

Barbara Stiebeis Systems that do not distinguish At and Pt h. NOM system i. Valency-sensitive A, Pi_ Ai Pi A,

p.

A,

P,

Conceptually implausible systems that do not distinguish At and P, j. A¡-prominent k. P¡-prominent 1. Pi Ai Pi Pi A,

p.

Conceptually implausible systems that distinguish At and P, m. n. o. Ai Ai Pi Pi 1 Ai 1 p. A, A, P. Θ

Pi

A,

P,

At

Pt

At

The default linker of each system is shaded. I assume that languages with an intransitive split make use of the feature [icontr] 'the situation is (not) controlled by the subject referent' or some similar feature. In languages without an intransitive split, no such feature is relevant. The four possible argument roles can thus be distinguished by the following features: A¡ [-hr,-lr,+contr], P¡ [-hr,-lr,-contr], A, [+lr,-hr,(+contr)], and Pt [+hr,-lr,(-contr)]. The systems in (13e/f) show a P-split or an Α-split besides an intransitive split; they may be regarded as transitional stages between an ACC language and an active language (vice versa) or between an ERG language and an active language (vice versa), respectively. Some Carib languages (Derbyshire 1999) show sparse relics of such systems. (13g) is morphologically maximal, exhibiting four morphological linkers, which, however, is not attested, whereas (13h) represents a system without morphological linkers. The system in (13i) uses two linkers according to the valency of the verb: one linker in intransitive verbs, the other in transitive verbs. (13j/k) single out A¡ or P¡, respectively. All four systems in (13i-l) represent inefficient morphological linking systems in that they admit UNIQUENESS violations in favor of a minor argument role distinction; these systems are not attested - as expected. (13j-o) are conceptually implausible because the default linker subsumes contradictory information regarding the feature [contr]. After this informal discussion I will turn to a formal account of linker inventories. I will provide a factorial typology in the CT-based version of LDG. I have presented the evaluation of linking patterns already in (7): the correct distribution of the linkers is determined by MAX and IDENT constraints. With respect to the linker inventories, the factorial typology must account for the actual linkers of a given language on the one hand and the overall preference for ACC, ERG, ACC/ERG and active systems on the other hand. For the former, UNIQUENESS, MAX and markedness constraints suffice; however, for the latter I have to make use of meta-constraints which show that the non-attested linker inventories suffer from unwanted lexical entries of the linkers. I regard the opti-

Linker inventories, linking splits and lexical economy

223

mization of linker entries as a subcase of 'lexicon optimization' (Prince & Smolensky 1993). 12 The evaluation of the linker entries regarding lexicon optimization is based on two assumptions: (i) the first is taken over from Minimalist Morphology (Wunderlich & Fabri 1996) and states that morphological items are positively specified; negative feature values generally represent defaults and should be avoided in the lexical representation of the linkers - if possible. The following constraint formalizes this assumption: (14) *[-]: Avoid negative specifications in lexical entries. (ii) Disjunctive specifications, which reflect an unsystematic syncretism of forms and are equivalent to multiple independent lexical entries, enlarge the burden of lexical storage and should be avoided for the sake of economy of representation. The following constraint penalizes disjunctive specifications: (15) *[v]: Avoid disjunctive specifications in lexical entries. The table in (16) lists the lexical entries for the linking systems in (13), (16a) for (13a), and so on. Some linking systems allow several lexical specifications in terms of [lr] and [hr]: the tripartite system (13d), for instance, could be characterized by the linker sets {[-hr,-lr], [+hr], [ ]}, {[-hr,-lr], [+lr], [ ]} or {[+hr], [+lr], [ ]}. The last specification does not yield a violation of *[-] and, hence, performs best. In the following table I have always chosen the specification of linkers that incurs the least constraint violations in terms of *[-] and *[v]. Under the assumption that negative specifications should be tolerated in order to avoid disjunctive specifications, a system such as (13i) is given the lexical entries in (16i), in which {[—hr,—lr], [ ]} is chosen instead of {([+hr] ν [+lr], [ ]}). The feature [contr] is only instantiated in systems with an intransitive split. The overall preference of ACC, ERG and active systems can be explained by means of a constraint ranking in which the meta-constraints *[v] and *[-] are generally highranked. UNIQUENESS is also more important than the markedness constraints *[+hr], *[+lr] and *[+contr] 'avoid a linker that is specified for [+contr]'. The various systems are ranked (last column) according to their constraint violations.

12

Unlike Prince & Smolensky (1993), who introduced the concept of lexicon optimization first, I make use of additional meta-constraints that are confined to the determination of optimal lexical entries.

224

Barbara Stiebeis

(16) Optimal lexical entries for the linking systems in (13) Linker entries a. b. c. d. e. f. g· h. i.

*M

UNIQUE * H

[+hr], [ ] [+lr], [ ] [+contr], [ ] [+lr], [+hr], [ ] [+contr], [+hr], [ ] [+contr], [+lr], [ ] [+contr], [+lr], [+hr], [ ]

[]

*

_

*

**

*

**

[ hr,-lr], [ ] f-hr,-lr,+contr], [ ] j· k. [-hr,-lr,-contr], [ ] [-hr,-lr,+contr], [-hr,-lr], [ ] 1. m. t-hr,-lr,-contr], [+lr], [ ] n. [-hr,-lr,+contr], [+hr], [ ] 0. ([-hr,-lr,+contr]v[+hr]), [ ]

*

***

*

**** *** **

*

**

*[+hr] *[+lr] *[+contr] Rank * 1 * 1 * 1 * * 2 « * 2 * * 2 * * * 3 4 6 * 7 8 * 9 * 7 * * 5 * * 10

As the table shows, (16o) is the only linking system that necessarily requires a disjunctive specification in a linker, and in this respect fares worst among all systems because it reflects an unsystematic syncretism of linkers. All other systems can be simply accounted for by the use of an underspecified linker. There are 7 systems that neither violate UNIQUENESS nor *[-], among them the four systems mostly found in the languages of the world: (16a) ACC languages, (16b) ERG languages, (16c) active languages, and (16d) tripartite systems. A system that violates UNIQUENESS in 2-place verbs fails in the most important function of linking systems, which is to distinguish the two arguments. Therefore, systems such as (16h-l) are quite unlikely to occur. The systems (16i-o) cannot circumvent negative specifications in their lexical entries. Systems with multiple violations of *[-] are extremely rare (e.g., [-hr,-lr] linker in Chorti, see footnote 13) and may represent intermediary stages of linker inventories in language change. Note that the tripartite system (16d) and the P-split system (16e) have a default linker that does not subsume contradictory specifications - in contrast to the A-split system (16f), in which the default linker subsumes [-hr] and [+hr] for P¡ and Pt. The systems (16e/f) are more complex than the tripartite system in that they show splits along two dimensions: control and valency; tripartite systems only encode a split along the valency dimension. This might be the reason that tripartite systems are attested to a certain degree, the systems (16e/f) only in some sparse relics. The conceptual implausibility of (16j-o) could be accounted for by an additional constraint that rules out inconsistent feature specifications in the default linker regarding the feature [contr].

Linker inventories, linking splits and lexical economy

225

I simplify the possible linker inventories in several respects here: (i) the tableau (16) only shows the evaluation within one linker type: case or agreement. Even if two arguments are not distinguished by agreement, they may be distinguished by case. Therefore, case and agreement (and position) need to be evaluated separately, (ii) The tableau (16) does not deal with languages whose linking system is not primarily defined in terms of [hr] and [lr], e.g., the Algonquian inverse systems. Here, the linkers are more or less defined in terms of a language-specific Salience Hierarchy (see Fabri 1996 and Wunderlich 1996 for an analysis), (iii) Certain types of linking splits, especially those conditioned by tense or aspect, show an alternation between linking systems: the Mayan language Chorti (see Quizar & Knowles-Berry 1988), for instance, uses an ERG system in perfective aspect, and some kind of Α-split system in imperfective aspect. 13 (iv) Additional morphological constraints can enforce the presence of a highly marked linking system, which occurs especially in agreement linking systems (see, for instance, Wunderlich 2000a for Dalabon and Wunderlich 2000c for Yimas). In order to account for the actual linkers of a particular language, e.g., ERG vs. ACC, constraints such as the MAX and markedness constraints come into play. For the sake of simplification, I will only consider systems without an intransitive split (16a/b/d/h). The system in (16i) is left out from the discussion because it is conceptually implausible and inefficient due to the UNIQUENESS violation. ACC and ERG languages differ in the relevance of the MAX(+hr) and MAX(+lr) constraint and, accordingly, in the relevance of the corresponding markedness constraints *[hr] and *[lr]; tripartite systems tolerate an extra linker in order to avoid violations of the two MAX constraints, as the table in (17) shows, in which 2-place verbs are evaluated for the four linking systems; there are no constraint violations in intransitive verbs. In the following discussion of linking systems, I will always present a table that illustrates the constraint violations in the various systems first, and then derive the constraint rankings for the individual systems from this table. ( 1 7 ) Constraint violations with 2-place verbs in ERG, ACC, ERG/ACC and NOM systems; Input: X.y[+hr,—lr] λχ[-1ΐΓ,+ΐΓ]

y-x a.

ACC-NOM

b.

NOM-ERG

C. d.

MAX(+hr)

MAX(+lr) *

n+lrl *

* *

UNIQUENESS

*

*

ACC-ERG NOM-NOM

*r+hrl

*

*

* -

•

13 Note that the A¡ linker in the imperfective aspect in Chorti is not defined in terms of the feature [contr], but needs the specification [-hr,-lr]/-perf because the split is aspectually conditioned and does not depend on the control properties of the subject referent. The underspecified linker [ ] is used to link subjects of intransitive perfective verbs and objects of transitive verbs.

226

Barbara Stiebeis

On the basis of the constraint violations shown in (17), the following constraint rankings, which consist of at least two partial rankings, must be assumed in order to account for the presence and distribution of linkers in the respective systems; recall that the IDENT constraints are ranked on top: (18) a.

ACC language:

{MAX(+hr) » *[+hr]} & {*[+lr] » MAX(+lr)}

b.

ERG language:

{MAX(+lr) » *[+lr]} & {*[+hr] » MAX(+hr)}

c.

ERG/ACC language:

{MAX(+lr) » *[+lr]} & (MAX(+hr) » *[+hr]}

d.

NOM language:

{*[+hr] » [MAX(+hr), UNIQUENESS]}

&

{*[+lr] » [MAX(+lr), UNIQUENESS]} In languages without morphological linking by case or agreement, the markedness constraints must be ranked higher than the corresponding MAX constraints and UNIQUENESS. In the other three systems, the ranking of UNIQUENESS does not play a role, at least with 2-place verbs. The fact that both rankings of *[+hr] and *[+lr] relatively to each other occur falsifies Woolford's (1999) claim that there is a universally fixed ranking of markedness constraints (*DAT/*ERG » *ACC » *NOM).14 Note that the constraint rankings in (18) subsume a set of rankings that are all compatible with the respective linking system. For instance, a language comes out as ACC language as long as MAX(+hr) dominates *[+hr] and *[+lr] dominates MAX(+lr).

4.2

Ditransitive verbs

Given the four systems in (17), one may ask how the medial arguments of 3- and 4place verbs are linked. This section deals only with languages that allow 3 and more structural arguments in a verb and discusses whether the medial argument should be linked by structural DAT, ACC, ERG or ΝΟΜ. It will become evident that the UNIQUENESS constraint plays an important role in the organization of linker inventories. The linking patterns attested so far are listed in (19).15 The typology of attested case patterns has been taken partly from Kiparsky (1998); I have listed them according to their decreasing frequency. I do not claim that the corresponding agreement patterns are exhaustively considered. Languages with three or more agreement markers are much more seldom than languages in which three or more structural arguments are marked by case (see Gensler 1998 for a study of languages with three agreement markers on the verb).

14 In LDG's terminology, Woolford claims that *[+Ir] » *[+hr] is universally valid, which would clearly favor ACC systems and could not account for some linking patterns in ERG languages (e.g., (25a)). 15 ERGr is restricted to the highest argument, ACCr to the lowest argument.

Linker inventories, linking splits and lexical economy

227

( 19) Attested linking patterns 16 ditransitive verb: λζ [+hr,-lr] Xy[+hr,+lr] λχ[-1ΐΓ,+ΐΓ] intransitive verb: λχ [-hr,-lr] Case

Agreement

ditransitive (z-y-x)

intransitive

DAT/ACC/NOM

German

Choctaw

ACC-DAT-NOM

NOM

DAT/ERG/NOM

Basque

Basque

NOM-DAT-ERG

NOM

ACC/NOM

Quechua

Nahuatl

ACC-ACC-NOM

NOM

ERGR / ACC/NOM

Wangkumara

ACC-ACC-ERG

NOM

ERGR /NOM

Pitjantjajara

NOM-NOM-ERG

NOM

ERG/NOM

Kabardian

DAT/ERG/ACC/NOM

Thangu?

? Chinook

? Yukulta

(X)

NOM-ERG-ERG

NOM

ACC-DAT-ERG

NOM

Note that the Kabardian data need to be verified because it is unclear whether ERG is a structural case here. In Colarusso (1992), this case is characterized as an oblique case. Up to now, no language with a fully structural DAT/ERG/ACC/NOM case system has been attested, which indicates that linguistic economy strongly disfavors such morphologically complex linking systems. Thangu (Schebeck 1976) would be a potential candidate like some other Australian languages; however, the role of DAT as a structural linker needs to be verified. A four-way linking system can be found in Yukulta (Keen 1983) for agreement markers of 1st and 2nd person. Starting with ACC systems, in which *[+lr] and MAX(+hr) are ranked higher than other MAX and markedness constraints, one must consider three subsystems: the medial argument is linked with ACC, NOM or DAT. The following table shows the constraint violations yielded by these systems. In order to distinguish these systems from one without morphological linkers (NOM system), I have included its evaluation in the last line of table (20). (20) Constraint violations with 3-place verbs in ACC systems Input: Xz[+hr,-lr] λy[+hr,+lr] λχ[-1ΐΓ,+ΐΓ] z-y-x a.

16

MAX(+hr)^ MAX(+lr)

ACC-ACC-NOM

b.

ACC-NOM-NOM

C.

ACC-DAT-NOM

d.

NOM-NOM-NOM

#

**

*[+hr]

**

**

**

*

*

**

**

*[+lr]

UNIQUENESS * *

* **

See Joppen-Hellwig (1999) for Basque, Davies (1986) for Choctaw, Silverstein (1976) for Chinook, Bowe (1990) for Pitjantjatjara, Wunderlich & Lakämper (1999) for Quechua, Launey (1979) for Nahuatl.

228

Barbara Stiebeis

The following constraint rankings can be derived from (20). (21) a.

ACC/NOM:

{*[+lr] » [ΜΑΧ(+1Γ), UNIQUENESS]} & {MAX(+hr) » * [ + h r ] }

b.

