Semi-lexical Categories: The Function of Content Words and the Content of Function Words [Reprint 2013 ed.] 9783110874006, 9783110166859

The distinction between functional categories and lexical categories is at the heart of present-day grammatical theory,

222 88 13MB

English Pages 562 [564] Year 2001

Report DMCA / Copyright

DOWNLOAD PDF FILE

Table of contents :
Semi-lexical categories
Part I. Semi-lexicality and Syntactic Projection
The flat structure economy of semi-lexical heads
Heads and selection
Children’s semi-lexical heads
Semantically empty lexical heads as last resorts
Part II. Semi-lexicality in the Nominal Domain
Numeral/Quantifier-Classifier as a complex head
Classifiers and semi-lexicality: Functional and semantic selection
Agreement patterns of Czech group nouns and quantifiers
Part III. Semi-lexicality in the Verbal Domain
On the (semi)lexical status of light verbs
“Semi-lexical” motion verbs in Romance and Germanic
Underspecification in serial verb constructions
Part IV. Semi-lexicality of Adpositional and Adposition-like Elements
Semi-lexical heads in a semantically charged syntax
As for as / for, they are semi-lexical heads
Lexical particles, semi-lexical postpositions
Index
Recommend Papers

Semi-lexical Categories: The Function of Content Words and the Content of Function Words [Reprint 2013 ed.]
 9783110874006, 9783110166859

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

Semi-lexical Categories

W DE G

Studies in Generative Grammar 59

Editors

Henk van Riemsdijk Harry van der Hulst Jan Köster

Mouton de Gruyter Berlin · New York

Semi-lexical Categories The Function of Content Words and the Content of Function Words edited by

Norbert Corver Henk van Riemsdijk

Mouton de Gruyter Berlin · New York

2001

Mouton de Gruyter (formerly Mouton, The Hague) is a Division of Walter de Gruyter GmbH & Co. KG, Berlin.

The series Studies in Generative Grammar was formerly published by Foris Publications Holland.

© Printed on acid-free paper which falls within the guidelines of the ANSI to ensure permanence and durability.

Library of Congress Cataloging-in-Publication

Data

Semi-lexical categories : the function of content words and the content of function words / edited by Norbert Corver, Henk van Riemsdijk. p. cm. - (Studies in generative grammar ; 59) Includes bibliographical references and index. ISBN 3110166852 1. Grammar, Comparative and general — Function words. 2. Grammar, Comparative and general — Syntax. 3. Semantics. I. Corver, Norbert, 1963II. Riemsdijk, Henk C. van. P283 .S46 2001 415-dc21 2001037016

Die Deutsche Bibliothek — Cataloging-in-Publication

Data

Semi-lexical categories : the function of content words and the content of function words / ed. by Norbert Corver ; Henk van Riemsdijk. — Berlin ; New York : Mouton de Gruyter, 2001 (Studies in generative grammar ; 59) ISBN 3-11-016685-2

© Copyright 2001 by Walter de Gruyter GmbH & Co. KG, D-10785 Berlin. All rights reserved, including those of translation into foreign languages. No part of this book may be reproduced in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Printing & binding: Hubert & Co., Gö ttingen. Cover design: Christopher Schneider, Berlin. Printed in Germany.

To Haj Ross supreme grammarian and grand master of the subtle distinction

Contents

Semi-lexical categories Norbert Corver and. Henk van Riemsdijk

1

Parti Semi-lexicality and Syntactic Projection The flat structure economy of semi-lexical heads Joseph Emonds

23

Heads and selection Hubert Haider

67

Children's semi-lexical heads Susan M. Powers

97

Semantically empty lexical heads as last resorts Carson T. Schütze

127

Part II Semi-lexicality in the Nominal Domain

Numeral/Quantifier-Classifier as a complex head Tanmoy Bhattacharya

191

Classifiers and semi-lexicality: Functional and semantic selection Elisabeth Löbel

223

Agreement patterns of Czech group nouns and quantifiers Ludmila VeselovsM

273

VIII Contents

Part III Semi-lexicality in the Verbal Domain On the (semi)lexical status of light verbs Miriam Butt and Wilhelm Geuder

323

"Semi-lexical" motion verbs in Romance and Germanic Anna Cardinaletti and Giuliana Giusti

371

Underspecification in serial verb constructions Tjerk Hagemeijer

415

Part IV Semi-lexicality of Adpositional and Adposition-Iike Elements Semi-lexical heads in a semantically charged syntax Kristin M. Eide and Tor A. Äfarli

455

As for as / for, they are semi-lexical heads Joan Rafel

475

Lexical particles, semi-lexical postpositions Jochen Zeller

505

Index

551

Semi-lexical categories Norbert Corver and Henk van Riemsdijk

1.

On function words and content words

Ever since the earliest research on language, syntactic categorization of lexical items has played an important role in linguistic description and theorizing. A central dichotomy in the categorization of syntactic categories is that between content words (also called: lexical or substantive categories) and function words (also called: functional categories). Content words are often characterized as being those lexical items which have a relatively 'specific or detailed' semantic content and as such carry the principal meaning of the sentence. They name the objects (N), events (V), properties (A) and locations/directions (P) that are at the heart of the message that the sentence is meant to convey. As opposed to content words, function words have a more 'non-conceptual' meaning and fulfill an essentially 'grammatical' function; in a sense they are needed by the surface structure to glue the content words together, to indicate what goes with what and how. The abstract meaning of the functional domain comprises such properties as: tense, modality, definiteness, number, degree, interrogativity, etcetera. Although there is, of course, an intuitive plausibility underlying this major distinction in the system of syntactic categories, one of the aims of linguistic theory should be to make this dichotomy more precise, i.e. to define what the 'functional properties' are that make a lexical item into a function word and what the 'content/lexical properties' are that characterize a lexical item as belonging to the class of content words (i.e. lexical categories). An often referred to property distinguishing the two types of categories is that of openness of membership. Content words belong to open classes of words; that is, new members can freely be added to this class. Function words, on the other hand, have a fixed roster, and once it has been learned, few other members are ever added. As a consequence, the membership is much more restricted; the set of lexical items falling under some function word X forms a closed class.

2 Norbert Corver and Henk van Riemsdijk

Besides the two above-mentioned properties (lack of semantic/descriptive content and restrictive class membership), a number of other properties that differentiate functional categories from lexical ones have been given in the literature (cf. among others, Emonds 1985 and Abney 1987). A morpho-phonological characteristic of function words is that they are generally phonologically and morphologically dependent. They are typically stressless, often clitics or affixes, and sometimes even null. A selectional characteristic is that functional elements typically combine with a phrase of a specific categorial class. D, for example, combines with NP, I with VP and C with IP. Lexical categories, on the contrary, often take different types of syntactic categories as their complement. A verb like believe, for example, combines with a noun phrase (DP), as in / believe that story, or a clause (CP/IP), as in I believe that he is ill and I believe him to be ill. One way of expressing this difference is to say that selection of a lexical category by a lexical head is top-down, while the selection relation between a functional head and a lexical element is bottom-up. In the former case, we tend to say that a verb, say, selects certain types of direct objects and does not select certain others. On the other hand, the specific form of a determiner, its gender for example, is determined by the noun it determines rather than the other way round. Another (lexical) property differentiating between lexical and functional categories is that the former enter into theta marking, while the latter do not (cf. Ouhalla 1991; Chomsky and Lasnik 1993). Content words like friend (N), dead (A) and to sleep (V), for example, take an argument-category which receives a thematic role from the content word; such function words as the (Determiner), how (Degree word) and that (Complementizer) do not assign a thematic argument role to some element in their vicinity. Another grammatical property distinguishing function words from content words relates to movement/displacement. As noted in Abney (1987), functional elements are usually inseparable from their complement. Complements of lexical heads, on the contrary, can often be moved away from the lexical head. So, while the clausal complement to the verb believe can be topicalized (cf. (lc)), it is impossible to topicalize the IP-complement of the complementizer that (cf. (lb)). For the nominal domain (i.e. DP), we can observe that the NP-complement to D cannot be proposed (2b), while preposing is permitted with (certain) complements to Ν (cf. (2a)). (1)

a. I don't believe [cp that [n> Mary hates soccer] ] b. *[n> Mary hates soccer\ I don't believe [cp that tj] c. [CP That Mary hates soccer]i I don't believe tj

Semi-lexical categories

(2)

3

a. [Of whomh did you see [several pictures ti\? b. *[Red car]j John bought [dp the ii]

These facts suggest that content words, as opposed to function words, are able to license empty categories in their complement position. This asymmetry used to be captured in terms of the notion of proper government (cf. Chomsky 1981; Rizzi 1990): lexical categories (i.e. theta-role assigners) are able to license the trace of a displaced constituent, functional categories are not.1 Another distinguishing feature relating to the displacement property of human language concerns the landing site of displacement operations. According to recent theories on displacement, displaced elements typically move (overtly) to some position (i.e. specifier or head position) in the functional domain (cf. Chomsky 1993, 1995). Or to put it differently, it is only the feature system of functional categories that triggers displacement of constituents. The feature system of lexical categories (i.e. content words) typically does not attract moved constituents. In Chomsky (1995: 232, (1)), this contrast is formulated as follows: "If F is strong, then F is a feature of a nonsubstantive category and F is checked by a categorial feature." Thus, lexical (i.e. substantive) categories like Ν and V differ from functional (i.e. nonsubstantive) ones (e.g. D, T) in not being able to attract a moved constituent. 2.

On the content of function words and the function of content words

The distinction between content words and function words is a central one in studies on the syntactic categories of natural language. And, as pointed out in the previous section, a number of characteristic properties have been identified, which make it possible to classify some lexical item as belonging to the class of content words or that of function words. But as with all types of categorization, there are elements, which cannot be put straightforwardly under one of the two classes. Certain lexical items display ambiguous behavior: they share properties with lexical categories and at the same time they display functional characteristics. Prepositions (or more generally: adpositions) are a well-known case of uncertainty. Although they seem less functional in a sense than determiners, they are more "grammatical" than Ν, V and A. For one thing, the syntactic category Ρ does not really constitute an open class. And even

4 Norbert Corver and Henk van Riemsdijk

though it makes sense to associate the lexical property of having an argument structure with locative and directional prepositions (see e.g. (3a)), such a lexical property possibly does not extend to more grammatical Ps (like on in (3b)). (3)

a. b.

The money was on the table John counted on Bill

In view of this ambiguous behavior of the category P, it has been proposed that a distinction should be made within the class of prepositions between lexical ones and grammatical/functional ones (cf. e.g. Van Riemsdijk 1990; Zwarts 1992; Van Riemsdijk and Huijbregts, forthcoming). This implies that the property of being functional or being lexical is not or not necessarily an absolute property associated with some syntactic category (i.e. P). Ross (1972, 1973) already argued that the traditional view of syntactic categories as being discrete elements that can be rigidly distinguished from one another is incorrect. Instead of a fixed, discrete inventory of syntactic categories, he proposes a quasi-continuum, where the distinction between one category and another one is not discrete, but "squishy," i.e. one of degree. According to this view, syntactic categories are not, or not necessarily, distinguished from each other by the absolute presence versus absence of some property X. Categories rather differ from each other in relative terms: that is, category α (say V) may have property X to the highest degree, category β (say A) to a lesser degree, while for category γ (say N), this property is nearly absent or maybe even completely so. Ross surveys a number of properties with respect to which adjectives are intermediate between verbs and nouns. One example comes from the distribution of "expletive" it in complement position. He observes that this element appears after a number of verbs, and after only the adjective aware, but after no nouns: (4)

a. b. c.

I hate/(dis)likeAove/?regret it that he talked so much. I am ?aware/*convinced/*certain (of it) that we may have to pay more. My regret (*ofit) that he talked so much is well-known.

So, the distribution of expletive it can be characterized as a gradient phenomenon: it is most typically found in verbal contexts, less so in the adjectival environment, and impossible with nouns. The adjectival category

Semi-lexical categories 5 is intermediate between Verbs and Nouns, as far as this distributional property is concerned. Another argument Ross gives in support of the category squish concerns the phenomenon of Raising, the rule which makes the subject of the complements of certain verbs a clause member of the next clause up. Ross assumes that raising can be of two types: Raising to subject (cf. (5a)) and Raising to object (i.e. so-called Exceptional Case Marking constructions; cf. (5b)): (5)

a. b.

[ s Oliver seems/turns out/happens/etc. [s — to like walnuts]] [s We knew/showed/proved/believed/etc. Oliver [ s — to like Walnuts]]

While the embedded subject, Oliver, can be raised to become the surface subject of certain adjectives (cf. (6a)), there is no adjective which allows raising to produce a derived object (cf. (6b)): (6)

a. b.

[s Oliver is sure/certain/likely/etc. [s — to like walnuts]] *[s I am afraid/ready/willing/etc. (of) Oliver [ s — to like walnuts]]

Ross further points out that Raising typically does not apply in nominal contexts. (7)

a. *[John's belief [s — to be a fool]] b. *[my belief {of) John to be a fool]

In short, Raising applies "more" to V than to A. And it is least applicable to nouns. An important outcome of Ross's research on what could be called "gradience" in syntactic categorization (see Anderson 1997) is that some property X that is most centrally associated with some category α does not necessarily exclusively hold of α but may also apply to a lesser degree to some category β. This way, category β has α-like properties. And the more α-like properties β displays, the closer it is to α on the "category scale". Ross speaks in this respect about degrees of 'nouniness' and 'verbiness'.2 The question of whether the inventory of syntactic categories is gradient rather than discrete and fixed could of course also be raised for the dichotomy between content words and function words. Is this traditional distinction rigid and clear-cut? Or are there also borderline cases on this dimen-

6 Norbert Corver and Henk van Riemsdijk

sion of syntactic categorization? That is, are there function words with lexical (i.e. content word) properties and content words with functional properties? Or, to put it in Ross's terms: Does a squish also cut across the functional/lexical boundary? In Emonds (1985: 162-191), this question of gradience on the lexicalfunctional dimension is addressed. He points at the existence of closed classes of grammatical formatives that are subclasses of the lexical categories Ν, V A and P. In informal terms, these subclasses can be characterized as the most frequently used and least semantically specific members of each lexical category. Emonds calls these "in between" subclasses "grammatical nouns, verbs, adjectives and prepositions." Some examples of each subclass are given in (8): (8)

a. b. c. d.

Grammatical nouns one, self, thing, place, time, way, body Grammatical verbs be, have, get, do, come, make, let, want, say Grammatical adjectives other, same, different, such Grammatical adpositions out, up (i.e. particles)

Let us illustrate the "in between" nature of these grammatical (lexical) categories by focussing on one subclass, viz. that of grammatical nouns. The noun-like nature of these categories is suggested by the appearance of plural morphology (-s) on the noun (cf. (9)) and by their relationship to their lexical equivalents (cf. (10)-(11)): (9)

a. b.

John's good ones They hate themselves

(10)

a. b.

/ bought something (thing as a grammatical N) I bought a nice thing (thing as a lexical N)

(11)

a. b.

Have you seen anybody? (body as a grammatical N) Who discovered the body? (body as a lexical N)

Clearly, the grammatical nouns in (9)-(ll) are semantically less specific than such lexical nouns as car, woman, house, etcetera. The latter typically

Semi-lexical categories 7 refer to some (concrete or abstract) object or individual. They have the property of referentia(bi)lity. Emonds further argues that one property of grammatical as opposed to lexical heads is that certain types of transformational operations only apply to the former. As a result of this, grammatical (lexical) heads differ in their distribution from their (truly) lexical equivalents. One of the illustrations he gives comes from such composite pronouns as some-one, some-body, anybody, some-thing, etcetera. He argues that these compound pronouns are derived by moving syntactically the grammatical noun (one, body, thing) to the quantifying element (some, any). One distributional phenomenon quite clearly shows that the grammatical noun has raised to a higher position, viz. the fact that the composite pronoun must precede simple adjectives; see the examples in (12), drawn from Emonds (1985). Observe that truly lexical nouns do not display such behavior. (12)

a. b. c.

[Somebody clever—] is invited *Clever somebody is invited *Housemates clever can be fun Clever housemates can be fiin Some clever fellows are invited *Clever some fellows are invited

As Emonds points out, the lexical status of these categories (e.g. the N-like status of body in (12a)) is often hard to discern after movement has applied to it. This is why Emonds also refers to these moved grammatical (lexical) heads as disguised lexical categories. Another illustration of a class of lexical items that is on the borderline of the lexical-functional dichotomy is given in Van Riemsdijk (1998). He discusses so-called Direct Partitive Constructions, like (13) from Dutch (see also Vos 1999 for extensive discussion of this construction): (13)

a. b.

eenplakkaas a slice cheese een snee brood a slice bread

In these nominal constructions, the two nouns (N1 and N2) that are in a partitive relation to one another are directly juxtaposed. One of the most significant facts about these DPCs is that they show the behavior of single projections rather than dual projections. That is, N2 does not behave like a

8 Norbert Corver and Henk van Riemsdijk separate syntactic projection (say, an argumental DP) in the complement position of Nl. There are a number of phenomena that support this single projection hypothesis. First of all, selection is between the predicate, say the verb governing the DPC, and either Nl or N2. In (14a), for example, it is Nl (schaal) that the verb enters into a selectional relationship with; in (14b), on the other hand, the verb selects N2 (gebakjes). (14)

a. b.

Zij hebben [een schaal gebakjes] omgestoten They have a tray pastries turned-over Zij hebben [een schaal gebakjes] gegeten They have a tray pastries eaten

Another argument for the proposal that DPCs are single extended nominal projections comes from the distribution of case. As shown by the following facts from German, which has morphological case realization, both Nl and N2 can bear the case that has been assigned by the governing verb or preposition: (15)

a. b.

nach [zwei Flaschen rotem Wein] after two bottles-DAT red-DAT wine-DAT Ich habe [eine Kiste Kubanische Zigarren] bestellt I have a-ACC case-ACC Cuban-ACC cigars-ACC ordered

Of course, if the two nouns are part of one and the same extended nominal projection, we expect all heads on the projection line to agree in case. A final argument that shows that N2 is not the head of an independent nominal projection comes from the distribution of determiners and quantifiers, i.e. elements that typically close off an independent nominal projection. As shown by the Dutch examples in (16), N2 cannot take any functional heads of the D/Q type: (16)

a. b.

een fles (*deze) wijn a bottle (this) wine een Stapel (*alle) publicaties van Halle

a pile (all) publications by Halle The above considerations suggest that DPCs constitute single projections. Crucially, N2 does not constitute an independent nominal projection which stands, for example, in some argumental relationship to Nl. But if both Nl and N2 are part of one and the same extended nominal projection, the

Semi-lexical categories 9 following question arises: What is the categorial status of each of the two nouns? An analysis according to which each of the two nouns is a "truly" lexical noun is implausible; there can only be one lexical head Ν on the projection line of an extended nominal projection. A way out might be to say that Nl, the highest noun on the projection line, is part of the functional domain of the extended nominal projection. In other, words Nl would be a functional head. One argument in support of the "functional" nature of Nl is the fact that the class of nouns that can appear as Ν1 in these DPCs is restricted. Although the lexical item pan ('pan') can appear as Nl, the lexical item sauciere ('saucer') cannot. The set of nouns that can designate the container part of DPCs is limited; it does not constitute a truly open class. (17)

a. b. c.

een pan soep a pan soup een sauciere soep a saucer soup een sauciere met soep a saucer with soup

Even though lexical items representing Nl in DPCs are members of a class which is not really open, it is clear that they are less "function word-like" than, for example, determiners (D). As a matter of fact, lexical items instantiating Nl have a number of "content word-like" (i.e. noun-like) properties. Morphologically, Nl has the nominal property of being able to carry plural morphology (e.g. twee flessen wijn; two bottle-PL wine). Van Riemsdijk further points out that Nl has the "lexical noun" properties of (i) being able to function as the antecedent of a relative clause (cf. (18a)) and (ii) being modified by an attributive adjective (cf. (18b)). (18)

a. b.

een bus toeristen [die in de sneeuw was blijven steken] a bus tourists that in the snow had remained stuck een bruine kist Cubaanse sigaren a brown case Cuban cigars

Evidently, Nl in DPCs is a borderline case: it is neither a truly lexical head (i.e. lexical N), nor a truly functional head. Being somewhere in between the core functional categories and core lexical categories, Van Riemsdijk refers to them as semi-lexical heads.

10 Norbert Corver and Henk van Riemsdijk

3.

Semi-lexical categories: further explorations

Although the distinction between content words and function words is a useful one, it seems too coarse. There are content words with a degree of "functionalness" and there are function word having a degree of "lexicalness". Although there is a certain intuitive appeal to the existence of such semi-lexical categories, it is clear that for a better understanding of semilexicality a great deal of in-depth research is required. The purpose of this volume is to further our still very limited understanding of this borderline area within our system of syntactic categories. Various questions need to be answered to come to a better understanding of this notion of semi-lexicality. For example: What types of semi-lexical nouns, verbs, adjectives and prepositions can be distinguished? What distinguishes them from truly lexical categories and in what sense are they different from truly grammatical functors? Is this distinction expressed in terms of their lexical featurecomposition, and if so, what features are involved? What is their assembling property; i.e. how do they combine in syntactic structure and how do they project syntactically? Are they involved in idiosyncratic displacement phenomenona, and if so, what feature of the semi-lexical head triggers this? What is the licensing function of semi-lexical heads? What makes them interpretable at the interface levels? From the point of view of language acquisition, the question arises whether semi-lexical categories are found in child language. Besides the question of how semi-lexical heads behave in the syntactic component, questions arise about their behavior in other components: Are there morphological processes characteristic of semilexical heads? As concerns the lexicon, the question arises how they are stored in the lexicon and what distinguishes their lexical entry from that of truly lexical categories. These are only a few of the many questions that could be raised. Some of these questions, but also other ones, are addressed in the papers in this volume, hi what follows we will give a short description of each of the papers. Those papers that have a common theme are discussed together in a subsection.

3.1.

Semi-lexicality and syntactic projection

Emonds' article investigates the syntactic projection of closed classes of grammatical categories Ρ, A, Ν and V which modify open class elements of the same category. As his starting point he takes Van Riemsdijk's (1990)

Semi-lexical categories

11

Categorial Identity Thesis (CIT), which states that in the unmarked case, the lexical head and the corresponding functional/grammatical head have the same categorial features (see also Grimshaw 1991). Emonds argues that the CIT results from a particular conjunction of general phrase structure principles which gives rise to "flat structures." These structures contain "multiple" heads (e.g. N-N, V-V, P-P), of which only one, the least peripheral, can be realized as an open class lexical item. The others are semilexical (i.e. grammatical) heads. Importantly, such structures do not conform to a widespread insistence on binary branching. By discussing various types of constructions (e.g. pseudopartitive constructions, verbal structures featuring restructuring verbs), Emonds gives substance to his proposal for the structural relation between grammatical heads and lexical heads. Haider's paper discusses the selectional property of heads and the way heads project in syntactic structure. It is claimed that (non-lexical) functional heads targeted by head movement are heads that precede their complements. In other words, the projection of such a functional head is universally initial. The functional category I(nflection), for example, is to the left of its VP-complement in all natural languages. Lexical functional heads (i.e. function words with lexical content), on the contrary, have a directionality option. Just like lexical heads of the major categories, they can be set as head-initial or head-final, depending on the setting of the value for the directionality of complement licensing. This directionality of complement licensing is a lexically specified property. The absence of head-final I(nflection) predicts the absence of head movement to a functional head on the right (e.g. movement of V to a right branch I-head, as was standardly assumed for Verb-final structures in languages like Dutch and German). Thus, the clause final finite verb in Dutch and German stays in situ, i.e. in the position where it occurs as a non-finite verb. The idea that lexical functional heads have the directionality of complement selection specified as a lexical property is extended to lexical complementizers. It is shown that even within one and the same language, the directionality value for the set of complementizers need not be uniform In Powers' paper, it is pointed out that although semantically such lexical items as more and no in early English child language equal adult functional heads, their syntactic distribution is radically different: while adult functional heads select unique syntactic complements, the elements more and no appear with a variety of complements in child language. Powers also points out that these elements seem to supplete functional heads in child grammars. For example, while not is generally missing in the early word combinations of English, the semantically similar semi-lexical item

12 Norbert Corver and Henk van Riemsdijk no is widely attested. The lack of functional morphemes and the appearance of such semantically similar (semi-lexical) elements in their stead suggests that it is not the features of functional elements per se which is responsible for their omission, but rather their functional status. Powers claims that in the lexicon such semi-lexical elements as more/no are represented as frames with an open slot, that is filled by some content word in the syntactic projection of the semi-lexical head. Thus, the difference between functional and substantive categories is reflected structurally in the lexicon. Closed class items are represented as frames but open class items (which fill in these frames) are not. Schütze argues that semi-lexical heads are fully lexical heads that lack intrinsic semantic content. He argues that these heads are used as last resort defaults to spell out syntactic positions whose presence is forced by formal (i.e. morphological or syntactic as opposed to semantic) requirements. The distribution of these last resort elements is best characterized as the complement set to the environments where nondefault elements (i.e. lexical heads) appear. Schütze discusses the following semi-lexical (i.e. last resort default) heads: one (Noun), so (Adjective), be (Verb) and with (P).

3.2.

Semi-lexicality in the nominal domain

There are three papers that consider the presence of semi-lexical heads within the nominal projection. Bhattacharya's paper discusses classifier constructions in Bangla, an Eastern Indo-Aryan language. He discusses such noun phrases as kO-Ta chele (some-CLA boy; 'some (of the) boys') and tin-Te chele (three-CLA boy; 'three (of the) boys'). It is argued that the sequences Q+Classifier and NUM+Classifier form a complex head displaying both lexical properties (e.g. agreement properties and the possibility of floating the complex head) and functional properties (e.g. lack of descriptive content). Given this in between behavior, it is concluded that the Q/NUM+Cla complex is a semilexical category. It is further shown that syntactically the complex head behaves as a unit, even though morphologically it is clearly composed of a NUM/Q-part and a Classifier-part. In other words, Q/Num and Cla are not visible to syntax, which implies that the complex can never be split up in two parts as a result of syntactic displacement. Löbel discusses the notion of semi-lexicality in the context of classifier constructions. Syntactically, classifier constructions contain two constituents. As far as their semantics is concerned , they have only one referent.

Semi-lexical categories

13

For her analysis of classifier constructions, Löbel assumes that besides socalled f(unctional)-selection, two kinds of s(emantic)-selection must be distinguished within the nominal domain, namely: (a) selection of a participant role; (b) selection of a non-participant role. This distinction was proposed by Rizzi (1990) for the verbal domain: the direct object (i.e. the referential argument) in He weighed apples carries a participant role, while in The apples weighed three pounds, the complement to V carries a nonparticipant role. More specifically, the verb selects a non-referential argument denoting a property which is predicated of the (referent of) the subject DP. Löbel claims that this distinction also holds for nouns: there are relational nouns denoting parts which select a referential argument (e.g. the end of a story), and other relational ones which denote a property which is either inherent to or ascribed to (the meaning) of their selected nonreferential arguments (e.g. a head of cattle versus three pounds of apples). This approach is applied to classifiers in Vietnamese. It is shown that the syntactic function of a classifier, namely to serve as a unit-counter in order to make the selected noun countable in the sense of combinable with numerals, is not confined to a class of nouns. Rather, the definition of grammatical nouns in contrast to lexical nouns as having not only semantic, but also syntactic features (see Emonds 1999) holds for an open class in Vietnamese, which is due to the lack of inflectional morphology in this isolating language Veselovskä's paper discusses nominal phrases containing quantifying elements in Czech. On the basis of a variety of phenomena (e.g. case and agreement patterns, relativization, ellipsis of the N™ within the extended nominal projection) she distinguishes three kinds of quantifying elements: (a) group nouns (QN) like trocha ('a bit of), fura ('a heap o f ) ; (b) existential quantifiers (QGEN) like mnoho ('many, much'), kolik ('how many'); this class also comprises most cardinals above 4; (c) universal quantifiers (QA) like vsichni ('all') and oba ('both') and the cardinals 2/3/4. Besides these three classes of quantifying expressions, she notes the existence of a class of lexical nouns with quantificational force (NQ). Veselovska proposes that quantifiers and cardinals are semi-lexical nominal or adjectival categories, which contrary to lexical nouns and adjectives, have a semantically restricted feature content. She further argues, following Emonds (1999), that the distinct properties of Czech quantifiers can be explained in terms of distinct levels of their insertion into a cyclic derivation.

14 Norbert Corver and Henk van Riemsdijk

3.3.

Semi-lexicality in the verbal domain

Cardinaletti and Giusti discuss the notion of semi-lexicality in the context of motion verbs such as go and come which enter a particular "inflected construction" across unrelated languages such as some Southern Italian dialects, American English and Swedish. They show that the motion verbs in this construction type share many properties with functional verbs (e.g. auxiliaries) but at the same time maintain their semantic content and a few other lexical properties. In view of this mixed behavior, these motion verbs seem to be candidates for the class of semi-lexical heads. Cardinaletti and Giusti argue, however, that crosslinguistically these "semi-lexical" verbs cannot be identified by a fixed set of properties. That is, there is crosslinguistic variation in die functional versus lexical properties displayed by these verbs. This disfavors the assumption that a third type of category, i.e. semi-lexical V, exists in addition to lexical (V) and functional categories (I). They claim that semi-lexical motion verbs are in fact lexical categories that are merged into a non-lexical (i.e. functional) projection. Consequently, these verbs display a functional usage in the inflected construction. This usage blocks a number of lexical properties that motion verbs normally display. Hagemeijer discusses Serial Verb Constructions (SVCs) in Säo Tomense, a Portuguese-based Creole language spoken on the island of Säo Tome in the Gulf of Guinea. He claims that SVCs as verbal complex predicates have significant effects on both verbs in these constructions. SVCs are combinations of lexically restricted classes of syntactic and semantic heads fusing two argument structures into a single one. The semantic head (V2) imposes s-selection restrictions on the syntactic head (VI), which Hagemeijer argues to be semantically unspecified to some degree. A parallel is drawn between those Vis in SVCs and so-called light verbs. Hagemeijer argues that inside the verbal, complex predicate, the verb in second position (V2) contributes semantic features lacking on the VI. It is also shown that in some cases the V2 slot displays hybrid categorial behavior between V and P. In those cases, the elements in the V2 slot are underspecified for the [V] feature in the lexicon. Butt and Geuder consider light verb constructions in Urdu and Wagiman. They analyze light verbs as doubly headed structures (i.e. complex predicates): the full verb and the light verb form a unit. The light verb patterns in many respects with fully lexical elements, not with functional heads. This distinguishes light verbs syntactically from auxiliaries, in that auxiliaries typically involve VP-embedding constructions and do not form

Semi-lexical categories 15 part of the lexical core of the clause. As regards the lexical semantic properties of light verbs, they argue that light verbs, though being lexically defective, are not completely empty elements. A light verb conveys a special kind of lexical meaning which consists in a modulation of the event description provided by the full verb that it is in construction with. They point out that all light verbs in Urdu are paired with formally identical full verb uses. The relation between the light verb use of a verbal item and its use as a full verb is identified as a case of lexical polysemy (not grammaticalization). On the other hand, it is shown that in spite of the wellmotivated conceptual connections between full and light verb meanings, a marked difference in semantic type appears with the light verb variant.

3.4.

Semi-lexicality of adpositions and adpositions-like elements

Eide and Äfarli argue in their paper that the semi-lexical nature of certain functional heads stems from a general mechanism of natural language, and that 'hybrid heads' actually constitute not the exception, but the rule. They propose that syntactic representations are structured objects essentially consisting of functional operators (such as tense, mood, aspect and predication operators) that are made visible by various types of lexical elements, via insertion or alternatively by movement. Thus, the visible heads are semi-lexical in virtue of supporting an underlying functional operator. This approach towards the syntactico-semantic content of functional projections is illustrated on the basis of phenomena of predication within the Norwegian clause. It is proposed that clauses contain a Pr(edicate)-projection, which can be made visible by a variety of elements, e.g. som ('as'), copulas, prepositional elements like til ('into') and for ('for'). Rafel's paper also discusses the notion of semi-lexicality in the context of predicative constructions. More in particular, he examines the behavior of such particles as for and to in such constructions as They took John for a fool and They regard John as smart. It is shown that these items display both functional and lexical characteristics. As such, they can be characterized as 'semi-lexical'. Rafel further argues that the syntactic structure corresponding to the predicative relationship is that of a Complex Small Clause. In his paper about the properties of postpositions and particles in German, Zeller claims that semi-lexicality can be defined in morphological terms. He assumes that a semi-lexical head is half lexical and half functional. It is a morphologically complex element that consists of a lexical

16 Norbert Corver and Henk van Riemsdijk

node and a functional suffix. Postpositions, he claims, are semi-lexical elements. They are derived from lexical prepositions via suffixation of a zero-operator that alters the thematic properties of the P-element. Since this operator is a functional element, the derived postposition is a complex functional head. The postpositional element inherits the semantics from the lexical preposition from which it is derived. Therefore, postpositions are semi-lexical elements; they have the semantic content of a lexical element, but the categorial properties of a functional suffix. Zeller further argues that the particle status of a prepositional element is caused by its syntactic environment. If a lexical Ρ is structurally adjacent to a functional head F, it is a regular locational preposition; however, if it is structurally adjacent to a verb, it is a particle. Thus, particle status is contextually determined. 4.

Conclusion

The aim of this volume is to further our insight in syntactic categorization by examining the properties of a class of categories that are neither fully lexical nor fully functional. We labeled these items as semi-lexical. In the collection of papers brought together in this volume, different but also converging views are presented on this notion of semi-lexicality. A variety of definitions and characterizations of the notion 'semi-lexical head' is presented. To give a few examples: • • • •

Semi-lexical heads are those N,V,A and Ρ which have no purely semantic features; Semi-lexical heads are morphologically complex elements that consist of a lexical node and a functional suffix; Semi-lexical heads are functional slots carrying (moved) lexical material or lexical heads without distinct argument selection; Semi-lexical heads are fully lexical heads that lack intrinsic semantic content and which are used as a last resort spell out for syntactic positions present in the syntactic structure.

Some researchers consider semi-lexical categories to be a class of syntactic categories distinct from the purely functional ones and the truly lexical ones. Others assume that semi-lexicality is a derivative notion; e.g. semilexicality is due to a merge of a lexical category and a functional position. Even though the views may differ on the exact "implementation" of the

Semi-lexical categories 17

notion of semi-lexicality, there is one common outcome, viz.: there is more between function words and content words than was hitherto assumed.

Acknowledgements The present collection of articles grew out of a workshop on Semi-lexical heads which took place at Tilburg University in May 1999. We gratefully acknowledge the financial support of the Netherlands Organization of Pure Scientific Research (NWO), Tilburg University and the Beiport Familienstiftung. Anonymous thank-you's are due to our anonymous referees. Finally, our thanks go to Carol McGregor, Nicole van Eijndhoven, Anneke Smits and Hans Broekhuis for their assistance in preparing the camera ready copy.

Notes 1. It should be noted, however, that the situation appears to be the reverse for the licensing of elliptic lexical heads. This is exemplified by cases of VP-ellipsis (*John began singing a song before Bill began [e] vs, John began to sing a song before Bill began to [e]) and N-bar-ellipsis (Iprefer those (*green) [e]). Cf. Lobeck (1995). 2. See also Corver (1997) for discussion of the verbal versus nominal orientation of adjectives. 3. In Corver (2000) it is proposed that the element mente in such Italian adverbial adjectives as estremamente ('extremely') is a semi-lexical noun which function as the external argument of the adjective estrema. The surface order is derived by application of predicate displacement to the adjectival element estrema. This displacement operation is restricted to these adverbial contexts, which is in line with Emonds* (1985) observation that semi-lexical heads are involved in displacement phenomena that do not hold for truly lexical heads.