ACC r /NOM:

{ * [ + l r ] » UNIQUENESS » * [ + h r ] » MAX(+hr)}

&

{ * [ + l r ] » MAX(+lr)} c.

DAT/ACC/NOM:

{UNIQUENESS » * [ l r ] » M A X ( + l r ) }

&

{MAX(+hr) » * [ + h r ] } d.

NOM:

{ * [ + h r ] » [MAX(+hr), UNIQUENESS]}

&

J*[+lr] » [MAX(+lr), UNIQUENESS]}

These rankings show that without modification of the ranking of ACC systems in (18a), the medial argument should be linked with ACC, as shown in (20a)/(21a); in this case, the UNIQUENESS constraint must be ranked below *[+lr]. DAT is only enforced if the 17 UNIQUENESS constraint is ranked above *[+lr], as in (21c). Note that the ranking *[+lr] » MAX(+lr) must be preserved in (21c) because otherwise, an ACC-DAT-ERG pattern would be admitted. The system (21b) with a restricted accusative (ACCR: [+hr,-lr]) is not attested yet, which can be explained by the fact that in such a system, ACC is enforced by the high ranking of the UNIQUENESS constraint and not by MAX(+hr), the common trigger for ACC, which is absolutely irrelevant in this system. 18 Since both markedness constraints are ranked highest in a NOM language, it could never have a DAT/NOM linker inventory, unless MAX(+hr,+lr) dominates *[+hr] and [+lr], which are co-ranked and dominate MAX(+hr) and MAX(+lr). With respect to constraint violations, ERG systems can be regarded as a mirror image of ACC systems, as the following table shows.

17

In some languages, DAT may be enforced by MAX(+hr,+lr), which is evidenced by the passive data of ditransitive verbs: if DAT cannot be absorbed in passive, MAX(+hr,+lr) must be a highranked constraint, as illustrated in (10). As Wunderlich (2000b) shows, passive data may yield evidence for a high ranking of both MAX(+hr,+lr) and UNIQUENESS. Generally, the linking patterns resulting from voice operations show that the constraint rankings proposed here must be further differentiated.

18

Note that with the constraint ranking in (21b), a language must develop a restricted ACC. A linking pattern NOM-ACC-NOM, which would also be possible with the given ranking, is excluded by the linking mechanism: a non-restricted ACC must be used for both [+hr] arguments. An ACC that is restricted to medial arguments of ditransitive verbs is per definitionem a DAT. Hindi (Mohanan 1994) shows a NOM-ACC-NOM pattern in the imperfective aspect because ACC is restricted to definite or animate objects, and a specific instance of UNIQUENESS penalizes doubling of ACC (Wunderlich 2000b).

229

Linker inventories, linking splits and lexical economy (22) Constraint violations with 3-place verbs in ERG systems z-y-x a. b.

MAX(+hr)

NOM-NOM-ERG NOM-DAT-ERG

d.

NOM-NOM-NOM

*[+lr]

T+hrl

*

* *

NOM-ERG-ERG

C.

MAX(+lr)

* *

*

UNIQUENESS

*

*

* *

*

* *

*

* *

* *

* *

In order to derive the respective linking systems, the following constraint rankings must be assumed: ( 2 3 ) a.

ERG r /NOM:

{ * [ + h r ] » UNIQUENESS » * [ + l r ] » MAX(+lr)}

&

{ * [ + h r ] » MAX(+hr)} b.

ERG/NOM:

C.

DAT/ERG/NOM:

{ * [ + h r ] » [MAX(+hr), UNIQUENESS]} & {MAX(+lr) » * [ + l r ] } {UNIQUENESS » *[hr] » MAX(+hr)}

&

{MAX(+lr) » * [ + l r ] }

As in ACC systems, DAT must be enforced by UNIQUENESS on top of *[+hr], as shown in (23c). The unmodified constraint ranking of ERG systems in (23b) yields a linking pattern in which the medial argument is linked by ERG. A system with a restricted ERG requires a high-ranked UNIQUENESS constraint as in (23b) and is not triggered by MAX(+lr), which does not play a crucial role in this system. The picture changes with tripartite systems, in which the two MAX constraints are ranked above the corresponding markedness constraints. Here, an additional markedness constraint, *[+hr,+lr], a local conjunction of *[+hr] and *[+lr], comes into play. Consider the following table. (24) Constraint violations with 3-place verbs in z-y-x a.

MAX(+hr) MAX(+lr) :*r+hrl * [ + l r l

ACC-ERG-ERG

C.

ACC-DAT-ERG

d.

ACC-NOM-ERG

**

*

!

*

* *

;

* *

* *

;

*

*

*

ACC-ACC-ERG

b.

systems

ERG/ACC

*

*

*

* [ + h r , + l r ] ¡UNIQUE

*

:

*

;

*

! ¡

The respective linking patterns result from the following constraint rankings: ( 2 5 ) a.

ERG r /ACC/NOM:

{ * [ + h r , + l r ] » MAX(+hr) » ( * [ + h r ] , MAX(+lr)) » * [ + l r ] } & { * [ + h r , + l r ] » UNIQUENESS}

b.

ERG/ACC r /NOM:

{ * [ + h r , + l r ] » MAX(+lr) » ( * [ + l r ] , MAX(+hr)) » * [ + h r ] } & { * [ + h r , + l r ] » UNIQUENESS}

c. d.

DAT/ERG/ACC/NOM:

{MAX(+lr) » * [ + l r ] } &

ERGr / ACCr /NOM :

{ * [ + h r , + l r ] , UNIQUENESS} » ...

{MAX(+hr) »

*[+hr]}

Unlike in ACC and ERG systems, DAT does not need to be enforced by the UNIQUENESS constraint, but is already required by the highest-ranked MAX constraints, as (25c) shows. Hence, the ranking of UNIQUENESS does not a play a role in that system - at least for 3-place verbs. DAT must be excluded by the complex markedness constraint

230

Barbara Stiebeis

*[+hr,+lr] in languages with a restricted ERG (see (25a/d)) or a restricted ACC (see (25b/d)); *[+hr,+lr] must be ranked above UNIQUENESS and above the ranking that enforces ERG or ACC. The non-attested case, in which the medial argument is linked by NOM, results from a ranking in which *[+hr,+lr] and UNIQUENESS are ranked above the other constraints (see (25d)). 19 In the linking systems considered so far, UNIQUENESS plays a crucial role in five of them (see (21b/c), (23a/c), and (25d)). Like the MAX constraints, UNIQUENESS determines the number of structural linkers in a language depending on its ranking relative to the markedness constraints. However, unlike the MAX constraints, UNIQUENESS does not determine which linkers occur in the respective language. The table of attested case patterns in (19) shows that languages tolerate UNIQUENESS violations in favor of a smaller linker inventory. Cross-linguistically, UNIQUENESS violations with [+hr] arguments are much more frequent than UNIQUENESS violations with [+lr] arguments, a case for which only Kabardian has been cited. The reason that UNIQUENESS(+lr) is often high-ranked results from the strong preference for a designated argument, which is singled out from all other arguments and mostly coincides with the semantically highest argument. Among the 15 logically possible linking patterns for ditransitive verbs, 4 have not been considered so far. These systems are schematized in (26), where S denotes the subject of an intransitive verb (subsuming A¡ and P¡) and A, P, and R denote agent, patient, and recipient of a ditransitive verb. (26)

a.

R

S

A

P

b.

S

Α

Ρ

R

c.

S

A

P

R

d.

S

R

Α

Ρ

The system in (26a) corresponds to a DAT/NOM system. The other three systems fail in not being characterizable with the common linker inventory, as shown in (27m-o): these inventories contain one linker with an uncommon characterization ([-hr,-lr]) and thus fare badly in lexicon optimization; in addition, the system in (26d/27o) needs a disjunctive specification in the lexical entry of the linker, which is an indication of an unsystematic syncretism. The evaluation of the various linker inventories in (27) shows that ACC, ERG and tripartite systems perform best if they are extended by DAT, so that UNIQUENESS violations are avoided. The distribution of attested linker inventories suggests that *[-] must be split up in two subconstraints: *[-lr] 'avoid [-lr] specifications in linker entries' and *[-hr] 'avoid [-hr] specifications in linker entries'. *[-lr] dominates *[-hr], which 19 The ranking in (25d) would also allow a system with the linking pattern NOM-ACC-ERG. This pattern cannot be derived by the linking mechanism because ACC is per definitionem not restricted to a medial argument. Yimas (Foley 1991), however, displays such a pattern with lst/2nd person recipients, which results from the specific linker inventory: a DAT/ERG/NOM inventory for 3rd person and an ERG/ACC/NOM inventory for lst/2nd person. In other languages, the NOM-ACC-ERG pattern (e.g., in Hindi perfective aspect) is enforced by additional constraints (see Wunderlich 2000b).

Linker inventories, linking splits and lexical economy

231

might be explained by the fact that [-hr] singles out the lowest argument, whereas [—hr] singles out the highest argument. Since languages show a strong tendency to single out the highest argument, a [-lr] specification should be avoided. All systems attested so far do not violate *[-lr]. (27) Optimal lexical entries of linking systems for ditransitive verbs Linkers

*[V]

*[-lr]

UNIQUE-

*[-hr] *[+hr] *[+lr] Rank

NESS a. b.

DAT/ACC/NOM

**

*

1

DAT/ERG/NOM

*

**

1

**

**

2 3 3 4 5 6 7 8 9 9 10 11 12

C.

DAT/ERG/ACC/NOM

d.

ACC/NOM

*

e.

ERG/NOM

*

f.

DAT/NOM

g.

NOM

h.

ERGR /NOM

i.

ERGR/ACC/NOM

* *

*

*

*

** *

*

*

*

* *

*

j.

ACCr/NOM

*

k.

ERG/ACC R /NOM

*

ERG R /ACC R /NOM

*

*

*

*

m.

DAT/NOM/[-hr,-lr]

*

*

*

*

*

η.

NOM/[-hr,-lr]

*

**

•

*

*

*

*

*

I.

o.

N O M / ( [ - h r , - l r ] ν [+hr,+lr])

*

*

*

*

*

*

The constraints used in (27) do not suffice to account for the lack of DAT/NOM inventories, which have rank 4.20 Such an inventory has an unusual markedness increase: the inventory has a DAT but no ERG or ACC, which already instantiate either [+lr] or [+hr], A high-ranked constraint that penalizes such a markedness increase might account for the lack of such an inventory; no other inventory would violate this constraint. The rare occurrence of systems with a maximal inventory (DAT/ERG/ACC/NOM) might be attributed to the highest number of markedness violations.

20

One might argue that languages such as Spanish have a DAT/NOM inventory. Such an assumption, however, implies that the feature [+lr] is either lexically designated in 2-place verbs with definite human objects, or that DEP(+lr) '[+lr] in the output should have a correspondent in the input' can be violated in Spanish under certain conditions. I assume that Spanish has an ACC that is restricted to definite human objects.

232

Barbara Stiebeis

5. Linking splits Languages may exhibit alternating linking patterns, for instance, an ACC/NOM inventory for pronominal elements and an ERG/NOM inventory for nouns, or an ACC pattern in the imperfect and an ERG pattern in the perfect (see, e.g., Silverstein 1976, Hopper & Thompson 1980, DeLancey 1981, Givon 1984, Dixon 1994). These linking splits pose a challenge to those theories of argument linking that do not consider all alternating patterns as instances of structural linking; they have to assume that there is an alternation between structural and lexically marked (inherent) linking. I will show in the following that LDG can treat most types of linking splits as instances of structural linking; a central assumption hereby is that DAT and ERG are (potential) structural cases. According to Dixon, linking splits may be triggered by semantic properties of the NP/DP to be linked, by tense/aspect/mood of the verb, by the polarity of the sentence or by the distinction of main vs. subclause. In this paper, I will only deal with linker inventory splits (see Stiebels in prep, for aspectually conditioned splits). I will follow Dixon's (1994) and Kiparsky's (1999) assumption that linking splits reflect different preferences for ACC or ERG systems. Especially, I will reconstruct Dixon's informal account of linker inventory splits in a CT-based approach. Moreover, I will show in section 5.2 that linker inventory splits do not necessarily affect economy. Whereas the linking splits themselves are instances of semantically conditioned structural linking, their origin often lies in conceptual preferences and markedness principles, as I will show below. Situations are generally conceptualized and, hence, lexicalized from the perspective of the more agentive and discourse-prominent entity; therefore, languages do not have simple verbs with the meaning 'be hit by someone'. Dowty's (1991) conception of Proto-Agent and Proto-Patient is an attempt to account for the asymmetry of arguments in verbs: in polyadic verbs, the highest argument prototypically exerts control and refers to the causer if the verb expresses a cause-effect relation, whereas the lowest argument is prototypically affected, undergoes a change of state or is controlled.

5.1

Linker inventory splits

Depending on their sortal and referential properties, DPs differ in how likely they denote causers or effects, i.e., occur as highest or lowest argument. According to Silverstein (1976) and Dixon, NPs/DPs on the left of the scale in (28) are more likely to show agent or control properties, i.e., constitute the semantically highest argument ([-hr]) in a polyadic verb, whereas NPs/ DPs on the right are more likely to show patient properties, i.e., constitute the semantically lowest argument ([-Ir]) in a polyadic verb. However, the cut-off point varies from language to language, as Silverstein has shown.

Linker inventories, linking splits and lexical economy

233

(28) NP/DP scale 1 > 2 > 3/DEM > proper nouns > human > animate > inanimate nouns

< [-lr];[+hr] [-hr];[+lr] N-prominent D-prominent In the following, I will call elements to the left of the hierarchy D-prominent elements and elements to the right of the hierarchy N-prominent elements. In terms of referentiality, D-prominent elements are strongly referential, whereas N-prominent elements tend to be weakly referential. The scale in (28) is thus associated with the referential scale [+def(inite)] > [+spec(ific)] > [-spec]. From the perspective of linguistic economy, it is reasonable to assume that overt morphological marking is confined to informative features, i.e., to unexpected properties; predictable information is default information and should remain unmarked. This corresponds roughly to Dixon's motivation of splits. Given the hierarchy in (28), a 1st person pronoun, for instance, will prototypically refer to an agent or controller and, hence, should receive the default linker (NOM) as [-hr] argument, disfavoring ERG, whereas it should be marked as [+hr] argument. Therefore, one expects an ACC pattern in pronouns. Likewise, an inanimate noun is unmarked as [-lr] argument, thus prefers NOM to ACC, but should receive a marked linker as [+lr] argument. Therefore, one expects an ERG pattern. The following lexical entries for 1st person pronouns and inanimate nouns are thus preferable, yielding a split on the NP/DP level: 21 (29) a. b.

[+l,+hr] [-anim,+lr]

ACC ERG

Consequently, the global markedness constraints introduced so far must be further differentiated. *[+lr] is more closely correlated with elements on the left of the hierarchy in (28) (e.g., *[+lr]/+l), whereas *[+hr] is more closely correlated with elements on the right of the hierarchy in (28) (e.g., *[+hr]/-anim]). I will abrevíate the contexts as D and Ν and propose the markedness constraints *[+lr]/D and *[+hr]/N, which I will derive from the scale in (28) by means of Harmonic Aligment (Prince & Smolensky 1993). Note that the range of D and Ν is determined language-specifically and that D and Ν may be complementary to each other (D = ~N, Ν = ~D), creating a binary contrast as in (30a), or non-complementary, creating a ternary contrast with a category that is neither Ν nor D (~N/~D) as (30b) or both Ν and D (N/D) as in (30c).22

21

22

The Turkish object split (Kornfilt 1996), in which only specific objects are marked by ACC, whereas unspecific objects are linked with the default linker (NOM), confirms the prediction that weakly referential elements show a strong disposition for the [+hr] argument role; hence ACC should be avoided with elements of weak referentiality. represents negation.

234

Barbara

Stiebeis

( 3 0 ) Possible scale segmentations Ν

~N

~D

D

Ν

~N ~D

Ν

D ~N

~D

D

According to Dixon ( 1 9 9 4 : 1 0 9 ) the following types o f inventory splits are attested: ( 3 1 ) Types o f inventory splits D a.

~D/~N

Ν

ACC/NOM

NOM

English

b.

NOM

ERG/NOM

Burushaski

C.

ERG/ACC/NOM

ERG/NOM

Waga-Waga

d.

ACC/NOM

ERG/ACC/NOM not attested

e.

ACC/NOM

f.

ACC/NOM

ERG/ACC/NOM ERG/NOM

ERG/ACC

Yidin y

g.

ACC/NOM

NOM

not attested

ERG/NOM

Dyirbal

Before I will introduce my proposal, I would like to comment on Kiparsky's ( 1 9 9 9 ) approach. 2 3 Kiparsky has made a proposal for inventory splits, in which he introduced contextualized MAX constraints. Kiparsky's constraints are the following: MAX(+hr)/D requires that [+hr] is made visible in D-elements (pronouns) through a corresponding linker, whereas MAX(+lr)/N requires that [+lr] is made visible in N-elements through a

23

Aissen (1999) has also made an interesting proposal for inventory splits. The factorial typology she has proposed is in so far convincing in that she derives her constraints via Harmonic Alignment of independently motivated scales (among them Silverstein's scale). However, as she herself admits, her proposal makes the wrong prediction for ERG languages that the subject of an intransitive verb should be linked with ERG. A fundamental problem of her approach is the fact that her typology is based on grammatical functions: on the one hand, the notion of subject is not easily applicable to all languages; on the other hand, grammatical functions are defined independent of morphological linkers so that no concrete predictions concerning the actual linker can be provided. Aissen makes no claim as to which actual linker should appear in a particular language. Finally, her account for the presence or absence of morphological linkers is unnecessarily complex (see Stiebels in prep, and Stiebeis & Wunderlich 2000).

Linker inventories,

linking splits and lexical

235

economy

corresponding linker. 24 The language-specific cut-off point of the linking split is accounted for by different categorizations of nominal elements as D or Ν elements. Splits can only surface if the contextualized MAX constraint is ranked higher than the corresponding markedness constraint and the global MAX constraint, e.g., MAX(+lr)/N » *[+lr] » MAX(+lr). Only this constraint ranking ensures that some instances of [+lr] are obligatory (in N-prominent elements), whereas other instances of [+lr] are blocked (in those elements that are not N-prominent). If both the global and the contextualized MAX constraint were ranked above or below the corresponding markedness constraint, no split could arise, as (32) illustrates: if MAX(+lr) is ranked highest, a [+lr] argument must be linked by ERG or DAT, independent of its category. If *[+lr] dominates both MAX(+lr) and MAX(+lr)/N, a [+lr] argument must be linked by NOM, again independent of its category. (32) Constraint violations in the linking of a [+lr] argument Input N: R+lrl

Linker MAX(+lr)/N NOM ERG DAT

* R+lrl MAX(+lr)

*

* * * *

~N: r+lrl NOM ERG DAT

* *

On the basis of the constraint rankings for ERG, ACC, NOM and ERG/ACC languages given in (18), the addition of the contextualized MAX constraints can effect a split in ERG, ACC and NOM languages; ERG/ACC languages have already high-ranked global MAX constraints. As (33) illustrates, NOM languages may show three split types, depending on the contextualized MAX constraint on top of the general constraint ranking. (33) a. b. c. d. e.

ACC language: ERG language: NOM language 1 : NOM language 2: NOM language 3:

{MAX(+lr)/N {MAX(+hr)/D {MAX(+lr)/N {MAX(+hr)/D {MAX(+hr)/D {MAX(+lr)/N

» » » » » »

*[+lr] » MAX(+lr)} & . *[+hr] » MAX(+hr)}

&

*[+lr] » MAX(+lr)} & . *[+hr] » MAX(+hr)}

&

*[+hr] » MAX(+hr)} & *[+lr] » MAX(+lr)} & .

For Kiparsky, Dyirbal would be a NOM language of type 3; both contextualized MAX constraints are active, i.e., effect a split. (34a) illustrates the evaluation of the linking patterns of a 2-place verb in Dyirbal with a D-subject and an N-object. The double line

24

I have reformulated Kiparsky's constraints according to LDG's interpretation of the features [hr] and [lr]. Kiparsky has dubbed these constraints MAX(-HR)/D and MAX(-LR)/N. Moreover, I have replaced his markedness constraint *[F], which penalizes both positive and negative feature specifications, by *[+hr] and *[+lr]. In addition, I have extended his proposal to all linking systems I have discussed so far.

Barbara Stiebeis

236

separates the two independent partial constraint rankings. Since both contextualized MAX constraints are irrelevant, Dyirbal exhibits a NOM-NOM pattern in this case. (34) ERG/ACC split in Dyirbal in the sense of Kiparsky a. Input: Xy[+hr,-lr]/N Xx[-hr,+lr]/D y-x β"

M A X ( + h r ) / D *r+hrl

ACC-NOM NOM-NOM

MAX(+hr) M A X ( + l r ) / N *r+lrl

*

*

*!

*!

ACC-ERG

*!

Input: λγ[+1ΐΓ ,-lr]/D Xx[-hr,+lr]/N y-x ACC-NOM

Β*

* *

NOM-ERG

b.

MAX(+lr)

*!

M A X ( + h r ) / D T+HRL MAX(+hr) |ΜΑΧ(+1Γ)/Ν *R+LRL MAX(+lr) * * *!

NOM-NOM

*!

*

NOM-ERG

*!

*

*

*

ACC-ERG

*

*! *

With N-subjects and D-objects, however, the two contextualized MAX constraints come into play, thus enforcing an ACC-ERG pattern as illustrated in (34b). Generally, Kiparsky would predict the linking splits presented in (35). The second column lists the inventories in the split context, including the category that instantiates the split. The third column lists the general inventory of the languages. (35) Possible linking splits in Kiparsky's approach a. b. c. d. e. f. g.

ACC language ERG language NOM language NOM language NOM language NOM language NOM language

1 2 3a 3b 3c

Inventory of split context

General inventory

Ν : ERG/ACC/NOM

ACC/NOM

D : ERG/ACC/NOM

ERG/NOM

Ν : ERG/NOM

NOM

D : ACC/NOM

NOM

Ν : ERG/NOM, D : ACC/NOM Ν : ERG/NOM, D : ACC/NOM

NOM (~N,~D)

Ν : ERG/NOM, D : ACC/NOM

ERG/ACC/NOM ( N / D )

Depending on the scale segmentation, i.e., whether Ν and D categories overlap or not, three subcases must be regarded for NOM languages of type 3. If Ν and D are complementary (type 3a), the underlying NOM pattern does not surface in any category. If Ν and D are not complementary and do not overlap, the category that is neither Ν nor D shows the NOM pattern. If Ν and D overlap, the category that is both Ν and D has a tripartite inventory. Recall that type 3c is attested (e.g., Yidin y ), whereas 3b is not. Generally, Kiparsky assumes that the underlying inventory (the 'elsewhere case') is impoverished, while the categories that instantiate a split have an elaborated inventory. This assumption is problematic from the view of markedness theory. Conceptually, an unmarked context (to be identified with the elsewhere case) should be able to show more differentiation than a marked one: singular contexts, for instance, show more case or gender distinctions than plural contexts. Thus, Kiparsky's approach, claiming that

Linker inventories, linking splits and lexical economy

237

the elsewhere case is less expressive than the marked case, is counter-intuitive in the spirit of markedness theory. Moreover, the relevant splits seem to be triggered by surface features (pronominal vs. nominal arguments) rather than by faithfulness conditions, which evaluate to which extent the underlying categorial structure is realized. Concerning faithfulness, one expects that all surface categories behave in the same way. A further empirical problem is the fact that the non-attested ternary split 3b is predicted to be more natural than the attested ternary split 3c because there is a category whose inventory corresponds to the underlying pattern, which is not the case in the split 3c. As alternative to Kiparsky's approach I would like to propose an analysis that (i) considers the markedness facts more seriously, (ii) formalizes Dixon's idea that the marked linker is used for the unexpected argument role, and (iii) is additionally motivated in terms of Harmonic Alignment (Prince & Smolensky 1993), as proposed by Aissen (1999). 25 The two linking features [hr] and [lr] are not fully symmetric. Recall that *[-lr] dominates *[-hr] (see (27)) because languages strongly tend to single out the highest argument ([-hr]), not the lowest argument ([-lr]). [+hr] thus separates all internal arguments from the highest argument and is therefore preferred to [+lr], which separates the nonlowest arguments from the lowest argument. These facts suggest the preference hierarchy [+hr] > [+lr] in LDG. The Harmonic Alignment of this hierarchy with Silverstein's hierarchy (simplified as D > N) yields the hierarchies in (36) and the constraint rankings in (37). (36) a. b.

[+hr]/D > [+hr]/N [+lr]/N > [+lr]/D

(37) a. b.

*[+hr]/N » *[+hr]/D *[+lr]/D » [+lr]/N

According to (37), *[+hr]/N and *[+lr]/D are the highest-ranked markedness constraints. Note that *[+lr]/D subsumes all markedness constraints that dominate MAX(+lr) (e.g., *[+lr]/l » *[+lr]/2 » *[+lr]/3 » ...), whereas *[+lr]/N subsumes all markedness constraints that are dominated by MAX(+lr). This holds analogously for *[+hr]/N and *[+hr]/D. I assume that in split languages, the lower-ranked constraints *[+hr]/D and *[+lr]/N are generally co-ranked with the corresponding global markedness constraint. According to the constraint rankings for ACC, ERG and ERG/ACC languages in (18), the following additions of contextualized markedness constraints to the relevant partial rankings enforce a split:

25

Harmonic Alignment relates a binary scale (X > Y) and a multi-dimensional scale (a > b > ... > z) to the two Harmony scales X/a > X/b > ... > X/z and Y/z > ... > Y/b > Y/a. These two Harmony scales can be interpreted as markedness hierarchies: *X/z » ... » *X/b » *X/a and *Y/a » *Y/b » ... » *Y/z (see also Aissen 1999).

238

Barbara

(38) a. b. c. d. e.

ACC language: ERG language: ERG/ACC language 1 : ERG/ACC language 2: ERG/ACC language 3:

{*[+hr]/N {*[+lr]/D {*[+hr]/N {*[+lr]/D {*[+lr]/D {*[+hr]/N

Stiebeis

» MAX(+hr) » *[+hr]} & ... » MAX(+lr) » *[+lr]} & ... » MAX(+hr) » *[+hr]} & ... » MAX(+lr) » *[+lr]} & ... » MAX(+lr) » *[+lr]} & » MAX(+hr) » *[+hr]} & ...

The markedness constraint *[+lr]/D could not have any effect in an ACC language, whereas *[+hr]/N could not have any effect in an ERG language because the corresponding global markedness constraints are ranked above the MAX constraints. Likewise, NOM languages, in which both *[+hr] and *[+lr] dominate the MAX constraints, cannot develop an inventory split. In ERG/ACC languages, five types of split can arise depending on the presence or absence of the contextualized markedness constraints and the categorial differentiation as Ν or D category. The rankings in (38) account for the following linking splits: (39) Possible linking splits in markedness approach Inventory of split context

General inventory

a. b.

ACC language ERG language

N: NOM

D: NOM

ACC/NÖM ERG/NOM

c.

ERG/ACC l a n g u a g e 1

Ν: ERG/NOM

ERG/ACC/NOM

d.

ERG/ACC l a n g u a g e 2

D : ACC/NOM

ERG/ACC/NOM

e.

ERG/ACC l a n g u a g e 3a Ν : ERG/NOM, D : ACC/NOM

f.

ERG/ACC l a n g u a g e 3 b N : ERG/NOM, D : ACC/NOM ERG/ACC/NOM ( ~ N , ~ D )

g.

ERG/ACC l a n g u a g e 3c N : ERG/NOM, D : ACC/NOM NOM (N/D)

As desired, the unmarked context has an elaborated inventory, whereas the marked context has a reduced inventory. Moreover, the ternary split in Yidin y , in which categories that are neither Ν nor D show a tripartite pattern follows naturally from the fact that Yidin y is analyzed as an ERG/ACC language of type 3b. The non-attested split in (39g) is predicted to be unusual because no category would instantiate the underlying inventory. Despite the fact that the two approaches identify opposite linker distributions for the marked vs. unmarked context, they agree in all of the predicted splits. The central difference lies in the underlying inventories. Whereas Kiparsky assumes that languages are economic regarding linker inventories and only extend their inventories in certain categories, the markedness approach takes linker inventories to be expressive underlyingly: languages tend to make the linking features [hr] and [lr] visible but reduce the inventory of categories that show a strong disposition for a certain argument role. The empirical differences between the two approaches become especially evident in 3-place verbs, which I want to illustrate for languages with an ACC/NOM inventory. Such a language displays an ACC-ACC-NOM pattern in the unmarked context. In Kiparsky's approach, MAX(+lr)/N could effect a split, yielding an ACC-DAT-ERG pattern for Ν categories, as (40a) illustrates. In the markedness approach, *[+hr]/N could effect a split, yielding a NOM-NOM-NOM pattern for Ν categories, as (40b) shows.

Linker inventories, linking splits and lexical economy

239

(40) Linking patterns of 3-place verbs in ACC/NOM language Input: Xz[+hr,-lr]/N λγ[+1ΐΓ,+ΐΓ]/Ν Xx[-hr,+lr]/N a. Kiparsky's approach MAX(+lr)/N *[+lrl MAX(+lr) ||MAX(+hr) *r+hrl **| ** I ** ACC-ACC-NOM **1 ** **t NOM-NOM-NOM z-y-x

b.

ACC-DAT-ERG

**

ACC-ERG-ERG

**

*

*!

Markedness approach z-y-x

w

**

ACC-ACC-NOM NOM-NOM-NOM

*r+hrl/N **|

MAX(+hr)

n+hrl ln+lrl ** ι

**

ACC-DAT-ERG

**|

ACC-ERG-ERG

*!