References Abney, Steven P. 1987 The english noun phrase in its sentential aspect. Ph.D. Dissertation, MIT. Anderson, John M. 1997 A notional theory of syntactic categories. Cambridge: Cambridge University Press.

18 Norbert Corver and Henk van Riemsdijk

Chomsky, Noam 1981 Lectures on government and binding. Dordrecht: Foris 1993 A minimalist program for linguistic theory. In: Kenneth Hale and Samuel J. Keyser (eds.), The view from Building 20, 1-52. Cambridge, Mass.: ΜΓΓ press. 1995 The Minimalist Program. Cambridge Mass MIT Press, Cambridge. Chomsky, Noam and Lasnik, Howard 1993 The theory of principles and parameters. In: Joachim Jacobs, Arnim von Stechow, Wolfgang Sternefeld and Theo Vennemann (eds.), Syntax: An international handbook of contemporary research, 506-569. Berlin and New York: Walter de Gruyter. Corver, Norbert 1997 The internal syntax of the Dutch extended adjectival projection. Natural Language and Linguistic Theory 15.2: 289-368. 2000 Degree adverbs as displaced predicates. Italian Journal of Linguistics, 12.1: 155-191. Emonds, Joseph E. 1985 A unified theory of syntactic categories. Dordrecht: Foris. 1999 How clitics license null phrases: A theory of the lexical interface. In: H. van Riemsdijk (ed.), 291-367. Fujimura, Osamu (ed.) 1973 Three dimensions of linguistic theory. Tokyo: TEC. Grimshaw, Jane 1991 Extended projection. Unpublished Manuscript, Brandeis University. Lobeck, Anne 1995 Ellipsis: Functional Heads, Licensing, and Identification. Oxford, Oxford University Press. Pinkster, Harm and Inge Gen6e 1990 Unity in diversity: Papers presented to Simon Dik on his 50th birthday. Dordrecht: Foris. Riemsdijk, Hendrik C. van 1998 Categorial feature magnetism: The endocentricity and distribution of projections. Journal of Comparative Germanic Linguistics 2: 118. 1990 Functional prepositions. In: Harm Pinkster and Inge Genee (eds.). 1999 Clitics in the Languages of Europe. Vol. VUI of Language Typology. Berlin: Mouton de Gruyter Riemsdijk, Hendrik C. van and Riny Huijbregts forthcoming Location and Locality. In: Marc van Oostendorp and Elena Anagnostopoulou (eds.) 0/

one

This description can be implemented in the present framework by stating that a has a null allomorph that must be used in the specified context, i.e., when the following word is pro-form one.7 Note the contrast with the sequence the one, which though ungrammatical in many contexts (e.g., *First the dog barked and then the one bit me) is possible in postcopular position, as we will see in (22), therefore not ruled out by a restriction of the kind in (4). A restriction parallel to that in (2) and (4) does occur with the numeral one. As (5a) shows, the numeral and the empty Ν are compatible in general, but when they would have been linearly adjacent only one instance of one surfaces (5b). (5)

a. b.

I'll take one blue one and one green one. I'll take (*one) g one with blue stripes and (*one) one with green stripes.

Thus, along with (4) we need an implementation of (6): (6) one 0/ one Given the very close relationship between determiner a and numeral one (cf. Epstein (1999) and sources cited there), these are presumably collapsible into a single rule; indeed, Perlmutter (1970) argues that a is the weak form of the numeral one? Having dealt with this complication, let us return to the paradigm in (1) and ask why an empty Ν is required in these environments. I argue that these data implicate two separate requirements: restrictive modifiers must have a Ν head to modify, and clitic determiners must have a Ν to cliticize to. I motivate these claims concurrently. English provides an alternative to the use of one for avoiding repetition of a content noun: under certain circumstances a Ν head of DP can be gapped (Jackendoff 1971; cf. Abney 1987), in very much the same way as

132 Carson Τ. Schütze the more familiar process that gaps verbs. The (b) examples in (7) and (8) show Ν gapping at work. (7)

a. the letter for b. the letter for c. *the letter for d. *the letter for

my broker and the one for my accountant my broker and that for my accountant my broker and the for my accountant my broker and that one for my accountant

(8)

a. the letters for b. the letters for c. *the letters for d. *the letters for

my broker and the ones for my accountant my broker and those for my accountant my broker and the for my accountant my broker and those ones for my accountant

The (a) and (b) sentences in each group are entirely synonymous. At first blush that would seem to be a problem for my analysis: if one is a last resort default, it should appear only when its presence is demanded; given the (b) alternatives without one, why are the (a) versions with one also possible? The answer is demonstrated by the ungrammaticality of the (c) examples: with the clitic determiner the, Ν gapping is impossible and one is obligatory. The (b) examples involve nonclitic determiners that and those, so they are grammatical without one. Furthermore, as the (d) sentences show, the nonclitic determiners are impossible with one as paraphrases for the (a) and (b) sentences; this is the evidence that one is indeed a last resort: when not needed to support a clitic, it must be omitted. Notice additionally that we cannot replace ones with a contentful Ν in the (d) sentences and get a grammatical paraphrase either: (9)

*the letters for my broker and those letters for my accountant

The reason is that those has the wrong meaning here - specifically, when it co-occurs with an overt Ν it must have demonstrative meaning, whereas in the (b) sentences it has the same semantics as the. Therefore, that and those, in addition to their usual function as demonstratives, are also the nonclitic allomorphs of proclitic the. (The claim that they are not demonstratives in this context is further supported by the observation that their proximate demonstrative counterparts this and these are completely impossible in (7b) and (8b) on the relevant reading.) I assume that the clitic/nonclitic status of these words is not relevant in the syntax; the (a) and (b) pairs can then be derived from the same syntactic input. In the (b) examples, the Ν is gapped (i.e., marked as unpronounced) before vocabu-

Semantically empty lexical heads as last resorts

133

lary insertion occurs. When it comes time to choose the appropriate allomorph of the determiner, since the context is now one where no Ν host for a clitic is available, the nonclitic variant (that or those) must be inserted. By contrast, if Ν had not been marked as gapped then the unmarked clitic form of the determiner would be inserted, and the last resort circumstance for insertion of one would obtain: without pronouncing one in that Ν position, the clitic would be stranded, leading to a PF crash.1 ' n Given that Ν can be gapped, we must return to (1), (2) and (5) and explain why one is obligatory in those examples, that is, why gapping is not an option. There are three subcases here. First, compare the ungrammatical examples in (10) to the grammatical ones in (7b) and (8b): (10)

a. *the happy student and the lonely b. *We met students from Belgium and they met

from

Sweden.

The reason gapping is impossible here is that insufficient material is left in DP after deleting the head N. Specifically, Ν gapping is like V gapping - and unlike ellipsis - in requiring remnant material on both sides of the gap site (Jackendoff 1971; Lobeck 1995); for valid Ν gapping, there must be overt material within DP on both sides of the gap: (11) *They talked about people from Boston and about from New York. (cf. ...andabout some from New York) The second subcase is illustrated in (12). (12) *I like the student from Mexico, and I adore that from Canada. Here we have satisfied the remnant requirement, but the intended meaning is not available because that as a nondemonstrative allomorph of the is [-human]; compare (7b). (The fact that its demonstrative homophone that lacks this restriction is further evidence for separating them.) The plural those does not share this restriction, so the plural version of (12) is fine: (13)

I like the students from Mexico, and I adore those

from

Canada.

The third subcase of ungrammaticality also satisfies the requirement for two remnants:

134 Carson Τ. Schütze

(14) *the attractive candidate from Mississippi and {the/that} ugly from Missouri This differs minimally from grammatical examples like (7a,b) in having a pre-nominal adjective. As is well known, the pre-N modifier position is highly restricted in English, which has led to the suggestion that modifiers can arrive there only by head-adjunction to Ν (e.g., Sadler and Arnold 1994). This provides a natural account of (14): it would involve gapping just one half of a complex head, which is evidently disallowed. Before examining other formal restrictions on the distribution of one, let us consider its semantic restrictions.

2.2.

Semantic restrictions

The classic semantic restriction on one is that it cannot co-occur with arguments (Lakoff 1970, crediting Baker). (15)

a. *the top of the hill and the one of the mountain b. *tops of hills and ones of mountains c. *He managed to refute the claim [that Fred was in jail] but not the one [that Sue was on the loose].

Traditionally this was explained by stipulating that one is an N-bar proform; if all arguments are under N-bar, then none of them can remain behind when we substitute one for the N-bar (Jackendoff 1977). This analysis is not available to me because I claim that one is an N°. Furthermore, under a theory such as Bare Phrase Structure (Chomsky 1995a) there is no way to directly refer to the single-bar level. (There are more recent analyses treating one as a functional head, e.g., Kester (1996) and Campbell (1996); I will not review these here.) Instead of the N-bar stipulation, then, I explain the restriction in (15) by proposing that semantically empty Ν cannot project a complement position in the syntax, because by definition it has no inherent semantics (i.e., no 12

lexically specified meaning), hence no argument structure. Modifier PPs and APs can co-occur with one because modifiers do not need to be syntactically projected by the elements they modify; this is in some sense the very definition of the argument/modifier contrast. Now it may appear that I have created another problem: in (16) one seems to have an N-bar meaning, not an meaning.

Semantically empty lexical heads as last resorts 135 (16)

I met the member of parliament from Peel, and the one from Brighton, (one = member of parliament)

To explain this fact I appeal crucially to different levels of representation. Specifically, I claim that in the syntactic structure of (16) the DP one from Brighton contains an Ν and a modifier PP, but no structure corresponding to an argument (i.e., sister) of N. Rather, one manages to mean member of parliament because at the level of interpretation (say, LF) it copies its meaning from its antecedent. The situation is completely analogous to the fact that he can take as antecedent a complex DP like the man standing over there, even though no one would claim that in the syntax he actually contains all the internal structure that its antecedent has. The only trick is to ensure that the antecedent meaning that is being copied is semantically compatible with the position into which it is being copied; that is, one and member of parliament need to be of the same semantic type, as far as their external behavior is concerned. But this is surely true: both of them are nominals with no unbound argument positions, i.e., one-place predicates. Thus, the key to explaining how one can have N-bar meanings while being just an Ν syntactically is the claim that empty elements such as one receive their interpretation only at LF, and relations of antecedence at LF are sensitive to semantic type but not to syntactic category. Putting it another way, member of parliament is a possible N° meaning - indeed, in Canadian English the noun parliamentarian has this meaning - and so it can serve as antecedent to a semantically empty Exactly the same approach will cover cases where the semantic antecedent of one is bigger than Ν + arguments, but includes modifiers as well: the addition of restrictive modification does not alter the semantic type of the expression. Notice that an example like (15a) ought to be interpretable at LF under this view, if one takes as its antecedent the meaning top(x), i.e., something with an open argument position into which the mountain could be slotted. But (15a) is already ungrammatical because the syntax cannot generate its surface string: a PP like of the mountain can only be an argument, and one does not project any argument positions. There is another, rather different kind of semantic restriction on empty one: its antecedent must be a count noun, not a mass or abstract noun: (17)

a.

*Joe wanted the expensive furniture but Jan wanted the cheap one. b. *We sympathize both with the despair caused by the drought and with the one caused by the tornado.

136 Carson Τ. Schütze Is this restriction compatible with the claim that one is semantically empty? Recall that semantic emptiness was defined (following Emonds) as absence of purely semantic content, that is, bits of meaning not relevant for syntax. Clearly, the difference between count and noncount nouns is relevant for syntax: among other things, the determiner a, unlike the, is compatible only with the former. So the easy answer is to say that one is empty in the intended sense, but happens to have one syntactico-semantic feature in its lexical entry. A potentially more interesting avenue to pursue would be to ask whether the restriction is more than a lexical idiosyncrasy, that is, whether noncount nouns might actually not be possible antecedents for Ν in general. (The fact that gapping is allowed, as in the despair in Guatemala and that in Venezuela, is irrelevant, since the unpronounced noun in the second conjunct could be despair, in which case at LF no copying of meaning is required.) It is also important to consider restrictions that are not placed on the meaning of empty one. In particular, we have seen that it is not restricted to [+human] antecedents. As a result, I differ from Emonds (1985) and Epstein (1999) in not treating the word everyone as containing an instance of the empty Ν one. The reason is that everyone can only be [+human], and this restriction holds neither of one nor of every; I conclude that everyone (and similarly anyone, etc.) must have its own lexical entry.

2.3.

Distributional restrictions

We return now to distributional restrictions on one. If one is no more or less than a semantically empty N, we need an explanation for situations where contentful Ns occur but one does not. Here are some typical cases: (18)

a. I bought some/several pizzas. b. I bought some/several. c. *I bought some/several ones.

(19)

a. Bill's car arrived today. b. Bill's arrived today. c. * Bill's one arrived today.

(20)

a. Some dogs don't like chicken, but all dogs like beef. b. ...but all like beef. c. *...but all ones like beef.

Semantically empty lexical heads as last resorts 137 As we see, certain DP-initial material allows optional omission of a contentful noun, and disallows one in place of a contentful noun. Note that the omissions cannot be the result of gapping, given the previously motivated constraint on it: there is no remnant following the (potential) gap site in DP. It will be useful to have a structure for DP to refer to in this subsection; I shall use the schematic tree in (21) for this purpose. Space restrictions do not permit any justification of the assumptions embodied in (21), and specific details may not be crucial to the analysis, but I hope that the general picture I have in mind is sufficiently clear and plausible. Among the many sources relevant, the reader may consult Abney (1987), Ritter (1992), Epstein (1998, 1999), Longobardi (in press), and works cited there; my tree is modeled very closely on Epstein's, but I distinguish Quantifier Phrase (QP) from Quantity Phrase (QtyP).

(21)

DP Spec all

D' D the this that these those

QP QtyP

Q

some every each any no

Qty many few several two one a

NumP Num [sg] [pl]

NP NP

CP

N'

RelCl

Ν Ν Returning now to the badness of the (c) examples in (18)-(20), I propose that the relevant generalization is that one(s) is ruled out because it would 13

be redundant, in a certain formal sense. First, consider the structure of the DP in an example like I bought several. There is no overt Ν in this DP; if Ν

138 Carson Τ. Schütze

could not have been gapped, and if Ν positions are filled by one in the absence of any other N, then such examples must imply that the Ν position has been completely omitted from the structure. The possibility of a DP containing no Ν position is not novel: this was Abney's (1987) proposal for analyzing "intransitive determiners" such as this in a sentence like I like this. Suppose the grammar makes the option of N-less DPs available in general, not just with determiners like this but with other material above Ν in DP, including quantity expressions like several. Then (18a) and (18b) involve different syntax, (18a) with an Ν (projection) and (18b) without. It is clear that we could not generate (18c) from a structure without an N, since one(s) is an Ν and nothing else. Why can we not generate (18c) from the structure with an N, substituting ones for pizzas'? The answer I put forward is that the relevant syntactic structure is ruled out by a ban on vacuous projection. Consider what information is contributed to a DP by its Ν head. In the normal situation, this is information about the kind of entity to which the DP refers - pizza in (18a), for example. With the kind of articulated DP structure generally assumed today, all other information about the meaning of the DP originates outside N(P). In particular, number, specificity, etc. are encoded in functional heads above NP. What that means is that a DP whose Ν head is empty, as would be required to generate (18c), has no use for that Ν - it contributes no meaning whatsoever. Since DPs can exist without Ns (e.g., (18b)), the empty Ν could just as well be omitted. My proposal is that it must be omitted under these circumstances, because it constitutes structure that has been projected for no reason. Specifically, I posit a ban on vacuous projection that states that contentless structure must be omitted when possible (cf. Speas 1994). This will of course never happen when a semantically contentful Ν is part of DP, but it is always in danger of happening when Ν is empty.14 We have accounted for the pattern in (18)-(20), but do we now predict that some grammatical uses of one discussed earlier should be ungrammatical? The crucial point is that arguments and restrictive modifiers occur inside NP, as is clear from their meaning. Therefore, whenever any of these is present, failing to project NP will not be an option. That is, NP is by definition contentful if it contains an argument or a restrictive adjectival, PP, or relative clause modifier; when any of these elements is present, NP must be projected, and we expect its head Ν to surface, unless the conditions for gapping it are met.

Semantically empty lexical heads as last resorts 139 This correctly rules in all of the grammatical uses of one illustrated above; an additional case that might now be problematic is exemplified in (22).15 (22)

a. b. c.

These are the ones. He's the one. My mother is a doctor, and my father is one too.

According to what has been said so far, we now have two problems: (22a,b) should be ungrammatical, and their meanings should be expressed by (23a,b).16 (23)

a. b.

*These are those. *He's that.

I defer the second problem. The first prediction is almost correct, actually the one(s) is not a grammatical DP in any position other than postcopula: (24)

a. *The ones have arrived. b. */ saw the ones.

This suggests that (22) calls for some special exemption to the ban on vacuous projection. It may be that bare (the) one(s) is possible only as a predicate NP and not as a referential DP (cf. Dahl 1984).17 Supposing this is correct, one might speculate that this counteracts the ban on vacuous projection in the following way. Predicate nominals differ from referential nominal expressions in lacking the functional projections/features having to do with reference; as a result, failing to project NP is not an option because there would not be enough material left above it to constitute an interpretable expression. (The analysis of the and a in English predicate nominals is unclear, but we do know that they do not have the same semantics as in referential DPs, and in closely related languages determiners do not appear in this environment.) Let us now consider some further consequences of the proposed ban on vacuous projection. It correctly rules out (25a) since the presence of N(P) is forced neither by a contentful Ν nor by restrictive modifiers.18

140 Carson Τ. Schütze

(25)

a. */ bought ones. b. *I bought. c. I bought some. d. */ bought some ones.

Unfortunately, we seem to predict that the grammatical counterpart to (25a) should be (25b), whereas it is actually (25c). (We do nonetheless correctly predict that (25d) is not an alternative to (25c).) To address this shortcoming I invoke a principle that can override the effects of the ban on vacuous projection, a principle of recoverability that requires that the presence of DP must be signaled by some phonological material; in the absence of a contentful Ν or modifier, this will force the choice of a nonnull allomorph of one of the functional heads in DP. This does not literally entail overriding the structural impact of the ban on vacuous projection; rather, it constrains how the functional structure can be pronounced when the lexical NP has not been projected. This is therefore a Spellout condition on recoverability of deletion, in keeping with more standard appeals to this notion. Thus, the steps of deriving (25c) are conceptually as follows: the meaning to be expressed in the DP consists of indefinite quantity {some) of some unspecified entity (to be interpreted via an antecedent at LF). With no lexical content for the potential NP, it cannot be projected. This leaves just the functional structure, with the indefinite quantity meaning. At Spellout, we find that that meaning can be expressed by either the word some or a null allomorph thereof. Since choosing the null allomorph would leave no pronounced material in DP, that option is ruled out, so the only realization that survives is to pronounce some. In contrast to the paradigm just above, there are environments where one(s) appears to be in free variation (with varying degrees of preference) with a higher functional head:19 (26)

a. b.

I promise to buy some that you 7I enjoy. I promise to buy ones that you 71 enjoy.

This is once again a prima facie problem for my last resort claim. Notice that unlike (25), however, (26a) is a potential gapping structure, since there is both pre- and post-head overt material within the DP. That explains why these two options exist, but not why the structure equivalent to (26a) before gapping is not also grammatical. (27)

?

*/promise to buy some ones that you'll enjoy.

Semantically empty lexical heads as last resorts 141

This is not the same case as (25d) because the relative clause demands that NP be projected. I suggest treating (27) in a manner analogous to the the/that alternations discussed in (7) and (8) above: in this case, the allomorphs are some and 0, with 0 the default and some being forced in (26a) by the recoverability condition. The latter requires a refinement now: since we cannot say *I promise

to buy that you'll enjoy on the relevant reading,

recoverability is evidently not satisfied by modifiers: it must demand an overt element in either the head Ν or the functional material above it (which makes some intuitive sense, modifiers not being an integral part of 20

the phrase that contains them). There are numerous other small details to be taken care of as part of a full account of the interaction between one(s) and the determiners and quantifiers above it (see Epstein (1999) for relevant observations), but for reasons of space I do not explore those here, turning instead to one final sort of case. 1 As we saw above, except in certain predicative contexts a DP cannot consist exclusively of a definite determiner plus one(s); what surfaces instead is a pronoun (cf. Jackendoff 1968; Postal 1969). (28)

a.

*/ saw the one. (cf. I saw the guy.)

b. c. d.

I saw him/her/it. */ saw the ones. I saw them.

This is another application of the ban on vacuous contains no meaning and cannot be projected. What intransitive determiner, which surfaces as a pronoun The same explanation applies to the following sort digm:23 (29)

a. b. c. d.

projection: the NP is left is simply an (cf. Abney 1987).22 of possessive para-

*my one(s) mine *her one{s) hers

We can now explain the badness of (23), deferred above. Nonclitic nondemonstrative determiners that and those are used only as transitives, i.e., they must take an NP complement (whose Ν may eventually be gapped); (23) cannot contain a gapped Ν due to the lack of a following remnant.

142 Carson Τ. Schütze Syntactically intransitive D must be a pronoun. Thus, while (22a) cannot be paraphrased as (23a), it can be colloquially paraphrased as (30): (30) 3. 3.1.

These are them. Empty A: So Core cases

The examples in (31) demonstrate environments where so is freely interchangeable with an adjective, and may not be omitted. Nonomissibility is of course what we expect from a last resort default. (31)

a. b. c.

Mary is quite intuitive, but Sue is more *(so/intuitive). John seems exhausted, and. Mary seems *(so/exhausted) too. John's work is not yet world-class, but it will no doubt become *{so/world-class) eventually.

Although cases like (31) are greatly outnumbered by cases where this intersubstitutability between so and contentful A does not literally hold, I aim to show that treating so as being of category A (and no other - see section 3.4) is correct. Except where noted, I assume the analysis of English adjectival phrases presented by Corver (1997), developing insights from Bresnan (1973); many of my examples are adapted from Corver as well. As was the case with one in section 2, so substitutes not just for bare adjectives but also for adjective phrases: (32)

a. b. c. d. e. f. g.

John seems fond of Mary, and Bill seems {so/fond of Mary) too. I know a man very eagerfor work and a woman equally {so/eagerfor work). John seems too tall to serve on a submarine, and Bill seems {so/too tall to serve on a submarine} too. John is fond of Mary. Bill seems much less [so/fond of Mary}. Of all the careless people, no one is more so/careless than Bill. Brett's work is not yet consistent in style and quality, but will no doubt become {so/consistent in style and quality}. Fred may not be the most reliable employee we have, but he is reasonably so/reliable.

Semantically empty lexical heads as last resorts 143

I analyze this possibility for complex antecedents just as I did for one: as long as the meaning supplied by the antecedent of so is of the same semantic type, the fact that one is a complex phrase while the other is just a head is not a problem.24 Note also that (32b) illustrates the use of so as a restrictive modifier within DP, whereas the other examples have been predicative uses (for reasons that will become clear).

3.2.

Semantic restriction

The pro-form so shares with one the restriction that it is incompatible with an overt argument of A (33), but compatible with modifiers (examples above): a. ?*Fred is very proud of his son, but he is less so of his daughter. b. *John was entirely convinced of my sincerity, but only partly so of my determination. Again, the analysis proposed for Ν carries over to A, as does the alternate strategy of gapping the head, in which case argument PPs may be stranded:25 (33)

(34)

3.3.

John is proud of his father, and Fred is

of his mother.

Distributional restrictions

The examples so far have been craftily selected to conceal the fact that so is probably not in the same surface position as its contentful counterparts. The most straightforward evidence involves the position of enough·. (35)

a.

b.

John is good at mathematics. He seems {?enough so/*so enough} to enter our graduate program, (cf. good enough, *enough good) Mary isn 't very efficient, but Sue is more than {enough so/*so enough}, (cf. more than efficient·, enough tj)

Following Corver (1997) and work cited there, I assume the position of so in (35) reflects the underlying position of adjectives; contentful adjectives obligatorily undergo fronting across enough, but so evidently cannot. Corver argues that the order good enough is derived by head movement of

144 Carson Τ. Schütze the adjective to adjoin to the quantifier. Intuitive evidence for this is the fact that the semantic constituency of more than efficient enough is clearly [more than enough] [efficient].26 Corver's explanation for the failure of so to undergo this process is that it is an unanalyzed phrase, not a head, therefore not eligible for head adjunction. Obviously this account is not available to me since I treat so as a genuine A head; for the moment I will simply stipulate that A-to-Q raising is restricted to contentful adjectives; I return to this point in the Conclusion. A second consequence of so's inability to head-raise is that when it combines with certain degree words, a dummy form much must be inserted: (36)

a. b. c.

John is fond of Mary. Maybe he is too much so. John is fond of Mary. Maybe he is as much so as Bill. The weather was hot in Cairo - so much so that we stayed indoors all day. d. Uohn told me he was afraid of spiders, but I wonder how much so he really is.

Not all degree words behave this way, however: looking back at (31a) and (32d), we see that more and less co-occur with so without much·, indeed, insertion of much is not even an option: (37)

a. *more much so b. Hess much so

A third class of degree words are referred to by Corver as adjectivals: reasonably, terribly, completely, extremely, highly, unbelievably, etc. These behave like more/less and unlike so/too etc.: they co-occur with so without dummy much: (38)

a. b.

reasonably {*much) so completely (*much) so

Corver argues that these three classes of modifiers occupy three different structural positions, and I will assume that this is correct. In (39) I give a tree that schematizes my adaptation of Corver's structures and the elements that can appear in each of the positions in the tree; aside from treating so as a head, this faithfully encodes his proposals, except that I have been more literal in the treatment of comparative and superlative morphology.

Semantically empty lexical heads as last resorts 145

(39)

DegP Deg'

Spec (adjectival degree modifiers)

far sufficiently DegP

Deg as how 27 SO too that

QP Spec

Q'

(adjectival degree modifiers)

extremely i/te[superl] sufficiently very well ,28 much far rather DegP

Q

more -er most -est enough less least much 0

AP

|

A intelligent smart aware different so

The tree shows that so/too etc. are Deg(ree) heads, which Corver characterizes as determiner-like. More, less, and a couple of others are what he describes as quantifier-like; they are hosted by a Q head. Finally, the adjectival degree modifiers appear in Spec-QP. The generalization about the behavior of so can provisionally be stated as follows: if Q and/or SpecQP are occupied by a degree expression, so is licensed without much·, if only Deg and higher positions are occupied, dummy much must be inserted in Q. Put another way, either the head or the specifier of Q must always be occupied. Corver argues that this pattern derives from the need for a very local relationship between Deg and an element that it modifies: that element must be in the closest head below Deg. As a result, contentful adjectives must raise from A to Q in expressions like too tired. But when empty so is used, recall that it cannot undergo this raising, for whatever reason. Then the only way to satisfy the need of Deg for a modifiee in the Q position is to insert dummy much, which will inherit or copy the relevant part of the meaning of the adjective (its degree argument slot, on Corver's analysis). For our purposes, the motivation for A-to-Q is not so crucial; what matters is that except for not undergoing this process, so is an ordinary adjective syntactically.

146 Carson Τ. Schütze There is a further wrinkle, however. There is another set of degree words, including far, well, and much·, as indicated in (39), Corver places these in the same position as the -ly adjectivals, since they occur to the left of comparatives: (40)

a. far more intelligent b. Iwell sooner c. much less wealthy

For descriptive convenience I refer to this class as the short modifiers. If this treatment is correct, we would predict that the following are grammatical: (41)

a. *far so b. *well so c. *much so

Their ungrammatically might suggest that the short modifiers are patterning more like the Deg words (*too so, etc.), but unlike those, they are also no good with dummy much inserted: (42)

a. *far much so b. *well much so c. *much much so

Corver suggests that the difference between -ly modifiers and the short modifiers is lexically idiosyncratic, in the sense that semantically similar 2Q words differ on their compatibility with so: (43)

a. extremely different b. far different c. extremely so d. *far (much) so

(44)

a. greatly different b. much different c. greatly so d. *much (much) so

Semantically empty lexical heads as last resorts

(45)

147

a. entirely aware b. well aware c. entirely so d. *well (much) so

He notices that none of the more restricted short modifiers is marked with -ly, and tentatively suggests an account based on Case, which I do not review here. As Corver acknowledges, it is not at all clear what is wrong with (43d), (44d), and (45d) when one considers that the presence of a comparative in Q often rescues the short modifiers (46) - so far, the difference between contentful versus dummy Q has made no difference to so. (46)

a. b. c. d. e.

far more so far less so much more so much less so 77 well more so

I suggest, also tentatively, that the fact that it is specifically comparatives in Q that rescue these modifiers is crucial. Notice that, contra what the tree in (39) might suggest, there is another asymmetry between the two classes of elements in Spec-QP: while the -ly words combine quite freely with bare adjectives, the short modifiers are highly restricted: (47)

a. b. c. d. e. f. g.

{extremely/sufficiently/reasonably/?highly/completely} happy/smart *well happy/smart, *far happy/smart, *much happy/smart well aware, *far aware, llmuch aware *well tall, *far tall, *much tall *well different, far different, much different Iwell discussed, *far discussed, much discussed Iwell maligned, *far maligned, much maligned

Contrariwise, with comparatives the short modifiers are quite productive, while the -ly modifiers become more restricted: (48)

a. b.

far/much {happier/smarter/more aware/more different/better situated/taller] well more intelligent, well more aware, well taller

148 Carson Τ. Schütze (49)

considerably taller, Reasonably safer, *extremely smarter

What all this suggests to me is that the short modifiers are semantically licensed only if they co-occur with a comparative or something similar to it. (E.g., far different is possible because different is inherently comparative.) Obviously there are many cases not covered by this vague hunch, but let us suppose there is some such dependency. The explanation for the contrast between *far so and far more so would then go as follows: modifiers need to be semantically licensed, these particular modifiers need a special kind of licenser (roughly, a comparative element), and so is literally semantically empty - it gets interpreted at LF by copying the meaning of an antecedent, but before that point in the derivation it has no semantic content. Clearly, then, so cannot license the short modifiers, because it lacks the crucial semantic feature(s), and the Q much cannot provide those features since it too is semantically empty, but more possesses the relevant semantics, so once it is in place, the (lack of) meaning of so is no longer relevant, and it becomes compatible with these modifiers. Another seemingly troublesome discrepancy between the distributions of so and contentful adjectives is that so can never occur in pre-N position: (50) *a happy man from Canada and a so woman from Mexico As noted above in (32b), DP-intemal attributive use of so is in general possible; the problem is specifically the A-N order. As in section 2,1 will take the common view that adjectives are underlyingly to the right of Ν in English, with the A-N order derived by head movement of the adjectival material, which explains the very tight restrictions on what can accompany the preposed adjective. Corver could presumably explain the gap in (50) by 30

once again appealing to the irreducibly phrasal nature of so. My proposed alternative is to tie this restriction back to the absence of A-to-Q raising of so within the adjectival projection: although details remain murky, it seems not unreasonable to suggest that the head raising to Ν targets the highest head in the extended projection of an A, which in the absence of an overt Deg head would be Q. Furthermore, there is evidence that elements lower than the Q head can be stranded when Q head-adjoins to N, from examples like (51), where the meaning makes the necessary movement quite transparent, and from a puzzling construction mentioned by Corver (52).

Semantically empty lexical heads as last resorts 149 (51)

a better, book [oegp... tj than this ]

(52)

the cheapest ticket possible to New York

The underlying structure of (52) is (53), whence cheap head-raises to -est forming cheapest and then adjoins to the noun ticket on the left, stranding possible in its base position to the right. (53)

31

[qp [q -est ... possible ... [ap cheap ]]]

If it is indeed Q that head-raises out of the extended adjectival projection to adjoin to N, then clearly so will never be a candidate for this raising; what is perhaps still in need of explanation is why the Deg and/or Q heads cannot raise and strand so\ (54)

a. *a more person so (cf. a more intelligent person) b. *a too train extremely so, *a too extremely train so (cf. Ία too extremely fast train)

However, there must independently be some constraint against stranding most sorts of material post-nominally when a head left-adjoins to N: (55)

a. *a proudiman ... [ap t, of his son] b. *a too; train [negp tj terribly fast] (cf. a train too terribly fast, Ία too fast train)

Thus, I find it not unreasonable to assume that the fact in (50) can be made to follow from the assumptions I have already made and does not challenge the status of so as an adjective. While there remain numerous puzzling restrictions on so, I hope to have at least shown that the hypothesis that so is an A is a plausible one.

3.4.

Other uses of so: Is it more than an A ?

Given the close, productive relationship between adjectives and adverbs (they are argued by Abney (1987) to be the same category, in fact), it is no surprise that so also behaves like the default adverb:

150 Carson Τ. Schütze (56)

a. b. c.

The police searched the big room carefully, but the small room less so. John drove somewhat erratically, and Mary drove even more so. They did run quickly, but not enough so to escape from the bear.

This I do not take as a challenge to so being fundamentally an adjective. However, it has been claimed (e.g., by Corver) that so is much more than that: a general pro-form for predicates, including NPs and PPs as well as APs. Despite appearances that quite favor this possibility, I shall argue against it. The relevant data serve to highlight the distinction I have already been crucially drawing on, that between surface distribution, which is sensitive to syntactic category, and interpretation, which cuts across syntactic categories to a certain degree. Let us begin with the non-AP predicates in (57). (57)

a. b.

John was somewhat [under pressure], and Fred was even more so. John was [a fool], but Bill was even more so.

Here the literal antecedents for so are the bracketed PP and NP. Does it follow from this that so must belong to multiple categories, or even be category-neutral? I believe the answer is "No", for reasons parallel to those given above for why one need not be an N-bar when it takes an N-bar antecedent. I propose that at LF so can inherit the meaning of a PP or NP if these match its logical type, i.e., a one-place predicate. Thus, there is so far no evidence for treating so as anything other than an A. What about the following uses of the word so: (58)

a. b.

John is a fool, and so is Mary. Mary is out of control, and so is John.

It is easy to see that so here is not a pro-form of any kind, but rather just an adverbial (as noted also by Quirk et al. (1985)). First of all, substituting the potential antecedent for so is ungrammatical: (59)

a. b.

*John is a fool, and a fool is Mary. *Mary is out of control, and out of control is John.

Semantically empty lexical heads as last resorts

151

Secondly, as Emonds (1985) points out, clause-initial so is equivalent to clause-final too: (60)

a. b.

John is a fool, and Mary is too. Mary is out of control, and John is too.

This intuitive parallel suggests an alternative account of (58), namely that it involves VP-ellipsis, with so as a kind of focus marker that triggers inversion when fronted:32 (61)

Mary is also a fool. —> Sot is Mary t; a fool.

Another construction adduced by Corver involves the pro-form do so: (62)

Fred takes long lunch breaks, and we do so as well.

The proposed analysis alluded to by Corver is that so stands in for the predicate take long lunch breaks, and do is a dummy inserted to support tense. Neither half of that treatment can work, however. First of all, do so contains main verb eventive do, not the dummy/auxiliary, as is clear from (63a), where a modal auxiliary precedes it, and from (63b), illustrating that it is incompatible with stative verbs. (63)

a. Fred will take a long lunch break, and we should do so too. b. *Fred feels tired, and we do so too.