** **

** *

MAX(+lr)

*

I

**| 1**1

Kiparsky would thus predict that the ACC/NOM inventory is extended by ERG and DAT in the split context. The markedness approach predicts a reduction to ΝΟΜ. While the two approaches predict some equivalent split patterns for 3-place verbs, again with a difference in the proposed underlying system, there are splits that do not have an equivalent in the two approaches: for instance, while Kiparsky would predict that a language with an underlying ERGr /ACC/NOM inventory could exhibit an ACC-DAT-ERG pattern with Ν categories, the markedness approach could not account for such a split and thus predicts that it does not occur. Consequently, 3-place verbs are the major test case for the decision as to which approach is preferable on empirical grounds. 26 The markedness approach also makes interesting predictions for splits that are related to referential properties. Recall that weakly referential DPs tend to be linked by NOM. The ACC/NOM alternation in Turkish objects (Kornfilt 1996), in which ACC is restricted to specific objects, can be accounted for by a contextualized markedness constraint such as *[+hr]/-spec. This constraint also explains clitic doubling effects (Suñer 1988) that are related to the specificity of the object: *[+hr]/-spec excludes the doubling with unspecific objects. Object agreement that is restricted to animate or definite objects results likewise from markedness constraints such as *[+hr]/-anim or *[+hr]/-def. Possessor splits and splits conditioned by tense, aspect or mood can also be accounted for by means of such contextualized markedness constraints (Stiebels in prep.).

26

Dative splits (e.g. in Yimas, where only 3rd person affixes exhibit a DAT linker) can be derived from the preference scale [+hr] > [+hr,+lr]. Harmonic Alignment yields the markedness hierarchy *[+hr,+lr]/D » *[+hr,+lr]/N (see Wunderlich 2000c for Yimas).

240

5.2

Barbara

Stiebeis

Linker inventory splits and economy

In this section, I will deal with the question of how expensive an inventory split is in terms of storage. Generally, split systems do not differ from non-split systems in terms of economy if they use the same number of linkers, i.e., if the split is encoded by the linkers already present in a non-split context. This is typically the case in intransitive splits, which are very often encoded by an alternation in the linkers used for the subject and object of a transitive verb. Linker inventory splits represent another type of economic split. Since many languages have case-marked forms of pronouns that cannot be derived transparently (e.g., German ich lSG.NOM, mich lSG.ACC and mir lSG.DAT) and, therefore, have to be listed, a split in the linker inventories of pronouns and nouns does not increase the number of lexical entries in this domain, in comparison with a non-split system. Likewise, case-agreement splits, in which case linkers display an inventory different from that of agreement linkers, are unproblematic in terms of lexical economy. A typical example of a case-agreement split is given in (41). (41) Kalkatungu (Blake 1994:52) a=ni ingka? a. nyini you.NOM PURP=2SG.N go 'are you going?' a=ni nuwa? b. nyin-ti you-ERG PURP=2SG.N see 'do you want to see her?' a=kin nuwa c. nyini you.NOM PURP=2SG.A see 'she wants to see you' Kalkatungu has an ERG-based case system and an ACC-based agreement system. The reverse pattern (ERG-based agreement and ACC-based case system) is not attested (Dixon 1994), a fact that is predicted by the markedness approach (and Kiparsky's approach), given that agreement affixes are pronominal in nature and can be analyzed as D-elements. Hence, case-agreement splits are analyzed as a subcase of inventory splits. Given the lexical specification of the linkers in (42a), no additional mechanism is needed in LDG to derive the split and to account for the actual linking patterns shown in (42b): nyini [+2] nyinti [+2,+lr] λχ Xs GO(x)(s) -hr -lr =ni nyini

=ni =kin λΥ +hr -lr =kin nyini

[+2] [+2,+hr] λχ λβ SEE(x,y)(s) -hr +lr =ni nyinti

Since case markers and agreement affixes need independent lexical entries in any case, both a homogeneous and a split case-agreement system are equal in their expense of lexical storage. Consequently, such a split does not affect economy. Syntactic approaches in which the various structural cases are associated with distinct syntactic positions/projections, i.e., approaches within Government and Binding

Linker inventories, linking splits and lexical economy

241

or the Minimalist Program, have problems to account for inventory splits: why should a pronominal subject marked by NOM target a different syntactic position than a nominal subject marked by ERG, especially, if the agreement pattern does not differ in the two cases? Two solutions seem possible: (i) The proponents of a syntactic approach could appeal to abstract case thus ignoring the morphological linkers, so that they actually fail in explaining the linking patterns, (ii) It could be assumed that the pronominal/nominal features to be checked trigger different syntactic movements, thus yielding different linking patterns. The most ambitious approach in this respect is presented by Ritter & Rosen (to appear), who try to account for the full range of inventory and aspectually conditioned splits. Although Ritter and Rosen's approach seems to be empirically adequate, it only works under the assumption that in linker alternations, there is a structural pattern, resulting from movement, and a lexical pattern, in which the linker is inherently determined (e.g., ERG). Moreover, their approach describes morphosyntactic-semantic correlations, but it does not motivate them: why should ACC be restricted to definite objects or delimited events?27

6. Conclusions I have suggested in this paper that functional considerations of economy vs. expressivity of a language can become substantial if they are based on articulate representations of lexical items and their argument structure, as provided by LDG, as well as on a congruent conception of morpheme-based morphology, in which the linker inventories are adequately specified. The framework of CT allows us to consider the theta-structure of verbs (and nouns) as the input, and the surface realization of arguments as the output. One important aspect of expressivity can, then, be formulated by means of MAX constraints, which check whether the hierarchy of argument roles is adequately expressed, and a likewise important aspect of economy can be formulated by means of markedness constraints, which check the amount of information on the surface. Another surface constraint is UNIQUENESS, which checks the potential ambiguity in co-occurring linkers, and thus contributes to the evaluation of explicitness. I have shown that by studying the interaction of these constraints, one can evaluate the possible linking systems for intransitive, transitive and ditransitive verbs, and can thus derive a principled typology of these systems. Substantial predictions can be made as to which argument linking systems are expected and which are not. These predictions have turned out to be valid in many instances. Another question I have dealt with concerns linker inventory splits, which many languages exhibit and whose appearance seem to influence language change from one system to another. I have shown that a reconstruction of Dixon's insights into linking splits by means of markedness constraints is superior to a proposal by Kiparsky, who 27

Ritter & Rosen unnecessarily distinguish between structural and default NOM, the latter being a NOM not licensed by AgrSP. In LDG, both instances of NOM would be structural; no distinct mechanisms are needed.

242

Barbara Stiebeis

assumes MAX constraints to effect the split. There is a deeper insight from this study, namely that expressivity plays a minor role in splits found in transitive verbs. These splits are mainly triggered by economy. In contrast, intransitive splits are quite often triggered by expressivity because they make sortal and control properties of the argument referent visible. The number of insights to be gained by this procedure also show that the theory advocated here is on the right track. Not only is the decomposition of argument structures with just two relational features ([hr] and [lr]) an indispensable (and fruitful) premise of this approach, but it is also the particular characterization of ERG and DAT by structural rather than semantic means that allows us to get such a coherent picture. Thus, indirectly, the success of this approach supplies evidence for its crucial assumptions. Besides several possibilities to increase the empirical adequacy of this approach by additional constraints, there remains one major open question, namely why ACC and ERG systems, if they are mirror images to each other in terms of the constraints, do not distribute equally among languages. In both frequency and homogeneity, ACC languages clearly outrank ERG languages. The answer lies outside the domain of this paper: there is a strong tendency in languages to qualify some argument as the most prominent one for the discourse; if there is no specific topic device, the semantically highest argument can best function for this purpose. Therefore, the single argument of an intransitive verb (S) and the highest argument of a polyadic verb (A = [-hr]) often coincide in their discourse function. It is this discourse aspect that favors ACC systems. Syntactically, ERG systems nevertheless turn out to be ACC systems in most cases, which means that the syntactic topic function often overrides morphological facts. The amount of information to be overridden in ACC systems is less than that in ERG systems; this, then, establishes the observed asymmetry. In this paper, however, I was only concerned with the morphological facts.

Appendix A ABS ACC ADE APPL ASP AUX CL DAT DEM DET ERG

'accusative' agreement absolutive accusative adessive applicative aspectual marker auxiliary class marker dative demonstrative determiner ergative

F GEN LOC

M Ν NOM PAST PL PROP PURP SG SUP

feminine genitive locative masculine 'nominative' agreement nominative past tense plural proprietive purposive singular superessive

Linker inventories, linking splits and lexical economy

243

References Aissen, Judith (1999). Markedness and subject choice in Optimality Theory. Natural Language and Linguistic Theory 17 (4), 673-711. Bierwisch, Manfred (1983). Semantische und konzeptuelle Repräsentation lexikalischer Einheiten. In Untersuchungen zur Semantik, Rudolf Ruzicka and Wolfgang Mötsch (eds.), 61-99. Berlin: Akademie-Verlag. Bierwisch, Manfred and Ewald Lang (eds.)(1987). Grammatische und konzeptuelle Aspekte von Dimensionsadjektiven. Berlin: Akademie-Verlag. Blake, Barry J. (1994). Case. Cambridge: Cambridge University Press. Bowe, Heather J. (1990). Categories, constituents and constituent order in Pitjantjatjara. London: Routledge. Chomsky, Noam (1991) Some notes on economy of derivation and representation. In Principles and Parameters in comparative grammar, R. Freidin (ed.), 417-454. Cambridge, Mass.: MIT Press. Chomsky, Noam (1995). The Minimalist Program. Cambridge, Mass.: MIT Press. Colarusso, John (1992). A grammar of the Kabardian language. University of Calgary Press. Davies, William David (1986). Choctaw verb agreement and universal grammar. Dordrecht : Reidel. DeLancey, Scott (1981). An interpretation of split ergativity and related patterns. Language 57 (3), 626-657. Dench, Alan & Nicholas Evans (1988). Multiple case-marking in Australian languages. Australian Journal of Linguistics 8, 1-47. Derbyshire, Desmond C. (1999). Carib. In The Amazonian languages, R. M. W. Dixon and Alexandra Y. Aikhenvald (eds.), 23-64. Cambridge: Cambridge University Press. Dixon, R. M. W. (ed.)(1976). Grammatical categories in Australian languages. Canberra: Australian Institute of Aborginal studies. Dixon, R. M. W. (1994). Ergativity. Cambridge: Cambridge University Press. Dowty, David (1991). Thematic proto-roles and argument selection. Language 67, 547-619. Fabri, Ray (1996). The inverse morphology of Plains Cree. In Yearbook of Morphology 1995, Geert Booij and Jaap van Marie (eds.), 17-41. Dordrecht: Kluwer. Foley, William A. (1991). The Yimas language of New Guinea. Stanford: Stanford University Press. Gensler, Orin D. (1998). Object ordering in verbs marking two pronominal objects: Nonexplanation and explanation. Ms., Max-Planck-Institute of Evolutionary Anthropology, Leipzig.

244

Barbara Stiebeis

Givon, Talmy (1984). Syntax: a functional-typological introduction, Vol. I. Amsterdam: Benjamins. Hopper, Paul J. & Sandra A. Thompson (1980). Transitivity in grammar and discourse. Language 56, 251-99. Joppen, Sandra & Dieter Wunderlich (1995). Argument linking in Basque. Lingua 97, 123-169. Joppen-Hellwig, Sandra (1999). Verbklassen und Argumentlinking: Nicht-kanonische Argumente, Expletiva und vierstellige Kausativa in Ergativ- vs. Akkusativsprachen. Dissertation, University of Düsseldorf. Keen, Sandra (1983) Yukulta. In Handbook of Australian languages, Vol. 3, R. M. W. Dixon & Barry J. Blake (eds.), 191-304. Amsterdam: Benjamins. Kibrik, A. E. (1979). Canonical ergativity and Daghestan languages. In Towards a theory of grammatical relations, Frans Plank (ed.), 61-77. London: Academic Press. Kimenyi, Alexandre (1980). A relational grammar of Kinyarwanda. Berkeley: University of California Press. Kiparsky, Paul (1992). Structural Case, Ms., Center for Advanced Studies, Berlin. Kiparsky, Paul (1998). Cases as complementizers. Paper presented at the conference 'The effects of morphological case', Utrecht, August 1998. Kiparsky, Paul (1999). Analogy and OT: Morphological change as emergence of the unmarked. Paper presented at the 21st annual meeting of the German Linguistics Society (DGfS), Constance, February 1999. Klamer, Marian (1998). Kambera intransitive argument linking. Studia Linguistica 52 (2), 77-111. Kornfilt, Jaklin (1996). Naked partitive phrases in Turkish. In Partitives: Studies on syntax and semantics of Partitive and related constructions, Jacob Hoeksema (ed.), 107-142. Berlin: de Gruyter. Launey, Michel (1979). Introduction à la langue et à la littérature Azteques, Vol. 1: Grammaire. Paris: L'Harmattan. McCarthy, John and Alan Prince (1993). Prosodie Morphology I: constraint interaction and satisfaction. Ms., University of Massachussetts, Amherst and Rutgers University. McCarthy, John and Alan Prince (1995). Faithfulness and reduplicative identity. In 'Papers in Optimality Theory', University of Massachusetts Occasional Papers 18, Jill N. Beckman, Laura Walsh and Suzanne Urbanczyk (eds.), 249-384. Mohanan, Tara (1994). Argument structure in Hindi. Stanford: CSLI publications. Ortmann, Albert (1998). Where plural refuses to agree: feature unification and morphological economy. To appear in Acta Linguistica Hungarica. Prince, Alan and Paul Smolensky (1993). Optimality Theory: constraint interaction in generative grammar. Ms., Rutgers University, New Brunswick and University of Colorado, Boulder. Quizar, Robin & Susan M. Knowles-Berry (1988). Ergativity in the Cholan language. IJAL 54(1), 73-95.