Secondly, in an example like (63a) we cannot replace so with the antecedent predicate: (64) *Fred will take a long lunch break, and we should do take a long lunch break too. Thus, although do so as a whole somehow takes on the meaning of an (eventive) VP (see section 4.6), the contribution of so to this process cannot be unified with its adjectival uses. Finally, consider one more set of facts that seem to suggest that so is more general than an A. The antecedent for so can be bigger than an AP: in the following sentences the antecedent is a QP (65a,b) or DegP (65c).

152 Carson Τ. Schütze (65)

a. b. c.

John seems [more intelligent than Bill] and Mary seems so too. John seems [less conscious of the consequences than Sue] and Bill seems so too. John seemed [too proud of his accomplishments], and Fred became so too.

Does this imply that so itself can be a QP or DegP? Facts that we have already seen indicate that it cannot. Specifically, if so really could be a QP by itself, it would become mysterious why it cannot play that role in the presence of a Degree head: (66) *John seems mildly agitated, but Mary seems too so. Corver notes this problem for his analysis and tentatively suggests an ad hoc solution.33 In contrast, under my account the badness of (66) is directly predicted: Deg must take a QP complement, so is only an A(P), so there will be an illegally empty Q head in (66), calling for dummy much. This remains consistent with the ability of so to take a DegP or QP antecedent: once again, we must distinguish the syntactic category of so, which governs its surface distribution, from its semantic category, which is compatible with antecedents of type AP, QP, or DegP, because these are all one-place predicates.

4.

Empty V: Be

In this section I seek to defend the strong claim that be is the default verb of English and nothing more, that is, that be is devoid of semantic content in all of its uses, that it is always a verb and never an INFL head, and that it is used only when some formal requirement demands it. The relevant formal requirements are greater in number and diversity than those we have examined above, which results in a blatantly heterogeneous distribution that is the hallmark of a default. The idea that be is "empty" and merely "supportive" is of course not new: some of the major recent works that explore it include Bach (1967), Schölten (1988a, 1988b), Dechaine and Tremblay (1998), Ouhalla (1991), Heycock (1994), and Rapoport (1987). However, to my knowledge the range of data discussed here and the three aspects of the default claim have not been brought together before.

Semantically empty lexical heads as last resorts

4.1.

153

Semantic emptiness

Support for the claim that be has no inherent meaning comes from the mundane observation that there are numerous environments where its omission and insertion apparently make no difference to the meaning of the sentence. Potential examples of such environments are given in (67); I will return shortly to the fact that in some of these cases, notably (67d), (67e), and (67f), one's intuition may be that there is some subtle change in implications when be is inserted/removed. Still, I think it will be agreed that if one were asked what be itself means in any of these sentences, no answer would come readily to mind. (67)

4.2.

a. b. c. d. e. f.

With John (being) so tired, we decided to stop for the night. You described her as (being) fitssy. I consider him (to be) a fool. The medicine made her (be) tired all the time. I saw her (be) rescued twice. What?? Him (be) three hours late?! Impossible!

Formal requirement 1: Providing Case

Interestingly, optionality of this type is typically not fully general. Environments like (67) vary in this regard, as a function of the contents of the subordinate clause; I exemplify three of the constructions here, but the others work similarly. (68)

a. b.

I consider this crowd (to be) too big for the elevator to hold. I consider there *(to be) too many people in this elevator.

(69)

a. b.

John was considered (to be) the winner. The winner was considered *(to be) John.

(70)

a. b.

I would like to see fewer mistakes (be) made next game. I would like to see there *(be) fewer mistakes made next game.

(71)

a. b.

With John (being) so tired, we had to stop. With there *(being) no more daylight, we had to stop.

This is evidence for the first of the formal requirements that can force be to appear: the need for referential DPs, but not predicative nominals, to get

154 Carson Τ. Schütze

Case, combined with the ability of be to assign Case (cf. Rapoport 1987, for Hebrew; Lasnik 1992, 1995; Heycock 1995; Campbell 1992; Law 1996; Schütze 2000). That plays out in the examples in the following way. In (68), (70), and (71) the (a) sentences contain a predicate following (optional) be, whereas the "associate" of existential there is a referential DP in the (b) sentences and therefore demands Case. In (69), John is a referential DP while the winner is a predicate, so (69b) is an instance of Predicate Inversion (Moro 1997) followed by raising, and John would lack Case if be were omitted. See Rothstein (1987) and work cited there for tests of predicatehood/referentiality. See Lasnik (1995) for the conclusion that there is no Case "transmission" between there and its associate, so that the associate must have its own direct source of Case. Alternative accounts of some of these facts can be found in works cited in this paragraph and in Den Dikken (1994), inter alia. Of course, all this is helpful in explaining (68)-(71) only if be really can assign Case. This has often been assumed not to be true, but largely for irrelevant reasons based in the failure to separate abstract Case and morphological case; only the former is relevant here. More recently, Lasnik (1995) has developed proposals by Belletti (1988) to the effect that be assigns "partitive" Case, with partitivity intended to capture the (in)definiteness effect in existentials such as (68b). I find that untenable as a general claim; as the reader will have gathered, I propose that be assigns Case to John in (69b), in which John clearly does not receive a partitive or indefinite interpretation (see also Law (1996)). Furthermore, the notion that be could be specified as assigning a particular Case is incompatible with my claim that be is the "emptiest" verb. Rather, I would like to say that whatever syntactic properties be has, including the ability to assign Case, must be properties of the category V itself (or functional elements projected by it), and therefore not specified in individual lexical entries.34 There is in fact independent evidence that Case assignment to the "associate" of an existential there clause is occurring, based on Case adjacency effects (72) (cf. Stowell 1981). The important points about this paradigm are that: the form of be is nonfinite, to control for the fact that finite be raises to Tense in English; the adverb that blocks Case assignment is semantically compatible with the sentence (72b,e); the same adverb can intervene between a verb and its non-DP complement (72c,f). (72)

a. b. c.

*There will be regularly a man in the garden. There will regularly be a man in the garden. John will give regularly to his favorite charity.

Semantically empty lexical heads as last resorts 155 d. *There will be frequently a shirt on the floor. e. There will frequently be a shirt on the floor. f. We will speak frequently with the landlord. The same sort of paradigm can be used to show that predicate nominals do not require Case (73), while a postverbal referential DP after be does (74). In (74) I use embedding under seem to make the examples sound more plausible. (73)

a. (7)John has been consistently a nuisance. b. *John has done consistently his homework. c. Mary has been proudly a Marine for ten years now. d. *Mary will join proudly the Marines ten years from now.

(74)

a. (?)Superman seems to be secretly a reporter. b. *Superman seems to be secretly that guy over there. c. ΊSpeedmaster Jen will be easily the winner of the race. d. *The winner of the race will be easily Speedmaster Jen.

Thus, we have independent evidence that the postverbal nominals in (68b)(71b) require Case, and that the one in (69a) does not. The logic of my account for (68)-(71) is therefore as follows. In the (b) variants, but not the (a) variants, a referential DP winds up in a nonsubject position, where it will lack Case in the shorter versions of these sentences - there is either no Case assigner available, or it is not sufficiently local. The way to save the sentence is thus to provide a Case assigner for that DP, while not introducing any new meaning. Since verbs are inherently Case assigners (I claim), this can be accomplished by inserting an empty V head in the embedded clause; empty V is realized as be. (I am of course not proposing this as the actual sequence of steps in a derivation, and V cannot be tossed into sentences willy-nilly; see below.) That this Case-based approach is on the right track is suggested by a further fact, noted by Heycock (1994): when a copular clause contains two referential DPs, and is therefore identificational rather than predicational, omitting be is impossible in both orders, because the second DP will always be in need of Case. (75)

a. b.

I consider the man in the corner *(to be) John. I consider John *(to be) the man in the comer.

156 Carson Τ. Schütze If this treatment is correct, it implies that there can be no merely unpronounced counterpart to be in small clauses such as the shorter version of (68a); that is, small clauses are not uniformly headed by (possibly silent) verbs.

4.3.

Formal requirement 2: Event binding

There are of course many environments where be is always required. The largest class of these arise from a higher head requiring a V when the intended content is expressible without one; elements that do this include many forms 1NFL in various guises (76) (finite indicative, to-infinitive, mandative subjunctive, etc.) and also verbs like let (77). (76)

a. b. c. d. e. f.

Mary can *(Jbe) helpful. Fred wants to *(be) in control. Sue will *{be) a success. Jane might *(be) dancing. It is important that John *{be) happy/the winner. John will *(be) humiliated.

(77)

a. b.

Mary let John *{be) in charge. Mary let John *(be) humiliated by the crowd.

We can see from (76d,f) and (77b) that participles do not satisfy the need for a V that is imposed by INFL. One might think this is because the driving force behind that need is the requirement for affixal INFL to be affixed to a verb, since participles cannot take tense/agreement affixes in English. This is incorrect. Even when an INFL affix has an appropriate host in the form of dummy do, INFL is not satisfied without a genuine verb below it in the clause (78), except on a VP-ellipsis reading, which the examples are designed to make implausible (Rapoport 1987). (78)

a. *John does (not) tired. b. *John doesn't with glee.

As noted in section 1 and elaborated below in section 4.6, dummy do is not of category V in my analysis; rather it can be thought of as an empty modal. Participles (perfect, progressive, passive) are also not of category V - they take their category from participial suffixes, which are acting as

Semantically empty lexical heads as last resorts 157

derivational morphemes in this regard, determining the category of the derived form. Indeed, when finite, be and dummy do are essentially in complementary distribution: do is followed by verbs, be by predicates of other categories (including participles). In fact, INFL of this type has two separate requirements: it morphologically requires an appropriate host for its affixal form, and it semantically requires a V (details below). The two requirements can be satisfied by different words: for example, in the sentence John does not swim, do serves as host for the tense/agreement suffix, but it is swim that satisfies the need for a V. I follow Pollock (1989), Ouhalla (1991), and Rothstein (1999) in assuming that the reason INFL requires a V is that V provides an event variable for Tense to bind. Whether this is a purely formal requirement, as opposed to a semantic one, might be debated. However, I take it as formal on the grounds that languages differ as to the circumstances under which a given Tense meaning needs to be expressed morphologically at all, as well as on whether there must be a verb present to enable a Tense meaning to be available. One possible instantiation of the formal requirement suggested by the cited authors is that Tense is an operator that must bind a variable, otherwise it will violate the ban on vacuous quantification. See Rothstein (1999) for evidence that even stative verbs introduce a Davidsonian event (or eventuality) argument, in contrast to predicates of other categories (she discusses only adjectives), which express genuine states. She characterizes the difference between these two semantic types as follows: eventualities are temporally locatable count entities, whereas states are mass entities that are not temporally locatable. The event binding requirement will now explain some of the facts in (67). Note that these examples are superficially problematic for a default account of be: be should be either required or prohibited, but not optional. The solution is of course that lexical heads can be compatible with complements of more than one syntactic type. For example, consider may take either a small clause or a ίο-infinitive (67c): once the second option has been chosen, though, the needs of INFL (to) demand a V, hence be if 35

the contentful predicate is not a V, and similarly for make in (67d). The real optionality lies in the lexical entry of the selecting verb. The alternation in (67a,b) arises from the fact that various syntactic categories can function as predicates: here, gerunds are alternating with APs. Using a gerund entails the INFL element -ing, which once again demands a verb for event binding; thus, once one fixes the choice of a gerund, V will actually be obligatory. (See section 4.7 for further discussion.)

158 Carson Τ. Schütze 4.4.

Formal requirement 3: Providing a modifiable verbal predicate

So far we have seen two triggers for be: requirements of Case and event binding. A third trigger, less well understood (at least by me), appears to be the fact that certain adverbiale can modify only verbs, not other categories, as described by Law (1996). As a result, an environment where be is otherwise optional ((79a), (80a), (81a)) can be made into an obligatory be environment by the use of one of these adverbials: (79)

a. b.

The professor doesn 't seem {to be) tired/intelligent/out of control. The professor doesn't seem *(to be) around/here/abroad/ away/out of the country.

(80)

a. b. c.

The concert seems (to be) unusually early. The concert seems *(to be) at 8:00. The concert seems V.(to be) over already.

(81)

a. b.

The trip appeared (to be) in jeopardy. The trip appeared *(to be) by ferry.

(82)

The cake seems *(to be) for the children.

(83)

The cop doesn't seem *(to be) alone.

36

The range of examples allows us to rule out some alternative explanations. Specifically, the lexical category of the modifier is not relevant - PPs and adjectives can go both ways. Thematic content also does not make the needed distinction: locatives and temporals go both ways. Most or all of the specially restricted modifiers are stage-level, but so are many that do not share this restriction. Whatever is going on here, it seems unlikely that something about the meaning of be is rescuing these adverbs; its lexical category, however, is not an unreasonable candidate. (All of the special adverbials can of course modify verbs other than be as well.) Suggestive evidence to that effect comes from the fact that none of the predicates in question can occur as prenominal modifiers in DP (this is of course interesting only for the adjectives, not the PPs):

Semantically empty lexical heads as last resorts 159 (84)

a. b. c. d. e. f. g.

*the around professor *the here professor *the abroad professor *the over concert *the alone cop *the away neighbors *the out of the country professor

This indicates that these predicates require the mediation of some other element in order to be predicated of a DP denotation. Evidently V fits the bill, perhaps because at least some of the modifiers describe events rather than entities.37

4.5.

Formal requirement 4: Supporting an affix

There are environments within the INFL system of English where forms of be appear but their presence cannot be reduced to the need of Tense to have a V for event binding. These include the following: (85)

a. b. c.

Mary was being complimented. Mary had been complimented. Mary had been being complimented.

In these sentences, the finite auxiliary is a V, not a modal, so the event binding requirement of Tense is already met. The passive and progressive meanings are carried by -ed and -ing, respectively, and the perfective by have and -en. What do we need the boldfaced bes for, then? I claim that what is forcing be to be inserted in (85) is the need to morphologically support the verbal affixes -en and -ing. These affixes can in general be carried by contentful verbs (Mary was complimenting herself, Mary was complimented), but this is impossible in (85) because the main verb already bears one affix, and in English that is the maximum number allowable. These are therefore genuine instances of fe-support, providing a host for (participial) affixes. Be does not need to be listed as part of the lexical entry for passive or progressive, because its distribution is fully predictable from the rest of the clause structure, given that be is the default V. The scheme I have in mind is the following. All it takes to make a clause progressive, for example, is to insert the -ing morpheme in the appropriate functional head of the INFL system, call

160 Carson Τ. Schütze

it Aspect. We now have an affix in need of a host, but since ^-insertion is a last resort, we first explore other ways to rescue this affix, and we may find that we can do so by raising the main verb to it. (If we cannot, because the main verb already has a participial affix, as in Mary is being teased, then fee-support will apply.) Head raising and adjunction of a main verb to ing creates a complex head with the category of the affix, i.e., progressive participle, crucially not V. As such, it cannot satisfy the need of Tense for event binding, so as a last resort we insert an empty V to perform this function, which will again be spelled out as a form of be, e.g., Mary is dancing", the fact that be happens to bear tense/agreement morphology in this example is a side effect. How do we get have to be inserted in similar fashion just in a perfective clause? I propose that the semantics of perfect includes a temporal priority relation that is encoded by means of a Ρ head that takes participial -en as its complement (see section 5.5). There will again be no V to satisfy the event binding requirement of Tense at the point when we try to combine Tense with the subtree constructed so far, whose highest head is (empty) P, with a perfect participle below it. We insert empty V as before, and by the same (mysterious) mechanism that forces these two empty heads to unite in general (cf. (1 lOd), section 5.4 below), they must unite here too, yielding have.

4.6.

Why do and do so are not the default V

Let us return to (78) and ask why do does not satisfy the event binding requirement itself. There are two possible answers, which I shall not choose between here. One is that the position where do is inserted is actually above the Tense/Agreement position(s) and thus not a candidate for event binding by Tense. The second is that dummy do, in contrast to main verb do (which occurs also in pro-forms such as do so), is not a V at all but rather a modal. As is well known, standard English modals (can, will, etc.) have neither the distribution nor the full inflectional paradigm of true verbs. A detailed account along these lines is presented in Schütze (1997: §5.2), where I suggest that INFL contains an additional head, Mood, under which modals and dummy do are inserted. Indeed, the two solutions I suggested could now both be correct, if Mood is above Tense, as seems plausible. What we would then say about the requirements of INFL (specifically Tense) is that it needs a true V for event binding, but when affixal it morphologically requires something that can bear a Tense inflection, namely a V or Mood oo head. Thus be differs from dummy do by (always) being categorially a V.

Semantically empty lexical heads as last resorts

161

Regardless of the analysis of (78), there are further reasons for rejecting the idea that do rather than be might be the default V of English. One important difference between do and be is that do-support applies only in indicative and imperative moods, and is impossible, for example, in mandative subjunctives (86a), Mad Magazine sentences (Akmajian 1984; Schütze 1997: inter alia) (86b), and gerunds (86c,d,e), even for emphasis; these are all environments where be does occur (87). (86)

a. It is important that John (*do) not (*do) sing. b. What—Him (*do) not pick up the kids ?! c. *[Him doing not respect Mary] was upsetting. d. *[Him not doing respect Mary] was upsetting. e. [Him not respecting Mary] was upsetting.

(87)

a. b. c.

It is important that John (not) be early. What—Him (not) be on time?! [Him (not) being on time] was upsetting.

The restriction on do is consistent with analyzing it as a Mood head, and would be puzzling if it were treated as an ordinary V, since there is obviously no anti-verb restriction in the environments of (86)-(87). At the very least, the elements that do supports are clearly not the same ones that be supports. Let us now consider pro-form do so, typically treated as a VP-anaphor (Lakoff and Ross [1966] 1976; Jackendoff 1977); much the same logic applies to variants such as do the same. These expressions clearly involve the main verb do, since they co-occur with dummy do: (88)

John did not do so.

When it is used as a contentful verb, do assigns an agent 0-role to its subject (Schölten 1988b) and some sort of obligatory theme role to a complement, and it represents an eventive predicate by the standard progressive diagnostic: (89)

a. b.

I'm doing it now. I do it now.

Crucially, these properties carry over to its pro-form use: do so cannot stand in for noneventive predicates (90a), nor eventive predicates without

162 Carson Τ. Schütze an agent subject (90b), and it cannot occur without some sort of complement (90c). (90)

a. *A book lies on the sofa, and a magazine does so too. b. *John is being punched, and Fred is doing so too. c. *Mary is mowing the lawn, and Sue is doing too. (cf. ...and Sue is doing so/doing the same/doing that too.)

It is clear, then, that the lexical entry for pro-form do is far from empty: it contains at least a 0-grid including Agent and Event slots, and perhaps a subcategorization frame. It could conceivably be the default agentive event verb in English,39 but be beats it hands down for emptiness.40

4.7.

Further semantic considerations

As noted in connection with (67), in environments where be is syntactically optional its presence/absence nonetheless seems to induce subtle semantic shifts. This would be problematic for my claim that be is always semantically empty, unless one can explain these semantic effects with reference to the category V itself, rather than the specific verb be. Fortunately, Rothstein (1999) has already done this for a large range of cases, developing the basic insight attributed to her above: verbs always represent eventualities, and eventualities are temporally locatable countable objects.41 For space reasons I cannot recap the arguments she develops, I can only catalogue the range of facts that she deals with and assert my belief that certain further cases will succumb to the same approach 42 Environments where Rothstein shows that the apparent semantic contribution of be is no more than that of V in general include: • complements of consider-type verbs; • complements of control verbs like want, prefer, etc.; • complements of make-type verbs and perception verbs; • the apparently agentive progressive form of be, e.g., Jane is being polite/a nuisance (cf. Partee 1977); • identificational/equative be, which Rothstein assumes is special only in that it necessarily induces type lifting of a DP complement to something that means roughly 'the state of being DP' (cf. Partee 1986). Rothstein claims that the be that introduces progressives (Mary is dancing) cannot be assimilated to all the other bes because its complement, a VP according to her, is of the wrong semantic type (namely, an eventuality). I

Semantically empty lexical heads as last resorts

163

think this is overly hasty. In particular, since I treat progressive as containing a head -ing that takes a VP complement, I can attribute to -ing the semantic work that she assumes progressive be must carry out. Having done so, I can retain my proposal that this be takes a participial complement whose semantic type is in relevant respects identical to that of the adjectival complements that are the focus of Rothstein's investigation. Consider finally another environment where be is apparently optional but induces a subtle semantic change, namely Mad Magazine sentences (MMs). (91)

a. John drunk at 3:30 last night??!! I can't believe it! b. John be drunk at 3:30 tomorrow night?! Very unlikely! c. *John be drunk at 3:30 last night?!

(92)

a. b.

(93)

a. b.

What? ? Clinton assassinated?! I can't believe it! What? ? Clinton be assassinated?! That would never happen! Me tall? Compared to whom??! Me be tall?

11

Akmajian (1984) noted that MMs headed by a verb have to be interpreted as irrealis (94a); as a consequence, using an adverb to force a verbal MM to refer to the past is strongly degraded (94b). (94)

a. b.

John tell a lie?! Never! ' John tell a lie yesterday?

This is not true with nonverbal predicates ((91a), (92a), (93a)), which need not be taken as irrealis. Crucially, however, when be is added it reimposes an irrealis interpretation ((91b), (92b), (93b)). I have tried to bring out this contrast in the examples, whose paraphrases would be Ί can't believe X is true' (without be) versus Ί can't imagine X would/could ever be true' (with be). If Akmajian's generalization is correct, then we are dealing once again with a property of verbs in general, i.e., of the category V, and not one specific to be. To minimize the confound from switching predicates, we can confirm this by looking at near minimal pairs of a verb and a noun or adjective derived from it:

164 Carson Τ. Schütze (95)

a. b. c.

John dance (?*last night)?! (irrealis) John dancing (last night)?! (realis possible) John (still) a dancer?! (realis possible)

(96)

a. b.

John hopeful that there will be peace?! (realis possible) John hope that there will be peace?! (irrealis)

Thus, it is the categorial status of the predicate in a MM that determines whether it must be interpreted as irrealis; nothing special need be said about be. While I cannot engage in a full analysis of MMs here, it seems promising to try to relate Akmajian's generalization to Rothstein's proposals about the semantic nature of V. Specifically, it would make sense that predicates that are temporally locatable are just the ones that can be forced not to refer to the past or present. 5.

Empty P: With

Notwithstanding the fact that the category Ρ might not be a fully lexical category at all, I follow Emonds (1985) in discussing it in the context of the previous three undeniably lexical categories. The central claim of this section is taken from Tremblay (1996), who argues that with (and its counterpart avec in French) is a semantically empty preposition whose only meaning is that which is inherent to the category P, namely the expression of a (necessarily binary) relation. I begin by summarizing her central arguments. The reader is referred to Tremblay's paper for discussion of the nature of this relation; she argues that with can serve to link a variety of elements in its complement to elements in the rest of the sentence.

5.1.

Semantic emptiness

First, the meaning conveyed by with in one sentence can be conveyed without it in a closely related sentence. For example, the instrumental role of a soccer ball is the same in (97a) and (97b), though only the former uses with. Thus, that DP cannot be relying on with for its 0-role in (97a). (97)

a. b.

A kid broke the window with a soccer ball. A soccer ball broke the window.

The same is true for the absolutes in (98) and the comitatives in (99).

Semantically empty lexical heads as last resorts

(98)

(99)

a. b.

He walked leisurely to the store with a hat on his head. He walked leisurely to the store, a hat on his head.

c. d.

With John being in town, it's hard to get any work done. John being in town, it's hard to get any work done.

a. b.

He is playing tennis with his wife. He and his wife are playing tennis.

165

A further argument from (99) concerns what the semantic contribution of with could be, if it were not empty. Tremblay notes that (this kind of) comitative with cannot be viewed as assigning a 0-role of its own, but rather it just links its complement (his wife) to another DP (he), with which it will then share a 0-role. Compare (99a) with (100). (100) Mary threw out the candles with the cake. Here the cake is a theme of throwing out, whereas in (99) his wife is an agent of playing tennis. Thus, with is not providing the θ-role to its complement, the main predicate is doing so.43

5.2.

Formal requirements

What is the nature of the formal requirement that demands the presence of with? In many of these examples, it is simply the fact that verbs in English cannot generally express two complements as bare DPs; the exceptions, double object verbs, are tightly constrained in meaning. More generally, Case Filter violations would ensue if with were removed from these sentences. The clear exception is (98c,d); I have only a speculative answer to offer as to the source of this optionality. A suggestive contrast arises when the absolute phrase is not sentence initial: (101) a. b.

It's hard to get any work done with John being in town. It's hard to get any work done, John being in town.

I use the comma in (101b) to indicate that in the absence of with, a strong prosodic boundary must separate the absolute from the main clause, whereas this is clearly not required (in fact disfavored) in (101a). (Though subtler, this effect may well obtain even for sentence-initial absolutes.) One

166 Carson Τ. Schütze

might suggest that with is marking the presence of a relation (in this case, contemporaneousness) between the two clauses, thereby allowing them to phrase as one prosodic unit, whereas in the absence of with, they must be treated as formally unconnected to each other, their juxtaposition within one sentence serving to imply that some relationship must be intended.

5.3.

Why as and of are not the default Ρ

Emonds (1985) discusses a wide range of facts about as that could be construed as arguments that it is actually the default Ρ of English. For example, it is the counterpart to be in certain nonverbal contexts: (102) a. b.

John is our friend. With John as our friend, ...

I reject this conclusion on the grounds of distribution, taking it as a fundamental property of the category Ρ that it can take only DPs as complements. (Emonds does not share this view.) That restriction holds of all uses of with,44 but clearly does not hold of as: (103) People often view/describe John as {crazy/under the spell of witches/jaded by his youth/knowing too much}. It appears that as selects for a predicate, independent of syntactic category. Therefore, it could not be an empty P, though it certainly could be an empty functional head (cf. the distinction between be and do in section 4). What about of, most of whose uses certainly seem to be semantically empty? Here again, my reason for not analyzing it as the default Ρ has to do with syntactic category; specifically, I claim that in the relevant uses, of is not a true preposition but rather a K(ase) head (cf. Lamontagne and Travis 1986; Travis and Lamontagne 1992). Among the evidence for this is the fact that o/-PP complements to nouns behave like bare DP complements to verbs, and not like any other PP complements, in terms of their availability for predication (Neeleman 1997; Wechsler 1995). This can be explained by assuming that Κ is part of the extended projection of N/D, and therefore does not serve to block command, whereas a DP complement to a true Ρ is more deeply embedded. In addition to the category label, Tremblay argues that there is a structural difference between Ρ and Κ that corresponds to a difference in mean-

Semantically empty lexical heads as last resorts 167 ing: P, being inherently relational, must always project two argument positions (in Spec and Comp, under her analysis), whereas Κ does not express a relation and takes only a complement. Tremblay further points out that Κ heads are used to express (semantically) compulsory arguments, e.g., the destruction of the city, whereas with can add an argument: the destruction of the city with bulldozers. Perhaps the most telling set of facts illustrating that with is not (just) a dummy Κ head are those that show a split in behavior among uses of with itself. Tremblay argues that, in environments other than those discussed in section 5.1, with is purely a Κ head (a dummy Case assigner, in her terminology): (104) a. b. c. d. e.

Anne presented Mary with a gift. Everything is right with the world. The apple is rotten with worms. Into the dungeon with that traitor! To hell with this assignment!

There is a clear distributional difference between with as Ρ and with as K: only the former can be negated as without·. (105) a. b. c. d. (106) a. b. c. d. e.

He mounted the picture with(out) a hammer. He walked to the store with(out) a hat on his head. He is playing mixed doubles with{out) his regular partner. With(out) John being in town, it is hard to get any work done. *Anne presented Mary without a gift. * Everything is right without the world. *The apple is rotten without worms. *Into the dungeon without that traitor! *To hell without this assignment!

It makes sense to think that a relation can be negated, i.e., one can refer to the absence of a relation, but it makes no sense to negate a Case marker. Tremblay observes that one use of with that she analyzes as an empty Ρ does not allow negation using without: (107) My keys are with( *out) my wallet.

168 Carson Τ. Schütze She does not offer an explanation for this fact. My suggestion is that this with, meaning 'in proximity o f , actually requires lexically specified content. (Since be is meaningless, we cannot treat this with as a dummy Κ because there would then be no predicate in the sentence at all.) A suggestive fact that points in this direction is that the so-called PP specifier right (Jackendoff 1977) is compatible with this use of with but not with those in (97)-(99): (108) a. b.

(109) a. b. c. d.

Your keys are right with your wallet. The trouble started right with the first play of the game. (temporal proximity sense) *He mounted the picture right with a hammer. *He walked to the store right with a hat on his head. *He is playing mixed doubles right with his wife. *It is hard to get any work done right with John being in town.

The explanation would be that one cannot intensify an unspecified meaning - the ungrammatical sentences would have to mean 'X is in more of a certain relation with Υ'. I will provide further evidence for the claim that this one use of with is not a default in the next subsection.

5.4.

Empty V+P: Have

Building on a rich tradition in the literature (cf. Benveniste 1966; Fillmore 1969; Freeze 1992; Kayne 1993; Gueron 1995; Dechaine, Hoekstra and Rooryck 1995; Ouhalla 1991; Den Dikken 1997; Moro 1997: inter alia), I combine two of my claims above and propose that have is the combination of empty V and empty P, that is, be + with. This accounts for its semantic lightness and for numerous facts about its distribution; I take these in turn. Semantic lightness is often thought to be the reason why be and auxiliary have differ from all other finite verbs of English in raising to Tense, hence undergoing Subject-Aux inversion etc. A natural explanation for the correlation between lightness and V-raising may lie in the proposal that be and have (on some uses) are generated higher in the INFL system than contentful verbs (cf. Ouhalla 1991). This would follow from the claim that they are defaults - the element that demands a V in most types of clauses is Tense, so it is only at the height of Tense in the tree that the need for a V can be determined.

Semantically empty lexical heads as last resorts

169

Turning now to distribution, the following familiar paradigm is derived on the assumption that have, as the default spellout of V+P merged into one head position, blocks insertion of the separate heads be with, which would otherwise be expected. (110) a. b. c. d.

the boy from Oklahoma the boy who is from Oklahoma the boy with the bat the boy who has the bat (cf. *the boy who is with the bat)

This analysis accounts in a natural way for the very wide range of diverse uses shared by have and with: have is simply the verbalization of with. ( I l l ) a. b.

a woman who has blue eyes a woman with blue eyes

c. d.

a woman who has great influence a woman with great influence

e.

That guy has {his face painted blue/his hands in his pockets/Mary as his escort/his shoes all muddy). a guy with {his face painted blue/his hands in his pockets/Mary as his escort/his shoes all muddy}

f.

g. h.

That student has {an essay to write/her homework finished}. a student with {an essay to write/her homework finished}

i. j.

The table has a lamp on it. the table with a lamp on it

Note that all these uses of with can be negated as without. Interestingly, although the sequence be with is ungrammatical as an alternative to have in the environments of (111) (e.g., *The woman is with blue eyes), this is not true for the proximity use of with mentioned in the previous subsection, which we saw could not be negated as without: (112) a. My keys are with my wallet. b. *My keys have my wallet.

170 Carson Τ. Schütze This contrast with (111) supports two claims: first, that proximity with is not empty, and second, that have as a replacement for be with crucially depends on the emptiness of those heads - it is not simply an alternative spellout of two words under adjacency. Yet another parallel in the behavior of have and with shows up in free adjuncts. Stump (1985) notes that adding V+ing to a bare predicate in an adjunct changes its interpretation with respect to a matrix modal: the verbless predicate has an 'if reading (which Stump labels "weak"), while the verbal counterpart has only a 'because' ("strong") reading: (113) a. b.

On a horse, John could get there in time. (= If he were on a horse...) Being on a horse, John could get there in time. (= Because he is on a horse...)

This pattern is exactly paralleled in pairs involving with and have, showing once again that have is the verbalization of with: (114) a. b.

With three-inch claws on each paw, my cat could terrorize the whole neighborhood. (= If it had...) Having three-inch claws on each paw, my cat could terrorize the whole neighborhood. (= Because it has...)

There is one further alternation that is correctly predicted when we couple the claims of this and the previous subsection: if we combine empty V, empty P, and negation, there will be two ways of spelling this out, depending on which pair of heads is combined first: V + (P+Neg) or (V+P) + Neg: (115) a. b.

They have been without food for three days. They have not had (any) food for three days.

c. d.

He is without faults. He does not have faults.

e. f.

Sarah is without equal. Sarah has no equal.

Semantically empty lexical heads as last resorts 171 5.5.

Further semantic considerations

We have seen that there is at least one use of with, the proximity sense, that appears to involve some semantic content. The reader has probably already arrived at the idea that the same needs to be said for have. For example, causative have (/ had him dance) has no with counterpart and intuitively feels more substantive than the purely relational uses in (111). Also, while aspectual have may not be very contentful, it is hard to see how its meaning could be derived from any of the meanings of with. And there are more senses of have: obligation (have to), consumption (have a drink), and experiential (/ had a bee sting me).45 These do all have exactly two arguments, however, suggesting that a treatment in terms of an underlying Ρ across the board might still be maintained. For example, aspectual have might contain a Ρ meaning something like before, on a view of tenses as two-place predicates relating events. When we take into account the further fact that not all languages that have a verb have for expressing possession can use it in all of the above extended senses, it seems that what we need is an underspecification treatment. That is, given a range of meanings associated with members of category P, a language can choose to verbalize any number of those meanings with unique verbs, but any leftover cases for which there is no lexical entry for V+P will use the default spellout, have (cf. Becker 1997). The Elsewhere Principle will guarantee that have is not generally an alternative to the more specific verbs. To complete the picture, we must return to the allegedly contentful proximity with: I am compelled to posit a lexical entry for this item, separate from the entry for default with. We saw that it cannot be verbalized as have, but there is also no verb that seems to correspond exactly to the proximity meaning of be with, so what blocks the use of have as its verbalization? We know independently that prepositions can differ idiosyncratically with regard to how tightly they combine with verbs; for example, pseudopassivization seems to depend on the arbitrary fact that some Ps can be reanalyzed as part of their governing V while others cannot. The same mechanism will provide the desired result here: proximity with is lexically specified as not incorporable into a governing V; as a result, no complex V+P head can be built from it, and have will not be a candidate for insertion.

172 Carson Τ. Schütze

6.