Linker inventories, linking splits and lexical economy

245

Ritter, Elizabeth and Sara Thomas Rosen (to appear). Event structure and ergativity. In Events as grammatical objects, James Pustejovsky and Carol Tenny (eds.). Stanford: CSLI publications. Schebeck, Bernhard (1976). Thangu and Atjnjamathanha. In: R. M. W. Dixon (ed.), 516-550. Silverstein, Michael (1976). Hierarchy of features and ergativity. In: R. M. W. Dixon (ed.), 112-171. Smolensky, Paul (1995). On the internal structure of the constraint component Con of UG. Paper presented at UCLA. Stiebels, Barbara (in preparation). Ökonomie und Expressivität des Argumentlinkings. Habilitation thesis, University of Düsseldorf. Stiebels, Barbara and Dieter Wunderlich (2000). On inverse systems: A reply to Aissen. Ms., University of Düsseldorf. Suñer, Margarita (1988). The role of agreement in clitic doubled constructions. Natural Language and Linguistic Theory 6, 391-434. Wilder, Chris and Hans-Martin Gärtner (1997). Introduction. In Chris Wilder, HansMartin Gärtner and Manfred Bierwisch (eds.), 1-35. Wilder, Chris, Hans-Martin Gärtner and Manfred Bierwisch (eds.)(1997). The Role of Economy Principles in Linguistic Theory. Berlin: Akademie-Verlag. Woolford, Ellen (1999). Case patterns. To appear in Optimality-theoretic syntax, Geraldine Legendre, Jane Grimshaw and Sten Vikner (eds.). Cambridge, Mass.: MIT Press. Wunderlich, Dieter (1996). A minimalist model of inflectional morphology. In Chris Wilder, Hans-Martin Gärtner and Manfred Bierwisch (eds.), 267-298. Wunderlich, Dieter (1997). Cause and the Structure of Verbs, Linguistic Inquiry 28, 27-68. Wunderlich, Dieter (2000a). A correspondence-theoretic analysis of Dalabon transitive paradigms. To appear in Yearbook of Morphology 1999. Wunderlich, Dieter (2000b). Optimal case. Ms., University of Düsseldorf. Wunderlich, Dieter (2000c). How gaps and substitutions can become optimal: the pronominal affix paradigms of Yimas. To appear in Transactions of the Philological Society, special issue on lexical paradigms, James P. Blevins (ed.). Wunderlich, Dieter & Ray Fabri (1996). Minimalist Morphology: An Approach to Inflection. Zeitschrift fur Sprachwissenschaft 14, 236-294. Wunderlich, Dieter & Renate Lakämper (1999). On the interaction of structural and semantic case. To appear in The effects of morphological case, Helen de Hoop et al. (eds.). Lingua, special issue.

Predicate composition and argument extension as general options - a study in the interface of semantic and conceptual structure Dieter

Wunderlich

1. Introduction

*

Complex predicates and their potential to license arguments have become an important issue in the last few years (see Alsina, Bresnan and Sells 1997); they are interesting for the study of the syntax-semantics interface as well as for the debate between lexicalist vs. nonlexicalist accounts. Lexical Decomposition Grammar (LDG) has been proposed as a lexicalist theory of argument structure (Wunderlich 1997a, among others): it assumes that the addition of arguments is triggered by a semantic extension of verbs, and thus tries to account for just the kind of phenomena that are labeled by the term 'complex predicate'. In this paper, I will mainly consider two constructions: resultative extensions and possessor extensions. 1 Examples are given in (1) to (3); in addition, (4) shows a (resultative) verb compound. (1)

Resultative construction a. The children ran (*the lawn). b. The children ran the lawn flat. c. *The stone fell the flowers flat.

(2)

Possessor extension (German) a. Die Brille zerbrach. 'The glasses broke' b. Mir zerbrach die Brille. me.DAT broke the glasses 'My glasses broke'

* The analysis presented in this paper was developed in the project on verb structures, which is part of the Sonderforschungsbereich 282 'Theory of the Lexicon', supported by the German Science Foundation (DFG). I would like to thank Manfred Bierwisch, Thomas Gamerschlag, Ingrid Kaufmann, Chris Piñón, and Barbara Stiebels for many valuable comments. 1 Originally included in this paper was a third type of construction, (resultative) verb compounds in Chinese and Japanese; it has been removed for the sake of brevity. I owe the remaining examples of Chinese to Yi-chun Yang.

Dieter

248 (3)

Possessor extension combined with resultative (German) a. Die Kinder liefen (*mir). 'The children ran (*me.DAT)' b. Die Kinder liefen mir den Rasen platt. 'The children ran my lawn flat'

(4)

Verb compound (Chinese) a. Lisi zui-le. Lisi inebriate-PERF 'Lisi became drunk' b. *Nei-ping jiu zui-le that-CL

c.

w i n e inebriate-PERF

Wunderlich

Lisi. Lisi

'That bottle of wine inebriated Lisi' Nei-ping jiu zui-dao-le Lisi. that-CL wine inebriate-fall-PERF Lisi 'That bottle of wine inebriated Lisi so that he fell down'

These examples invite us to make some initial observations. First, the (agentive) intransitive verb in (la) becomes transitive if, simultaneously, a resultative (adjectival) predicate is added, as in (lb); however, a (change of state) intransitive verb does not undergo such a construction, see (lc). Second, the (change of state) intransitive verb in (2a) allows the addition of a dative NP with possessor interpretation in (2b); however, the (agentive) intransitive verb in (3) only allows possessor extension if, simultaneously, a result is added. Finally, the (change of state) intransitive verb in (4a) becomes transitive with an additional causer only if, simultaneously, the verb is compounded with another (change of state) intransitive verb. These three constructions have different complexity; none of them is morphologically marked on the verb. What is visible on the surface is the result of some hidden derivation. For each of these constructions syntactic accounts have been proposed in the literature. 2 In the present paper, I will consider these constructions as the result of lexical extensions. Under certain conditions, the Semantic Form (SF) of a primitive verb may be extended by an additional predicate, which is B E C O M E ( Q ( Z ) ) with Q = F L A T in (1), POSS(z,x) in (2), the combination of these two in (3), and C A U S E R ( z ) , together with F A L L . D O W N ( x ) , in (4). Additional arguments of a verb must be licensed semantically by some predicate. Therefore, as I will claim, argument extension of a given verb always results from predicate composition, though, predicate composition itself does not necessarily lead to additional arguments. On the contrary, predicate composition often requires arguments to be shared by the individual predicates, and may also involve implicit arguments, as well as implicit predicates. It is these aspects of semantic construal that will be high-

2 There is a tradition that classifies the argument of change of state intransitives (the so-called unaccusatives) as VP-internal, while, in contrast, the argument of agentive intransitives (the socalled unergatives) is classified as VP-external. According to syntactic accounts, resultatives result from the addition of a small clause, possessor extensions result from possessor DP raising, and compounds result from head raising.

Predicate composition and argument extension as general options

249

lighted in this paper. Consider the following resultatives with transitive verbs; in all of them the internal argument (direct object) of the base verb cannot be expressed, but must be reconstructed in order to get the full reading of these sentences. 3 (5) a. Er trank mir den Kühlschrank leer, he drunk me.DAT the fridge empty 'He drunk (so much from my fridge) that as a result my fridge got empty (of beverages)' b. Sie fállten eine Schneise in the Wald, they felled a lane into the woods 'They felled (so much trees) that as a result a lane came into existence in the woods' c. Markus stellte den Keller voll. Markus put the cellar full 'Markus put (so much things into the cellar) that as a result the cellar got full' Why is it that the resultati ve object must be expressed, but the things that are primarily affected by the action cannot? Lexical Decompositional Grammar (LDG), which has adopted important ideas from Bierwisch (1989b) and Kiparsky (1989, 1992), attempts to account for these different constructions in a uniform way. In section 2, I will briefly introduce this framework (for more detailed information, the reader is referred to Joppen and Wunderlich 1995, Kaufmann 1995, Wunderlich 1997a, 1997b, 1999, Stiebeis 1997, this volume, Kaufmann and Wunderlich 1998), illustrating it with two resultative examples. Resultatives are considered in more detail in section 3, and possessor extensions are studied in section 4.

2. The theoretical framework Lexical Decompositional Grammar (LDG) provides a principled account for phenomena in which predicates and/or arguments are added to a base verb functioning as a lexical head. LDG assumes four levels of representation, each having its own structural properties: 4 Conceptual Structure (CS), Semantic Form (SF), Theta Structure (TS), and Morphology/Syntax (MS), and a set of principles that constrain the mappings between these levels. The distinction between SF, a partial semantic representation which is part of the grammar, and CS, a richer semantic representation which is part of the extralinguistic conceptual system, goes back to at least Bierwisch (1983). 3 I owe the examples (5b,c) to Ingrid Kaufmann. 4 A possible objection would be that TS is not a separate level because its main function is to interface between SF and MS. However, it includes information not present in SF, and also serves as the underlying structure for MS. (See also below.) In a correspondence-theoretic approach to argument structure (Stiebels, this volume), TS features form the input for MS.

250

Dieter

Wunderlich

This four-level architecture of LDG is illustrated in (7), representing the ditransitive verb geben 'give', which is canonically realized by the pattern in (6). (6)

(7)

a.

(als) der Torwart dem Jungen den Ball gab

b.

( w h e n ) the goal-keeper the b o y the ball g a v e [DPXNOM [DP Y DAT [DPZACC geb-AGRX ]]]

The four levels of LDG TS

SF

λζ

λΥ

λχ

+hr -Ir

+hr +lr

-hr +lr

i

1

1

DAT

AGR NOM

ACC

Xs

{ACT(x) & BEC POSS(y,z)}(s)

CS x=Agent or Controller y=Recipient z=Patient or Affected Causal event: ACT(x)(si) Result state: POSS(y,z)(s 2 )

MS The fact that the argument variables x, y, and ζ belong to different, more atomic predicates is expressed in SF. The corresponding list of λ-abstractors (theta or argument roles), each of which encoded by means of abstract case features ([+hr] for 'there is a higher role', [+lr] for 'there is a lower role'), is expressed in TS. The linking of these theta roles to agreement and case morphology is expressed in MS. Finally, the individual arguments can be characterized by thematic or eventive roles, and the whole situation (s) can be decomposed into relevant subevents, all of which is expressed in CS. SF plays the most crucial role among these four levels: the SF of a lexical item is a partial (or minimal) semantic representation, formulated by means of a binary categorial syntax (with the bracketing [A [& B]] for the conjunction &); it belongs to the grammar of a language insofar as it determines morphological and syntactic properties of that item (and for this reason may involve partial decomposition into more atomic predicates 5 ). All predicates used in SF can be explicated by bundles of conceptual conditions in CS; for instance, the predicate BEC(OME) has been explicated as a control condition for a change of state within Dynamic Event Semantics (Naumann, to appear). Everything that can be inferred by general means is not part of SF itself, although it may be incorporated in CS, which is a more elaborated semantic representation that includes all conceptual refinements explicitly. Notions such as implicit argument, subevent, thematic or eventive roles, which are often relevant in more detailed semantic analyses, do not belong to SF itself, but rather to CS. For instance, a one-level semantic representation of 'give' would have to include the predicate CAUSE for connecting the agent's action (performed on some object) with the result (see, among others, Jackendoff 1990, Pustejovsky 1991). However, it is not necessary to represent both the causal 5 Other reasons for decomposing a lexical predicate have to do with the scope of adverbials (von Stechow 1996); they do not play a role in LDG. I believe that the scope of adverbials depends on the semantic type on which they operate, and has to be fixed in CS rather than SF (see Wunderlich 1997a).

Predicate composition and argument extension as general options

251

factor (the agent's action) and the causal relation, because the latter can be inferred (see below). The motivation to assume SF as a separate level of representation is minimality,6 but also the possibility to express generalizations for the mappings between semantic and morphological/syntactic structure. The mapping between SF and CS is constrained by one general principle and two subconstraints. POSSIBLE VERBS (first proposed by Kaufmann 1995) demands every lower predicate in SF to be in some way more specific than the higher predicate. CONNEXION regulates the necessary connection between the predicates in terms of argument sharing, while COHERENCE regulates a coherent temporal/causal structure of the overall situation denoted by the complex verb. A predicate can specify another predicate only if these predicates share at least one argument and form a coherent temporal or causal structure, so CONNEXION and COHERENCE are necessary preconditions for POSSIBLE VERBS (see Kaufmann and Wunderlich 1998 for a more elaborate discussion). (8) POSSIBLE VERBS. In a decomposed SF representation of a verb, every more deeply embedded predicate must specify the higher predicate or sortal properties activated by the higher predicate.7 Every predicate ranges over a certain domain of arguments which can be characterized by some sortal property; specification of these sortal properties takes place if the lower predicate provides more information in the same domain. (9) CONNEXION. In a decomposed SF structure, each predicate must share at least one argument with another predicate, either explicitly or implicitly. (10) COHERENCE. Subevents encoded by the predicates of a decomposed SF structure must be connected contemporaneously or causally. CONNEXION ensures that the individual predicates form at least an argument chain and thus can characterize one and the same complex event; however, it is possible that implicit participants come into play, i.e., arguments that are only inferred in CS. COHERENCE ensures that the subevents encoded by the individual predicates either share their temporal structure or form a cause-result (accomplishment or achievement) complex; therefore, the causal connection in resultatives can be inferred and need not be explicitly represented in SF.

One reason for assuming a binary branching for SF is that this allows us to establish some interesting structural properties, which is in conformity with the overall-attempt of LDG to apply structural notions in the study of meaning as far as possible. A binary 6 In Wunderlich (1997a), I assumed CAUSE to be part of SF, but departed from this view already in Wunderlich (1996), where I tried to defend a minimal SF structure for the purpose of specifying the semantic ingredients for argument structure. Note that also the decomposed Logical Form proposed by von Stechow (1996), which is similar to SF in many respects, dispenses with representing the predicate cause. 7 The architecture of LDG is not restricted to verbs. POSSIBLE VERBS can easily be generalized to prepositions, nouns, and adjectives. Since in this paper I am concerned with verbs, there is no need to rename this constraint for more general purposes. Note that the intuitive notion of embedding can be refined by means of L-command, see (12) below.

252

Dieter Wunderlich

branching structure uniquely determines the ranking of arguments in SF (which in turn determines the way in which the arguments are projected into MS). Most relevant are the lowest and the highest argument, whereas some arguments in between may be 'wrongly placed' and therefore suppressed from being projected into MS. The mapping between SF and TS is governed by two constraints: ARGUMENT HIERARCHY, which ensures that the semantic ranking of argument variables is preserved, and STRUCTURAL ARGUMENT, which selects the arguments that can be realized in MS by agreement or case morphology. ( 1 1 ) ARGUMENT HIERARCHY. The list of λ-abstractors in TS corresponds to the

depth of embedding in SF, with the lowest argument to the left (first subjected to Functional Application), and the highest argument to the right. Correspondingly, the lowest argument (of a polyadic verb) is designated as [+hr,-lr], and the highest argument as [-hr,+lr], whereas all medial arguments are designated as [+hr,+lr], (12) STRUCTURAL ARGUMENT. An argument is structural only if it is either the lowest argument or (each of its occurrences) L(exically)-commands the lowest argument; so every internal (non-highest) argument of a nonfinal predicate in SF is nonstructural (Wunderlich 1997a,b).8 L-command is defined for the nodes in SF, which represent logical types, as follows: α L-commands β if the node γ, which either directly dominates α or dominates α via a chain of nodes type-identical with γ, also dominates β. [This constraint accounts for the fact that the transitive verbs in (5) cannot realize their internal arguments.] TS is considered to be an independent level of representation for the following reasons: (i) The default designations on the basis of ARGUMENT HIERARCHY can be lexically overridden, which happens in all instances of quirky case or dative experiencers. (ii) Again lexically determined, it is possible that improper theta roles (expletive arguments) appear, which do not have a thematic correspondent in SF although they participate in morphological case, (iii) Argument demoting operations such as passive, antipassive, or lexical reflexive can best be conceived of as operations on TS: passive blocks the highest theta-role and antipassive the lowest theta-role from realization, whereas a lexical reflexive binds a lower theta-role to the highest one. Finally, the mapping between TS and MS, which is in the center of argument linking, can best be captured in a correspondence-theoretic scenario, in which TS is the input and MS the output. Faithfulness constraints require each abstract case feature in TS to be realized in MS, and, conversely, each morphological agreement or case feature in MS to be based on a corresponding feature in TS. Consequently, both morphological agreement and case are characterized by the same features [hr] and [lr], which have already been introduced to encode the ranking in TS: dative is [+hr,+lr], accusative is [+hr], ergative is [+lr], and nominative is unspecified ([ ]). Feature checking, 8 For independent reasons, the referential (situation) argument of verbs is nonstructural (and only affected by aspect, tense, and mood), whereas the same type of argument becomes structural in event nominalizations (see Bierwisch 1989a).