Conclusion

Let us step back from the details of the preceding analyses and consider some general properties of semantically empty lexical heads, assuming the proposals presented above to be on the right track. The first property is a standard property of defaults in (especially inflectional) morphology: since their distribution is defined as the complement to the environments for which all other items in a paradigm are directly specified, defaults are expected to appear in a heterogeneous set of environments that do not look like a natural class in the usual sense. I believe that the four (or five, counting have) default lexical heads that I have discussed clearly show this property, as the reader can verify by trying to encompass their full distribution with a unified statement. A second property, which has really only been hinted at here but is pursued in depth by Emonds (1985), is that semantically empty heads may be subject to grammatical rules in a slightly different way from contentful heads. We saw two examples of this: the empty adjective so fails to undergo a head raising operation that all contentful adjectives undergo, and the empty verbs be and have, on the standard account (cf. Pollock 1989), undergo a step of head movement that contentful verbs do not undergo. To this we might need to add that empty with obligatorily head raises to an immediately governing V when possible, in order to explain why be with is not in free variation with have. I hardly considered head movement properties of N, but it is possible that one is also special in this regard (see note 9). So, there does appear to be something in need of explanation here. If semantic emptiness really correlates with atypical extent of head movement, this would suggest that at least some head movement is actually governed by considerations of meaning, rather than just by arbitrary formal features. Indeed, Corver's proposal for A-to-Q raising has exactly this character: it is motivated by the need to discharge a semantic role in a local structural configuration. If it is true that default heads such as so are devoid of semantic content (thematic roles) within the (pre-LF) syntax, then their failure to undergo head movement is exactly what we should expect. It is a little more difficult to see why the opposite situation, extra movement by empty heads, should obtain, but see section 5.4 for a suggestion. As I have articulated it so far, this view may seem to contradict the architectural assumption with which we began, namely that purely semantic (nongrammatical) differences among heads are not visible in the syntax. In fact, it is easy to make the two ideas compatible. The simplest solution is to posit, based on evidence such as that assembled by Emonds, that the differ-

Semantically empty lexical heads as last resorts

173

ence between [+content] and [-content], so to speak, is a grammatically relevant feature and is therefore accessible to the syntax. This seems to be needed in any case, in order to apply the ban on vacuous projection within DP proposed in section 2.3. This might be manifested in terms of presence versus absence of a 0-grid, for example. A possibly more conceptually appealing alternative is available if we assume a copy theory of movement, as in the Minimalist Program (Chomsky 1995b). Suppose we implement this as follows: in the syntax all copies of a given element are identical; the Spellout component must choose among them the one that is to be pronounced. Since pronouncing a head involves an actual word, this means that the difference between red and so, for example, is independently needed at the point in the derivation when the choice must be made as to which position to spell out. Thus, the architecture already allows for the possibility that Spellout could choose to phonologically realize the higher copy of a contentful head but the lower copy of an empty head. (Some of the analyses in section 3 might need to be re-thought under this second approach.) One final point concerns the nature of the emptiness that I have been attributing to the default lexical heads. As suggested also by Tremblay (1996), studying empty heads should be a good route to understanding the fundamental nature of lexical categories, though I can do no more than hint at that enterprise here (see also Dechaine and Tremblay (1998)). From that perspective, there is an important distinction among the four heads that we have examined that I have not brought to the fore, namely, how they are eventually interpreted. The Ν one and the A so must take an antecedent, that is, the meaning associated with their position is copied from elsewhere in the sentence or discourse. This is clearly not true of the V be and the Ρ with, whose meanings I have claimed are nothing more than the inherent nature of their respective lexical categories; unlike one and so, they remain empty even at the level of interpretation. What is this difference telling us? In the case of with, the possibility of an antecedent is perhaps excluded on the grounds that Ρ is not a truly open class. This implies that any expressible meaning is already available as part of a single existing word, so there would be no economy gained by replacing one Ρ word with another. Still, even empty Ρ necessarily encodes a relation between two entities, while Ν and A both essentially encode properties of entities. In contrast, it appears that V need not do anything of either sort at all; sometimes it contributes nothing more than serving as a host onto which to hang aspectual morphology, for example. It seems that the fundamental purpose of category V, unlike the other three categories, is often a formal rather than a semantic

174 Carson Τ. Schütze

one: to be the syntactic workhorse of the sentence, to provide the spine that holds the other elements together. An element of this kind clearly does not admit of taking antecedents that could range across all its possible uses, because they would have no semantic type in common. The nature of do so - the closest thing to an anaphoric V in English - is thus rather telling: it is restricted to an open role slot of just one type (agent) and a relation among its participants of just one type (event). Returning to the point with which this article began, have we indeed found that presence/absence of semantic substance is not a basis for distinguishing lexical from functional categories? I think the answer is "Yes, with a qualification". For each of the lexical categories, there are certainly words that can fill a surface slot of that category that are devoid of semantic content. However, if the question is whether there are lexical head positions whose interpretation lacks semantic substance, the answer seems to be "No", except for verbs. Slicing the pie a different way, one might say (following Marantz 1997) that the substantive roots of a language are category neutral, at least among Ν, V, and A -a push, to push, and pushy consist of a single root in form and meaning, combined with three different derivational affixes (two of them zeroes). However, given the extremely tight restrictions on the amount of affixal material that a verb can bear in English - maximally one suffix - a main verb based on a contentful root simply cannot support all the verbal morphology that a clause can contain. As a result, an extra dummy element (or two, if have is to be seen as separate from be) that can never be used as an Ν or A root must be added to the inventory of verbs. It would be wrong to think of be as "a functional head in a lexical head's distributional clothing", however. Functional heads carry meaning, just meaning of a grammatical as opposed to encyclopedic sort (in Distributed Morphology terms). The dummy use of be, however, carries no meaning whatsoever. If language is a system striving for perfection, as Chomsky (1995b) would have it, then such an ungainly beast should truly be a device of last resort. 7.

Acknowledgements

I would like to thank the participants of the workshop on semi-lexical heads, one anonymous reviewer, and an audience at Stanford for their feedback, and Misha Becker, Melissa Epstein, Christine Erb, Edward Garrett, Christine Gunlogson, Amalia Llombart, Alec Marantz, and Pam Munro for

Semantically empty lexical heads as last resorts 175

extensive discussions of this material. Standard disclaimers apply. This research was supported by a UCLA Academic Senate grant.

Notes 1. Emonds (1985) explicitly argues to the contrary that it is precisely the semantically empty items that show special behaviour. I believe that many of his examples should be treated in other ways (e.g., everyone does not contain the empty noun one - see section 2.2 for justification). Nonetheless, my claim in the text may be slightly too strong. We shall see evidence below that, in at least one case (the adjective so), the contentful versus empty distinction seems to have consequences for head movement. I return briefly to this in the Conclusion. 2. As this paper was being completed, an analysis of one very similar in spirit to the one I develop here was independently proposed by Gunlogson (2000). 3. If this sentence sounds odd out of context, consider a discourse such as (i). (i) How do farmers decide which apples to use for what? - Well, they often use ones that have fallen to the ground to make cider. Contra Kester (1996), I find bare ones with a relative clause generally acceptable in the right sort of context. 4. One class of apparent counterexamples discussed extensively by Kester (1996) is characterized by her as [+human, +plural, +generic]: DPs like the rich, the homeless, the blind, the French, etc. She argues that the words rich etc. are not behaving syntactically as nouns here, and posits a default, phonologically empty pro as the Ν head of these expressions; at least for English, pro can have only the triplet of feature values specified above, which would account for why one cannot use just any adjective this way. I could in fact adopt this analysis, such that this pro can occupy only empty Ns in the context of [+human, +plural, +generic], while in all other contexts the default one will come into play. It is important to note that these DPs are not synonymous with their counterparts with one, indeed pro-form one never receives a generic interpretation: the French ones must be interpreted as a specific definite. These special cases also have no indefinite counterpart: *a homeless, *a French, *some rich, etc. Thus, they do not represent an instance where the meaning normally expressed with one(s) is expressed with a phonologically zero Ν instead, and hence do not challenge the last resort status of one. It is however interesting that this alleged pro shares with the semantically empty Ν one(s) the feature [+count]. This might support the view that [+count] is the default interpretation for category Ν and need not be stipulated in the lexical entry for one·, rather, any Ν is interpreted as [+count] unless something explicitly identifies it as [+mass].

176 Carson Τ. Schütze

5.

6.

7.

8. 9.

10.

There are sporadic cases noted in the literature where an adjective can be stranded without a head noun (Sadler and Arnold 1994; Dahl 1984; Kester 1996): (i) a. The pleasant rays are actually less dangerous than the unpleasant. b. This recipe requires plain flour, not self-raising. c. There are good people and bad in every country. The example in (ib) may be representative of the behaviour of modification of mass nouns more generally; this is not problematic for my account since one would not be an option here anyway. I have nothing to offer concerning (ia) and (ic), beyond the observation that they feel stylistically marked (perhaps more common in writing), and Dahl's (1984) comment that this option degrades if one tries to strand more than a single bare adjective. I therefore do not take these as counterexamples to the general pattern described in the text. For pro-forms we can replicate this pattern also with a preceding adjective: (i) We certainly like (the) green ones best. (i) is acceptable with focus on green and deaccenting of ones, but unacceptable with focus on ones. This is of course also the behaviour of it, the only unambiguously weak pronoun in English. This generalization is apparently violated (Epstein 1999) in the expressions many/not/nary/such/hardly a one, too good a one, what a one, you ARE a one! Although I believe that some explanations may be found for these exceptions, they all have a literary/dated flavour to them that makes it not a great embarrassment to leave them outside the scope of the analysis. Perlmutter (1970) notes that some of these expressions take primary stress on one, suggesting they do not actually involve the pro-form. According to the prosodic test just proposed, (5b) allows either numeral or proform one to surface. One might explore a head movement account of the *a one and *one one restrictions. It is clear that a is considerably lower in the DP than the (e.g., Epstein 1999), so the contrast with the one could be captured by saying that one differs from contentful nouns in undergoing one step of head movement when possible. The ungrammatical sequences are banned because the indefinite determiner/numeral position cannot be spelled out when an Ν head has moved into it, since a and one are not affixes. This N-raising would be blocked by an intervening adjective, perhaps because pre-N adjectives themselves form complex heads with Ν (see below). The potential hosts for clitic the are evidently quite restricted. Specifically, it can be hosted by a Ν immediately to its right, or any material that can precede the head noun, such as an adjective and its modifiers, but not by a PP or relative clause post-N modifier. It is not clear whether these restrictions are prosodic or syntactic in nature. A prosodic account is conceivable if there is necessarily a prosodic domain boundary between a head Ν and post-head modifiers.

Semantically empty lexical heads as last resorts 177

11. An anonymous reviewer suggests that the badness of (8c) has nothing to do with the clitic status of the, since the following examples are also ungrammatical: (i) a. *no from Canada b. * every from Canada The reviewer goes on to propose that a superior explanation of the badness of (8c) and (i) is the fact that these determiners are simply lexically specified as transitive (cf. Abney 1987). However, attempting to unify the cases in (i) with the behaviour of the in this way would miss the fact that the meanings of (7c) and (8c) are precisely expressible as (7b) and (8b), whereas the meaning of (ib) has no direct counterpart with gapping; it is also left as a coincidence that the contains a reduced vowel while no and every do not. My suspicion is that (ia) is blocked by the existence of the allomorph none. The badness of (ib) and related examples with every requires a stipulation in almost all theories, though see Lobeck (1995) for a more principled attempt. 12. As this paper was being completed, I discovered that this was proposed independently by Sadler and Arnold (1994). 13. For cases like (18), this claim was made independently by Kester (1996). 14. In order to guarantee that one(s) would contribute no semantics, we must assume that its restriction to count noun meanings is not relevant for purposes of this determination. Therefore, I am forced to assume that the count/noncount distinction is encoded semantically in one of the functional heads of DP; Ν will be required to be compatible with that functional specification, but the count/noncount status of the DP denotation is encoded apart from N. This view seems necessary anyway, since mass nouns (e.g., rice) can be coerced into count meanings (two rices = two kinds of rice) by functional material. Note further that the ban on vacuous projection crucially operates on syntactic features, since no vocabulary items have yet been inserted. Therefore, the fact that some in (18b) is ambiguous/homophonous between count and mass does not prevent the NP from being omitted: the distinction exists in the functional features of the DP, it just happens to be collapsed in the vocabulary. 15. Although prosody makes it clear that one in (22c) can be the pro-form, attempts to use it in the plural yield bizarre results, for reasons I do not understand. (i) ' My parents are both doctors, and my sisters want to be ones too. 16. It is important to keep in mind that the relevance of (23) is that these sentences are bad with the intended meanings of (22). Those meanings crucially are not demonstrative. 17. Thanks to the anonymous reviewer for suggesting this. 18. As before, I must assume that the plurality marked on ones is effectively a concord marking, the actual semantic number of the DP being established in NumP. 19. (26a) has a reading on which it is not synonymous with (26b): with focussed some, it could be continued ...but others that you won't. Only when some is

178 Carson Τ. Schütze

deaccented is it fuily equivalent to (26b); that is the reading of concern in the text. 20. The anonymous reviewer points out that the recoverability requirement could still be satisfied simply by having some random substring of a DP pronounced. To prevent this I must assume that PF deletion is not free, and that any lexical item with phonological content must trigger pronunciation of that content unless a specific PF deletion rule can delete it. These rules must in turn be constrained appropriately. 21. A systematic pattern that deserves further attention involves the ability of certain pre-N material to be optionally followed by one when there is no other material in the NP; that is, each item in (i) can be a complete DP: (i) each (one), another (one), which (one(s)), either (one), neither (one), the other (one), these (%ones), those (Woones) (Cf. Nerbonne, Iida and Ladusaw (1989) and sources cited there.) Superlatives and definite comparatives also follow this pattern: the best (one), the taller (one) (cf. Dahl 1984). This is in contrast to the pattern predicted by the ban on vacuous projection and exemplified in (18)-(20), wherein one(s) is ungrammatical; in the latter a Quantity or Quantifier head acts as a DP by itself, as in (ii). (ii) most (*ones), a few (*ones), three (*ones), both (*ones), more (*ones), many (*ones), any (*ones), some (*ones), all (*ones), several (*ones) The division between (i) and (ii) is clearly not arbitrary, so lexical stipulation would be an unattractive solution. I must leave the matter for future research, but see Lobeck (1995) for some potentially relevant considerations. 22. Jackendoff (1968) treats the fact that *a one surfaces as one as parallel to *the one surfacing as it, thinking of one as an indefinite pronoun. 23. The anonymous reviewer raises the question whether mine is in the same structural position as my, citing its posthead position in examples like relatives of mine. I think this question can be addressed by considering a related pair of DPs like Mary's (relatives) and relatives of Mary's-, here, 's appears prehead in the first expression and posthead in the second, but this would not necessarily force us to a nonuniform structural analysis o f ' s . Regardless of the apparent synonymy, the posthead Mary's is already not in the same structural position in the DP headed by relatives as it is in the prehead version; that by itself will explain how 's winds up on opposite sides of relatives, regardless of its position within the smallest DP containing it; perhaps the same could go for mine. 24. There is of course the traditional alternative (assumed by Corver) that so is inherently an unanalyzed AP and not a head. Counterevidence to this may be hard to find in the adjective domain, but since the facts otherwise parallel one it is more appealing to maintain a uniform treatment. 25. The anonymous reviewer suggests that (34) is not an instance of gapping, but rather of ellipsis after extraposition of the PP. However, according to the applicable tests catalogued by Lobeck (1995) for distinguishing gapping from ellipsis, (34) patterns as gapping. The reviewer also notes that we must be

Semantically empty lexical heads as last resorts 179

concerned about how the requirement for a remnant to the left of the gap site is satisfied in (34), since in section 2 we found that the left-hand remnant had to be within DP, i.e., the maximal extended projection of the gapped N. Whether this is a problem depends on what the extended maximal projection of the adjective is in a structure like (34), where the adjective heads the main predicate of the clause. Criteria for answering this question are lacking. 26. The anonymous reviewer questions this constituent structure, citing the following example: (i) a. He's efficient enough. b. He's more than that! While I agree that the structure relevant for (i) would be [more than] [efficient enough], this is not the same meaning as in (35b), where more than enough so can mean only efficient to a more than satisfactory degree. 27. This is not the pro-form so. 28. This is not dummy much but its contentful homophone - see Corver. 29. Although Corver places very in Spec-QP along with these two classes of modifiers, it patterns like neither of them: it cannot take bare so but it can be rescued by much: (i) a. *veryso b. very much so From the perspective of our concerns here, this seems less problematic than the short modifier class. 30. The anonymous reviewer questions the existence of a general ban on phrasal constituents as pre-head modifiers, citing examples like a very intelligent man and a difficult to understand sentence. Following Sadler and Arnold (1994) and sources cited there, I maintain that such cases actually support the restriction, since they do not generalize: *an intelligent beyond belief man, *a glad to accompany us guide, *a difficult for us to understand sentence, etc. This is not to deny that one would hope for an independently motivated theory of how complex a complex A head can be. 31. Corver places possible as the head A of a "modal AP" between Q and the head A cheap. If a head movement analysis of these facts is on the right track, however, the proposal in the text would lead to a violation of the Head Movement Constraint; I have to claim instead that possible is in a specifier or adjoined modifier position. 32. An apparently unrelated use of so involves a clausal antecedent (perhaps restricted to ίΛαί-clauses): (i) a. Is it true that Fred's getting married? b. Jane said so. c. So said her father as well. But will it be a big wedding? d. So it seems. Whatever the exact nature of this so, it is clearly not a predicate, hence not a candidate for unification with adjectival uses of so.

180 Carson Τ. Schütze

33. The badness of (66) cannot be blamed on a failure of head raising, since Q does not raise to Deg in English in general: (i) a. so extremely smart b. *so smart extremely c. *smart so extremely 34. Associating Case with V in general may appear problematic, since there are supposed to be verbs that cannot assign Case. An uninteresting technical solution to this would be to allow lexical entries that exceptionally mark verbs as [- Case assigning]. However, I believe a more promising avenue is to try to derive that property, insofar as it is real, from other considerations; see Schütze (2000). 35. The range of complement types that a given predicate will allow is to some extent a matter of dialectal variation. For example, in Pittsburgh (among other places in the United States) (ib) is an alternative to (ia), while for most English speakers (ib) is strongly ungrammatical. Similarly, British English allows (iib) while American English generally allows only (iia). (i) a. My car needs to be washed. b. %My car needs washed. (ii) a. John seems to be a fool. b. %John seems a fool. This is further evidence that there can be no deep semantic explanation for some of these cases of optionality. 36. Omission of to be is possible on the 'lonely' sense of alone, but not the 'unaccompanied' sense. 37. Post-N modification in DP is sometimes possible because the Ν itself denotes an event: (i) a. the concert at 8:00 b. the trip by ferry However, this does not explain why we can say the cake for the children (cf. (82)), a matter I do not explore here. 38. It is of course true that, because of the way V-raising operates in English, be's surface distribution is not identical to that of contentful verbs. However given how finite verbs behave in French (Pollock 1989), this is clearly no challenge to the claim that be is always a V. See Schütze (2000) for how the English Vraising facts can be naturally captured in the present framework; see also the Conclusion. 39. Indeed, we can apply the account of nominal and adjectival pro-forms developed above to explain the classic observation that do so cannot co-occur with an argument PP. As before, we need not appeal to a stipulation that do so is categorially a VP, for example. We simply note that main verb do does not include a second (PP) complement in its β-grid, so even when it is used as a pro-form there will be no way for it to project a position for such an argument.

Semantically empty lexical heads as last resorts 181

40. The situation is evidently more nuanced in British English, which has a use of do that appears to involve a genuine V that lacks the eventivity restriction (examples from Hudson (2000)): (i) a. He didn 't call today, but he may do tomorrow. b. Does he like her? Yes, he must do - just look how he talks to her. The sentences in (i) are not available in North American English, and I do not pursue the matter here. 41. A further point Rothstein (1999) makes that is consonant with my approach is that be (at least when it combines with an AP) differs from ordinary verbs by having no inherent aspectual class (i.e., state, activity, achievement, or accomplishment). 42. Stump (1985) presents an extensive discussion of the semantics of be that I do not explicitly address here. Although he motivates four different lexical entries for English be, I believe his observations can be accommodated under the view I present here, by taking more careful account of what triggers the presence of be in the various environments he examines. This is left for future work. For suggestions on how to derive the meaning of be while viewing it as essentially empty, see also Goossens (1992). Note that in earlier work, Rothstein (1987) herself posited multiple bes, some with crucial lexical content such as 0-role assignment. 43. The anonymous reviewer suggests that the facts in (99) ought to lead to the conclusion, following my reasoning, that and in (99b) is semantically empty. This is true in one sense and false in another, in just the same way for (97) as it is for (99). The sense in which it is true that both with and and are contentless here is that they do not contribute their own 0-role. The sense in which it is false is that they obviously perform some function that one might loosely call "semantic", namely they link a DP (a soccer ball or his wife) to its 0-role assigner. 44. This is not obvious for the absolute with construction; see Schütze (1997). 45. There is further evidence that the basic possession use of have differs from these extended uses in a syntactically important way. In British English, possessive have patterns as an auxiliary in that it raises overtly to Tense, while causative and obligational have, for example, do not (Ouhalla 1991).

182 Carson Τ. Schütze References Abney, Steven Paul 1987 The English noun phrase in its sentential aspect. Ph.D. dissertation, Massachusetts Institute of Technology. (Distributed by MIT Working Papers in Linguistics.) Akmajian, Adrian 1984 Sentence types and the form-function fit. Natural Language and Linguistic Theory 2: 1-23. Bach, Emmon 1967 Have and be in English syntax. Language 43: 462-485. Baker, Carl Leroy 1978 Introduction to Generative-Transformational Syntax. Englewood Cliffs, NJ: Prentice-Hall. Becker, Misha 1997 Acquiring a grammar with covert head incorporation: The case of have as a complex verb. M.A. thesis, UCLA. Belletti, Adriana 1988 The case of unaccusatives. Linguistic Inquiry 19: 1-34. Benveniste, Emile 1966 Problemes de Linguistique Generale. Paris: Gallimard. Bresnan, Joan W. 1973 Syntax of the comparative clause construction in English. Linguistic Inquiry 4: 275-343. Campbell, Richard 1992 Serial verbs and unaccusativity. In: Chris Collins and Victor Manfredi (eds.), Proceedings of the Kwa Comparative Syntax Workshop, MIT Working Papers in Linguistics 17: 41-51. 1996 One(s): The lonely number. In: Kiyomi Kusumoto (ed.), Proceedings of the North East Linguistic Society 26: 43-55. Chomsky, Noam 1995a Bare phrase structure. In: Gert Webelhuth (ed.), Government and Binding Theory and the Minimalist Program: Principles and Parameters in Syntactic Theory, 385-439. Cambridge, MA: Basil Blackwell. 1995b The Minimalist Program. Cambridge, MA: MIT Press. Corver, Norbert 1997 Λ/wc/i-support as a last resort. Linguistic Inquiry 28: 119-164. Dahl, Deborah A. 1984 The structure and function of one-anaphora in English. Ph.D. dissertation, University of Minnesota. (Distributed by IULC, 1985.)

Semantically empty lexical heads as last resorts 183

Ddchaine, Rose-Marie, Teun Hoekstra and Johan Rooryck 1995 Augmented and non-augmented HAVE. In: L6a Nash and Georges Tsoulas (eds.), Actes de Colloque Langues and Grammaire 1. Universite Paris 8. Dechaine, Rose-Marie and Mireille Tremblay 1998 On category features. Paper presented at GLOW 21, Tilburg. Dikken, Marcel den 1994 Predicate inversion and minimality. In: Reineke Bok-Bennema and Crit Cremers (eds.), Linguistics in the Netherlands 1994, 1-12. Amsterdam: John Benjamins. 1997 Introduction: The syntax of possession and the verb 'have'. Lingua 101: 129-150. Emonds, Joseph E. 1985 A Unified Theory of Syntactic Categories. Dordrecht: Foris. 1987 Parts of speech in generative grammar. Linguistic Analysis 7: 3-42. Epstein, Melissa A. 1998 The distribution of one in English. Manuscript., UCLA. 1999 On the singular indefinite article in English. In: Gianluca Storto (ed.), Syntax at Sunset 2, UCLA Working Papers in Linguistics 3: 14-58. Fillmore, Charles J. 1969 Toward a modern theory of case. In: David A. Reibel and Sanford A. Schane (eds.), Modern Studies in English: Readings in Transformational Grammar, 361-375. Englewood Cliffs, NJ: PrenticeHall. Freeze, Ray 1992 Existentials and other locatives. Language 68: 553-595. Goossens, Louis 1992 Graded predicationality, semanticization and be in a functional grammar of English. Belgian Journal of Linguistics 7: 53-73. Gu6ron, Jacqueline 1995 On HAVE and BE. In: Jill N. Beckman (ed.), Proceedings of the North East Linguistic Society 25, vol. 1: Papers from the Main Sessions, 191-206. Gunlogson, Christine 2000 The categorial status of anaphoric one. Paper presented at the Linguistic Society of America Meeting, Chicago Halle, Morris and Alec Marantz 1993 Distributed morphology and the pieces of inflection. In: Kenneth Hale and Samuel Jay Keyser (eds.), The View from Building 20: Essays in Linguistics in Honor of Sylvain Bromberger, 111-167. Cambridge, ΜΑ: ΜΓΓ Press.

184 Carson Τ. Schütze

1994

Some key features of distributed morphology. In: Andrew Carnie, Heidi Harley, with Tony Bures (eds.). Papers on Phonology and Morphology, MIT Working Papers in Linguistics 21: 275-288. Heycock, Caroline B. 1994 Layers of Predication: The Non-lexical Syntax of Clauses. New York: Garland. 1995 The internal structure of small clauses: New evidence from inversion. In: Jill N. Beckman (ed.), Proceedings of the North East Linguistic Society 25, vol. 1: Papers from the Main Sessions, 223-238. Hudson, Richard 2000 Grammar without functional categories. In: Robert D. Borsley (ed.), The Nature and Function of Syntactic Categories, Syntax and Semantics vol. 32: 7-35. San Diego: Academic Press. Jackendoff, Ray S. 1968 Quantifiers in English. Foundations of Language 4: 422-442. 1971 Gapping and related rules. Linguistic Inquiry 2: 21-35. 1977 X-bar Syntax: A Study of Phrase Structure. Cambridge, MA: MIT Press. Kayne, Richard S. 1993 Toward a modular theory of auxiliary selection. Studia Linguistica 47: 3-31. Kester, Petronella Maria 1996 The nature of adjectival inflection. Ph.D. dissertation, Utrecht University. Lakoff, George 1970 Global rules. Language 46: 627-639. Lakoff, George and John Robert Ross 1976 Why you can't do so into the sink. In: James D. McCawley (ed.), Notes from the Linguistic Underground, Syntax and Semantics, vol. 7: 101-111. New York: Academic Press. Originally published as "A criterion for verb phrase constituency", Report NSF-17 of the Aiken Computation Laboratory of Harvard University to the National Science Foundation on Mathematical Linguistics and Automatic Translation [1966]. Lamontagne, Greg and Lisa deMena Travis 1986 The case filter and the ECP. McGill Working Papers in Linguistics 3.2: 51-75. Lasnik, Howard 1992 Case and expletives: Notes toward a parametric account. Linguistic Inquiry 23: 381-405. 1995 Case and expletives revisited: On Greed and other human failings. Linguistic Inquiry 26: 615-633.

Semantically empty lexical heads as last resorts 185

Law, Paul 1996

Remarks on the verb be and the expletive there in English. Linguistische Berichte 166: 492-529.

Lobeck, Anne 1995 Ellipsis: Functional Heads, Licensing, and Identification. New York: Oxford University Press. Longobardi, Giuseppi in press The structure of DPs: Some principles, parameters and problems. In: Mark Baltin and Chris Collins (eds.), The Handbook of Contemporary Syntactic Theory. Blackwell. Marantz, Alec 1997 No escape from syntax: Don't try morphological analysis in the privacy of your own lexicon. In: Alexis Dimitriadis, Laura Siegel, Clarissa Surek-Clark and Alexander Williams (eds.), University of Pennsylvania Working Papers in Linguistics 4.2, Proceedings of the 21st Annual Penn Linguistics Colloquium, 201-225. Moro, Andrea 1997 The Raising of Predicates: Predicative Noun Phrases and the Theory of Clause Structure. Cambridge: Cambridge University Press. Neeleman, Ad 1997 PP-complements. Natural Language and Linguistic Theory 15: 89137. Nerbonne, John, Masayo Iida and William Ladusaw 1989 Running on empty: Null heads in head-driven grammar. In: E. Jane Fee and Katherine Hunt (eds.), Proceedings of the Eighth West Coast Conference on Formal Linguistics, 276-288. Ouhalla, Jamal 1991 Functional Categories and Parametric Variation. London: Routledge Partee, Barbara H. 1977 John is easy to please. In: Antonio Zampolli (ed.), Linguistic Structures Processing, 281-312. Amsterdam: North-Holland. 1986 Ambiguous pseudoclefts with unambiguous be. In: Stephen Berman, Jae-Woong Choe and Joyce McDonough (eds.), Proceedings ofNELS 16: 354-366. Perlmutter, David M. 1970 On the article in English. In: Manfred Bierwisch and Karl Erich Heidolph (eds.), Progress in Linguistics: A Collection of Papers, 233-248. The Hague: Mouton.

186 Carson Τ. Schütze

Pollock, Jean-Yves 1989 Verb movement, Universal Grammar, and the structure of EP. Linguistic Inquiry 20: 365-424. Postal, Paul M. 1969 On so-called "pronouns" in English. In: David A. Reibel and Sanford A. Schane (eds.), Modem Studies in English: Readings in Transformational Grammar, 201-224. Englewood Cliffs, NJ: Prentice-Hall. Quirk, Randolph, Sidney Greenbaum, Geoffrey Leech and Jan Svartvik 1985 A Comprehensive Grammar of the English Language. London: Longman. Rapoport, Tova R. 1987 Copular, nominal, and small clauses: A study of Israeli Hebrew. Ph.D. dissertation, Massachusetts Institute of Technology. (Distributed by MIT Working Papers in Linguistics.) Ritter, Elizabeth 1992 Cross-linguistic evidence for Number Phrase. Canadian Journal of Linguistics 37: 197-218. Rothstein, Susan 1987 Three forms of English be. In: M.A. Browning, Ewa CzaykowskaHiggins and Elizabeth Ritter (eds.), MIT Working Papers in Linguistics 9: 225-238. 1999 Fine-grained structure in the eventuality domain: The semantics of predicative adjective phrases and be. Natural Language Semantics 7: 347-420. Sadler, Louisa and Douglas J. Arnold 1994 Prenominal adjectives and the phrasal/lexical distinction. Journal of Linguistics 30: 187-226. Schölten, Tineke 1988a The expletive verb be. In: Hagit Borer (ed.), Proceedings of the Seventh West Coast Conference on Formal Linguistics, 305-318. Stanford University: Stanford Linguistic Association. 1988b Principles of universal grammar and the auxiliary verb phenomenon. Ph.D. dissertation, University of Maryland. Schutze, Carson T. 1997 INFL in child and adult language: Agreement, case and licensing. Ph.D. dissertation, Massachusetts Institute of Technology. (Distributed by MIT Working Papers in Linguistics.) 2000 A restrictive theory of overt movement for Case in English. Paper presented at the workshop "Case Theory", Annual Conference of the German Society for Linguistics (DGfS), Marburg, March.

Semantically empty lexical heads as last resorts 187

Speas, Margaret 1994 Null arguments in a theory of economy of projection. In: E. Benedicta and J. Runner (eds.), Functional Projections, University of Massachusetts Occasional Papers 17: 179-208. S to well, Timothy 1981 Origins of phrase structure. Ph.D. dissertation, MIT. Stump, Gregory T. 1985 The Semantic Variability of Absolute Constructions. Dordrecht: Reidel. Travis, Lisa and Greg Lamontagne 1992 The Case Filter and licensing of empty K. Canadian Journal of Linguistics 37: 157-174. Tremblay, Mireille 1996 Empty prepositions and UG. In: Jose Camacho, Lina Choueiri and Maki Watanabe (eds.), The Proceedings of the Fourteenth West Coast Conference on Formal Linguistics, 557-570. Wechsler, Stephen 1995 The Semantic Basis of Argument Structure. Stanford, CA: CSLI Publications.

Part II Semi-lexicality in the Nominal Domain

Numeral/Quantifier-Classifier as a complex head Tanmoy Bhattacharya

This paper1 claims that the complex consisting of the quantifier or the numeral in combination with the classifier in the Eastern Indo-Aryan language Bangla (Bengali)2 may be considered as a semi-lexical head in the sense that the complex exhibits several properties which are divided between lexical and functional heads. The thrust of this claim is most clearly visible in the analysis of NP movement inside the DP in section 3 which involves the complex as a whole and not some smaller part of it. In the first section of the paper, the semi-lexical nature of the complex head Numeral/Quantifier-Classifier (Num/Q-Cla hereafter) is considered followed by a brief discussion of five Zwicky criteria of head determination in section 2. 1.

Semi-lexicality of the Num/Q-Cla complex

In Bhattacharya and Dasgupta (1996: 73) it was proposed that South Asian languages are to be typologically bifurcated into Gender languages (typified by Hindi which is a classifier-less language) and Class languages (typified by Bangla which is a gender-less language). Classifiers in Bangla include the default classifier Ta - the main variety to be discussed in this paper, the collective classifier gulo, the human classifier jon, the inanimate count classifier khana, the inanimate mass classifier khatii, the numeral absorbing human collective classifier ra etc. (see Dasgupta 1983). Examples (1) to (5) illustrate the various uses of these classifiers.

192 Tanmoy Bhattacharya

(1)

a.

b.

c.

d.

chele-Ta boy-CLA 'the boy' du-To tWO-CLA 'two' bon-Ti sister-CLA '(affectionate) sister' tin-Te three-CLA 'three'

(2)

chele-gulo boy-CLA 'the (group of) boys'

(3)

du-jon lok two-CLA person 'two persons'

(4)

du-khana baRi two-CLA house 'two houses'

(5)

Onek-khani doi a lot-CLA yoghurt 'a lot of yoghurt'

Note that the default form of the common classifier Ta has various allomorphs governed by phonological conditions. Te occurs with 'three' and 'four' as in tin-Te 'three-CLA', car-Te 'four-CLA' - historically car is derived from Icaril with the terminal high vowel which raises Ta to Te, in free variation with Ta in ei/oi-Ta/Te, 'this/that-CLA' where the exact transcription should be ey/oy for the demonstrative, y denoting a high glide. The allomorph To occurs only with 'two', again, explained in terms of vowel harmony. Ta occurs with the rest of the numerals and with other nouns. The example in (lc) additionally shows that the form of a particular classifier may also be governed semantically, in this case, the intended

Numeral/Quantifier-Classifier

as a complex head 193

sense of affection is encoded by the use of the classifier Ti instead of the usual, Ta. The examples above also show that a numeral or a quantifier in Bangla must be followed by a classifier, that is, they form a complex structure. This fact of the language describes the abridged title of the paper and in what follows, I will discuss several criteria for considering this complex as a semi-lexical head.

1.1.

Criteria for Semi-Lexicality of the Num/Q-Cla Head

The criteria to be discussed in this section are based on the notion of disguised X° or semi-lexicality in Emonds (1985). He looks at syntactic categories outside the core consisting of the non-phrasal categories Ν, V, and A plus any material in their specifiers (notated as SP(X) in Emonds) and concludes that "... all grammatical formative categories are either "disguised" instances of X° or SP(X) themselves, or are sw^-categories (= features) of X° and SP(X)" (Emonds 1985: 158). Thus for Emonds, the three closed subsets of open categories are "grammatical" Ns, Vs and As. Some examples of closed categories that are reduced to SP(X) are as follows: (6)

SP(N) = this, that, these,... all, both, each, which, what, etc. SP(V) = will, can,... SP(A) = -er, -est SP(P) = right, clear

Some closed categories, on the other hand, are reduced to X°s and are called disguised X°s. Some typical examples follow (Dis here is meant to stand for disguised): (7)

Dis(N) = one Dis(V) = auxiliaries Dis(A) = hard, fast, long (Adverbs), other, same, different, such (nouns), many, few, much, etc. Dis(P) = as

In what follows, I discuss some of the criteria for determining what is a disguised or a semi-lexical category according to Emonds and conclude that the Num/Q-Cla complex is one such category.