Predicate composition and argument extension as general options

253

including also further morphological features for person, gender, and number, by means of faithfulness constraints is a more powerful device than unification, which is problematic for incorporating the concept of specificity in morphological paradigms (Barg u.a. 1996).9 Although LDG can be translated into a HPSG-like format (Rumpf and Stiebels 1999), it has a more sophisticated, and at the same time more elegant, architecture to express generalizations on argument structure (see also Stiebels, this volume). In this paper, however, I presuppose the LDG conception of argument linking and will not discuss it any further. It is important to notice that LDG is a strictly lexical account: The appearance of additional arguments (such as possessors, beneficiaries, or affected objects) is only licensed by a predicate that is added to the base SF. Furthermore, the formation of complex predicates is mostly triggered by some property of the lexical head. This is true for morphologically marked alternations such as causatives, the assistive in Quechua (van de Kerke 1996), or the affective in Basque (Joppen and Wunderlich 1995), which add a highest argument; here, the respective morpheme is both the head and the functor that selects a verb. It also holds for other complex predicate formation such as resultatives, particle-verb combinations, and Chinese verb-verb compounds, which may add a lowest argument; here, the verb is the lexical head that is qualified for taking another predicate. 1 So-called constructional meanings (Goldberg 1995) are always analyzed as being determined by the lexical head. In particular, every verb may undergo the operation ARG, by which some predicate Ρ is added as a further complement to the verb; Ρ must be predicated of the same situation s as the verb (see Wunderlich 1997b).11 (13) Argument extension ARG: ... Xs VERB(...)(s) =>

λ Ρ ... λ8{νΕΚΒ(...)(8) & P ( s ) }

which can be abbreviated as {VERB(...) & P}(s) ARG can only be iterated if the variable Ρ has been instantiated. This follows from the following SF constraint, which is motivated by the idea that the individual arguments can only be realized after the complex predicate is established. 12 9

A theta role specified as [+hr,+lr] can be unified with dative, accusative, or nominative, although only the most specific linker should be linked. Barg et al. propose to determine the possible linker candidates by means of a unification test, whereas the actual linker has to meet further conditions. With faithfulness constraints, the most specific linker incurs the least number of MAX violations.

10

There may also be mismatches between morphology and semantics: although the Bantu applicative is the morphological head, the verb is the functor that undergoes ARG, see (13). Similarly, in the right-headed verb-verb compounds of Japanese the non-head is the functor that undergoes ARG (Gamerschlag 1999). There is a variant of ARG in which Ρ predicates of a non-temporal object rather than a situation. In this case, Ρ must share its argument with some other argument of the verb, and it must be relativized to s. Examples are depictive predicates and poss, adding a possessor to the verb, see below. As I will show in section 4 , P R E D A R G also makes interesting predictions for restructuring a derived SF.

11

12

254

Dieter

Wunderlich

(14) PREDICATIVE ARGUMENTS (PREDARG). A predicate variable must occupy the lowest position in SF. (There can be only one at the time.) If Ρ is instantiated by a multi-valent predicate constant, all further arguments of it are inherited by Functional Composition. An example is shown in (15a) (from Stiebels 1996:67). Here, the verb schreiben 'write' is composed with the prefix er-, which contributes a resultative possessor meaning. In the resulting prefix verb, represented in (15c), the internal argument of schreiben cannot be realized, due to the restriction by STRUCTURAL ARGUMENT, and therefore is existentially bound. Since no lexical marking appears, all arguments are canonically realized (with the reflexive in dative position). (15) a.

Sie erschrieb sich den Pulitzer-Preis. She er-wrote herself the Pulitzer prize. 'She won the Pulitzer prize by her writing' b. schreib-' write': λ y λ χ Xs WRlTE(x,y)(s) ARG{schreib-y.

λ Ρ Xy λχ Xs {WRITE(x,y) & Ρ}(s)

er-:

λ ν λ ϋ Xs' BECOME POSS(U,V)(S')

c. er-schreib-·. λν λυ λχ λβ 3y {WRITE(x,y)(s)& BECOME POSS(u,v)(s)} The assumption is that, in principle, the composition with another predicate by means of ARG is free, but has to be checked in CS. Whether the particular composition is allowed depends on whether the SF-CS interface constraints can be satisfied (which often is the case only by argument identification due to CONNEXION), and, of course, also on the lexical and morpho-syntactic resources of that particular language. The semantic apparatus assumed in LDG is summarized in the diagram (16). (16) Overview of the SF-CS interaction Theta Structure Structural arguments and other complements that have to be realized morpho-syntactically

Semantic Form A partial semantic representation of a lexical predicate

Situation variable One complex event

A SF-CS constraints Conceptual Structure A more elaborated semantic representation. The inferred information includes: implicit arguments, > more predicates, > temporal/causal structure, sortal restrictions ¡ argument identification, ¡ subevents, ¡ thematic roles ¡ eventive roles This framework is illustrated with the example in (17a,b), in which tense is disregarded. The complex predicate run flat is derived in three steps: first, ARG is applied, then Ρ is instantiated by the resultative BECOME Q(z), and finally Q is instantiated by the adjectival complement flat, as shown in (17c).

Predicate composition and argument extension as general options (17) a. b. c.

255

The children ran the lawn flat. 3s {RUN(the children) & BECOME FLAT(the Iawn)}(s) run: ARG(rMrt): RES(rwn): run flat:

λ χ Xs RUN(x) (s) λΡ λχ Xs {RUN(x) & P}(s) XQ λζ λ χ Is {RUN(x) & BECOME Q(z)}(s) λζ λχ Xs {RUN(x) & BECOME FLAT(z)}(s)

Run flat is predicated of one complex event ('s'), consisting of a process (of running) and an achievement (of becoming flat). These subevents belong to event types with a different temporal structure; therefore, according to COHERENCE, a causal structure is established. Moreover, CONNEXION requires some (at least implicit) argument shared by the two predicates. Since the concept of running implies a certain ground on which this activity takes place and exactly this sortal requirement is met by 'lawn', this argument can indeed function as the connecting one. (Other potential implicit arguments are the feet or the shoes of the runner.) In contrast, running around usually does not imply any relation to spectators; thus, a sentence such as (18a) is odd for semantic reasons, and probably the same is true for (18b). However, Chinese verb compounds verbs seem to be more liberal (18c). (18) a. ?? The children ran their grand parents speechless. b. The children cried their grand parents helpless. c. xiao-hai ku-huang-le tarnen de yeye-nainai. small-child cry-get.confused-LE their grandpa-grandma 'The children cried so that their grand parents became confused' Although spectators of a running event may become speechless, this result usually does not specify anything which is involved in the process of running itself. Thus, (18a) - simply extending the running event - is ruled out by POSSIBLE VERBS, and similarly (18b). It is an open question whether the Chinese verb compound (18c) is possible because two independent verbs are combined, or because 'cry' is understood as a communicative utterance that implies a relation to some audience. Another, more complicated, example is (19) - repeated from above - , in which a transitive achievement verb undergoes resultative extension. (19) Markus stellte den Keller voll (mit Möbeln). 'Markus stood/put the cellar full (with furniture)' One can paraphrase this sentence as follows: 'By Markus' acting, some furniture (= y) became localized in a standing way, such that the cellar became full (of y)'. The argument y itself can only be realized as an adjunct, by means of the oblique preposition mit 'with'. This is predicted by STRUCTURAL ARGUMENT, because y, an internal argu-

256

Dieter Wunderlich

ment of stellen, becomes nonstructural in the derived SF of vollstellen. The full interpretation of the complex verb is schematically shown in (20).13 (20) Analysis of example (19) TS

stellen

λ ζ λχ Xs

3y {[ACT(x) & BECOME STAND(y)]

conceptually inferred predicates argument identification required by CONNEXION temporal/causal structure required by COHERENCE

strong resultative extension: voll & BECOME FULL(z)}

(s)

LOC(y,AT(u)) SUPPORT(u,y)

CONTAIN(z,v) u= ζ y= ν subevents: ACT & BEC STAND & BEC FULL

si.i causes s 1.2 si causes

S2 s2

S2=Si-2

STAND(y) implies that y is located at some surface u (or, that this surface u supports y), while FULL(z) implies that ζ contains a certain amount of v. CONNEXION requires some argument sharing. Indeed, the things ('y') which come to stand are identical with the things that are contained ( V ) , and the surface that supports these things ('u') is identical with the thing that becomes full ('z'). Furthermore, COHERENCE requires some causal relationship: (i) That y comes to stand on u is caused by some activity of x, and (ii), that ζ becomes full of ν is caused by the preceding achievement if a certain amount of things is moved. These two transitional events (come to stand and become full) are in fact only the two sides of one and the same event, in which two objects change some property: the things being put (into the cellar) ('y'), and the cellar ('z'), being occupied by these things. The following sections study the effects of the S F - C S interface constraints (CONNEXION and COHERENCE) in interaction with particular constructions and particular types of verbs in greater detail.

13 Stellen itself requires a directional PP, which is represented by the extension P(y). However, such a (weak resultative) extension is complementary to a strong resultative extension (see section 3); only one of them can be realized because of PREDARG. This example moreover shows that ARG can apply even if the base verb calls for a predicative argument. Thus, ARG is really a free option, and only restricted by further constraints. Other examples with resultative AP rather than PP complements are jemanden ruhig stellen 'put someone quiet', etwas flach legen 'put something flat'.

257

Predicate composition and argument extension as general options

3. The variance of resultative constructions Cross-linguistically, two types of resultative extension can be identified (Washio 1997, Kaufmann and Wunderlich 1998). • Weak resultatives, in which a result state already implied by the verb is specified more narrowly. • Strong resultatives, in which some result state predicating of one of the involved participants of a process is added. These two types of resultative construction are illustrated in (21) and (22). (21) Weak resultatives a. The children ran into the woods. b. Peter cut the meat into slices. c. The vase broke into several pieces. (22) Strong resultatives a. The children ran the lawn flat. b. John drank the guests under the table. c. The guests drank the wine cellar empty. d. He ran himself tired. If a language admits strong resultatives, it also admits weak resultatives. Strong resultatives are possible in the Germanic languages (German, English, and Dutch), as well as in Chinese (verbal compounds, i/e-construction). Other languages, such as the Romance languages and Japanese, only allow weak resultatives, whereas Hungarian seems to be in the process of developing strong resultatives. The properties in which strong and weak resultatives differ are summarized in (23). (23) The contrast between strong and weak resultatives • A new individual argument is introduced: • AP result predicates are possible: • The result predicate can specify a change which is not inherent to the meaning of the base verb: • An independent subevent is added:

strong resultatives weak resultatives yes no yes

no

yes

no

yes

no

All these differences can be predicted from the way in which the result predicate is added to the verb (Kaufmann and Wunderlich 1998). The formation of weak resultatives is a subcase of the introduction of secondary predicates, which itself is a subcase of ARG; it can be represented by the template in (24).14

14 Weak resultatives are often lexicalized, for instance, in the causative position verbs setzen, stellen, legen 'put', which require a directional PP (see footnote 13).

258

Dieter

Wunderlich

(24) Weak resultati ves: A predicate Q predicating o f the lowest argument o f the verb is added. λ y ... Xs VERB(...,y)(s)

=>

XQ Xy ... Xs {VERB(...,y) & Q ( y ) } ( s )

If the added predicate Q is realized by a stative AP (such as naked in (25a) and raw in (26a)), a depictive reading arises; therefore, adjectives usually do not qualify as result predicates in such a case. 5 Only if Q is realized by a directional PP (such as into the woods in (25b) and into slices in (26b)), a weak resultative reading arises, because the directional PP encodes some change in a relevant dimension specified by the verb, which is location o f χ in (25b), and constituency o f y in (26b). In none o f these cases is an independent subevent introduced: That the children reach the wood is a direct result o f their running, and that the meat is sliced is a direct result o f someone's cutting the meat. (25) a. b. c.

The children ran naked. The children ran into the woods. XQ Xx Xs {RUN(X) & Q ( x ) } ( s ) Q ( x ) = NAKED(X), Q ( x ) = CHANGE(L0C(X, INT(the w o o d s ) ) )

(26) a. b. c.

Peter cut the meat raw. Peter cut the meat into slices. XQ Xy Xx Xs {CUT(x,y) & Q ( y ) } ( s ) Q ( y ) = RAW(y), Q ( y ) = CHANGE(LN-SLICES(y))

In contrast, all these features change with strong resultatives, as indicated in the summary (23). The general template that allows the formation o f strong resultatives is shown in (27); it is a subcase o f ARG, with P= BECOME Q(z), and has already been illustrated by the examples in (17) and (19) above. (27) Strong resultatives: BECOME Q(z) predicating o f a new argument is added. ... Xs VERB(...)(s)

=> XQ Xz ... Xs {VERB(...) & BECOME Q ( z ) } ( s )

Here, the predicate BECOME explicitly adds an achievement, which constitutes an independent subevent. Hence, independently o f whether Q is realized by a PP or an AP, in any case a result reading arises. This extension is only constrained by CONNEXION, which requires the new argument to be at least an implicit participant o f the base verb. Moreover, an internal argument o f the base verb becomes nonstructural, so it cannot be expressed directly - an example is (28b), where the stuff being drunk remains implicit. As (28c) shows, the newly introduced argument can be realized by a reflexive pronoun, so it is possible that the result predicates o f the same argument as the base verb. (28) a.

The children ran the lawn flat. XQ Xz Xx Xs {RUN(x) & BECOME Q ( z ) } ( s )

b.

XQ Xz Xx Xs {DRlNK(x,y) & BECOME Q ( z ) } ( s )

15

Q ( z ) = FLAT(z)

The guests drank the wine cellar empty. Q ( z ) = EMPTY(z)

There are some exceptions, in which an adjective may be related to a result; they have been discussed in Napoli (1991), Washio (1997), and Kaufmann and Wunderlich (1998).

259

Predicate composition and argument extension as general options c.