194 Tanmoy Bhattacharya (i) A theory neutral reason for semi-lexicality of the complex is that it is used quite generally although semantically it is less transparent. For example, the classifiers Ta/To/Te/Ti (to be identified by Ta henceforth) in Bangla have widespread use but semantically they are non-transparent (see (8))· (8)

a.

chele-Ta aSbe boy-CLA come.will 'the boy will come.'

b.

jOl-Ta

gOrom holo

water-CLA hot happened 'the water became hot.'

c.

e.

radha-r baRi aSa-Ta Radha's home coming-CLA 'Radha's coming home.' bon-Ti amar4 sister-CLA mine 'my (affectionate) sister!'

(ii) Yet, there is a small class within this closed class which can be distinguished from each other where each classifier can have a unique usage and meaning (see (9)). According to Emonds (1985: 168), this is an indication that this subclass is a disguised X class or semi-lexical. (9)

a.

b.

c.

e.

du-jon chele two- CLA boy 'two boys' du-khana ruTi two- CLA bread 'two (pieces of) bread' du-joRa juto two- CLA shoe 'two pairs of shoes' du-gocha phul two- CLA flower 'two bunches of flowers'

(iii) Universally classifiers are derived from nouns. Also (as per Klein (1980) mentioned in Emonds (1985: 163)) certain quantifiers like many, few etc., although behave as disguised Adjectives, are more like Spec(N)

Numeral/Quantifier-Classifier as a complex head 195 material, that is, they exist in the domain of NP. It is possible therefore to see the Num/Q-Cla complex as a disguised noun. In the rest of the paper, I will consider the numeral as occupying the same slot as the quantifier given that numerals in some sense quantify the following complement. (iv) The Num/Q-Cla complex also exhibits properties of a closed class in being limited in productivity or possessing a small number of members and not encouraging novel coinages. Refer to the table in (16) which depicts the limited membership of the class as a whole. (v) Emonds (1985: 160) also observes that semi-lexical categories cannot be expanded. In other words, they cannot be used as the left-hand item of a syntactic re-write rule or further sub-divided. I have discussed this in detail in section 3 where I have argued against splitting this complex into two separate heads Q and Cla, i.e. Q—>Q+Cla is not possible. As mentioned at the end of section 1, a Q or a Cla alone cannot be used with a noun5: (10)

a.

*(du)-To boi dao two-CLA book give.2 b. *(kO)-jon chele eSechilo some-CLA boy came

(vi) Another property of these semi-lexical categories that Emonds (1985: 165) mentions is their Unique Syntactic Behaviour. Although I do not see a direct application of this property in the present context, a related property of semi-lexical categories is that a whole subclass of a particular category may be affected by a particular syntactic operation. This can be shown for the Bangla quantifier SOb 'all' and other quantifiers (which I call Non-All Quantifiers or NAQs) as explored in section 36. Thus, if we consider Q as a subclass of the category Num/Q-Cla, and if it is true that well defined syntactic rules do apply to this subclass, then the category Num/Q-Cla as a whole is semi-lexical by this criterion. (vii) Lastly, the duality of status in terms of whether a particular head is functional or lexical makes a case for the semi-lexical nature of the complex head. I will discuss this in detail in the next section.

1.2.

Functional or lexical head

In this section, I discuss in the following two subsections the functional or the lexical status of the semi-lexical head under consideration. Facts from the phenomenon of stranding suggest that the complex can be considered

196 Tanmoy Bhattacharya

lexical (section 1.2.1) but by certain other criteria, it may be considered functional (section 1.2.2).

1.2.1.

Num/Q-Cla as a lexical head

Stranding is discussed further in section 3 drawing on Bhattacharya (1998a, 1999a). In this section, I discuss briefly the phenomenon of quantifier-float inside the DP in Bangla. This phenomenon will be shown not to be restricted to the universal quantifier but is obtained with any other quantifier as well. First, let us see how the Num/Q-Cla complex occupies the space1 between the heads D and N. This space is shown to be uniquely occupied by a Quantifier Phrase QP in Löbel (1989) who observes that the relation between the quantifier and the noun is that of countability or rather the function of the category quantifier is to ensure the countability of the NP. For a [+Count] noun, the Q οhead is morphologically realised as a plural suffix in English and German: (11)

a. b.

drei [Q 0 ] Bäum-e three tree-s drei [Q Stück] Wild-0 three head game 'three head of game'

In ( l i b ) Wild is a non-discrete substance and the measure noun Stück is inserted to quantify over the noun whereas for Bäum-e the countability is marked by a suffix. Löbel shows that Q as a functional category has the status of a head (Löbel 1989: 151). She also mentions that in numeralclassifier languages, the Q is lexically realised as Num+Cla. In Bangla, for example, the examples in (13) are similar to the set in (12) if we consider that the verb in (13) governs a zero noun. More crucially, we see that in (12) the quantifier involved quantifies over nouns, whereas in case of (13) the quantifier quantifies over the zero noun. (12)

kichu-Ta/ SOb-Ta/ khanik-Ta/ Onek-Ta some-CLA/ all-CLA/ some-CLA/ a-lot-CLA 'some/ all of the / some/ a lot of yoghurt.'

doi yoghurt

Numeral/Quantifier-Classifier as a complex head 197 (13)

kichu-Ta/

SOb-Ta/ khanik-Ta/ Onek-Ta dekhechi

some-CLA/

all-CLA/ some-CLA/ a lot-CLA seen.l

Ί have seen some/ all/ some/ a lot of it.' The fact that Q-float in DPs follow the same pattern as in the clauses is shown in Shlonsky's (1991) work on Q-DPs in Hebrew. Now consider the following where the quantifier/numeral kO/ tin 'some/ three' both allow quantifier float shown schematically in (14b) (see section 3.1 for more examples): (14)

a.

kO-Ta/tin-Te

chele-Scheie

kO-Ta/tin-Te

ekele

some-CLA/ three-CLA bov 'some/ three (of the) boys'

b.

[DP (D) Q+Cla NP]

[DP (D) NPj Q+Cla ti ] 10

Notice that in this example, the object rather than the subject floats. This possibility is encouraged by Bobaljik's (1995: 131) claim that objectoriented floating quantifiers are possible in object-shift languages. I conclude that NP shift in (14) above leaves the quantifier complex stranded in the sense that it is followed by an empty NP position. Stranding therefore establishes the headedness of the complex. According to Abney (1987) one criterion for functional heads is that these heads are usually inseparable from their complements. In earlier work, I have shown (as a consequence of Linear Correspondence Axiom (or LCA) as operative in head-final languages) that NP movement (rather than noun movement in SVO languages) inside the DP is due to a feature of [SPECIFICITY] of the complex head. Thus, the fact that the configuration {Num/Q-Cla, t} (as in (14b)) is possible in the Bangla DP indicates that the complements of the complex are separable from it which makes the status of the complex head lexical. Although, unlike other lexical categories in other languages the complex does not inflect for number or gender, I will consider (here and in the rest of the paper) the possibility that the classifier is the remnant of agreement in a language without any noticeable agreement at the clause level.11 This approach, therefore, brings the complex category in line with other lexical categories. Thus (15) shows agreement between a mass noun and a mass classifier (15a), a count noun and a count classifier (15b), and a noun indicating pair and pair classifier (15c).

198 Tanmoy Bhattacharya (15)

a.

b.

c.

Onek-khani doi a lot-CLA yoghurt 'a lot of yoghurt' Onek-gulo chele a lot-CLA boy 'a lot of boys' kOek-joRa juto some-CLA shoe 'some pairs of shoes'

I will come back to this point in section 2.1 where we talk about agreement as a criterion for head determination.

1.2.2.

Num/Q-Cla as a functional head

The argument that functional heads define a closed class is shown to work to some extent for the complex. The following table shows the various combinatorial possibilities between a quantifier and a classifier (16) Table showing Q + Cla combinations: CLASSIFIERS QUANTIFIERS -Ta

SOb 'all' kOtok 'somewhat' kichu 'some' khanik 'a bit' Olpek- 'a little' prottek 'every other' Onek 'a lot' kOto 'how/so much' Oto 'so much' kOek 'a few' numerals num+Ek 'numorso'

• • • • • • • • • • X

Gulo

khana/khani

jOn

[+count]

[+count/mass]

[+human]

• • X X X X

• X X X X X • • • • • •

X X X X X X •

·/

• X X X

• • •

On the one hand, it shows that since Num/Q-Cla is a compositional head, it is less of a closed class than its individual components but on the other, it also shows the relative restrictions on combinatorial possibilities.

Numeral/Quantifier-Classifier as a complex head 199

Another criterion for a functional character of the complex is the lack of descriptive content. This holds as well for the complex as it does not pick out a class of objects (unlike demonstratives and possessives in Bangla) but elaborates some property of the complement noun. These two observations leads one to the conclusion that the complex may be a functional rather than a lexical head. As pointed out at the end of section 1.1 the duality of status in terms of whether a particular head is lexical or functional makes the nature of the head in question semi-lexical. The last two sections showed that while Qfloat and "agreement" make the complex lexical, its relative closed class properties and lack of descriptive content makes it functional. This ambiguity in its status confirms its semi-lexicality. 2.

Zwicky criteria for headedness

The following five criteria discussed in Zwicky (1985) are taken up in the subsection 2.1-2.5.12 (i) (ii) (iii) (iv) (v)

Agreement Obligatory constituent Distributional Equivalence Subcategorisation Governor

In the following discussion, it will become clear that (ii) and (iii) are vari13

ants of each other and that (iv) and (v) reduce to a single property when translated in terms of the concept of Merge within minimalism. However, whether reducible or not, these criteria establish the headedness of the Num/Q-Cla complex. 2.1.

Agreement

Zwicky uses the phrase Determination of Concord for this criterion which I have simplified here as agreement. He claims that the dependent always triggers agreement on the head. This is to be found in languages with object agreement (thus the V is the head of the VP). He further distinguishes determinant of control from governor as follows. In both these cases, the morpho-syntactic features of one element determines those of another but

200 Tanmoy Bhattacharya

in the case of concord the same features are involved in both the determiner and the determinant. Both Zwicky (1985) and Hudson (1987), who criticises and then redefines many of the criteria formulated by the former, also talk of another generalisation that holds for agreement, namely, that agreement morphemes agree with NPs. In the case of verbs and adpositions, the NPs act as the argument of the predicate but with modifiers the NP is the constituent containing the modifier. Croft (1996: 40) shows that in some languages the genitive modifier agrees with the head (Serbo-Croatian) and in some the head agrees with the modifier (Mam). Hudson (1987: 116) argues that the relevant features in a NP are fixed independently by the semantics of the noun, the features, therefore, always spread from the noun. In the context of Bangla, where agreement is not a pervasive clausal phenomenon (showing up only in person agreement), the shape of the classifier morpheme used in the DP can be considered as the only remnant of agreement, similar to the comments made at the end of section 1.2. Thus, in (17) below, the classifier chosen is determined by some feature of the noun; names of the classifiers are indicated in square brackets and the classifiers shown in boldface in the text: (17)

a.

b. c.

du-To chele/boi two-CLA boy/ book 'two boys/ books' du-jon chele/ *boi two-CLA boy/ book du-khana *chele/boi two-CLA boy/book

[General classifier]

[Human classifier] [Inanimate Count classifier]

Similar to the data set in (17) the classifier in Bangla can be shown to display number: (18)

a.

b.

du-To boi two-CLA book 'two books' dui-dOl hati two-CLA elephant 'two groups of elephants'

Numeral/Quantifier-Classifier as a complex head 201

c.

d.

kOto-gulo chele so many-CLA boy 'so many boys' Ek-gocha phul one-CLA flower 'a bunch of flowers'

If we consider that the noun selected (or the DP selected) has an abstract feature of plurality, then the classifier must be chosen accordingly. By Zwicky's definition, the classifier could be either the controller of concord or a governor (although Zwicky does not talk about Determiner + noun sequences for governor) but by Cann's (1993) formal definition this is a case of concord. This is also the case with Swedish where nouns are inherently marked for number and gender but the demonstratives/determiners are marked morphologically. The Bangla example in (18) therefore exhibits co-variance by concord. Therefore by Zwicky's criterion, the Num/Q-Cla complex, whose form is altered in some sense in (18), is the Head on which the dependent noun triggers agreement.

2.2.

Obligatory constituent

By this criterion, the head should be the obligatory constituent in the unit. By implication, non-heads are optional. Thus Peter played the game or Peter played are fine since the verb is the head but not *Peter game. Zwicky makes this criterion more restrictive by narrowing down the meaning of optionality: optionality that is due to ellipses is excluded from consideration. Thus in Swedish, the example (19b) is not considered for head determination since it involves an elliptical noun one. So the noun in (19b) is not optional and therefore D is not the head by this criterion. The adjective in (19c) is treated as a fully nominalised adjective as it has the narrow meaning typical of nominalised adjectives in Swedish. These considerations make the noun the head of the construction.

202 Tanmoy Bhattacharya (19)

a. b. c. d.

Dessa gamla kvinnor har det svärt these old women have it difficult Dessa har det svärt these have it difficult Gamla har det svärt old have it difficult Kvinnor har det svärt women have it difficult (Böijars 1998: 113)

In the following Bangla near equivalents of the Swedish example, the issue of ellipses is not relevant as the question of the demonstrative (not shown here) being a head or not is not relevant in the present context. See Bhattacharya (1999a) for a detailed discussion of the status of the demonstrative in the Bangla DP. (20)

a.

kOek-jon buRo lok aSbe some-CLA old man come.3 b. [QP kOek-jon] aSbe c. *[NP buRo lok] aSbe d. *[NP lok] aSbe14 e. *[ADJ buRo] aSbe15

The example above shows that without the obligatory presence of the whole complex (in this case kOek-jon), the supposed head, the constructions is unacceptable in the intended sense. Note that Hudson relates obligatoriness to subcategorisation by citing the following example: (21)

I got two books but I didn 't read either.

In (21) either subcategorises for a zero noun and therefore (21) is acceptable without a complement following either. As far as Bangla is concerned, the data in (22) shows that only numeral/ quantifier and the classifier together can act as a head by this criterion:

Numeral/Quantifier-Classifier

(22)

as a complex head 203

a.

*du/ *To/ du-To chele two/ CLA/ two-CLA boy 'two boys' b. *kO/ *jon/ kO-jon chele some/ CLA/some-CLA boy 'some boys'

This leads us to the conclusion that both the constituents of the Num/Q-Cla complex count as obligatory for the purpose of head determination.

2.3.

Distributional Equivalence

This Zwicky criterion states that a head is the constituent that belongs to a category with roughly the same distribution as the construct as a whole. This derives from 2.2 above since if the head is the obligatory constituent it is obvious that it will have roughly16 the same distribution as the construct, and certainly more than the dependent. It must be pointed out that Zwicky himself rejects this criterion as a relevant one. But consider the following: (23)

a.

du-jon chele khelo two-CLA boy ate 'two boys ate.' b. du-jon khelo 'two (persons) ate.' c. Ichele khelo 'boy ate.'

Note that the example in (23c) has some crucial information lacking, therefore semantically it is anomalous. But du-jon (of (23b)) is distributionally equivalent to the whole QP and thus must be considered the head. Other examples, establishing the same point are given below. (24)

a.

Onek-gulo baMdor a lot-CLA monkey 'a lot of monkeys' b. *Onek17/ *gulo baMdor

204 Tanmoy Bhattacharya

(25)

a.

SOb-gulo

chele

all-CLA boy 'all the boys'

b. *SOb/ *gulo chele

2.4.

Subcategorisation

The basic point about this criterion is that an element that requires a subcategorisation frame is a head and therefore this element needs to be listed in the lexicon. For example, in a V+NP construction, V requires a subcategorisation frame and is therefore the head. This requirement is satisfied by the same examples as (17) and (18) above if we consider that the NPs are selected by the Num/Q-Cla complex. To elaborate further, consider the example in (18b) repeated below: (26)

dui-dOl

hati

two-CLA elephant 'two groups of elephants'

As per the suggestion earlier, if we consider choosing a particular classifier as an "agreement" morpheme on the complex then the fact that the noun hati triggers agreement on the complex as a whole correctly predicts that the complex is the head. Also, as per the subcategorisation criterion, it can be shown that the noun hati cannot have a frame for the selection of the classifier dOl since it can combine with many other modifiers: (27)

a.

paMc-Ta

hati elephant Onek-gulo hati a lot-CLA elephant kOek-Ta hati some-CLA elephant five-CLA

b.

On the other hand, the classifier dOl must have a requirement (and therefore a subcategorisation frame) that the complement noun be able to create a collective. Since we have shown that the Cla is part of the complex, it is possible then to claim that the complex is the head of the phrase. However, according to Zwicky, the criteria of subcategorisation and governor need to be related to the semantic functor status of the element in

Numeral/Quantifier-Classifier as a complex head 205 question (and not its head status). In a way, this implies that an element can be the semantic head (two boys being types of boys rather than types of twos), but need not be the syntactic head. I am not sure if this argument can be transported to the domain of classifier expressions. For example, in (26) above, there is no rigorous way to tell whether the DP is about elephants or groups. One way out of this impasse is to consider the criterion set up by Cann (1993). He argues that if two elements are related by agreement then they have the spec-head syntactic relation, on the other hand, if they take restricted complements (i.e. needs a subcategorisation frame) they have a head-complement relation. In the examples above, since there is no strict agreement relation between the complex and the noun in the conventional 18

sense, the relation must be of head-complement rather than spec-head . However, given the ambiguous nature of the applicability of this criterion, I suggest that this requirement, together with the next criterion (see 2.5), falls out of the way Merge operates. I will therefore recast example (18), which is a valid example for this criterion, in terms of Merge. It may be noted that the earlier contention (in section 2.1) that a semantic feature of noun derives the shape of the classifier, is not in conflict with the syntactic requirement of the Num/Q-Cla head to select an NP subcategorisation frame. 2.5.

Governor

By this last criterion, the head of a construction is the constituent that governs the grammatical form of its sister constituent. Zwicky claims this to be different from subcategorisation as the form of the complement defined by government does not enter into semantic interpretation. Consider (28) in this connection. Cardinaletti and Giusti (1989) rule this out by suggesting that existential quantifiers assign partitive Case to their complements which cannot prepose to a subject position: (28) *Childreni are many tj noisy

[It.]

This is based on the observation that in Italian a partitive clitic is allowed since the clitic is not moved to a Case assigning position:

206 Tanmoy Bhattacharya (29)

Ne ho visti molti (I) of-them saw many Ί saw many (of them).'

If this is true then it shows that the quantifier indeed determines the form of the complement (in terms of assigning partitive Case) - a task usually performed by the governor rather than the element that needs subcategorisation.19 Given that in the minimalist framework there is no scope for a rule of lexical insertion based on subcategorisation frames, and because of the elimination of government as a grammatical concept, it is desirable to derive the criteria of subcategorisation and governor from some other source.

2.6.

Merge

Consider the fact that if a numeration Ν selected from the lexicon to construct a DP is (30), then a derivation as in (31a) crashes as the human classifier jon cannot be merged with a non-human noun; the derivation in (31b) which selects a human complement goes through: (30)

Ν = {du-jon 'two-CLA', boi 'book', chele-er 'boy's'}

(31)

a.

b.

{du-jon, boi} {chele-er, {du-jon, boi}}

20

* or * {du-jon, chele-er} {boi, {du-jon, chele-er}} or

A matching of features between the Num/Q-Cla complex and the following noun must be established for the derivation to proceed. Notice that a feature matching requirement for Merge is employed here. Let us see if that is a good strategy by itself. In the minimalist program, Merge is a basic operation (shown in (32)) whereby phrase structures are built up piece by piece as the computation proceeds. (32)

Merge (α,β) = {α,β}

Numeral/Quantifier-Classifier

as a complex head 207

Given a pair of syntactic objects (α, β) which are selected from the Numeration, the operation 'Merge' constructs a new syntactic object out of the pair (α, β) creating a single syntactic object (K). The operation Merge (α, β) is asymmetric, projecting either α or β. The element which projects becomes the label of the complex. In general, the syntactic object Κ must be of the form {γ, {α, β}}, where γ identifies the type to which Κ belongs, γ is called the label of Κ (not shown in (32) above). Notice crucially that the order of the merged elements is irrelevant in this proposal. The notation {α, ß} in (32) states precisely that. Collins (1997: 64) points out that this may not be sufficient as the operation fails to identify the head of the derived constituent. Collins rightly observes that the operation in (32) does not distinguish between segments and categories. Although finding the head is not an operation, Collins assumes that it is calculated automatically at the time the constituent is formed by Merge one simply finds the head of one of the daughters. Consider the following derivation from Collins (1997: 64) to see this more clearly: (33)

a. b. c. d. e.

Select V Select Ν Merge (N,V) = {N,V} Head ({N,V}) = V Select Agr0 Merge (Agr0, {N,V}) = {Agr0, {N,V}} Head ({Agr0, {N,V}}) = Agr0

If instead, at (33c), Ν was chosen as the head, at LF we would have an NP with a V complement. Collins concludes that it is reasonable to assume that the V will be uninterpretable at this position. Note that such an assumption rests on a grammar model with a look-ahead facility which is presumed to inflate the complexity of the computational component of the grammar.21 However, Collins proposes a principle of integration which responds to this. Consider the following partial derivation of John left: (34)

a. b. c.

Select John Select left Merge {John, left) = {John, left}

208 Tanmoy Bhattacharya

The question that we have been trying to answer is what motivates the Merge in (34c). It is unlikely that a feature of either John or left is being checked through Merge. One possibility is that in selecting either of the two lexical items, a property of the lexical item concerned is being satisfied, namely, the property of being taken out of the Numeration (and consequently its associated integer reduced by one). This is rejected by Collins on the grounds that if two phrases (and not lexical items) are merged, no appeal to the Numeration is made. He assumes the alternative that Merge of α and β, whether lexical or not, is driven on the basis of the fact that both must be integrated into the clause. He calls this trigger for Merge which involves no feature checking, Integration, and defines it as follows: (35)

Every category (except the root) must be contained in another category. (Collins, 1997: 66)

Collins further points out that Integration is conceptually related to the LCA since if a phrase is not integrated into a clause, its terminals will not be ordered with respect to other terminals of the clause. One possible way of looking at this relation is that Integration follows from the Linear Cor23 respondence Axiom (or the LCA) (Collins 1997: 69). In Bhattacharya (1999a), I have adopted this view and considered the LCA as the trigger for Merge.24 Although I argue against a selectional approach to Merge in Bhattacharya (1999a), it is nevertheless possible to construct a case in its favour based on a recent monograph by Chomsky. Chomsky (1998) distinguishes between set-Merge for merger by substitution and pair-Merge for merger by adjunction. Adjunction is inherently asymmetric (X is adjoined to Y) and leaves the category adjoined to, unchanged. So pair-Merge of α to β will project the target β. Set-Merge as an operation is symmetric, so either label may project. The result is either interpretable at LF or not. Such a formulation implies look-ahead as part of the language design since Merge proceeds in the manner dictated by the success of the derivation at LF. I will rejected this approach since increasing the complexity of the computation is undesirable. Chomsky sees a way out of this. Set-Merge also has an inherent asymmetry since α, β merge in order to satisfy selectional requirements of one of them (the selector) but not both. Chomsky observes that the selector is uniquely determined (emphasis mine). In particular he opts for a featural account for Merge triggers. A feature F of one of the merged elements in {α,β} must be satisfied for the

Numeral/Quantifier-Classifier

as a complex head 209

operation to take place. F is in the label of the selector and the label of the selector projects. Chomsky does not say which is the selector but I assume that the head is the selector. Since we have already proposed LCA as the trigger for Merge, the first Merge is decidedly {head, complement}. In conclusion, in the case of the asymmetric operation pair-Merge there is no selector whereas setMerge has a unique and obligatory selector which determines the label of the construction. Although many questions remain unanswered, it is nevertheless possible to derive the criteria of subcategorisation and governor from this feature matching requirement of Merge. 3.

Syntactic Evidence for the Semi-Lexicality of the Num/Q-Cla Complex25

In this section, I will provide some syntactic reasons for the complex headedness status of Num/Q-Cla. Recall from section 1, that according to Emonds (1985), one of the properties of semi-lexical heads is that they do not expand. In the context of the Num/Q-Cla complex, I have interpreted this as implying the impossibility of splitting up of the complex. In this section, I will first consider splitting up the complex head and show that it cannot be done for some well-formed syntactic reasons.

3.1.

Data on All and Non-All Quantifiers (NAQ)

Notice first the behaviour of the quantifier SOb 'all' in the following pair: (36)

a.

b.

SOb gulo chele aSbe all CLA boy come.FUT 'All the boys will come.' SOb chele gulo aSbe all boy CLA come.FUT 'all the boys will come.'

210 Tanmoy Bhattacharya

The difference between the two is that in (36b) SOb 'all' quantifies over a particular set of boys, a set which has a prior discourse reference. (36a) on the other hand is a quantification over an exhaustive set of boys. Additionally, (36b) shows, for the first time, that an NP can appear between Q and Cla. This would suggest that these two ought to be split up into two heads and that unlike -7a, the classifier gulo does not cliticise to the Num/Q. Before making any proposals, let us look at quantifiers other than all, which I identify, for purely mnemonic reasons, as non-all quantifiers (NAQs): (37)

a.

Onek gulo chele aSbe a lot CLA boy come.FUT 'a lot of boys will come.' b. *Onek chele gulo aSbe

(38)

a.

kOtok gulo chele aSbe some CLA boy come.FUT 'some boys will come.' b. *kOtok chele gulo aSbe

That is, in the case of NAQs, the leftward NP movement is disallowed. Recall that in connection with the data in (14) it was pointed out that the type of NP movement that (14) depicted involves topicalisation and is not the topic of discussion in this paper. Rather, the type of NP movement dealt with in this section are of the type shown in (39a) below where the NP moves between the Q and Cla rather than moving to some distant and outer spec position. This latter type of movement is probably triggered by some sort of a topicalisation feature. (39)

a. b.

[QP Q C L A ] NP [QPQCLA]NP

[QP Q NPj C L A ] tj NP;... [ Q P Q C L A ] ti

The same pattern of NAQs as in (37) and (38) as above is obtained with other classifiers and quantifiers. One possibility of accommodating the above data is by splitting the Num/Q-Cla into two separate heads Q and Cla:

Numeral/Quantifier-Classifier as a complex head 211

(40) spec

Q' ClaP

Q SOb

Spec A

Cla' Cla gulo

NP chele

The movement of the NP to [Spec,ClaP] would derive the order in (36b) whereas no movement is necessary for (36a). However, the above derivation is incorrect for at least three reasons: (i) Given the reasons for the headedness of the Num/Q-Cla complex and given the data in (37) and (38) above, it is likely that the Num/Q-Cla sequence is formed through head adjunction of Q and Cla. If that is the case then the derivation in (40) would give us the wrong order of [Cla-Q]. This is based on the reasoning that adjunction is always to the left. Although there are proposals in the literature in favour of a right adjunction at the word level, I will consider adjunction as always to the left for uniformity of analysis without committing myself one way or the other whether these movements are part of morphology or syntax. So the revised structure is as follows: (41)

ClaP spec

Cla'

Cla gulo

QP

φ Spec Q SOb

Q' NP chele

212 Tanmoy Bhattacharya This is the derivation for (36a) achieved through head adjunction of Q to Cla, but not (36b). One sensible possibility is to move the whole QP to [Spec.ClaP]. This will however not stop the derivation of the unwanted (37b) and (38b): (42)

* [ c i a P [QP

Onek [np chele]] gulo

tQP]

Secondly, this would imply that a feature of the Cla is responsible for the movement of the QP to its spec. This is not true since using any other classifier in place gulo in (37) and (38) would produce the same ungrammaticality. Based on these observations, I reject the head analysis of SOb. (ii) The structure in (40) cannot explain why the NP does not move in the case of NAQs. A closer inspection of the make-up of the quantifiers in the NAQ group reveals that all of them contain some indivisible version of the word for Ek 'one', at times morphologically unrecognisable:26 a. b. c. d. e. f.

Onek kOek khanik Olpek prottek kOtok27

'a lot' 'a few' 'a bit' 'a little' 'each one' 'a few'

I call this morpheme Vague-one since it gives a vague meaning to the numeral. The presence of this morpheme in some form bars the possibility of moving an NP between the Num/Q and the Cla. Thus, some feature of the quantifier decides on the NP movement noticed in (36b) and the lack of it in (37) and (38). (iii) The most serious problem with the derivation in (40) is its inability to distinguish between the two classes of quantifiers both of which are identified as Q heads in this structure. The difference between all and other quantifiers is well-established in the literature (e.g. Shlonsky (1991) for Hebrew, Giusti (1991, 1995) for Italian, among others). In connection with Bangla, one difference in their morphological make-up is immediately clear if we consider SOb in relation to the data in (43). SOb does not carry either a hidden or visible counterpart of the Vague-one morpheme shown in (43).29 Based on the discussion in this section, I conclude that SOb is an XP and is base-generated at [Spec.QP].

Numeral/Quantifier-Classifier as a complex head 213 3.2.

Revisiting the relevant data: back to Q and Cla as a fused head

Armed with the conclusion from the preceding section let us look at the relevant data presented in section 3.1 again. (44)

a.

b.

(45)

SOb gulo chele aSbe all CLA boy come.FUT 'All the boys will come.' SOb chele gulo aSbe all boy CLA come.FUT 'all the boys will come.'

a.

Onek gulo chele aSbe a lot CLA boy come.FUT 'a lot of boys will come.' b. *Onek chele gulo aSbe

It is clear from this data that SOb is different from NAQs in allowing the NP to appear between it and the classifier. Now with the conclusion that SOb is indeed different, I claim that the structure of the Bangla DP where the middle layer is a complex head (represented by) Q has a natural way of accommodating the data related to SOb. That is, the Q and the Cla should not be split into two separate heads. The headedness of the Num/Q-Cla, therefore, stands. The derivation for (44) (=(36)) (minus the verb) is shown in (46). Note that the Q in (46) hosts the classifier gulo. This need not be confusing since the Q node is a complex node demonstrated in this section as being Q+Cla internally, and since the internal Q is empty in this example, the only remaining element under the complex head Q must be the classifier. (46)

a.

^

^

Spec

(Represents (44a)) Q'

SOb Q gulo

NP chele

214 Tanmoy Bhattacharya b.

(Represents (44b))

QP Spec Q' SOb Spec Q' chele\ Q NP gulo ti

Apart from the fact that we do not require another head for the classifier, this analysis is desirable on three counts: (i) Note that the derivation in (46b) exhibits leftward NP movement inside the DP. This has been claimed to be the major thread of discovery in Bhattacharya (1998a et seq). The analysis of SOb therefore provides additional evidence towards this demonstration. (ii) Notice that the derivation in (46b) crucially depends on the availability of multiple specifiers. I claim that this is expected (a) given the minimalist framework adopted for this study and (b) confirms a crucial principle proposed in Bhattacharya (1999c), Tuck-in, based on Richards (1997) which predicts that later XP movements target inner specifiers. (iii) The analysis in (46b) provides an elegant solution to the puzzle of NAQs. Note that in (45b) (similarly for other NAQs) the NAQ Onek does not allow the leftward NP movement noticed with SOb. Recall one of the differences between the two types of quantifiers elaborated in section 3.1. NAQs were shown to embed a special morpheme -Ek 'one' which was missing in SOb. The analysis in (46b) has a natural way of incorporating the connection between this morphological observation and the lack of NP movement in NAQs as follows. DP-internal NP movement in Bangla is due to the presence of a feature of [SPECIFICITY] on the (complex) Q head. Similarly, the NP movement shown in (46b) above is also due to such a feature of the Q. In the case of NAQs, the -Ek morpheme makes the Q head non-specific. This is not unlikely, given that (at least) the Vague-one morpheme makes the meaning vague or non-specific. The derivation for NAQs, therefore proceeds as follows:

Numeral/Quantifier-Classifier as a complex head 215

(47)

QP Spec

(Represents (45a)) Q'

Q Onek-gulo

NP chele

The NP cannot move up because there is no attractor feature in Q. 4.

Conclusion

The sequence of the quantifier or the numeral followed by the classifier in Bangla is shown to behave like one unit. This behaviour is most pronounced descriptively, as shown in sections 1 and 2, and syntactically (section 3). Based on the criteria for semi-lexicality or the disguised nature of certain categories in Emonds (1985), it was shown that the complex is one such category. Furthermore, based on certain (revised) criteria of head determination in Zwicky (1985), it was shown that this complex indeed behaves as a head. Lastly, section 3 extended Emond's (1985) argument with respect to the expansion of a category in the realm of syntactic movement and concluded that for well-defined syntactic reasons the complex although morphologically visible as being composed of a Num/Q plus a Cla, cannot be further split into a Q and a Cla for the purpose of syntax.

Notes 1. I am thankful to Misi Brody, Ad Neeleman, Andrew Simpson and Neil Smith for comments and criticism of an earlier version of the paper and to the participants in the Semi-Lexical heads Workshop at Tilburg, in particular, to Norbert Corver and Elizabeth Löbel for comments and questions. Thanks also to an anonymous referee for raising some points which led to useful revisions. 2. For a general introduction to the language and its name, see Bhattacharya (forthcoming). 3. The transcription works as follows: Τ D R = Retroflex 4 r/; S = Palatoalveolar /J/; Ν = Velar /q/; = Ε Ο mid vowels /ae ο/; Μ indicates Nasalisation. 4. This is a case of Kinship Inversion (Bhattacharya 1998b) involving the affectionate classifier allomorph -Ti (also see (lc)).