John ran himself tired.16 XQ Xz λχ Xs {RUN(x) & BECOME Q(z)}(s)

Q(z) =

TIRED(z)

As pointed out above, SF itself neither represents subevents explicitly, nor any causal relation between subevents. However, considering the fact that BECOME predicates of a temporal situation one can infer that a new subevent is established, and that a causal relation between the subevents holds, according to COHERENCE. Several accounts of resultative constructions within a syntactic framework have been proposed (Hoekstra 1988, Wilder 1991, Winkler 1994, among others). None of these accounts is able to deal with examples such as (28b), in which a subcategorized direct object of a verb is replaced by a nonsubcategorized result object (see the discussion in Wunderlich 1997b). The lexical account of LDG does not have these problems, on the contrary, examples like those in (28b) turn out to be the standard case for resultative constructions with transitive base verbs.17 (For alternative lexical analyses, see Levin and Rappaport Hovav 1991, Legendre 1997). Not every verb allows a strong resultative: for instance, verbs that already encode BECOME (inchoative or causative verbs) are heavily restricted. German (29a) is deviant, while the English counterpart in (29b) is acceptable (although with some variation). I assume that the 'literal' representation in (29c) is ruled out by POSSIBLE VERBS: The second shift of property can specify the first one only if (i) the predicate Q entails the predicate encoded by the base verb (i.e., SOFT must entail NONSOLID, which is true), and (ii) if both predicates predicate of the same argument (i.e., ζ = χ). This identification of arguments is somewhat restricted; it is easier in English than in German. (29) a. ??Die Butter schmolz weich, b. The butter melted soft. C. XQ λζ λχ Xs { B E C O M E N O N S O L I D ( X ) & B E C O M E Q(z)}(s) Q(z) = SOFT(z) This difference between German and English has been analyzed in more detail in Kaufmann and Wunderlich (1998). The only possible interpretation for (29), where ζ = χ, shows that in fact only one transition takes place, hence, in this case the result AP does not constitute an independent subevent. Even German allows verbs that already encode B E C O M E to be extended by an AP result in some instances, as was illustrated with example ( 19) above. The analysis in (20) demonstrated that the interpretation of this example required both identification of argument variables and unification of the two potential subevents into a single event. Examples such as those in (19) and (29) thus show that the mechanism of strong resultative formation can also be used to introduce a subevent that is not independent. However, generally the addition of B E C O M E (which encodes a particular temporal structure) may introduce an independent subevent. In the next section I will show that 16

17

Some resultatives are only possible with a reflexive. For instance, 7John ran Anne tired is problematic because there is no obvious connection between John's running and a result state predicated of Anne. The same is true for 7He laughed me sick, in contrast to He laughed himself sick. One important motivation for developing LDG was the attempt to account for resultatives of this kind (Wunderlich 1992, Kaufmann 1995).

260

Dieter Wunderlich

also the addition of a Stative relation such as POSS is possible. Since a stative relation lacks any temporal structure, it cannot alter the event structure of the base verb. But such a relation must be integrated into the temporal/causal structure already expressed, thus, its actual reading must be derived conceptually, i.e., by semantic construal in CS.

4. Possessor extension on verbs: How semantic construal works In the following examples, a strong resultative licenses an additional possessor extension: a (reflexive) inalienable possessor in (30a,b), and an alienable possessor in (30c). There can be no doubt that the POSS relation holds between the dative possessor and the resultative object; it cannot be established without the resultative. (30) a. Ich lief mir die Füße wund. (*mir/*die Füße/*mir die Füße/*mir wund) I ran myself.DAT the feet sore Ί ran my feet sore' b. Die Demonstranten standen sich die Füße platt. the demonstrators stood themselves.DAT the feet flat 'The demonstrators stood their feet flat' c. Er trank mir den Kühlschrank leer. He drunk me.DAT the fridge empty 'He drunk my fridge empty' There are also underived transitive verbs that allow possessor extension, either inalienable (31a) or alienable (31b). (31) a. Er wusch mir die Füße. 'He washed my feet' b. Er putzte mir die Stube. 'He cleaned my room' But not every transitive verb allows possessor extension; the examples in (32) are rather odd. (32) a.

99

Er trank mir das Bier. 'He drunk my beer' 99 b. "Er betrachtete mir die Füße. 'He watched my feet' However, possessor extension is possible with (experiencer or change of state) intransitives (33), with stative locationals (34a), with intransitives subcategorized for a PP (34b,c), and with resultative particle verbs (34d). What is common to all these examples is the fact that the possessed object functions, in some way or other, as the affected object. (33) a. Mir schmerzt der Fuß. 'My foot hurts'

Predicate composition and argument extension as general options

261

b. Mir zerbrach die Brille. 'My glasses broke' c. Mir fiel die Tasche in den Fluß. 'My bag fell into the river' d. Mir fiel ein Blatt in die Suppe/ auf den Fuß. O n e of my leaves fell into the soup' or Ά leaf fell into my soup/onto my foot' (34) a. Er stand mir auf dem Fuß / ?auf dem Teppich. 'He stood on my foot / on my carpet' b. Eine Krankenschwester stach mir in den Arm. Ά nurse pricked my arm' c. Die Kinder trampelten ihr auf den Blumenbeeten rum. 'The children trampled on her flower beds' d. Er aß mir den letzten Apfel auf. 'He ate up my last apple' However, it is not yet clear whether all dative possessors have the same source. Note that the possessor is the highest argument in (33), according to several syntactic tests18, but it is a lower argument in (34). Interesting is the fact that the possessed thing in (33c,d) can be either an argument of the verb itself or an argument of the preposition, while it can be introduced only by the PP in (34a-c) - in all these instances I assume that the verb is subcategorized for a PP. Thus, whether the possessor becomes the highest argument of the verb is not determined by the occurrence of a PP or by the possibilities to identify the possessed thing, but rather by the respective type of verb. In any case, the identity of the possessed thing (i.e., the internal argument of the POSSrelation) must be reconstructed conceptually. Many languages with a rich case system including dative allow the introduction of a dative argument which functions in a broad sense as the possessor of some other argument of the verb. The interpretation of such an additional dative complement ranges from inalienable possessor over alienable possessor to recipient, beneficiary, or experiencer. The respective reading depends on several factors, for instance, whether a relational noun (encoding a part-whole, a kinship or a social relationship) is present and thus forces a strict possessor reading, and whether the verb has a more stative or a more transitional reading. Syntactic accounts only deal with the strict possessor readings by assuming possessor raising: the possessor phrase is moved from the complement of a DP into the complement domain of the governing verb. For instance, the account of Landau (1999) for similar examples of Modern Hebrew claims that a possessor phrase is moved from [Spec,DP] to [Spec,VP] in order to get dative case. (But why must it get dative case?) For empirical objections against the possessor raising account, see below.

18

For instance, the order of unspecific pronoun (weil einem der Fuß schmerzte 'because the foot of someone hurt'), VP topicalization (der Fuß geschmerzt hat mir aber nicht 'but my foot has not hurt'), and binding (weil jedem seine Füße schmerzten 'because everyone had his feet hurt'). For this type of possessors see also Blume (2000: 112ff).

262

Dieter Wunderlich

The lexical alternative assumes that possessors can be instantiated either on nouns or on verbs. The intention behind the concept of possessor raising concerns just one interpretation of POSS in verbs. Let us assume, similarly to the template of resultative formation, that there is another template that allows possessor extension, directly giving verbs the potential of having a possessor complement. As the examples above have shown, there must be an affected argument present in the verb, and the role of the possessor phrase depends on whether there is also an agent present or not. Thus, we have to distinguish two variants. (35) Possessor extension of verbs19 a. Topmost possessor: For intransitive verbs with an affected argument (regardless of whether they are subcategorized for a PP or not): λ χ λβ VERB(x)(s)

=>

λχ

λζ

Xs {POSS(z,u) &

VERB(x)}(s)

+hr Here, u is nonstructural, and its value must be reconstructed. If no PP is present, u=x because of CONNEXION. b. Otherwise (subordinated possessor): . . . λ β VERB(...)(s)

=> ... λ8 {VERB(...) & POSS(z,u)}(s)

The realization of both u and ζ depends on further properties of the verb. The elsewhere case (35b) is a particular instance of ARG, whereas the specific case (35a) is not; here, POSS rather behaves as a functor on the verb, similarly to the causative, the assistive in Quechua, or the affective in Basque. The designated feature [+hr] ensures that the possessor is realized as dative, because this theta role is also inherently [+lr]. The assignment of [+hr] thus allows the inference that 'there is some higher role which causes the event'; this is similar to other experiencer verbs with dative and is possible only if no agent is present; an agent would have to be realized as the higher role. Therefore, if one allows POSS to be a functor on verbs, it necessarily selects only intransitive verbs with an affected argument.20 The interpretation of POSS splits into inalienable (permanent) vs. alienable (temporary) possession, depending on how the possessed thing ('u') is fixed. In the more specific instance of inalienable possession, POSS(z,u) has the reading 'z is related to u via part-whole, kinship, or social function', while in the general case of alienable possession POSS(z,u) has the reading 'z has access to or some control over u'. Therefore, if a relational noun is present (such as Fuß 'foot' in (33d) above), it gets preference in the identification of the possessed thing. In any case, CONNEXION requires that the internal argument of POSS ('u') is identical with some other (explicit or implicit) argument of the verb. As the above examples show, u tends to be identical with the lowest argument of the verb (the subject of intransitive change of state verbs or the affected object of transitive verbs), but may also be identical with the object of a PP. Since in German the

19 The two options in (35) together correspond to the thematic hierarchy agent > possessor > affected object. 20 German also has a verb of possession, namely gehören 'belong', which exactly has the specification of the topmost possessor; consider mir gehört die Tasche 'The bag belongs to me'. We might assume that the topmost possessor is a covert variant of this verb.

Predicate composition and argument extension as general options

263

possessed thing must be affected, it can never be the case that u is identified with an agent. However, Serbo-Croatian also allows agents to be the possessed thing: 21 (36) Otac father.NOM

mu

cita

he.DAT read.PRES.3sg

novin-e. newspaper-ACC

'His father reads newspapers' or 'The father reads newspapers to him' In order to see how possessor extension works, let us first consider the elsewhere case (subordinated possessor). (37) a. Er wusch mir das Hemd. 'He washed my shirt' Xu λ ζ λ χ Xs{WASH(x,Y) & POSS(z,u)}(s)

b.

The nonstructural argument y is identified with u: y=u. Sie trampelten ihr auf den Beeten rum. 'They trampled on her flower beds' XQ λζ λ χ λβ{TRAMPLE(X) & POSS(z,u) & Q ( x ) } ( s ) -DIR +lr Q(X)=L0C(X,0N(V)); v=die Beete 'the beds'.

The nonstructural argument u is identified with v: u=v. The analysis of (37a) with a transitive verb is straightforward: the possessed thing becomes the lowest argument, the possessor the medial argument, and the object of the base verb nonstructural - this scenario predicts (i) that the structural arguments are canonically realized (NOM-DAT-ACC), and (ii) that it is the object of the base verb that is identified with the possessed thing. (Note that the Serbo-Croatian example (36) cannot be explained in this way.) The analysis of (37b) is more complicated. Let us assume that trampeln selects a nondirectional locative PP, and that a predicative argument must stay lowest, according to PREDARG.22 Then, POSS is integrated higher than the predicative variable, which is reflected in the assignment of [+lr] ('there is a lower role') for the possessor. Consequently, the possessed thing ('u') becomes nonstructural; it has to be identified with the object of the preposition because it cannot be identified with the agent. The analysis for the topmost possessor extension (35a), in which POSS is the highestranked predicate, is now obvious. The readings of the examples (33c,d) above can be captured by the representations in (38); [+DIR] classifies a directional PP. Recall that u is nonstructural in these representations; so u may be identified with some other argument, either with the argument of the base verb ('x') or with the argument introduced by the preposition ('v'). (38) a.

Mir fiel die Tasche in den Fluß. 'My bag fell into the river' XQ λ χ λζ Xs {POSS(z,u) & FALL(x) & Q ( x ) } ( s ) +DIR +hr Q(x)=CHANGE L O C ( X , INT(V)); v=der Fluß ('the river'); u=x.

21 22

I owe this example to Jelena Krivokapic. Alternatively one may say that possessor extension takes place after the PP is instantiated.

264

Dieter Wunderlich b.

Mir fiel ein Blatt in die Suppe. Ά leaf fell into my soup' XQ λχ λζ Xs {POSS(z,u) & FALL(x) & Q(x)}(s) +DIR +hr Q ( X ) = C H A N G E L O C ( X , INT(V)); v=die Suppe ('the soup'); u=v.

Notice that in all examples described in (37) and (38), a nonstructural argument of a higher predicate receives its value from a lower predicate, in accordance with POSSIBLE VERBS. The reader will now be able to check the following representation for an example in which resultative and possessor extension interact. (39) Er trank mir den Kühlschrank leer. 'He drunk my fridge empty' XQ λν λζ λχ Xs {DRiNK(x,y) & POSS(z,u) & BECOME Q(v)}(s); u=v. The order of derivation must be resultative-possessor because the possessed thing is identified with the resultative object. This shows that ARG can be applied even if a predicative argument is present. However, the predicative argument introduced by the resultative must stay lowest; therefore, some restructuring of SF must take place in order to satisfy PREDARG. Such a derivation is not fully monotonie; it resembles the well-known phenomenon of infixation in morphology, where an affix sometimes is infixed in order to satisfy higher-ranked constraints. In both types of possessor extension, and independent of whether alienable or inalienable possession is at stake, the possessed thing can be identified with the object of a preposition, which is only available if the PP is added, i.e., within the smallest VP. This fact is problematic for a syntactic account of possessor raising, because movement out of a PP violates the condition of proper government of the trace. Some instances of dative possessors in verbs also allow more than one interpretation, where a possessor raising analysis would have to assume ambiguity of derivation. It seems more adequate to assume ambiguity of semantic construal, which is part of the conveptual rather than the grammatical system. Moreover, some interpretations of a possessor phrase in verbs depend on whether the POSS-relation holds for the anterior or the posterior state, or for the whole situation including these states. The possessor raising analysis is unable to assume any temporal predominance; it cannot explain why the POSS-relation may be restricted to a certain phase (or subevent). The POSS-relation itself has no independent temporal structure; the only requirement is that it must be integrated into the temporal/causal structure of the verb - which is possible in several ways. Example (40) may have two readings. (40) Sie legte ihm das Buch auf den Tisch. 'She put him the book on the table' (i) A recipient reading: POSS is established by the action, i.e., POSS is in the scope of BECOME and only relevant in the posterior state. In this case, he (the referent of the dative DP) becomes the possessor of the book. ('She put the book on the table, and as a result he got the book')

Predicate composition and argument extension as general options

265

(ii) A beneficiary reading: POSS is established independently of the action, i.e., POSS can be relevant also in the anterior state, or in the whole event. In this case, he (the referent of the dative DP) may be the possessor of the book or the table, and she put the book on the table for his benefits. (41) Analysis of example (40) Row (a) represents the SF of the causative position verb legen, which is subcategorized for a directional PP (considered as a lexicalized version of weak resultati ve extension), row (b) represents the derived SF of the verb, and (c) the stage in which the directional PP is added. TS (a)

(b) (c)

legen

possessor extension

XQXyXxXs { [ A C T ( X ) & BEC LIE(y)] +DIR XQ Xy λ ζ λ χ Xs {[ACT(x) & BECLIE(y)] &POSS(z,u) λ y λ z λ x λ s {[ACT(X) & BEC LIE(y)] &POSS(z,u)

weak resultative extension &Q(y)} (s) &0(y)l & CHANGE LOC

00

(y,ON(d.Tisch))} (s) conceptually inferred predicates argument identification required

LOC(y,AT(w)) SUPPORT(w,y)

suPPORT(d.Tisch,y)

w - d.Tisch the possessed u = y or d.Tisch

b y CONNEXION

temporal/causal structure required by

as in ex. (20) above POSS is related to the anterior state of S2

(Ben)

COHERENCE

or to the posterior state of s 2

(Ree)

The predicate LIE(y) implies that y is located on some w, such that y is supported by w, while the result PP auf den Tisch 'on the table' implies that the table supports something. So we have to identify w=d.Tisch. However, the possessed thing u can be either y ('the book') or w ('the table'), which then determines the recipient or beneficiary reading. Let us now compare (40) with a sentence that displays the same structure but contains relational nouns. (42) a.

b.