216 Tanmoy Bhattacharya

5. However, consider the following counter-examples: (i) tin peala ca three cup tea 'three cup tea' (ii) kOto lok! some people 'so many people!' However, (i) is more like a measure phrase and less of a quantified expression. Similarly for (ii), as noted in connection with example (24), although classifier-less quantified DPs are more common than quantifier-less classified DPs, the former also has restricted and special uses (to form an exclamative in (ii) above). I will therefore consider these as marginal cases and not as the norm. 6. Although I finally conclude that all in Bangla is not a quantifier head but is an XP. 7. This space is identified as the middle layer in a three-layered DP structure in Bhattacharya (1998a, 1998b, 1999a, 1999b). 8. Emonds (1987) expressed similar views in his Invisible Category Principle which states that bright-er expresses inflectionally what more bright expresses in terms of a separate word. The pair *{the) bad student (count) versus bad students makes the same point that if a noun is capable of expressing plurality, in the case of count nouns, it is expressed as an inherent syntactic feature which constitutes a functional projection. 9. Note that the two orders have different specificity/definiteness reading. This fact is captured in the translation with the partitive/specific reading indicated within parentheses. 10. However, in section 3.2 it will be shown that quantifiers like kO do not indeed allow NP movement. The type of NP movement shown here must be distinguished from the NP movement of the sort analysed in section 3.2. It will become clear from the analysis in section 3 that I will be concerned with the latter type of NP movement which moves an NP to the specifier of the QP under discussion. The NP movement shown in (14) is of a topicalised variety. The most correct translation of (14) would therefore be 'as for the boys, there were some/ two'. This movement has been discussed to some extent in Sahoo (1999) for Oriya. 11. Recall in this connection that Bhattacharya (1995) and Bhattacharya and Dasgupta (1996) (mentioned in section 1) proposed a parameter between Hindi and Bangla which demonstrates the correspondence between the classifier of the Class languages and the gender/ number marking of the Gender languages. 12. Zwicky also discusses morpho-syntactic locus as another criterion by which an element bearing the mopho-syntactic markers which enable the constituent to link to a bigger constituent is identified as the head. However, in the context of DP-intemal material in a language without agreement, it is difficult to see the usefulness of such a criterion and I will, therefore, keep it out of the discussion. It has also been extensively argued (see Börjars 1998 for a review)

Numeral/Quantifier-Classifier as a complex head 217

that Zwicky's criterion of Functor and Argument cannot decide headedness one way or the other. 13. See Hudson (1987) for a similar argument. 14. This example is unacceptable in the intended sense, i.e., where the meaning intended is about the coming of a particular old man. The expression on its own is fine as an answer to a question. 15. If however buRo here is used as a proper name then this expression is acceptable. 16. See Croft (1996: 37) for questioning the use of this term (and other criteria of Zwicky) and for suggesting an alternative. 17. A classifier-less quantifier is apparently "more" acceptable as opposed to a quantifier-less classifier. However, the contexts under which this is true are identifiable as definite/specific or at least discourse anaphoric (e.g. in answer to a question). This restriction on the context can be read as the presence of an underlying classifier. Additionally, Quantifier+NP can also be used to make an exclamative or a generic existential or possession, making such uses as special cases. However, I leave the discussion of this possibility for future research. 18. Note that his observation is also in line with minimalist theory of agreement as checking between a specifier and a head. By the extension of agreement implied (as "agreement") in this paper, we can consider head-complement relation as establishing this latter extended version of agreement. However, given the speculative nature of this suggestion, a separate discussion of this possibility must be postponed to a later occasion. 19. As mentioned earlier, in Bangla however there is no restriction on NP movement across the Q head as long there is a classifier with a particular feature ([SPECIFICITY]). 20. Both orders may be produced depending on whether there is Move after the first Merge, I have ignored various details which are not relevant for the point being made. 21. See Chomsky (1998) for some relevant discussion on this point. 22. The problem with the definition of root (a category not contained within any other category) is not addressed in Collins. Without such a definition, Integration as stated above is not meaningful. One possible line of approach in defining the root could be in terms of look-ahead. If we say that the grammar needs look-ahead of some variety, contrary to the attempt in Chomsky (1998) of eliminating it, root could be the point where there is no more look-ahead. The asymmetry in Chomsky (1998) pointed out in Bhattacharya (1999a) regarding the reduction of complexity, indicates the possibility of incorporating a certain amount of look-ahead in the grammar. 23. In Collins (1997) he rejects this possibility based on the status and position of the Linear Correspondence Axiom (or LCA) discussed in Chomsky (1994). However, since the conceptual relation between Integration and LCA remains and because Collins (1997: 137) himself suggests the possibility of reducing

218 Tanmoy Bhattacharya

24. 25. 26. 27.

28.

29.

Integration to LCA, perhaps it is likely that some form of LCA is responsible for Integration and therefore, Merge. See, however, Uriagareka (1997) which derives LCA from basic minimalist assumptions. The evidence presented in this section is discussed at a greater length in Bhattacharya (2000). Notice the English glosses suggest a similar presence of one. In quantifiers without a visible -ek morpheme, we get either a reduced Whword (K-word) as in (ia,b) or a demonstrative particle (ic) in front: (i) a. kichu 'some' b. kOto 'how/so many' c. Oto 'so many' It is possible that all these indivisible particles contribute to the featural makeup of the Q head contributing towards a general notion of counting or enumeration. However, I have no idea if this connection between the -ek set and (i) is a robust one or whether it can be stated formally. In English too, this difference is reflected in the following minimal pairs: (i) a. All the boys b. *The all boys (ii) a. *Many the boys b. The many boys See Abney (1987) and Szabolcsi (1987) for some relevant discussion. In discussions by Shlonsky and Giusti on the phenomenon, it has been suggested that the QP embeds the DP. However, there is no evidence in Bangla to consider Qs as external to the DP. In particular, the demonstrative and the possessive which are independently shown to be higher specifiers of the DP in Bangla, always precede SOb: (i) a. ei SOb gulo chele this all CLA boy 'all the boys here' b. amar SOb gulo chele my all CLA boy 'all my sons' See Bhattacharya (2000) for further evidence.

Numeral/Quantifier-Classifier as a complex head 219

References Abney, Steven 1987 The English NP in its Sentential Aspect. Dissertation, Cambridge, MA: MIT. Bhattacharya, Tanmoy 1995 DPs in Bangla. Journal of the Maharaj Sayajirao University of Baroda 43.1 1998a DP-internal NP movement. University College London Working Papers in Linguistics 10: 225-252. 1998b Kinship inversion in Bangla. Papers in Linguistics from the University of Manchester, volume 7: 143-156. 1999a. The structure of the Bangla DP. Dissertation, London: University College. 1999b. Specificity in the Bangla DP. In: Rajendra Singh (ed.), Yearbook of South Asian Language and Linguistics Volume 2, 71-99. New Delhi/ London: Sage Publications. 1999c. Tuck-in: XP movement to multiple specifiers in DPs. Talk presented at University College London. 2000 In search of the vague 'One'. In: Tina Cambier-Langeveld, Anikö Liptäk, Michael Reford and Erik Jan van der Torre (ed.), Proceedings of ConSOLE 7 (Conference of the Student Organisation of Linguistics in Europe), 33-48. Leiden: SOLE. forthcoming Bangla; Bengali. In: Enclyopedia of World's languages: Past And Present. New York: W. W. Wilson. Bhattacharya, Tanmoy and Probal Dasgupta 1996 Classifiers, word order and definiteness. In: V.S. Lakshmi and A. Mukherjee (ed.), Word Order in Indian Languages, 73-94. Hyderabad: Booklinks. Börjars, Kersti 1998 Feature Distribution in Swedish Noun Phrases. Oxford: Blackwell. Bobaljik, Jonathan 1995 Morphosyntax: the syntax of verbal inflection. Dissertation, Cambridge, MA: MIT. Cann, Ronnie 1993 Patterns of headedness. In: G.G. Corbett, N.M. Fraser and S. McGlashan (ed.), Heads in Grammatical Theory, 44-72. Cambridge: Cambridge University Press. Cardinaletti, Anna and Guiliana Giusti 1989 Partitive ne and the QP hypothesis. University of Venice working Papers in Linguistics 91.

220 Tanmoy Bhattacharya

Chomsky, Noam 1994 Bare phrase structure, In: G. Webelhuth (ed.), Government and Binding Theory and the Minimalist Program, 383-439. Oxford: Blackwell. 1995 Toward a theory of optimal derivations. In: Robert Pensalfini and Hiroyuki Ura (ed.), Papers on Minimalist Syntax. MIT Working Papers in Linguistics 27: 65-103. Cambridge, MA: MIT 1998 Minimalist inquiries: the framework. Unpublished manuscript, Cambridge, MA: MIT. Collins, Chris 1997 Local Economy. Cambridge, ΜΑ: ΜΓΓ Press. Croft, William 1996 What's a head? In: J. Rooryck and L. Zaring (ed.), Phrase Structure and the Lexicon, 35-76. Dordrecht: Kluwer. Dasgupta, Probal 1983 On the Bangla classifier Ta, its penumbra, and definiteness. Indian Linguistics 44: 11-26 Emonds, Joseph 1985 A Unified Theory of Syntactic Categories. Dordrecht: Foris. 1987 Invisible category principle. Linguistic Inquiry 18.4: 613-632. Giusti, Guiliana 1991 The categorial status of quantified nominals. Linguistische Berichte 136: 438-452. 1995 Heads and modifiers among determiners: evidence from Rumanian. In: G. Cinque and G. Giusti (ed.), Advances in Roumanian Linguistics, 103-125. Amsterdam: John Benjamins. Hudson, Richard 1987 Zwicky on heads. Journal of Linguistics 23: 109-132. Klein, Sharon 1980 Determiner and the category Q. Unpublished manuscript. Löbel, Elizabeth 1989 Q as a functional category. In: C. Bhatt, E. Löbel and C. Schmidt (ed.), Syntactic Phrase Structure Phenomena In Noun Phrases and Sentences, 133-157. Amsterdam: John Benjamins. Richards, Norvin 1997 What moves where when in which Language? Dissertation, Cambridge, MA: MIT. Sahoo, Kalyanamalini 1999 The Syntax of Definiteness in Oriya. In: Rajendra Singh (ed.), The Year Book of South Asian Languages & Linguistics, 101-117. New Delhi, Thousand Oaks, California and London: Sage Publications.

Numeral/ Quantifier-Classifier as a complex head 221

Shlonsky, Ur 1991

Quantifiers as functional heads: A study of quantifier float in Hebrew. Lingua 84: 159-180. Szabolcsi, Anna 1987 Functional categories in the noun phrase. In: I. Kenesei (ed.), Approaches to Hungarian 2,167-190. Szeged: JATE. Uriagareka, Juan 1997 Formal and substantive elegance in the minimalist program (On the emergence of some linguistic forms). In: M. Bierwisch, H.M. Gaertner and C. Wilder (ed.), The Role of Economy Principles in Linguistic Theory, 170-204. Berlin: Akademie Verlag. Zwicky, Arnold 1985 Heads. Journal of Linguistics 21: 1-29.

Classifiers and semi-lexicality: Functional and semantic selection Elisabeth Löbel

1.

Introduction1

So-called pseudopartitive constructions equivalent in meaning to English a pound of fish or two lumps of gold are said to occur in all languages, whereas numeral classifier constructions which might be glossed in English as one living being fish or two fruit orange are confined to classifier languages which lack number distinction. In classifier languages, entitydenoting nouns meaning 'fish' or 'orange' are said to be 'transnumeral', which means that they are combinable with numerals or demonstratives only by means of a classifier. In this respect, they are similar to massdenoting nouns. Numerals constitute the paradigm case, and I will confine myself to these numeral classifier constructions, using data from Vietnamese which is unanimously regarded to be a typical classifier language. Below, the structural parallel between pseudopartitive constructions in (1) and classifier constructions in (2) is illustrated: (1)

a.

b.

mot cän cä one pound fish 'a pound of fish' hai cuc väng two lump gold 'two lumps of gold'

VIET

224 Elisabeth Löbel

(2)

a.

b.

möt con ca one living being fish 'a fish' (con 'living being; classifier for N[+animate]') hai qui cam two fruit orange 'two oranges' (qui 'fruit; classifier for fruits')

The terminology used in the literature on classifiers, namely "mensural classifiers" for nouns such as pound or lump in (la) and (lb) vs. "sortal" classifiers such as con 'living being' in (2a) or quä 'fruit' in (2b), suggests that, the differences notwithstanding, there is also a common denominator, since constructions such as *two gold (in contrast to two lumps or piles of gold) are as ungrammatical in English as in Vietnamese. On the other hand, in Vietnamese, there is an important difference within the class of mensural classifiers: Measure nouns or even container nouns such as bao 'bag' used as a measure noun in (3a) may be modified by a noun such as ri/di 'half, i.e. lexical material may intervene between the measure noun and the quantified noun. In contrast, this is allowed neither for quantitydesignating numerative nouns such as cuc 'lump' in connection with mass nouns in (3b) nor for classifiers such as quä 'fruit' in connection with entity-denoting nouns in (3c): (3)

a.

b.

c.

hai bao ri/di cam two bag half orange 'two and a half bags of oranges' (lit.: 'two bags and a half oranges') hai cuc väng rifdi two lump gold half 'two and a half lumps of gold' (lit.: 'two lumps of gold and a half) hai qui cam rifdi two fruit orange half 'two and a half oranges' (lit.: 'two fruits of orange and a half)

Classifiers and semi-lexicality: Functional and semantic selection

225

The contrast between (3a) on the one hand and (3b,c) on the other illustrates that, from a syntactic point of view, both mass- and entity-denoting nouns, with regard to their combinability with number (i.e. their non-countability) behave exactly alike. This phenomenon is interpreted in the literature on classifiers2 as there being no genuine syntactic distinction between these two types of nouns; they all behave somewhat like mass nouns. Therefore, the following questions arise for a classifier language such as Vietnamese: (i) Are there any syntactic criteria to distinguish between these two classes of nouns (count vs. mass), despite their identical syntactic behavior? (ii) Furthermore, and this question also applies to English, how can the relation between the quantity-designating noun and the quantified noun be characterized in syntactic terms? Both the Vietnamese examples and their English equivalents have in common that noun phrases such as (1) to (3) consist of two constituents; on the level of semantic interpretation, however, they have only one referent, i.e. they "show the behavior of single projections rather than dual projections" (Van Riemsdijk 1998: 13). The paper is organized as follows: In section 2, I will discuss some recent proposals concerning the analysis of classifier and measure phrases that I take as representative of this type of noun phrase; special emphasis is laid on the status of the quantity-designating nouns and classifiers, respectively. In section 3, I will present my own approach, where, apart from functional selection, two kinds of semantic selection are assumed, namely (i) selection of a participant role, and (ii) selection of a non-participant role. In section 4, this approach is applied to classifiers in Vietnamese. It is shown that the syntactic function of a classifier, namely to serve as a unitcounter in order to make the selected noun countable in the sense of combinable with numerals, is not confined to a closed class of nouns. Rather, the definition of grammatical nouns in contrast to lexical nouns as having not only semantic, but also syntactic features (Emonds [1995] 1999) holds for an open class in Vietnamese, which is due to the lack of inflectional morphology in this isolating language. 2.

Classifiers and quantity-designating nouns

There are several proposals to describe classifiers and quantity-designating nouns, which can be subsumed under the following aspects: These nouns are analyzed as (i) lexical instantiations of functional categories (Kitahara 1993; Li 1999; Löbel 1989,1997), (ii) purely syntactic items which are part of late lexical insertion (Kubo 1996), (iii) semi-lexical heads which exhibit

226 Elisabeth Löbel

both functional and lexical properties (Van Riemsdijk 1998), and (iii) basegenerated predicates (Corver 1998).

2.1.

Functional categories

For Kitahara (1993), classifiers in Japanese are functional categories, as it is illustrated in (4): (4)

a. b.

[DP [ D [NCP [NC NP NC]]D]] NCP SpecNCP

(Kitahara 1993: 176)

JAP

^NC^ NP

NC

(NC: numeral classifier)

hon-o san-satu book-Acc 3-Class 'three books' The case-marked NP is raised to the specifier position of the functional projection NC (for 'numeral classifier') in order to check the features of the classifier for compatibility in meaning, and then to the specifier of DP in order to check case. Numeral and classifier together constitute a zero-level category.3 This could be justified on the grounds that classifiers are bound morphemes in Japanese, but despite this morphological criterion, it is rather problematic to generate this combination as zero-level, as a classifier can be optional (5a), or may not occur at all (5b):4 (5)

a.

b.

sono heya-ni-wa isu-ga jyuunana-ko/0 aru. JAP the room-Loc-Top chair-Nom 17-CL exist 'There are 17 chairs in the room.' sono hon-ni-wa hanashi-ga jyuuni-0 aru. the book-Loc-Top story-Nom 12-CL exist 'There are 12 stories in the book.' (Zubin/Shimojo 1993: 499)

In her contrastive analysis of English and Chinese, Li (1999) assumes that singular and plural marking of a noun is analyzed as a functional category Num5 heading the NP. For English, she proposes the structure in (6) for the NP three students·.

Classifiers and semi-lexicality: Functional and semantic selection 227 (6)

NumP Spec

Num'

three Num PI

NP

(Li 1999: 86)

student

Although there is a correlation between the use of classifiers and the absence of plural morphology, she argues that a classifier language can have a plural morpheme, which is -men in Chinese: "Plurality in a classifier language (taking -men in Chinese as an example) shares with plurality in a non-classifier language (taking -s in English as an example) the position where they are generated: both are generated under the node Number" (Li 1999: 75). As the use of -men triggers a definite interpretation, which is illustrated by the contrast in (7), plurality is realized : by the element in Determiner in a classifier language" (Li 1999: 75): (7)

a.

wo qu zhao haizi-men

CHIN

I go find child-MEN Ί will go find the children.' b.

wo qu zhao haizi

I go find child Ί will go find the/some/ child/children.'

(Li 1999: 78)

The use of a classifier does not allow the use of -men in the construction 'Num+Classifier+Noun'; only pronouns and proper names can be suffixed with -men, and in this case, the classifier construction must be postnominal (8): (8)

a.

dui

ta-men sange

(ren)

to them three-Cl person '... to them three.'

(Li 1999: 83)

A classifier language such as Chinese, however, "has an additional Classifier projection between the Number projection and N" (Li 1999: 87):

228 Elisabeth Löbel DP

(9) D

'three students'

NumP

Spec

Num'

san Num PI

C1P Cl ge

NP xuesheng

Clf

Student

(Li 1999: 87)

What is not taken into account by Li's approach, however, is the difference between the use of Number as a functional category in English and Chinese. Num in English represents the grammatical category of number which is obligatorily realized and comprises both singular and plural. These two realizations of number are dissociated in Chinese. The functional category Num in Chinese is optional, it stands solely for the semantic notion of plurality, whereas the functional category Clf represents the syntactic function of unit-counting, i.e. singularity, which is obligatory in certain contexts.6 In Löbel (1989: 146-153), it is claimed that quantity-designating nouns in English and German pseudopartitive constructions are lexical instantiations of a functional category Q7, just as the article is the lexical instantiation of the functional category D: (10)

QP Num drei three

'three pounds of cherries' ^ Q ^

Q[+N] | Pfund pounds

NP | Kirschen (of) cherries

(Löbel 1989: 152)

The syntactic relation which prevails in pseudopartitive constructions, therefore, is functional selection, which stands in contrast to the relation between a thematic element and its complement: 'The syntactic relation between a functional element and its complement is f-selection. F-selection

Classifiers and semi-lexicality: Functional and semantic selection 229

corresponds semantically to the 'passing on' of the descriptive content of the complement" (Abney 1987: 56). A structure such as (10), however, is problematic for the following reasons: (i) in principle, quantity-designating nouns do not constitute a closed class; many container nouns may be used as measure nouns, and restrictions on the use of these measure nouns are based primarily on pragmatic or extralinguistic grounds; (ii) the element o/"is dangling somewhere in the nominal structure" (Corver 1998: 219); (iii) the most important objection is the following: The relation between a functional category and its selected complement is a non-thematic one; the reverse, however, i.e. every kind of selection which is non-thematic must be functional, is not compelling. I will return to this aspect below (section 3).

2.2.

Purely syntactic items

For Kubo (1996), classifiers in Japanese do not constitute functional categories; rather, they are analyzed as purely syntactic items which, in contrast to lexical items, do not contribute to Logical Form. In contrast to Kitahara (1993: 176), where the combination of 'numeral+classifier' is analyzed as a zero-level category, he claims that the numeral is adjoined to the classifier, as it is illustrated in (11): (11)

a.

NP Spec

b.

N' I (WP) Nc Ν I CD Nn Nx I I 5 mai CD 5-mai CD 5-elf 'five CDs'

JAP

(WP): complement Nc: noun (open class) Nx: classifier Nn: number NP = (nominal) QP

(Kubo 1996: 98)

On the basis of the division of among lexical items into purely semantic items, purely syntactic items, and cognitive items (Emonds 1999: 322-23), classifiers are said to be purely syntactic items which are inserted subsequent to any operation contributing to Logical Form. Kubo's criteria for

230 Elisabeth Löbel looking at classifiers as having no meaning are based on the following examples from Downing (1986: 348): (12)

a.

b.

c.

ume 1-ppon plum 1-cl (long, thin object) 'one plum tree/branch' ume 1-kko plum 1-cl (small, roundish object) 'one plum (fruit)' ume 1-rin plum 1-cl (flower) 'one plum flower'

JAP

According to Kubo, the possibility of having one noun with different classifiers is not to be interpreted in such a way "that the meanings of the classifiers add some meaning which nouns do not have originally, but rather that the classifiers' contextual features require certain features of nouns to come out" (Kubo 1996: 119). According to him, certain contexts can impute difQ ferent meanings parallel to (12) even without a classifier: (13)

a.

b.

Ano ume-wa dosshirishite iru. that plum-top massive be 'That plum tree is massive.' Ano ume-wa oishii. that plum-top tasty 'That plum fruit is tasty.'

(Kubo 1996: 119)

This phenomenon, however, is independent of the occurrence or non-occurrence of classifiers and their alleged lack of meaning, as it can be observed in other languages, too: (14)

a. b.

The cherries are in blossom. He doesn 't know whether to grow tomatoes or potatoes.

The assumption that classifiers have no meaning presupposes that for ume in Japanese, according to (13), (i) two different lexical entries are assumed (ume 'plum tree' and ume 'plum (fruit)'), which again reflects the common assumption that (ii) there is a one-to-one correspondence between the classified noun and the classifier. For Vietnamese, however, this criterion does not hold, as the following examples in (15) and (16) illustrate. Even the

Classifiers and semi-lexicality: Functional and semantic selection

231

most general classifier cäi 'thing' does have "meaning" insofar as there is a potential opposition to other classifiers with a more specific meaning. Therefore, even general classifiers, to a certain degree, also contain semantic features: (15)

a.

hai cai nhä two thing house 'two houses'

b.

hai ngdi

VIET

nhä

two place house 'two (large) houses' (16)

a.

{mot) cäy dao (one) stick knife 'a knife (with a handle)'

b.

(möt) cai

dao

(one) thing knife 'a (blade of) knife'

(möt) con

dao

(one) living.thing knife 'a (chopping) knife' For pseudopartitives in English, Kubo (1996: 106-114) proposes a structure which is similar to the classifier construction in (11). The general representation of pseudopartitives is represented in (17); Nc (the open lexical class) is inserted pretransformationally because it has purely semantic features, whereas Nx is inserted posttransformationally, as it has only contextual features; in contrast to Japanese where numbers are adjoined to Nx, they are base-generated in the specifier position in English: (17)

a.

NP Spec

N' Nx Nc

(WP)

(Kubo 1996: 107)

The assignment of Ns such as bunch, bottle, group, etc. to the closed class Nx, however, is not fixed. Number agreement and selectional restrictions

232 Elisabeth Löbel may be triggered either by Nx (bunch, bottle) or Nc {flowers, wine), as it is illustrated by the following examples from Selkirk (1977: 98-100): (18)

a. b.

A bunch of those flowers was/were thrown out on the back lawn. She broke/drank a bottle of good wine.

According to Kubo (1996: 113), this alternation can be accounted for if nouns such as bunch or bottle "can also have a fully semantic meaning, in addition to a usage corresponding in some way to classifiers; namely, words like bunch can optionally have purely semantic features. Then, depending on whether semantic features are present, Nx is inserted either by DL (Deep Lexicalisation, E.L.) or by PL (Phonological Lexicalisation, E.L.)·" In other words, the ambiguous nature of these nouns is accounted for by different levels of insertion. The parallelism between classifiers, which function as unit counters, and marking of grammatical number in languages such as English is accounted for by assuming that countable nouns in English are both Nx and Nc: "The only difference between NPs with count nouns and real pseudopartitives is just that a count noun plays the roles of both Nc and Nx, namely, an English count noun is first generated under Nc and then moves to Nx" (Kubo 1996: 115): (19)

NP SPEC five

^ N ^ Nx

tablets

Nc ti

(Kubo 1996: 115)

The plural suffix -s is base-generated under Nx, i.e. it is the lexical instantiation of Nx; this also holds for classifiers, as they are looked upon as purely syntactic items. There is, however, an important difference between classifiers and plural marking on the one hand, and the quantity-designating nouns which are used in pseudopartitive constructions such as bunch, herd, or group on the other: Apart from a few lexically marked exceptions such as five head of cattle, these quantity-designating nouns must themselves be marked for number, as *two bunch of flowers is ungrammatical. In other words, the

Classifiers and semi-lexicality: Functional and semantic selection 233 characterization of these nouns as lexical instantiations of Nx parallel to classifiers and to the plural marker -s proposed by Kubo does not take into account this important difference.

2.3.

Semi-lexical heads

In Van Riemsdijk (1998: 12-19), quantifier nouns such as number in a number of examples are considered to be functional heads, as they are closed-class items, whereas the other types of nouns which may be used in pseudopartitive constructions (measure nouns, partitive nouns, container nouns, collective nouns, and kind nouns) are semi-lexical heads: "This difference is further reflected in the role that these nouns play under subject verb agreement and under gender agreement with the determiner" (Van Riemsdijk 1998: 17). Especially agreement is taken as evidence that some measure nouns "may waver between functional and semi-lexical status" (Van Riemsdijk 1998: 17); in (20a), the measure noun kilo is functional, in (20b) semi-lexical: (20)

a.

Er zit drie kilo there sits three kilo b. lEr zitten meerdere there sit several

heroine in die zak DUTCH heroin in that bag kilo's heroine in die zak kilos heroin in that bag

These arguments are extended to "qualificational" constructions such as Germ.die Stadt Wien 'the city of Vienna'. For these NPs, the following structure is proposed, with NF for 'functional head', i.e. D(eterminer), N s for 'semi-lexical head', and NL for 'lexical head': (21)

NFP

NS

Nl

(Van Riemsdijk 1998:39)

Notice that number itself, which is said to be a functional noun (Van Riemsdijk 1998: 17), is used as such in (22a); in (22b), however, it is a semi-lexical head:

234 Elisabeth Löbel (22)

a. b.

a number of examples the number five

What seems to be decisive for the status of number as a functional head, therefore, is not the use of that noun as such; rather, it is its use as a relational noun (i.e. a number of examples). This argument also holds for container nouns such as bottle (of wine) as semi-lexical heads, as these nouns, more often than not, are also used in a non-relational fashion, i.e. they denote a concrete entity (a green bottle vs. *a green bottle of wine).10 Despite their diversity, the approaches presented above have in common that, in the case of English quantity-designating nouns, they presuppose a somewhat reduced lexical meaning in comparison to the quantified noun of which they are a sister. Another proposal which treats these nouns as predicates is the topic of the next section.

2.4.

Predicates

Based on recent studies by Moro (1997) for Predicate Inversion in equative sentences and nominal-internal predicate movement within nominal phrases by Den Dikken (1998) for qualitative predicates such as idiot11 in a Ν of a Ν construction in (23), Corver (1998: 216) extends this analysis to nouns denoting quantities or measures in pseudopartitive constructions such as (24): (23)

a. b.

that idiot of a man [dp D[ s c man [p^ idiot]]]

(24)

a. b.

a bunch of flowers [dp D[sc flowers [^bunch]]]

The analysis presented in (23b) and (24b) implies that the two nouns are related by the relation of predication. As it is generally assumed, the predication relationship is structurally represented in terms of a Small Clause (SC). For (25a) a bottle of water, therefore, the measure noun bottle is base-generated as a predicate whose subject is the measured nominal water, with the preposition o/being the nominal counterpart of the copula:

Classifiers and semi-lexicality: Functional and semantic selection

(25)

a. b.

a bottle of water Predicate Inversion: DP FP

a NP bottlei

235

(Corver 1998: 223) of: nominal counterpart of the copula bottle·, predicate nominal water: SC-subject

F XP

F+Xj of+0 NP water

X· X ti

NP ti

His analysis is further supported by a number of grammatical properties which are characteristic of predicate nominale, and which also hold for measure phrases (Corver 1998: 219-221): (i) measure phrases can be linked to the subject by a copular verb like become, (ii) they can be used as an answer to what...like questions, which typically ask for a property, and (iii) they can be coordinated with APs, which are typical examples of predicative phrases (26): (26)

a. b.

John is a judge and proud of it. The range of these guns is only 200 meters but nevertheless sufficient.

These observations are interpreted by Corver (1998: 220) as follows: "Given the descriptive generalization that the conjoined phrases must be of the same semantic category (e.g. arguments only with arguments, and predicates with predicates ..., we can conclude from the example in (20) [= 26b), E.L.] that quantity/measure nominals can function as predicates". The following example, however, illustrates that it is not the notion of predicate in contrast to argument which is decisive for conjoining predicatives. Example (27b) is semantically deviant: (27)

a. b.

John is happy and in a good mood. John is happy and in the library/outside.

??

236 Elisabeth Löbel

Rather, it is the notion of property in the narrow sense which is relevant for conjoining phrases, irrespective of their categorial realization as an NP, AP, or PP. These conjoined properties must be semantically compatible. In other words, the distinction between predicates and arguments is not finegrained enough. It is for this reason that I propose to dissociate the notions of predicate (vs. argument) and property, and to subsume the latter under the notion of non-participant roles such as Manner or Measure (Rizzi 1990), which is the topic of the next section. 3.

Participant and non-participant roles

The different proposals with regard to relational quantity-denoting nouns and classifiers as (i) instantiations of a functional category, (ii) purely syntactic items, (iii) full-fledged nouns which are base-generated as predicates, or (iv) semi-lexical heads which exhibit both functional and lexical properties makes it necessary to take a look at the kind of relation that exists between quantity-designating nouns and quantified nouns. The suggestion that quantity-designating nouns or classifiers are functional categories implies that they lack thematic grids, have selectional properties in terms of syntactic categories, and have grammatical features; this also holds for their status as purely syntactic items. Regarding these nouns as lexical categories implies that they have thematic grids, as they are relational, and have selectional properties in terms of semantic categories (Ouhalla 1991: 9-19). Base-generating these nouns as predicates implies that there is a relationship of predication within the two nominal constituents of a pseudopartitive construction. Especially the latter approach implicitly entails the common assumption that nominals are either arguments or predicates. Furthermore, arguments realized by NPs/DPs are implicitly understood as referential arguments bearing a thematic role with the notion of thematic role being understood (implicitly) as "participant role" (Agent, Patient, etc.). In what follows, I will argue for a refinement of the notions of argument and role, i.e. in elaboration of ideas developed in Rizzi (1990), I will distinguish two kinds of semantic selection, namely (i) selection of a (referential) argument bearing a participant role, and (ii) selection of a (non-referential) argument bearing a non-participant role. This distinction holds not only for verbs, which is the topic of the next section, but also for nouns (cf. section 3.2).

Classifiers and semi-lexicality: Functional and semantic selection 237 3.1.

Verbs

I will start with copular verbs such as become, remain, and be, which constitutes the paradigm case. Especially with regard to the latter, it is commonly assumed that be, in some sense, is "defective": It is not capable of assigning a theta-role, has no descriptive content, does not contribute to the overall meaning of the sentence, its main function being a bearer of tense, etc.12 In what follows, I will concentrate on the widely accepted SC analysis, taking Heggie (1988) as representative for the tremendous amount of literature on this subject.13 According to her, the primary function of be is to indicate a predicate in its postverbal position, not to select arguments. 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). The D-structure in (28) is proposed; 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), it is assumed that the index placed on the predicate spreads to the subject via predication: (28)

a. b.

Johni is [np tj [np a manj] IP [e]

VP V bei [XPi]

XP [XPi]

(Heggie 1988: 122)

This line of reasoning is based on the following assumption: The predicative NP a man is base-generated as a predicate, this being justified by the observation that a man, in fact, has "more meaning" than the copula which is supposed to be rather meaningless. With regard to adjectives, there is no apparent problem with such an analysis, as adjectives, irrespective of their predicative or attributive use, always express a property; in fact, they may be looked upon as the canonical lexical representation of a property. The crucial test case, however, are nouns or NPs/DPs, respectively. The SC analysis, as it is illustrated in (28), does not take into account that it is the copula itself which triggers the property-denoting reading of the NP/DP a man in John is a man in contrast to the individual-denoting reading in

238 Elisabeth Löbel sentences such as They saw a man. In other words, the base-generation of a man as a "predicate" is stipulated and not derived by any other criteria.14 Just as verbs such as meet in (29a) select an argument with the semantic type of individual (only such an argument will yield an interpretation that is consistent with this requirement), verbs such as be, become, remain select an argument which expresses not an individual, but a property (29b): (29)

a. b.

John met a nice girl. Barbara is a nice girl.

Both in (29a) and (29b), there are two NPs/DPs, a subject and an object DP (29a), and a subject DP and a postcopular NP/DP (29b); at the level of interpretation, however, there are two referents in (29a), but only one in (29b). This contrast between an NP/DP as an individual- or a propertydenoting expression can be interpreted in such a way that the copula, in contrast to common assumptions, does make a contribution to the overall meaning of a sentence.15 Properties themselves may be classifying (30) or identifying (31): (30)

a. b.

John was a (good) student. Paris is a beautiful city.

(31)

a. b.

John was the best student. Paris is the capital of France.

Simultaneously, the contrast of (30) and (31) illustrates that whether a NP/DP, depending on the choice of the (definite vs. indefinite) article, is "non-referential" or "less referential" is independent of this aspect (cf. footnote 14). Rather, the use of the definite vs. indefinite article, more often than not, correlates with a property as being interpreted as classifying (30), or as identifying (31). It is for this reason that, with regard to (in)definite NPs/DPs such as a nice girl, I do not assume that one of their readings, e.g. the reading as an individual-denoting expression, can be taken as basic or 'default'. The respective reading of a NP/DP is dependent on the selectional properties of the verb of which it is the complement, i.e. it is dependent on the lexicon}6 Therefore, I will not take the postcopular constituent as a starting-point, but the copular verb itself.

Classifiers and semi-lexicality: Functional and semantic selection 239 3.1.1.

Predicatives as complements

On the basis of minimal pairs such as (32), it is often argued that it is not the type of theta-role which is relevant in syntax, but rather the number of arguments, in (32a) Agent, in (32b) Patient: (32)

a. b.

John broke a vase. John broke a leg.

In contrast, Rizzi (1990) demonstrates that it is necessary to distinguish two kinds of theta-roles and arguments, namely "referential" and "non-referential" arguments and theta-roles. This is illustrated by the contrastive use of ambiguous verbs like weigh. In (33a), the agentive weigh takes a direct object, in (33b), the stative weigh selects a measure phrase: (33)

a. b.

John weighed apples. John weighed 200 lbs.

(Rizzi 1990: 78)

The two readings of weigh differ in their theta grid: (34)

a. b.

weigh weigh

(Agent, Patient) (Theme, Measure)

Patient as well as measure, according to Rizzi, are both theta-marked complements. Based on the observation that 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 'referentiality' 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 argumental or referential theta-roles (agent, theme, patient, experiencer, goal, etc.) and quasi-argumental, non-referential theta-roles (manner, measure, ...)" (Rizzi 1990: 85-86.). 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 (35), shows that referential indices, hence coindexation, are independent of whether a noun phrase has reference or not:

240 Elisabeth Löbel (35)

[Not a single guest]did I wonder why you invited tj.

As measure phrases are intrinsically "non-referential", Frampton proposes to derive sentences such as (36a) from a small clause structure. This structure implies that "the measure phrase is not an argument of any kind, but a predicate" (Frampton 1991: 39): (36)

a. b.

Jack weighed 100 pounds ec weigh [Jack 100 pounds]

(Frampton 1991: 38)

Both Rizzi's and Frampton's analyses suggest that there is a common denominator for verbs such as weigh, which select a "non-referential thetarole" (Rizzi) or a small clause (Frampton), and copular sentences, for which a small clause analysis is commonly assumed. This approach can be justified on the grounds that the measure phrase 100 pounds 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 John. From 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 (36) between an NP as a "non-referential" 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. 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 such as be, become, remain, and probably others, all have in common that they have an external, referential argument bearing the participant role of Theme, and an internal argument bearing a nonparticipant role, as outlined in (37):

Classifiers and semi-lexicality: Functional and semantic selection

(37)

(a) (b) (c) (d)

241

external argument internal argument participant role non-participant role Measure weigh, measure Theme Name Germ, heißen 'to be called' Theme Duration last Theme Property be, become, remain Theme

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 17

specified for a certain dimension (weight, measure, duration, name), 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: (38)

a. The meeting was at 10 o'clock, b. *The woman was at 10 o'clock.

In contrast to the lexical entry proposed by Heggie (1988: 122) in (39), I propose the following entry (40a) for copular verbs; for comparison, the lexical entry of a transitive verb such as hit is presented in (40b): (39)

be :[

(40)

a. b.

i-.XPi]

(Heggie 1988: 122)

be [XPi,XP2] [THEMEj, PROPERTYi ] 18 hit [DPi, DP2 ] [AGENS, PATIENS ]

In contrast to the participant role Agent which is canonically realized as a DP, the non-participant 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.19 In other words, it is not contingent to conclude from this potential categorial variation that it is the "predicate" 20

which is category-neutral. Furthermore, as it is indicated by the indices, the lexical entry in (40) entails that Property is to be understood as 'Property of Theme', as Property is a notion which is relational in itself. In fact, all the well-known peculiarities which "predicates" realized as NPs/DPs exhibit in contrast to

242 Elisabeth Löbel

argumental NPs/DPs (different interpretation of the (in)definite article, lack of scope, restrictions with regard to long movement, cf. Williams (1994: 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 (41) 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:2 (41)

Basic structure: VP subject,

V' V

predicative complement

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 types of roles can be distinguished: (i) selection of participant roles such as Agent, Patient, etc., and (ii) selection of non-participant roles such as Property, Measure, Name, etc. This distinction is summarized in (42): (42)

Selected Argument Agent, Patient, etc. + Property, Measure, etc. + Adjuncts -

Participant Role + -

Selected arguments of verbs such as hit, which bear 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 do not exhibit referential independence. In this sense, they are [-R]-expressions, regardless of whether they are indefinite (John is

Classifiers and semi-lexicality: Functional and semantic selection 243 a mayor) or definite (John is the mayor). 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. It is difficult to illustrate this in English, as English merely distinguishes natural gender, and this only on pronouns. Therefore, in a sequence such as The violet is a flower. It flowers in spring, one cannot decide what the pronoun it refers to. In German, however, as it is indicated in (43), it is by no means ungrammatical to use the feminine pronoun she in order to refer to the predicative NP/DP eine Blume: (43)

a.

b.

Das Veilchen ist eine Blume. The-neutr. violet-neutr. is 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.'

GERM

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 is German wiegen 'weigh', with the selected measure phrase being case-marked by the so-called 'adverbial 23

accusative' , or German heißen 'to be called', where the nominative is used as the default case (citation form). To summarize: There is a small subclass of verbs which select a nonparticipant role (Measure, Name, etc.); copular verbs, which select the nonparticipant role of Property, are part of this subclass. All these verbs exhibit the following characteristics: (i) 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·, as a corollary of (i), these verbs (ii) do not assign case, i.e. casemarking is triggered by agreement (syntactic condition) or is free (semantic condition). It is certainly not by mere chance that this small subclass of verbs dealt with in the preceding section is semantically related to measure phrases (weigh, measure) and restrictive appositives (to name, to be). In the next section, therefore, it is argued that the distinction between participant roles,

244 Elisabeth Löbel i.e. "referential arguments", and non-participant roles, i.e. "non-referential arguments", naturally extends to the lexical class of nouns.

3.2.

Nouns

The alternative to analyzing quantity-designating nouns and classifiers as either lexical instantiations of a functional category or as lexical categories correlates with whether or not they have a thematic grid. The observation that these nouns are relational is compatible with both alternatives. The next question to ask, therefore, is whether their relationality is to be defined by selectional properties in terms of syntactic categories, i.e. f-selection, or in terms of semantic categories, i.e. s-selection. I will start with non-relational nouns.

3.2.1.

Non-relational nouns

It is commonly agreed upon that derived nouns such as translation inherit the thematic grid of the verbs from which they are derived, i.e. these nouns assign the same number of theta-roles that the verb assigns (Ouhalla 1991: 11): (44)

a. b.

Mary translated the book. Mary's translation of the book.

As for non-derived nouns, one could "claim that they have the ability to assign a thematic role on the basis of the fact that they can function as heads of possessive constructions" (Ouhalla 1991: 11): (45)

a. b. c. d.

John's book The book's title The country's borders The President's secretary

(Ouhalla 1991: 11)

The thematic role assigned to the (genitive) subject is 'possessor', i.e. a referential participant role. "Given that virtually all nouns can head possessive constructions, it follows that non-derived nouns in general have the ability to assign a thematic role" (Ouhalla 1991: 11). What is decisive for virtually all non-relational nouns, however, is not assignment, but licensing,

Classifiers and semi-lexicality: Functional and semantic selection 245 i.e. these nouns license a possessor thematic role which, in fact, is an adjunct.24 In analogy, I claim that, in addition, these non-relational nouns also have the ability to license a non-participant role, which is 'Name'. Names, too, denote a property of the referent of the respective head noun, with which they are coindexed. In this sense, they are referentially dependent 25 and, therefore, [-R] expressions: (46)

a. b.

The book 'The Name of the Rose' Umberto Eco 's book 'The Name of the Rose' (Löbel 1991: 29)

Example (46b) illustrates that both kinds of roles, participant and nonparticipant roles, may be combined. Furthermore, names as such do not have a canonical lexical realization, as not only proper nouns (the poet Shakespeare) may function as names, but also other constructions (the film 'Gone with the Wind'), numbers (the report No. 26), and even non-linguistic items (the sign "+"). Parallel to the complements of verbs of naming, names in this function (i) do not occur with articles, (ii) are neither modifiable nor accessible by relative pronouns, i.e. they are syntactically inert, and (iii) are not Case-marked, with the nominative being used as the default case (citation form). It is difficult to illustrate this in English due to the lack of an elaborate agreement system, but the German example (47) may serve as an illustration:26 (47)

der Planet Erde, der/*die GERM the planet-masc. earth-fem. which-masc./*which-fem. um sein/ihr Überleben kämpft for its-masc./its-fem. survival struggles 'the planet Earth which struggles for (its) survival' (Löbel 1991: 16)

This is not to deny that nouns such as Earth do not have any reference at all. Similar to complements of copular verbs (cf. (43)), they exhibit the same behavior as R-expressions with regard to accessibility, i.e. they can be referred to anaphorically:

246 Elisabeth Löbel (48)

Der Planet Erde ist sehr bedroht. GERM the planet-masc. earth-fem. is very threatened. Er/Sie kämpft ums Überleben. it-masc./it-fem. struggles for survival. 'The planet Earth is very threatened. It struggles for survival.'

To summarize, non-relational nouns license the participant role of 'Possessor' as well as the non-participant role of 'Name'. Both kinds of roles are adjuncts, as they are merely licensed, i.e. they are not selected.

3.2.2.

Relational nouns

The distinction between participant and non-participant roles also holds for relational nouns and their selected complements. There are numerous relational nouns denoting parts such as end, head, chapter etc. which select a referential argument, i.e. a PP with a DP which expresses the whole of which the referent of these nouns is a part: (49)

a. b. c.

the end of a/the story the head of a/the department the top of a/the table

In contrast to possessive constructions such as John's book, these nouns denote something which is part of the lexical meaning of the head noun of the complement which denotes the whole. It is for this reason that the Part/Whole relation might be considered to be more fundamental than the possessive relation, as it is suggested by Barker and Dowty (1993: 55): "... we propose that there is indeed a set of proto-roles which are relevant for predicting nominal argument selection. ... We suggest ... that the complementary notions of Part and Whole could serve as nuclei for semantic proto-categories that may govern nominal argument selection." Similar to the complements of deverbal nominals (50), complements of nouns denoting parts cannot be separated by means of a copula (51): (50)

a. the destruction of the city b. *The destruction was of the city.

(51)

a. the top of the table b. *The top is of the table.

Classifiers and semi-lexicality: Functional and semantic selection

247

The complements obligatorily occur with an article, i.e. these nouns select a PP with of assigning Case to the selected DP. This correlates with the observation that a partitive construction "takes as its argument a bounded entity and addresses its internal articulation, picking out an identifiable bounded part" (Jackendoff 1991: 27, emphasis E.L.). At the level of interpretation, the two constituents of a partitive construction correlate with two referents, i.e. they are not coindexed. The notions of "bounded entity" or "Whole" suggest that the lexical head Ν of the complement contains the feature [-divisible]: (52)

Part-designating nouns: end, head, leg, etc. N PA rt,

PP[DP[+N[-divisible]]]

In contrast, quantity-designating nouns such as pound, lump, or grain "presuppose a noun that can be quantized" (Corver 1998: 219), i.e. the head noun of their selected complements contains the feature [+divisible]. These are mass nouns, plurals, and collectives such as cattle. Whereas entitydenoting nouns such as book or apple contain both form and substance as part of their lexical meaning, quantity-designating nouns lack substance as a meaning component. They merely express different dimensions of quantity {dozen, number), weight (pound), form (lump, grain), or configuration (bunch, group)·, substance is expressed by the head noun of their selected complements. Therefore, at the level of interpretation, pseudopartitive constructions only have one referent, although they consist of two constituents. Due to their "defective" meaning, quantity-designating nouns select a complement with which they are coindexed, hence obligatory non-occurrence of an article. In this sense, these complements are [-R], i.e. they are not referentially independent:27 (53)

Quantity-designating nouns: pound, lump, grain N[quantity] i.

PP[NPJ

[N[+divisible]]]

Despite their "defectiveness" in meaning, these nouns which head the first constituent of pseudopartitive constructions, more often than not, trigger the selectional restrictions on the verb and agreement in number (54), i.e. they are the head of the construction. Due to coindexation, however, there are also numerous cases which exhibit "semantic agreement" (55):

248 Elisabeth Löbel

(54)

a. b. c.

Two pounds of sugar *was/were strewn/thrown on the floor. Two lumps of sugar were *strewn/thrown on the floor. Two bottles of wine were smashed/spilled on the floor.

(55)

a. b.

A group of students was/were invited to give a talk. A large number of books was/were published last year.

Container nouns such as bottle may be used as a non-relational noun (a bottle with a long neck) or as a relational measure noun (two bottles of wine)·, other nouns, too, allow for both uses (e.g. number vs. number of). Another distinction is exemplified by the noun head, which may function as a relational part-denoting (56a) or quantity-denoting noun (56b): (56)

a. b.

head (of a tiger): head (of cattle)·.

NPArt,

PP[DP[+N[-divisible]]] +sg, NPj [N[+divisible]]

N[QUantity] i,

Quantity-designating nouns do not assign Case to their selected [-R]-complements. This can be illustrated in German, where the complement is Case-marked either by agreement (syntactic condition, cf. (57a)), or by 29 semantic conditions (cf. (57b) with the so-called "partitive' Genitive), or by default (Nominative, cf. (57c)): (57)

a.

b.

c.

mit einer Flasche rotem with a-dative bottle-dative red-dative

Wein GERM wine-dative (case agreement) mit einer Flasche roten Weines with a-dative bottle-dative red-genitive wine-genitive (semantic case) mit einer Flasche (Iroter) Wein with a-dative bottle-dative red-nom. wine-nom. (default case) 'with a bottle of red wine'

In the English equivalent with a bottle of red wine, however, the preposition of constitutes an apparent problem with regard to this analysis. Similar to German, however, where the Genitive may function either as a structural or a semantic case, the preposition of can be used either way. 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 (cf. 58a):

Classifiers and semi-lexicality: Functional and semantic selection

(58)

a. b. c.

The book was about Mary, the letter to John by himself The letter was to John.

249

(Williams 1994: 222)

In (58b), the preposition to functions as a grammatical preposition, in (58c) 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). Some prepositions are not ambiguous, such as the by Agent marker in (59a); in contrast, the "authorship" by is semantic and can appear in postcopular position (59b): (59)

a. b.

*The destruction of the city was by the enemy. The book was by Harry. (Williams 1994: 222)

In contrast, according to Williams, some prepositions are not ambiguous, such as the of Theme marker in (60a). The examples in (60b) and (60c), however, illustrate that of, indeed, may occur in postcopular position: (60)

a. b. c.

*The destruction was of the city. The picture is of John. This aspect is of great importance.

In sum, comparable to the German genitive Case, of can be used as a grammatical or a semantic preposition. With regard to the latter, it is capable of selecting either a complement bearing a participant role to which structural Case is assigned (60b), or a complement bearing a non-participant role to which no Case is assigned (60c). 0

3.3.

Summary

Based on observations made by Rizzi (1990), two kinds of semantic selection are distinguished: selection of (i) a referential argument bearing a participant role, and (ii) a "non-referential" argument bearing a non-participant role, where "non-referential" is to be understood as "not referentially independent". It is argued that this distinction holds not only for verbs (copular verbs, measure verbs, and verbs of naming selecting the non-participant roles of Property, Measure, and Name), but also for nouns (quantity-designating nouns selecting the non-participant role of "Substance"31). In contrast to participant roles, non-participant roles are not assigned Case; Case-

250 Elisabeth Löbel

marking is triggered either by (i) agreement (syntactic conditions), (ii) by semantic conditions, i.e. it is free, or (iii) by default. In fact, all the peculiarities which these complements exhibit in syntax may be derived by and attributed to this distinction which is located in the lexicon. In an isolating language like Vietnamese, notions such as Case-marking or agreement are not applicable. Not surprisingly, there are other criteria which distinguish quantity-designating nouns and classifiers from their selected entity- or mass-denoting nouns. This is the topic of the next chapter. 4. 4.1.

Classifiers in Vietnamese Introduction

As illustrated in the Introduction, Vietnamese nouns which correspond to count nouns in English are said to behave syntactically like mass nouns, i.e. it is claimed that there is no real difference between these two subclasses (cf. examples (1) and (2), repeated here for convenience under (61)):

(61)

a.

b.

{mot)

N[+unit]

Nftransnumeral]

cuc

väng

(one) lump gold 'a lump of gold' (lit.: 'a lump gold') (mdt) quä cam (one) fruit orange 'an orange' (lit.: 'a fruit orange') (quä 'fruit; classifier for fruits')

Example (61) contains two important aspects which will be relevant for the following discussion: (i) Unlike Japanese and Chinese, Vietnamese classifiers are not bound in their occurrence, the numeral möt 'one' is optional; (ii) nouns such as cam 'orange' are said to be transnumeral, they do not contain any notion of grammatical number; these nouns may be translated as either singular or plural, depending on the context, whereas väng 'gold' is analogous to English gold (cf. the English gloss of (61a)). In contrast, cuc 'lump' as well as quä 'fruit' denote exactly one unit ('unit counters'),

Classifiers and semi-lexicality: Functional and semantic selection 251

regardless of whether the numeral mot 'one' is present or not.32 Due to the feature [+unit], these nouns cannot be purely semantic items in Emonds' terms; what is more, according to Kubo (1996: 98), qui 'fruit' in (61b) is said to constitute a purely syntactic item (cf. section 2.2), i.e. a "placeholder" for countability. The next questions to ask, therefore, are (i) whether or not the feature [+unit] is part of the lexical entry of these nouns, and (ii) whether or not they constitute a closed class of nouns. In order to answer these questions, however, it is necessary to take a closer look at Vietnamese nouns and nominal subclasses. In languages such as English and German, apart from conceptual shifts, there is a certain correspondence between mass nouns such as gold and the referents of these nouns, i.e. substances, on the one hand, and count nouns such as orange and the referent of these nouns, i.e. entities, on the other. This is reflected by nominal features such as [+/- count], [+/- discrete], or, as it is suggested by Jackendoff (1991: 20), [+/- bounded]: (62)

a. b. c. d.

+bounded, +bounded, -bounded, -bounded,

-internal structure +internal structure -internal structure +internal structure

individuals groups substances aggregates

(a pig) (a committee) (water) (buses, cattle)

Individual objects are characterized as having an inherent shape, they are "the only subcategory that has physical boundaries" (Jackendoff 1991: 20). In contrast, substances, which are normally expressed by mass nouns, contain the feature [-bounded]. What they have in common, however, is the feature specification [-internal structure]. "Note: the value -i does not mean lack of internal structure, but rather lack of necessary entailment about internal structure" (Jackendoff 1991: 21, emphasis E.L.).33 In contrast, it is exactly the notion of being structured which plays a decisive role for Greenberg (1974) in his discussion of nominal subclasses in classifier languages in contrast to English: "If I cut a piece of meat in two, I have two pieces of meat, but if I cut a dog in two, I still have only one dog, a dead one. The property that distinguishes dogs and automobiles in these cases is evidendy internal organization into an integrated and organic whole, whether natural in the case of the dog or artificial in the case of the automobile. We might call this feature ±structured' (Greenberg 1974: 23, emphasis E.L.).

In the following section, it is argued that it is precisely this feature [± structured] which is decisive for the distinction between nominal subclasses

252 Elisabeth Löbel

in Vietnamese, i.e. it is argued that it is entailment vs. lack of entailment about internal structure which constitutes the main difference between entity-denoting nouns in Vietnamese and in English. 4.2.

Subclasses of nouns

Comparable to English, nouns such as väng 'gold' can be combined with numbers only by means of quantity-denoting nouns such as heap, pile, lump, grain (of gold), etc.; in both languages, combinations such as *hai väng *'two gold' are ungrammatical.34 Likewise, for a noun such as ch6 'dog', there are also numerous possibilities to combine this transnumeral noun with numbers ((63) and (64)). Notice that, with regard to noncountability and transnumerality of Vietnamese nouns, it is not relevant how dogs or cars are counted as such (extralinguistic aspect), but how the respective nouns which denote these entities behave (linguistic aspect, cf. Löbel 1999b: 274):

(63)

N[+unit] N[transnumeral] a. (mot) con ch6 (one) living-being dog 'a dog' (con 'classifier for N[+animate]') (lit. 'an animal dog' b. (mot) chän ch6 (one) leg dog 'a dog's leg / a leg of a dog' (lit. 'a leg dog') c. (mot) dau ch6 (one) head dog 'a dog's head / a head of a dog' (lit. 'a head dog') d. (mot) duöi ch6 (one) tail dog 'a dog's tail / a tail of a dog' (lit. 'a tail dog') etc. (Löbel 1999b: 274)

This example illustrates that the noun ch6 'dog' can be combined with numbers not only by reference to the whole (63a), but also by reference to

Classifiers and semi-lexicality: Functional and semantic selection 253

the parts which are part of the lexical meaning of ch The existence of real equative sentences is denied explicitly in Williams (1997). According to him, equative sentences reveal that there is no absolute notion of referentiality: "...the best we can say is that the predicate is less 'referential' than its subject..." (Williams 1997: 31). 15. For this reason, it seems to me problematic to regard the copula as a lateinserted item (Emonds 1999: 316). 16. The anonymous reviewer points out that the NP/DP a man can function as a predicate without the presence of the copular verb, e.g., in (i) / consider him a man, or (ii) John came out a wiser man. In section 3.1.1 below, I argue that the copula is a two-valued verb which selects a non-referential argument expressing a property. This analysis naturally extends to three-valued verbs such as consider which takes two referential arguments and a non-referential one (cf. Contreras 1995: 146-148), in contrast to give, which takes three

266 Elisabeth Löbel

(referential) arguments. The NP/DP a man in (ii) is an adjunct; following Moro (1997: 44), I argue that this adjunct is optionally introduced by as, which selects a non-referential argument expressing a property, where "nonreferential" is to be understood as "not being referentially independent" (cf. Löbel (1999a) for a more detailed discussion). 17. 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.). 18. This thematic grid is also assumed for other copular verbs such as become or remain. Similar to Heggie (1988: 122), the copula is listed in the lexicon as a "verbal element which must coindex a constituent", without following her assumption, however, that this constituent "becomes the predicate position of a small clause" (Heggie 1988: 122). The anonymous reviewer points out that coindexation of Theme and Property constitutes a problem to my analysis as it is not clear how this is done. I assume that coindexation constitutes a characteristic feature of non-participant roles, i.e. coindexation is part of the lexical entry of these verbs which select an argument which is not referentially independent. 19. Realization by different categories also holds for other participant roles. Theme, e.g., which is considered to be the default theta-role, may be realized by a CP (i), a PP (ii), or even a metalinguistic expression (iii): (i) That they didn 't come annoyed us. (ii) They talked about the book. (iii) You should abbreviate that is by "i.e.". 20. Cf. Larson (1988) where it is claimed that lexical properties of verbs should be expressed in terms of thematic grids instead of subcategorization frames. 21. 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) Johni became richv (ii) John·, [became richi ]i. "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 4 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). 22. Cf. Reinhart/Reuland (1993: 659-660) for discussion of the notion of Rexpression with regard to reflexives. There is, however, another interpretation of the term "non-referential", namely "having no referent at all" (impersonal it,

Classifiers and semi-lexicality: Functional and semantic selection 267

expletive there, idiom chunks, cf. Chomsky 1981: 101). A discussion of these crucial notions, however, would be far beyond the scope of this paper. 23. The adverbial accusative is also used for adjuncts: (i) Er arbeitet den ganzen Tag. he works the-Acc. whole-Acc. day-Acc 'He is working the whole day.' 24. According to Grimshaw (1990: 97), possessives such as John's in John's dog are modifiers which can be distinguished from complements "in a very clear way: only modifiers can be related to the head across a copula": (i) John's dog (ii) The dog is John's. 25. Cf. Löbel (1991) for a discussion of these constructions with regard to case assignment under adjacency. They have been a problem for X-bar syntax ever since, as it has been illustrated already in JackendofP s (1984) article "On the phrase 'the phrase the phrase"'. 26. Furthermore, examples such as these illustrate unambiguously that it is not the (semantic) head Erde 'earth' which is decisive for DP-internal agreement, but the noun Planet 'planet' with which it is coindexed; the latter one, therefore, constitutes the syntactic head. 27. In this respect, they are similar to "epithets" such as idiot, fool. The term "epithet" already suggests that they are property-denoting nouns. Therefore, sentences such as An idiot came in (in contrast to An idiot of a man came in) are somewhat deviant. For discussion of these nouns in terms of "pronominal R-expressions", cf. Lasnik (1991). 28. For an extensive analysis of these nouns and their selectional properties in German, cf. Löbel (1986). 29. In these constructions, the genitive is used as a semantic case (partitive genitive). 30. Cf. Jespersen ([1949] 1958: 331-342) for the various uses of of. 31. In Jackendoff (1991: 20-23), the term Material is proposed as a supercategory that contains substances (a drop of water) as well as aggregates (a stick of spaghetti). I leave this open for discussion. "Substance" is to be understood as a cover term for those (referents of) divisible NPs (mass, plurals, collectives) which indicate what (the referent of) the quantity-designating noun consists of. 32. The presence vs. absence of the numeral roughly corresponds to English an (orange) vs. one (orange). 33. The feature combinations in (62) show that [+ internal structure] holds for nouns that denote objects which consist of more than one entity (committee, buses, cattle), i.e. characteristics of one single entity, according to Jackendoff, are not relevant in English. 34. This section constitutes a summary of Löbel (1999b), to which the interested reader is referred to for more details. 35. This is consistent with the observation in Triicfng (1970: 259) that the classifier is not used in front of nouns which designate a part of an entity.

268 Elisabeth Löbel

36. Cf. Löbel (1999b: 306-312) for compounds with classifiers. 37. Cf. Löbel (1999b: 288-293) for a more detailed description. 38. In German, the selected argument is a NP which is case-marked either by agreement, by default or by a semantic case (partitive genitive).

References Abney, Stephen A. 1987 The English noun phrase in its sentential aspect. Ph.D. Dissertation, M.I.T. Barker, Chris and David Dowty 1993 Non-verbal Thematic Proto-Roles, NELS (Proceedings of the North East Linguistic Society) 23(1): 49-62. Cardinaletti, Ahna and Maria Teresa Guasti (eds.) 1995 Small Clauses. Syntax and Semantics 28. San Diego et al.: Academic Press. Carstens, Vicki 1991 The Morphology and Syntax of Determiner Phrases in Kiswahili. Ph.D. Dissertation, UCLA. Chomsky, Noam 1981 Lectures on Government and Binding. Dordrecht: Foris. Contreras, Heles 1995 Small Clauses and Complex Predicates. In: A. Cardinaletti and M.T. Guasti (eds.), Small Clauses. Syntax and Semantics 28, 135152. San Diego et al.: Academic Press. Corver, Norbert 1998 Predicate Movement in Pseudopartitive Constructions. In: A. Alexiadou and Ch. Wilder (eds.), Possessors, predicates and Movement in the Determiner Phrase, 215-257 Amsterdam/ Philadelphia: John Benjamins. Craig, Colette (ed.) 1986 Noun Classes and Categorization. Typological Studies in Language 7. Amsterdam/Philadelphia: Benjamins. Dechaine, Rose-Marie 1993 Predicates Across Categories. Ph.D. Dissertation, Amherst Den Dikken, Marcel 1998 Predicate Inversion in DP. In: A. Alexiadou and Ch. Wilder (eds.), Possessors, predicates and Movement in the Determiner Phrase, 177-214 Amsterdam/Philadelphia: John Benjamins.

Classifiers and semi-lexicality: Functional and semantic selection 269

Downing, Pamela 1986 The Anaphoric Use of Classifiers in Japanese. In: Craig (ed.): 345375. Emonds, Joseph 1999 How Clitics License Null Phrases: A Theory of the Lexical Interface. In: H. van Riemsdijk (ed.), Clitics in the Languages of Europe: Language Typology 10, 291-367. Berlin: Mouton de Gruyter. Frampton, John 1991 Relativized Minimality, A Review. The Linguistic Review 8: 1-46. Greenberg, Joseph H. 1974 Numeral Classifiers and Substantival Number. Problems in the Genesis of a Linguistic Type. In: L. Heilmann (ed.), Proceedings of the 11th International Congress of Linguists, 17-37 Bologna. Grimshaw, Jane 1990 Argument structure. Cambridge, Mass.: The MIT Press. Heggie, Lorie A. 1988 The Syntax of Copular Structures. Dissertation, University of Southern California. Iljic, Robert 1994 Quantification in Mandarin Chinese: Two Markers of Plurality. Linguistics 32: 91-116. Jackendoff, Ray 1984 On the Phrase The Phrase 'The Phrase'. Natural Language and Linguistic Theory 2: 25-37. 1991 Parts and boundaries. In: Beth Levin and Steven Pinker (eds.), Lexical and Conceptual Structure. Cambridge/Oxford: Blackwell. Jespersen, Otto 1958 A Modern English Grammar. Part VII Syntax. London/ Copenhagen: Allen and Unwin/Muksgaard. [1949] Kitahara, Hisatsugu 1993 Numeral classifier phrases inside DP and the specificity effect. In: S. Choi (ed.), Japanese/Korean Linguistics Vol. 3: 171-186. Stanford University: Stanford Linguistics Association. Kubo, Miori 1996 Some Considerations on Noun Classes and Numeral Classifiers: A Study of (Pseudo)partitives in Japanese and English. Keio Studies in Theoretical Linguistics I: 89-123. Larson, Richard K. 1988 On the double object construction. Linguistic Inquiry 19: 335-391.

270 Elisabeth Löbel

Lasnik, Howard 1991 On the necessity of Binding Conditions. In: R. Freidin (ed.), Principles and Parameters in Comparative Syntax. Cambridge: MIT. Levin, Beth 1993 English Verb Classes and Alternations. Chicago: Chicago University Press. Li, Audrey Yen-Hui 1999 Plurality in a Classifier Language. Journal of East Asian Linguistics 8 (1): 75-99. Löbel, Elisabeth 1986 Apposition und Komposition in der Quantifizierung. Syntaktische, semantische und morphologische Aspekte quantifizierender nomina im Deutschen. Linguistische Arbeiten 166. Tübingen: Niemeyer. 1989 Q as a functional category. In: Ch. Bhatt, E. Löbel and CI. Schmidt (eds.), Syntactic Phrase Structure Phenomena in Noun Phrases and Sentencesi, 133-158. Amsterdam/Philadelphia: Benjamins. 1990 Typologische Aspekte funktionaler Kategorien in der Nominalphrase. Zeitschrift für Sprachwissenschaft, 9: 135-169. 1991 Apposition und das Problem der Kasuszuweisung und Adjazenzbedingung in der Nominalphrase des Deutschen. In: Sascha W. Felix and G. Fanselow (eds.), Strukturen und Merkmale syntaktischer Kategorien, 1-32. Tübingen: Narr. 1997 Numerus: Funktionale Kategorie vs. syntaktische Funktion. In: E. Löbel and G. Rauh (Hgg.), Lexikalische Kategorien und Merkmale, 87-123. Tübingen: Niemeyer. 1999a Copular Verbs and Argument Structure: Participant vs. Non-Participant Roles. Theolex [= Arbeiten des SFB 282] No. 115: 1-27 Düsseldorf: University of Düsseldorf. 1999b Classifiers vs. Genders and Noun Classes. A Case Study in Vietnamese. In: Barbara Unterbeck and Matti Rissanen (eds.), Gender in Grammar and Cognition. I: Approaches to Gender. Trends in Linguistics, Studies and Monographs 124: 259-319. Berlin: Mouton de Gruyter. Moro, Andrea 1997 The Raising of Predicates. Predicative noun phrases and the theory of clause structure. Cambridge: Cambridge University Press. Ouhalla, Jamal 1991 Functional Categories and Parametric Variation. New York: Routledge. Reinhart, Tanya and Eric Reuland 1993 Reflexivity. Linguistic Inquiry 24(4): 657-720.

Classifiers and semi-lexicality: Functional and semantic selection 271

Riemsdijk, Henk van 1998 Categorial feature magnetism: The endocentricity and distribution of projections. Journal of Comparative Germanic Linguistics 2: 148. Ritter, Elizabeth 1991 Two Functional Categories in Noun Phrases: Evidence from Modem Hebrew. In: Susan Rothstein (ed.), Perspectives on Phrase Structure. Syntax and Semantics Vol. 25: 37-62. New York: Academic Press. 1995 On the Syntactic Category of Pronouns and Agreement. Natural Language and Linguistic Theory 13: 405-443. Rizzi, Luigi 1990 Relativized Minimality. Cambridge, Mass.: MIT Press. Rothstein, Susan 1995 Small Clauses and Copular constructions. In: A. Cardinaletti and Μ. T. Guasti (eds.), 27-48. Selkirk, Elizabeth 1977 Some remarks on noun phrase structure. In: P.W. Culicover, Th. Wasow and A. Akmajian (eds.), Formal Syntax, 285-316. New York et al.: Academic Press. Stowell, Timothy 1978 What was there before there was there. Chicago Linguistic Society 14: 458-471. Thompson, Lawrence C. 1987 A Vietnamese Grammar. Mon-Khmer Studies XIII-XIV, University of Hawaii Press. [1965] Williams, Edwin 1980 Predication. Linguistic Inquiry 11(1): 303-38. 1994 Thematic Structure in Syntax. Linguistic Inquiry Monograph 23. ΜΓΓ Press. 1997 The Asymmetry of predication. In: R. Blight and M. Moosally (eds.), Proceedings of the 1997 Texas Linguistics Society Conference. Texas Linguistic Forum 38: 323-333. Zubin, David A. and Mitsuaki Shimojo 1993 How 'general' are general classifiers? with special reference to ko and tsu in Japanese. In: Guenter, J.B.S., A. Kaiser, Ch. C. Zoll (eds.), Proceedings of the 18th Annual Meeting, 490-502. Berkeley Linguistic Society, Berkeley: Linguistic Society.

Agreement patterns of Czech group nouns and quantifiers Ludmila Veselovskä

1.

Introduction

The complex nominal structures which I am going to illustrate in this paper1 contain a nominal structure headed by a noun (=N) preceded by some quantifying element (=Q). Observing the properties of the structures I will concentrate on the following phenomena: 1. Possibility of relativization and of external reference, 2. Complexity of the Q and the following N™1*, 3. Ellipsis of the Nmax, 4. Feature Content of the Q, 5. Case and Agreement Patterns of the [Q+N™*]. The above criteria make it possible to distinguish the Q element from a lexical Noun followed by a DP complement. Moreover, they allow me to show three distinct groups of quantifying elements (Qs) in Czech. The main characteristics and labels used throughout the text are given below and in more details are illustrated and discussed in the following sections. A. Lexical Nouns: NQ Open class of nouns possibly interpreted as 'a group of or similarly, e.g. skupina 'group' rada 'row' and others illustrated in (3) and (11)7 The subscript in NQ suggests the similarity between these Ns and Qs. B. Group Nouns: QN Closed class nouns e.g. trocha 'a bit o f , spousta 'plenty o f , vetsina 'most o f , hromada 'a pile o f , rada 'a number o f , pär 'a few',/wra 'a heap o f , etc.5 Q n denotes 'a group of and the following N™" is singular (with mass nouns) or plural (with count nouns). For examples see (4) and (12). The subscript in QN suggests the similarity (above all morphological) between these Qs and Nouns. Both QN and the following group C:Q G EN semantically function as vague quantifiers.

274 Ludmila VeselovsM

C. Existential Quantifiers: Q G E N Quantifiers like mnoho 'many/much', mälo 'little/few', dost 'enough', hodne 'plenty', kolik 'how many' and the forms derived from these quantifiers.6 Q G E N also comprise most of the Cardinals higher that 4. With

η

respect to their semantic and selectional properties, Q G E N ' S are intersective and the Nmaxs, which follow Q E N quantifier, are either singular mass or plural count nouns. With Cardinals N™" must be a plural count noun. The noncategorial subscript in Q G E N suggests the idiosyncratic Case pattern of the following N™"" discussed in detail in section 6. For examples see (13). D. Universal Quantifiers: QA Quantifiers like vSichni 'all', oba 'both' and the Cardinals 2/3/4. QA are non-intersective quantifiers.8 The Noun which follows a QA quantifier is a plural count noun (with 'all' and cardinals), a singular mass noun (with 'all'), or set of two of a count noun (with 'both'). The subscript in QA suggests the similarity (above all morphological) between these Qs and Adjectives. For examples see (14). Assuming that DP dominates NP and Ν can take a DP complement in Czech marked by Genitive; the following (1) is a plausible structure for a Q complex 'double' DP in which the head Ν has a DP complement. G

(1)

a.

'double DP'structure

ID™"1

ID0

lN m x IN0

2Dmax 2D

2Nn

To make a clear-cut distinction between (1) and the structure of [QN+D/N™3*] i.e. a functional element QN forming a complex with a D/NP is not easy. The distinctions are often stated referring to interpretation which I do not discuss here in detail.10 Still, there are some distinctions evidenced in syntax, which I will provide in the following sections. In my analysis I will assume an endocentric phrase structure (see Chomsky 1986), but with multiple SPEC. 11 1 will follow Abney (1987) assuming that lexical (open class, major) categories are dominated by some functional category projections.

Agreement patterns of Czech group nouns and quantifiers

2.

275

Relativization and pragmatic reference

A syntactic distinction between NQ and all Q elements, which is discussed in detail in Selkirk (1977) and Jackendoff (1977), concerns the ability of the element to be postmodified by a relative clause and to serve as pragmatic antecedent to a pronoun. In Table 1 in (2) I first schematically demonstrate that these properties are able to differentiate NQ from any kind of Q. The examples follow. (2)

1 2

2.1.