Sie legte ihm die Hand auf die Schulter. she put him.DAT the hand on the shoulder. 'She put her/his hand on his shoulder' Sie legte sich die Hand auf die Schulter, she put refl.DAT the hand on the shoulder 'She put her hand on her shoulder'

266

Dieter

Wunderlich

In the presence of relational nouns, the preferred reading for a dative possessor is that the referent of the relational noun is inalienably possessed, so POSS is related to the overall situation s and not to a partial phase of it. The person referred to by ihm (he.DAT) in (42a) cannot be a recipient (unless the hand is understood as a detached object rather than a body part), so the only interpretation left is that the possessed thing is his shoulder. It cannot be her shoulder, unless this is explicitly expressed by a reflexive pronoun, as in (42b), because both ζ and χ are structural arguments and therefore distinct. However, the hand in (42a) can be her or his hand. In the latter case (u=w=y), the possessor ζ in fact possesses both w and y (the shoulder and the hand), which are the two arguments of the loc-relation; and, with the reflexive in (42b) it is even allowed that subject and object stand in a poss-relation. However, the preferred reading for the possessed thing is the prepositional object: it is only the body-part reading of 'hand' that requires further identification. The possessor extension itself never expresses a POSS-relation between the two arguments of a transitive verb, as shown by the oddity of (43a), in contrast to (43b,c). 2 The intended reading for (43a) would be (44) with z=x, u=y, whereas (43c) requires z*x, u=y. (43) a. b. c. (44)

??

Er hebt sich die Hand. Er hebt die Hand/seine Hand. 'He lifts his (own) hand' Er hebt ihr die Hand. 'He lifts her hand'

{LlFT(x,y)&POSS(z,u)}(s)

Above I have stated that possessor extension requires the possessor to be distinct from an agent; this suffices to exclude (43 a) with the intended reading. There seems to be no structural reason to exclude representations of the form {A(x,y) & B(x,y)}; for instance, Japanese verb compounds allow this transitive parity. 2 This means that the restriction is conceptually determined. Let us finally consider to what extent the identification of the possessed thing ('u') is structurally determined. In all examples we have seen, u receives its value from a lower predicate, unless it itself is the lowest argument and is projected into syntax; insofar the identification is structurally determined, but solely as a consequence of STRUCTURAL ARGUMENT. TO see whether, moreover, there is a structural antecedent relation between u and its value in SF, let us repeat representation (41c) from above, with some constant a for the prepositional object:

23

24

Interestingly, (43a) has the following possible reading: Assume that he cannot move his left hand because it is paralyzed, and so he uses his right hand on the left one in order to lift it. In this reading, he is possessor of the left hand but not able to control it; so he himself is affected by the action taken with the right hand. An example is the following (Gamerschlag 1999): (i) Rikisi ga aite o zimen ni osi-taosi-ta. sumo.wrestler NOM opponent ACC ground to push-topple-PAST 'The sumo wrestler pushed his opponent to the ground.'

Predicate composition and argument extension as general options (45) λ y λ ζ λ χ ks

267

{[ACT(x) & BEC LIE(y)] & POSS(z,u) & CHANGE.LOC(y, ON(a))}(s)

STRUCTURAL ARGUMENT predicts that only y, z, and χ are structural arguments of the verb, and ARGUMENT HIERARCHY predicts that they are mapped onto TS in that order. CONNEXION demands argument sharing, so u could be identical with y or a. If a=d.schulter(v) ('the shoulder of v') and y gets the value d.hand(w) ('the hand of w'), then both ν and w are inalienable possessors. Thus, identifying u=y implies the identification z=w. This crossing identification immediately makes clear that argument identification of this kind does not apply structurally on SF. There is further reason why this is excluded. Note that χ L-commands ζ in (45), so χ can anaphorically control the reflexive sich in (42b);25 but neither does y L-command u, nor does u L-command a in (45). Given the apparatus of LDG, one has to conclude that argument identification in connection with dative possessor readings is conceptually determined, not structurally (which makes the possessor raising account quite implausible).

5. Conclusion Two types of extension have been discussed: resultative and possessor extension, each showing some variation. Resultatives can be weak or strong; only the latter licenses additional arguments; but both can be seen as instances of ARG, the general option of predicate composition with the verb as the morphological head. Possessors can be added as the highest or a non-highest argument of the verb (topmost vs. subordinated possessor extension); only the latter is an instance of ARG, while the former is similar to a causative and, hence, can be regarded as a functor on verbs. There is one major difference between strong resultatives on the one hand and weak resultatives and subordinated possessor extensions on the other: strong resultatives add a predicate with a temporal structure of its own (thus constituting a subevent) to the verb, whereas both the weak resultative and the possessor extension add a predicate that does not have temporal structure of its own. Therefore, these predicates must be integrated into the temporal/causal structure provided by the verb. There is only one way to integrate the additional directional PP in weak resultatives, but there are often several ways to integrate the possessor (the predicate POSS) in possessor extensions. Moreover, weak resultatives require the added PP to predicate of the lowest argument of the verb, but (subordinated) possessor extensions allow several ways of identifying the possessed thing. Throughout the paper, it has been demonstrated that LDG is a successful tool to describe the minimal requirements for predicate composition and argument extension on the level of SF. In one direction, the interface constraints between SF and TS, and those between TS and MS, allow us to predict how the resulting argument structure is realized (and which arguments are suppressed from being realized). In the other direc25

I assume that the binding of reflexives is determined in SF.

268

Dieter Wunderlich

tion, the constraints proposed for the SF-CS interface help us to derive the full meaning of the construction: CONNEXION requires at least a partial identification of arguments, and COHERENCE requires the temporal integration of predicates. Needless to say, that there are open questions; however, LDG has proven sophisticated enough to make these questions precise. Finally, the interaction of these extensions has been studied. Weak and strong resultatives are mutually exclusive, and so are topmost and subordinated possessor extensions. Weak resultatives are often lexicalized, namely in verbs that are subcategorized for a directional PP; nevertheless, these verbs can undergo strong resultative extension by stripping off the requirement of a weak resultative - this might be studied within an optimality approach, which allows the optimal SF to violate lower-ranked constraints. If POSS is added to the SF of a resultative verb (either weak or strong resultative), it cannot be added as the lowest predicate, according to PREDARG. Therefore, one has to conclude that ARG itself is a free option, but may lead to results that must be improved, or restructured. As in other levels of grammar, the formation of SF may be subject to conflicting constraints, therefore, the optimal SF resulting from lexical derivation may differ from what is predicted by the principle of monotonicity. In this respect, it is enlightening to compare SF formation with PF (phonological form) formation, especially if SF and PF are considered to be the major levels of grammar interfacing with extralinguistic knowledge: SF with conceptual structure, and PF with phonetic structure. In both SF and PF formation, a minimal violation of monotonicity might be tolerated in order to enlarge expressivity. If there is pressure to realize an affix morpheme even if it does not conform to PF constraints, this morpheme may instead be realized as an infix, slightly passing over material of the base - a candidate is chosen that does not violate the relevant PF constraints. Similar things may happen in SF formation. If there is pressure to realize an argument extension even if it does not conform to SF constraints, it may yet be realized by passing over material of the base (i.e., by raising the added predicate into the base SF) in order to satisfy the relevant SF constraints. In principle, nothing hinders us from considering the optimal SF to be triggered in this way. What I wish to have demonstrated in this article is that if one concedes monotonicity violations in SF to be possible, the theory gains predictive force concerning properties of both the morpho-syntactic surface and the conceptual interpretation. If one does not accept the monotonicity violations in the analysis of examples such as (19), (37b), and (39) above, one would be unable to predict which arguments have to be realized, which have to be suppressed, and which have to be identified with another one.

Predicate composition and argument extension as general options

269

References Alsina, Alex, Joan Bresnan and Peter Sells (eds.)(1997). Complex Predicates. Stanford: CSLI Publications. Barg, Petra, Sandra Joppen, Ingrid Kaufmann and Suzanne Wolting (1996). Die Repräsentation von Argumentlinking in einem unifikationsbasierten Formalismus. Working Papers of the SFB 282 'Theorie des Lexikons ' 83. University of Düsseldorf. Bierwisch, Manfred (1983). Semantische und konzeptuelle Repräsentation lexikalischer Einheiten. In Untersuchungen zur Semantik, Rudolf Rüzicka and Wolfgang Mötsch (eds.), 61-99. Berlin: Akademie Verlag. Bierwisch, Manfred (1989a). Event nominalization: proposals and problems. Linguistische Studien 194, 1-73. Berlin: Akademie der Wissenschaften der DDR. Bierwisch, Manfred (1989b). Thematische Rollen und Dekomposition. Seminar Papers, Third DGfS Summerschool, Hamburg. Blume, Kerstin (2000). Markierte Valenzen im Sprachvergleich: Lizensierungs- und Linkingbedingungen. Tübingen: Niemeyer. Gamerschlag, Thomas (1999). Deriving argument structure in Japanese V-V compounds. Ms., University of Düsseldorf. Goldberg, Adele E. (1995). Constructions: A construction approach to argument structure. University of Chicago Press. Hoekstra, Teun (1988). Small clause results. Lingua 74, 101-139. Jackendoff, Ray (1990). Semantic Structures. Cambridge, Mass.: MIT press. Joppen, Sandra and Dieter Wunderlich (1995). Argument linking in Basque. Lingua 97, 123-169. Kaufmann, Ingrid (1995). What is an (im)possible verb? Restrictions on Semantic Form and their consequences for argument structure. Folia Linguistica 29, 67-103. Kaufmann, Ingrid and Dieter Wunderlich (1998). Cross-linguistic patterns of resultatives. Working Papers of the SFB 282 'Theorie des Lexikons ' 109. University of Düsseldorf. Kiparsky, Paul (1989). Agreement and linking theory. Ms. Stanford University. Kiparsky, Paul (1992). Structural case. Ms. Berlin: Insitute for Advanced Studies. Landau, Idan (1999). Possessor raising and the structure of VP. Lingua 107, 1-37. Legendre, Géraldine (1997). Secondary predication and functional projections in French. Natural Language and Linguistic Theory 15,43-87. Levin, Beth and M. Rappaport Hovav (1991). Wiping the slate clean: A lexical semantic exploration. Cognition 41, 123-155. Napoli, Donna Jo (1991). Secondary resultative predicates in Italian. Journal of Linguistics 28, 53-90. Naumann, Ralf (to appear). Verbs as programs - a dynamic approach to aspect. Journal of Semantics. Pustejovsky, James (1991). The syntax of event structure. Cognition 41,47-81.

270

Dieter

Wunderlich

Rumpf, Christof and Barbara Stiebeis (1999). Representing LDG in a constraint-based framework. Proceedings of the Workshop on lexical semantics and linking in constraintbased theories. 11. ESSLLI. Utrecht. Stiebels, Barbara (1996). Lexikalische Argumente und Adjunkte. Berlin: Akademie Verlag. Stiebeis, Barbara (this volume). Linking splits, linker inventories, and lexical economy. van de Kerke, Simon (1996). Affix order and interpretation in Bolivian Quechua. Phil. Diss. University of Amsterdam. von Stechow, Arnim (1996). The different readings of wieder 'again': A structural account. Journal of Semantics 13, 87-138. Washio, R (1997). Resultatives, compositionality and language variation. Journal of East Asian Linguistics 6, 1-49. Wilder, Chris (1991). Small clauses and related objects. Groninger Arbeiten zur Germanistischen Linguistik 34, 215-236. Winkler, Susanne (1994). Secondary predication in English: a syntactic and focus-theoretical approach. Phil. Diss. University of Tübingen. Wunderlich, Dieter (1992). CAUSE and the structure of verbs. Working Papers of the SFB 282 'Theorie des Lexikons ' 36. University of Düsseldorf. Wunderlich, Dieter (1996). Models of lexical decomposition. In Lexical structures and language use, Edda Weigand and Franz Hundsnurscher (eds.), 169-183. Tübingen: Niemeyer. Wunderlich, Dieter (1997a). Cause and the structure of verbs. Linguistic Inquiry 28, 27-68. Wunderlich, Dieter (1997b). Argument extension by lexical adjunction. Journal of Semantics 14, 95-142.

Addresses of contributors Stephen R. Anderson [email protected] Department of Linguistics Yale University Box 208236 Yale Station 320 York Street New Haven, CT 06520-8236 USA

James P. Blevins [email protected] Research Centre for English and Applied Linguistics University of Cambridge Keynes House, Trumpington St Cambridge, CB2 1QA United Kingdom

Gisbert Fanselow [email protected] Institut für Linguistik Universität Potsdam Postfach 601553 14415 Potsdam Germany

Sharon Inkelas [email protected] Department of Linguistics University of California at Berkeley Berkeley, CA 94720 USA

Elisabeth Löbel [email protected] Institut für Sprachwissenschaft Universität zu Köln 50923 Köln Germany

Cemil Orhan Orgun [email protected] Department of Linguistics One Shields Avenue University of California, Davis Davis, CA 95616 USA

Ronald Sprouse [email protected] Department of Linguistics University of California at Berkeley Berkeley, CA 94720-2650 USA

Janet Grijzenhout Martin Krämer Carsten Steins Barbara Stiebels Dieter Wunderlich

griizenh@... kraemer@... steins@·.. stiebels@..· wdl@...

.. [email protected] Seminar für Allgemeine Sprachwissenschaft Heinrich-Heine-Universität Universitätsstr. 1 40225 Düsseldorf Germany