Table 1

External Reference (Selkirk 1977) Independent relative clause Pragmatic antecedent to pronoun

No

QN

+ +

.

QGEN .

QA .

Relativization

In (3) I show that the nominal complement of both NQ and QN can be modified by a relative clause. The structure is overtly signaled by morphology because the Phi features (Number/Gender) of the head of the relative clause are in Czech reflected by the form of the relative pronoun and verbal agreement when the relative pronoun appears as subject of the relative clause. For example, in (3) all the NQ/Q elements are feminine/singular but the following Nouns are masculine/plural: the overt agreement on the relative pronoun as well as the verbal agreement are formed with the latter, i.e. in (3) it is 'boys-PM' that is relativized. All structures in (3) are grammatical and (3) therefore cannot be used to make the distinction between NQ and QN(3)

f

skupina (a group) fada (a row/a number Na ulici byla ^ hromada (a pile) ^ -SF chlapcii spousta (a lot) trocha (a bit) on street was-SF bovs-PM kteri se pohybovali... who-PM R E F L moving-PM... O n the street there was a group/ ... of boys who were moving...'

In (4) I show that not all elements preceding the head noun can be relativized themselves. Only lexical Nouns (i.e. NQ) but not the close class

276 Ludmila Veselovskä

QN S can serve as the head of a separate relative clause: the overt agreement on the relative pronoun as well as verbal agreement are formed with '-SF'. (4) shows that quantifiers QN are lacking some relevant quality that would allow them to be related to a relative clause. (4)

Pred

domem

byla

NQ: skupina NQ: fada QN: *rada NQ: hromada QN.· *hromada Qff: *spousta QN: *trocha

(a group) (a row) (a number) (a pile) (a pile) (a lot) (a bit)

-SF

in front of house was-SF chlapcii, kterd se pohybovala... bovs-PM, which-SF REIFL moving-SF... 'In front of the house there was a group/ ... of boys, which was moving...' In (5) I demonstrate a structure containing an existential quantifier Q G E N · Morphology shows that in the complex [QGEN+D/N™ 3 *] it is only the Ν which is separately postmodified by a relative clause since a relative pronoun as well as the verbal agreement in the relative clause can agree only in Phi features of the head Noun. (5)

[QGEN]

Pfislo mnoho /osm chlapcii, kteri /*-re came-SN many /eight bovs-PM who-PM/*SN meli /*-lo prijit. should-PM/*SN come 'There came many/eight boys who should come.' Notice that one cannot have recourse to a lack of Phi features of Q G E N . because the verbal agreement with the subject [QGEN+D/N™"'] complex in the main clause shows default 3SN and a relative pronoun could presumably use those default features as well. 13 Example (6) demonstrates the pattern of a universal quantifier QA. Because the complex [QA+D/Nmax] has a uniform set of Phi features, only one relative clause can be formed, i.e. one with a relative pronoun (and verbal agreement) agreeing in Phi features with the head Noun following the QA

Agreement patterns of Czech group nouns and quantifiers 277

(6)

[QA] a.

b.

Pfisli vsichni chlapci, kteri meli prijit. came-MP [all]-MP boys-MP who-MP should-MP come 'There came all/three boys who should come.' Prisla obe dve devcata, kterd mela prijit. came-FP [both two]-FP girls-FP who-FP should-FP come 'There came both/two girls who should come.'

To conclude, when compared with NQ, all Q elements lack some quality that allows separate relativization. I propose that this quality can be stated in terms of the syntactic presence of the functional projection D. That is, Q elements and the following NPs form a single DP, not a double DP.

2.2.

Pragmatic antecedent to a pronoun

The distinction between NQ and all kinds of Q is signaled also by their ability to serve as pragmatic antecedents to pronouns i.e. in contexts which use pronouns referring back to the complex structures containing N/Q. The following example (7) shows that N-complement structures [N(q) +D/Nmax] allow a pronoun in a following clause/sentence to separately refer back to either N(q) or to its complement. (7)

[Nq] a.

b.

Prisla skupina chlapcii α jä je /ji pozdravil. came-SF group-SF bovs-PM and I 3PM/3SF greeted Ά group of boys arrived and I greeted them/it.' Chystal-i /-a se vejtt dodomu. prepared-3MP/-3FS REFL enter to house 'It/They was/were getting ready to enter the house.'

The following (8) gives a similar example with QN. In contrast to (7), in (8) the pronoun in a subsequent clause/sentence can refer only to the Noun which follows the QN- The pronoun cannot refer to QN itself in spite of the fact that Qn has separate Phi features and these features are reflected by the V-agreement in the main clause.

278 Ludmila VeselovsM

(8)

[QN]

a.

b.

Pfisla spousta chlapcii α ja je /*ji pozdravil. came-SF plenty-SF boys-PM and I 3PM/*3SF greeted 'Plenty/Most of boys arrived and I greeted them/*it.' Chystal-i /*-a se vejit do domu. prepared-3MP/*3FS REFL enter to house 'They were getting ready to enter the house.'

In (9) a comparable QGEN structure shows that the pronoun can find its reference in only the Phi features of Ν and cannot reflect the (default) 3SN of the QGEN overtly reflected in verbal agreement in the main clause. (9)

[QGEN]

Prislo mnoho /osm chlapcii α jä je /*ho pozdravil. came-SN many /eight bovs-PM and I 3PM/*3SN greeted There came many/eight boys and I greeted them/*it.' As for Q a , it has no specific semantic features that could be relativized or referred to separately by following pronouns. Therefore in an equivalent to (7)-(9) the possible pronouns would refer back to the features of the Noun which are at the same time present in the morphology of QA. The above examples support the claim that the structures containing Q are simple in the sense that they contain only one element with the referential properties allowing relativization and pragmatic binding of pronouns. The relevant element can be a functional projection of D, which is plausibly required for relativization as suggested in section 1 and which can be also related to the presence of e.g. demonstratives, as discussed in the following section. 3.

Complexity of Q and the following NP

In this section I demonstrate the distinct ability of NQ and of Qs to be premodified by a separate field of adjectival premodifiers, possessives and demonstratives. A brief summary of these characteristics is given in Table 2 in (10), and examples follow.

Agreement patterns of Czech group nouns and quantifiers 279 (10) Table 2 Complexity of Q (and of the following NP) 3 Q and Ν™" have separate sets of DEM/POSS/ Amax agreeing premodifiers 4 The following DP can be independently quantified 5 Q has adverbial prequantifiers (nearly, exactly, etc.)

N0 + +

QN QGEN QA -

-

-

+

± +

First, in (11) I illustrate that common nouns can have their own demonstratives or/and possessive14 and can be followed by an independently quantified nominal complex. In other words, a double DP structure containing two head nouns can contain two separate fields of adjectival premodifiers, demonstratives or/and possessives. (11)

[N q ] Objevila se ta tvoje velM skupina appeared-SF REFL fthe your big group]-SF.NOM tech nasich peti kräsnych chlapcii fthe our five nice boys]-PM.GEN 'There appeared the/your big group of those/our five nice boys.'

The contrasting example in (12) shows that the [QN+D/N™3*] complex cannot contain two DEM/POSS fields, which are present in (11), and that the presence of Q N excludes the presence of another Q. 15 The unmarked position of the only DEM/POSS in (12) is following the Q N . 16 (12)

[Q N ]

Objevila se *ta /*tvoje/ IvelM spousta appeared-SF REFL i*the / *vour / ?big plenty]-SF.NOM *mnoha / tech / nasich / *peti kräsnych chlapcii r*manv / the / our / *five nice boys]-PM.GEN 'Such a huge amount of those/our nice boys appeared' (12) suggests that the domain of QN does not contain a separate field for demonstratives and possessives, the elements which have been since Abney (1987) related to the presence of the DP projection. Moreover, the [QN+D/N™2*] complex contains only one position for a quantifying element, this being the QN itself. The same conclusion can be made about structures containing QGEN demonstrated in (13). QGEN can be premodified only by a grading adverb

280 Ludmila

Veselovskä

and (13) contains only one demonstrative/possessive and one field of adjectives. As for the position of QGEN, (13a) shows that QGEN quantifiers are fully grammatical only preceding DEM/POSS.17 (13)

[QGEN]

a.

( *tech / *poslednich / *celych) velmi mnoho (*the / *last / *complete) very many tech vasich kmsnych let the your nice years 'very many of the nice years of yours'

(13b) demonstrates that the ability of premodification is the same with QGEN cardinals (capitalized in (13b)). These, however, can appear either preceding DEM/POSS or following DEM/POSS.18 b. i. (*tech / *vasich) celych pet tech vasich kmsnych let (*the / *your) complete FIVE the your nice years 'the full five nice years of yours' ii. tech vasich celych pet (*tech/ *vasich) krdsnych let the your complete FIVE (*the / *your) nice years 'the full five nice years of yours' In (14a) I demonstrate comparable examples of [QA+D/N™3*] structures concentrating on the premodifiers. The characteristics are the same as in (13): the structures containing Q A can have only one D E M / P O S S . (14ai) shows Q a with a mass noun, (14aii) QA with a count noun. The elements immediately preceding QA are restricted to adverbs and no element with A™" agreement is allowed. The only D E M / P O S S and semantic Amaxs are fully felicitous only when they follow the QA quantifier (but see footnote 16).

Agreement patterns of Czech group nouns and quantifiers

(14)

281

[QA]

a.

i. (* ten/ *väs/ *dlouhyi liplne vsechen (* the/ *your/ *long) entirely all 'entirely all the nice time of yours' ii. ( * takovä /* vase / * ύρΐηά) ύρΐηέ (* such /* your / * complete) entirely ta vase kräsnä devcata the your nice girls 'entirely all the nice girls of yours'

ten väs kräsnycas the your nice time vsechna all

(14b) demonstrates the pattern of the cardinals 2/3/4. As in (13b), the field that separates the 'higher' position of the cardinals from their possible 'lower' position is formed by the elements realizing DEM/POSS. (See footnote 17). b. i. (*ty) skoro ctyri ty vase posledni kräsne (*the) nearly FOUR the your last nice 'the last nearly four beautiful years of yours' ii. ty vase posledni skoro ctyri (*ty) kräsne the your last nearly FOUR (*the) nice 'the last nearly four beautiful years of yours'

roky years roky years

The criteria mentioned in (2) and (10) and the examples in sections 1-2 argue in favor of distinct analyses for [N(Q)+D/NIDax] and [QN+D/N1™*] structures, namely against the analysis of Qn as IN in (1). I propose that the absence of D (signaled by the impossibility of relativization and pragmatic reference and by the lack of DEM/POSS19) is one plausible distinction between Ν and a non-lexical (functional) head Q. Quantifier (Q) is a functional projection of a Noun. As a functional head, it does not project into another D. The scheme in (15a) gives the proposed structure for a double DP pattern compared with the structure of QP in (15b).20

282 Ludmila Veselovskä

(15)

a.'double DP'structure of Ν

b. QP

Based on the data above I propose that the subcategorization frame for Czech quantifiers (including group nouns QN) is Q, [ D]. Given that Q cannot be followed by another Q and cannot be preceded (in unmarked contexts and with the exception of cardinals) by a separate demonstrative and possessive, the following (16) hold for Czech. (16)

4.

a. b. c.

D, [_N/*Q] N/Q, [ D] N/*Q, [ Q]

Ellipsis

Another criterion allowing us to state the distinction between Ν and Qs (and among Qs themselves) concerns ellipsis. Table 3 in (17) gives a summary and the examples follow. (17) 6 7

Table 3 Ellipsis The following DP can be omitted The elliptically omitted DP in NOM/ACC must be represented by a GEN clitic

N0 + -

QN +

QGEN

QA

+

+ -

Czech nouns can be postmodified by a genitive DP. In (18a) I demonstrate that the presence of such DP is usually optional.21

Agreement patterns of Czech group nouns and quantifiers

(18)

283

[Nq] a. i. (Ία velkä) sklenice (vody) stäla na stole. [the big glass]-SF.NOM (water.GEN) stood on table 'The big glass of water was standing on the table.' ii. Videljsem (tu velkou) sklenici (vody). saw AUX-1S [the big glass ]-SF.ACC (water-SF.GEN) Ί saw the big glasses (of water).'

Head nouns in Czech can be elliptically omitted if their presence is signaled by one or more agreeing premodifiers (the most salient examples contain also an agreeing demonstrative) as in (18b). b. i. Ta velkä stäla hned vedle. [the big]-SF.NOM stood just next 'The big one was standing just next.' ii. Videl jsem tu velkou. saw AUX-1S REFL [the big]-SF.ACC Ί saw the big one.' (19) shows that the postnominal complement can also be represented by a GEN clitic (cl) which is optional to the same extent as the complement itself is optional.22 (19)

a.

b.

Veera (jf) stäla sklenice na stole yesterday (SF.GEN(cl)) stood glass-SF on table 'Yesterday a glass (of it) was standing on the table.' Videl jsem (jt) velkou sklenici. saw AUX-1S (SF.GEN(cl)) [the big glass ]-SF.ACC Ί saw a big glass (of it).'

In the contrasting examples (20) I demonstrate that the GEN clitic is obligatory with Qn- The GEN complement of QN must be either present overtly or it must be represented by an extracted GEN clitic.

284 Ludmila

(20)

Veselovskä

[Q n ] a.

b.

Spousta *(yody/ji) stäla na stole. plenty *(water/CL-SF. GEN) stood on table 'Plenty of water was standing on the table.' i. Videl jsem spoustu *{yody). saw AUX-1S plenty-ACC *(clear water-SF.GEN) Ί saw plenty of water.' ii. Videl jsem *(jf) spoustu. saw AUX-1S *(SF.GEN(cl)) plenty-ACC Ί saw plenty of it.'

In (21) I illustrate an ellipsis of a nominal complex following QGEN- It shows that QGEN is similar to QN: neither of them tolerates a complete ellipsis of the following structure and the presence of at least GEN clitic is obligatory in NOM/ACC contexts.23 (2D

[QGEN] a.

b.

Mnoho /Pet *(chlapcit/jich) neprislo. Many /Five *(boys /3P-GEN(cl)) not-came-3SN.PAST 'Many/Five of them did not come.' Vcera * (jich) videl mnoho/pet yesterday * (3P-GEN(clitic)) saw-3SM.PAST many /five. 'Yesterday he saw many/five of them.'

As for Q a in (18b) I demonstrated that a Czech noun can be elliptically omitted in the presence of some agreeing premodifier and in section 5 I will show that the morphology of QA is equivalent to the morphology of agreeing adjectivals. As expected then, in contrast to QN and QGEN. (22) shows that complete ellipsis of the following noun is possible with QA. Since a NOM clitic is missing in the inventory of Czech, in (22a) the presence of any clitic is ungrammatical. As for ACC contexts, an ACC clitic is optional only.

Agreement patterns of Czech group nouns and quantifiers 285

(22)

[Q A ]

a.

Vsichni/Tri

(*jich)

pfisli

brzy

ratio.

All /Three (*3P-GEN(cl)) came-3PM early morning. 'All/Three came early in the morning.' b.

Videl (je)

vsechny/ tri

saw (3P-ACC(cl)) all 'He saw all/three of them'

/ three-3P.ACC

The distinction between Qs with respect to ellipsis is plausibly related to their distinct feature content. The feature content in Czech is overtly reflected in their morphology. I will discuss this topic in the following section to provide more evidence for distinct analyses of QN, QGEN and QA. 5.

Feature content and morphology

In this section I attempt to show that the distinctions in the three kinds of quantifiers can also be stated in terms of complexity or 'richness' of their feature content reflected in the 'richness' of their morphology. Table 4 in (23) gives a schematic picture demonstrating a gradual loss of independent 'nominal' morpho-semantic properties of the distinct kinds of quantifiers. More detailed description of (23) follows together with examples. (23) 8

9 10 11

5.1.

Table 4 Feature Content of Q Semantic Gender Formal Gender (non-interpreted) Interpreted Singular/Plural Formal Singular/Plural (non-interpreted)

QN

QGEN

+ +

+

-

+

%

-

NO

±

QA agrees agrees agrees agrees

Feature content and morphology of QN

The morphology of elements labeled here as QN is traditionally labeled as nominal, i.e. having the morphology which reflects the typical nominal features. In Czech the morphology differentiates Number (singular vs. plural), Gender (masculine/feminine/neuter), and Case. As for Number, Czech QN can appear in both singular and plural forms. (24a) shows singular Number of NQ and QN.

286 Ludmila Veselovskä

(24)

a.

[NQ/Q N ]

Objevila se skupina/spousta appeared-SF REFL group/plentv-SF.NOM chlapcü / divek /vody boys-PM.GEN / gMs-PF.GEN / water-SF.GEN 'here appeared a group/amount of boys/girls/water.' (24b) illustrates plural Number of NQ and QN. Notice that unlike on the NQ, the plural on QN does not change the meaning of the complex (and is incompatible with cardinals). The plural is, however, formally marked in the morphology of both on NQ and QN and also reflected by plural morphology in V-Agreement.24 (24)

b.

[QN]

Objevily se dve skupinv/(*dve) spoustv appeared-PF REFL two groups/(*two) plenties-PF.NOM chlapcü /divek /vody boys-PM.GEN / girls-PF.GEN / water-SF.GEN 'There appeared two groups/(*two) amounts of boys/girls/water.' As for Case, Czech Case paradigms are divided according to Gender and each paradigm contains seven forms with many idiosyncrasies and syncretisms. In (25) the Case paradigms for QN show that QN has at least three distinct forms in both singular and plural. Notice that for nominal paradigms the suffixes are at least monosyllabic and can contain consonantal elements. The suffixes are capitalized in (25) and are the regular suffixes for the most productive feminine nominal paradigm in Czech. (25)

Table 4.1.

QN: 'plenty'

Gender Number NOM

Qa plenty Fem.

SG PL

spoust-A spoust-Y

(productive nominal

paradigm)

ACC

GEN

DAT=LOC INSTR

spoust-U spoust-Y

spoust-Y spoust

spoust-έ spoust-ÄM

Spoust-OU Spoust-AMI

To conclude, compared with standard Nouns in Czech QN formally reflects all the nominal features: Number, Case and Gender. Recall, however, that the variation between singular and plural is not semantically relevant. As for Gender, QN seems to appear prevailingly with feminine in Czech, which since quantifiers are inanimate is a purely formal marking as well.

Agreement

5.2.

patterns

of Czech group nouns and quantifiers

287

Existential quantifiers and cardinals higher than 4

In ( 2 6 ) I exemplify forms of a Q G EN quantifier mnoho 'many' and the Cardinal pet 'five' when followed by a mass/countable noun of masculine/feminine/neuter Gender. Notice that the form of the quantifier in (26) is the same irrespective of the Number/Gender of the following Noun. (26)

[QGEN]

a.

mnoho vody /casu much water-SF / time-SM (=mass Noun) b. mnoho chlapcit /divek /deti many boys-PM /girls-PF /children-PN (=count Noun) c. *pet vody /casu five water-SF /time-SM (=mass Noun25) d. pet chlapcu /divek /deti five boys-PM/girls-PF /children-PN (=count Noun) The morphology of QGENS is traditionally called frozen, fossilized, or numeral. The descriptive terms refer to the 'richness' of the Case paradigm. In ( 2 7 ) below I illustrate that no QGEN has more than two forms: one for NOM/ACC and 'the other' for all Lexical Cases. Moreover, with some QGENS (e.g. mälo 'few') some Lexical Cases are not felicitous at all. (27)

Table4.2.

QGEN: 'much', 'several',

QGEN

Gender Num.

NOM

ACC

GEN

much several five

M/F/N M/F/N M/F/N

mnoh-0 nSkolik

mnoh-0 n£kolik pit

mnoh-A mnoh-A n£kolik-A nSkolik-A pSt-I pet-I

SG/PL PL PL

'five'

(adverbialparadigms)

DAT

LOC

INSTR

mnoh-A mnoh-A n£kolik-A nSkolik-A pit-I pöt-I

Whether the poor frozen morphology of QGEN in ( 1 2 ) reflects Case is questionable: in (25) we see that the most marked cases (in Slavic e.g. Instrumental) are usually signaled by overt morphology containing consonants. In (27), however, the two endings which form the paradigm are only vocalic ^o / -a and - 0 M . Given that the quantifiers mnoho 'many/much', mälo 'few', kolik 'how many' are traditionally analyzed as adverbs, it is, in fact, surprising that they have any inflectional morphology at all. With adverbs related to adjectives the ending replaces the adjectival agreement as illustrated in (28). (28a) gives adjectives, which have a vocalic agreement ending reflecting the gender/number of the following noun. With the related adverbs in

288 Ludmila Veselovskä

(28b) the same position is occupied by -o. (28c) shows that is also a standard Czech linking vowel used in A+A compounds. (28)

a. b. c.

vysok-Ymuz/ vysok-Azena / vysok-E mesto tall-SM man / tall-SF woman / tall-SNcity vysok-0 polozene mesto high located town vysok-O-zdvizny vozik high - raising cart

Thus seems to be a standard ending for the part of speech of adjectives/adverbs in positions where the elements cannot agree in gender and number, i.e. for adverbs modifying either verbs or adjectives. The complementarity between ^o and ^a endings demonstrated above for quantifiers also occurs with some adverbs when they are used after prepositions, i.e. they presumably acquire a more 'nominal' characteristics and are expected to be case-marked. This is illustrated in (29) where the relevant preposition selects GEN. (29)

'adverbial' paradigm letel vysok-0 flewhighADv 'he flew high'

> spadl ζ vysok-A fell from high-GEN? 'he fell from high'

In section 6 I will argue that the morphology of adjectival agreement morphology in Czech.26

5.3.

QGEN

is neither Case nor

Definite quantifiers and the low cardinals 2/3/4 27

The morphology of Czech Q A 's is pronominal. The following examples in (30) exemplify the form of a QA quantifier vsichni/oba 'all/both' and the Cardinal ctyri 'four' when followed by a mass/countable noun of masculine/feminine/neuter gender. Notice that the form of the QA in (30) changes, reflecting the distinct features of the following N.

Agreement patterns of Czech group nouns and quantifiers 289

(30)

[Q A ]

a.

vsechna /* ctvri /* oba

b.

vsechni / ctvri / oba

[all [all c.

voda

/ * four / * bothl -SM-NOM water-SF-NOM chlapci

/four/bothl-PM-NOM boys-PM-NOM

vsechna / ctvri / obe

[all

devcata

/four / bothl-PF-NOM girls-PF-NOM

Contrary to QGEN, the morphology of Q A has special forms reflecting Case, Number and Gender. Contrary to QN, however, the morphology of QA does not reflect any intrinsic features of the QA itself, but copies the Phi features of the Noun, i.e. it is to be subsumed under the notion of NP-agreement. 28 The 'richness' of the paradigm of QA is seen in (31) Table 4.3. (31)

Table 4.3. QA: 'all', 'three', 'four' Gender Num. NOM ACC GEN Masc.

all

femin. neutr.

three

four

M/F/N M/F/N

SG PL SG PL SG PL PL PL

väech-EN v5ichn-I vSechn-A vSechn-Y vSechn-0 vSichn-A tf-I etyf-I

väech-EN vSechn-Y vSechn-U vSechn-Y vSechn-0 vSechn-A tf-I etyf-I

(pronominalparadigm) INSTR DAT LOC

vSe-HO vSe-CH vS-ί vSe-CH väe-HO vüe-CH tf-I Ctyf

vSe-MU v5e-M vS-ί vSe-M vSe-MU vSe-M tf-EM ötyi-EM

vSe-M väe-CH vJ-ί väe-CH v5e-M v§e-CH tf-ECH etyf-ECH

vSi-M väe-MI νδ-ί väe-MI vä-ίΜ vSe-MI tfe-MI etyf-MI

Table 4.3 demonstrates that QA morphology distinguishes between NOM/ACC and Lexical Cases, and for Lexical Cases it has at least two distinct forms containing consonantal suffixes. The Case paradigms for the Cardinals 'three' and 'four' are identical for all Genders of the following (plural, count) Noun, but they still reflect Case distinctions, containing four distinct forms with consonantal affixes. In (32) Table 4.4 gives the Case paradigms for 'both' and the Cardinal 'two', which are followed by a countable noun in the plural. (32) QA

both

Table4.4. Gender masc. F/N masc.

QA: 'both', 'two' NOM ACC ob-A ob-A ob-6 ob-δ dv-A dv-A

GEN ob-OU ob-OU dv-OU

dv-fi

dv-OU

two F/N

dv-E

(pronominalparadigm) DAT LOC obS-MA ob-OU obö-MA ob-OU dvö-MA dv-OU dv-OUM dvö-MA dv-OU

INSTR obö-MA obö-MA dvS-MA dvö-MA

290 Ludmila Veselovskä

The specific form of 'both/two' is one of the few remnants of the Old Slavic dual in Czech. Notice that a Gender distinction is visible only in NOM/ACC. As for 'richness' the paradigms in (32) still contain 3 forms and consonantal endings. QA has a significantly richer morphology than QGEN· Its morphology shows distinct forms for some Genders and relatively rich paradigms but contrary to QN is of a secondary character, i.e. it only copies the Phi features of a Noun. 6.

Case, NP Agreement and V-Agreement

In the following sections I will demonstrate the distinctions among the quantifying elements with respect to three morpho-syntactic phenomena: Case, NP-Agreement and V-Agreement. Based on parallel examples I will propose analyses which explain the distinctions and will refer to the properties of the Q elements illustrated in the above sections. My analyses will be based on definitions introduced in (33). Without further argumentation, I am going to assume that: (33)

a.

b.

c.

NP Agreement in Czech is a morphological signal of a SPEChead agreement, i.e. the features of the head X are reflected on the head Y of the constituent Y™" in the SPEC position with 29 respect to the head X, V-Agreement is an agreement between the head of the nominal 30

complex in subject position and a verbal predicate, Case is assigned by a head element X to the head of the maximal Dmax/Nmax which is sister to X. Case features become a part of the feature complex of the head D/N which is spread via NP Agreement.31

Table 5 in (34) gives a summary of the characteristics of Czech Qs with respect to these phenomena. Because (33) several times crucially refers to a notion of 'syntactic head' I will take the case agreement patterns as criteria showing the 'head' properties of the elements. (34) 12 13 14

Table 5. Head characteristics Case and Agreement Patterns Q's features trigger V agreement D/N"1" following the Q in NOM/ACC is in GEN D/N™" following the Q in Lexical Case is in GEN

NQ

ON

QGEN

+ +

+ + +

+

+

± -

QA

Agreement patterns of Czech group nouns and quantifiers 291 6.1.

Homogeneous Double Agreement Pattern of QN

With respect to the properties observed in this section there is no distinction between NQ and QN. The example (35a) demonstrates the complex [NQ/QN+D/N110*] structure in positions where the complex is assigned NOM, and (35b) shows a structure assigned a Lexical Case (INSTR is due 32 to the preposition). (35)

[Nq/QN] a.

b.

Objevil -a se skupina/spousta appeared-SF REFL group/plenty-SRNOM yysokych chlapcii [tall bovsl-PM.GEN 'There appeared a group/plenty of tall boys.' ... se skupinou/spoustou γysokych chlapcii ... with [group/plenty]-SFJNS [tall bovsl-PM .GEN ... with a group/plenty of tall boys.'

In (36) I give a descriptive summary of the observed patterns referring to Qn-

(36)

6.1.1.

Homogeneous Double Agreement Pattern of QN: • The Case and Agreement pattern of QN is comparable in both NOM/ACC and Lexical Case contexts. • Qn is followed by N™"" in the Genitive. • The V-Agreement reflects the Phi features of the QN

Analysis of Qn

The agreement principles (33) and the subcategorization feature Q, [ D] 33

proposed in (16) are the bases for explaining the characteristics of QnThe Homogeneous Double Agreement Pattern of the QN described in (36) exhibits characteristics of Czech nominals postmodified by GEN prepositionless DP complements as schematically demonstrated in (1) and (15b). The main similarity is that Qn as well as Ν assign GEN to a sister DP, and their Phi features are reflected by V-agreement if they appear in the subject position. By (33) both Ν and Q n are the heads of the complex

292 Ludmila

Veselovskä

[N/Qn+D/N™*] and enter into the external syntactic relations observed in (36). The scheme in (37) uses structure (l/15b) to analyze the pattern of Q N . (37)

[Nq/Q n ]: Homogeneous (Continuous) Double Agreement Pattern: 'There arrived plenty of those talented students of yours' QN"

DN

0 7 SP(D) DEM/POSS

N" SP(N) Λ max

N' N°

Pfisla

arrived-SF Qn field: a) marked by relevant case b) Q n enters V-Agreement

D/NP field: a) marked by GEN b) Phi features and case of Ν

The subcategorizations D, [_*Q] and Q N) [ D] given in (16a/b) prevent structures with Q N from containing two separate functional levels which could host DEM/POSS and have separate references required for a relative clause and pragmatic reference of pronouns. Therefore Q N is interpreted at LF only as a quantity-related head with scope over the following N""". 34

6.2.

Homogeneous single agreement pattern of the QA

Because it shows the pattern most distinct from Q N illustrated above in (35)-(37), I am now going to demonstrate the pattern of QA- In (38a) I give the complex [QA+D/Nmax] structures in subject position (NOM), and in (38b) in a Lexical Case position (here INSTR).

Agreement patterns of Czech group nouns and quantifiers

(38)

293

[Qa] a.

Vsechen

ten cas

ubehl

jako

[all the time]-SM.NOM ran-SM 'All the time passed away like water.' b.

... s(e)

temi

vsemi/tremi

kräsnymi

voda.

like water

devcaty

... with [the all /three beautiful girls]-PF.INS ... with such/last three nice girls' Compare the pattern of QA described in (39) with (36) for QN. (39)

The Homogeneous Single Agreement Pattern of the QA: a. The Case and Agreement pattern of the QA is comparable in both NOM/ACC and Lexical Case contexts, b. Q a reflects the Phi features and Case of the following N1™", the V-Agreement reflects the Phi features of the Noun.

The properties of QA in (39) do not show any of the head properties demonstrated by QN in (36): QA does not assign GEN to the following nominal complex and the features reflected by V-Agreement are the Phi features of the Noun. On the other hand, QA reflects the features of Noun in the same way as all adjective premodifiers. Therefore following (33a) I propose a structure as in (40), which locates QA in the SPEC position with respect to the only lexical head in the structure, i.e. N.

294 Ludmila Veselovskiι

(40)

[QGEN]

Homogeneous Single Agreement Pattern of QA : 'There arrived all those talented students of yours'

is the only lexical head in (40), and therefore the resulting structure acts as a homogeneous single field throughout the complex whole. Similar position (i.e. SPEC position) for QA and adjectival A™"1 premodifiers in Czech is also supported by the ability of both of these elements to license ellipsis of the NP complex without a GEN clitic, which was discussed above in section 4, and by their identical morphology discussed in section 5.35

6.3.

Heterogeneous double agreement pattern for QGEN

The agreement facts are the most complex for the QGEN elements. In this section I will show that in NOM/ACC, QGEN acts like QN, while in other cases QGEN is more like QA. The following example (41a) demonstrates the complex [QGEN+D/N™3"] structures in both subject position (with N O M case) and a position requiring ACC (in Czech ACC appears with direct objects and after prepositions requiring ACC). Notice that the presence of GEN case following QGEN makes (41a) comparable with the pattern of Q N demonstrated in (35a), and distinct from QA illustrated in (38a). 36

Agreement patterns of Czech group nouns and quantifiers 295

(41)

a.

[QGEN]

i.

Ubehl-o mnoho kräsnych tydnii passed - 3SN.PAST many-NOM [beautiful weeks]-PM.GEN 'Many beautiful weeks passed.' ... za tech poslednich pet ... after [the last]-P.GEN five -ACC kräsnych tydnii [beautiful weeks]-PM.GEN ... 'after the last five beautiful weeks'

ii.

NOM/ACC: (= 35a; * 38a)

In (41b) I give examples of [ Q G E N + D / N ™ " 1 ] in Lexical Case positions. Those are cases other than NOM or ACC, which in Czech can be prepositionless Lexical (Oblique) Cases and/or cases after prepositions which select anything other than ACC.37 Notice that there is no GEN following QGEN in (41b) and in contrast to (41a). (41b) is distinct from the pattern of QN illustrated in (35b), but similar to the pattern of QA given in (38b). (41)

b.

[QGEN]

i.

Projel mnoha kräsnymi zememi traveled-3SM.PAST [many beautiful countries] - PF.INS 'He traveled through many beautiful countries.' ...pred temi necelymi peti tydny ... before [the incomplete five weeks] -PM.INS ... 'before the not quite five weeks'

ii.

non-NOM/ACC: (*35b; = 38b)

In (42) I give a brief description of the pattern of Q G E N which, due to its mixed characteristics, is labeled as 'heterogeneous'.38 (42)

Heterogeneous Discontinuous Agreement Pattern of the

QGEN

The Case and Agreement pattern of the QGEN depends on the case assigned to the complex [ Q G E N + N]. a.

In NOM/ACC, QGEN shows some characteristics of the Homogeneous Double Agreement Pattern of QN described in (34).

296 Ludmila Veselovskä However, i.

the V-Agreement with QGEN is 3rd singular neuter, i.e. a kind of default form, and ii. any agreeing elements preceding QGEN cardinals copy thefeatures of the Noun, i.e. their presence results in discontinuous agreement, 39

b.

In Lexical Case structures, QN shows some characteristics of the Homogeneous Single Agreement Patters of QA described in (35). However, i.

QGEN does not reflect the Phi features of the Noun, and ii. Q's morphology is highly syncretic, frozen.

Given the mixed properties of the pattern of QGEN. it is difficult to straightforwardly subsume its analysis under either the structure of QN (37), or the structure of QA (40). In the following section I will propose an analysis which explains the mixed properties of QGEN in terms of distinct levels of a derivation. 7.

Levels of derivation

Recall that the distinction between the agreement patterns with QGEN depends on the kind of case assigned to the whole structure. This characteristic has been used for analyses of QGEN in some works mentioned in the following section, which I will to some extent adopt.

7.1.

Syntactic Case hierarchy

According to the division used in Babby (1987) all the structures used above exemplify syntactic Case, i.e. a case which is supposedly determined by some structural relation of the nominal phrase to other categories in the sentence or phrase. Among syntactic Cases, Babby distinguishes the two groups which have been demonstrated separately in the sections above: Lexical Cases, assigned to the D™ by a specific lexical item40, and Configurational Cases which are assigned to the D™"1 on the basis of its overall structural environment.41

Agreement patterns of Czech group nouns and quantifiers 297 Babby (1987) states the distinction between the Lexical vs. Configurational Case in terms of precedence, which provides a basis for the resolution of possible case conflict. The Syntactic Case Hierarchy he states for Russian claims that (43)

Syntactic Case Hierarchy (Babby 1987: 116) Lexical Case takes precedence over a Configurational Case.

Babby motivates the Case Hierarchy by a general principle that requires the subcategorial frame of the lexical element to be satisfied. This concept is basically adopted also in Franks (1995) who takes the descriptive hierarchy for epiphenomenal and derived from the distinction between inherent (=Lexical, =+Oblique) case and Configurational Case. Franks states the distinction following Chomsky (1981, 1986) as in (44). (44)

Case Hierarchy (Franks 1995: 95-96) a. the +Oblique Cases are assigned at D-Structure, b. the -Oblique Cases are assigned at S-Structure.

The distinction between (43) and (44) is not relevant for the moment. As for the terminology, I will go on using the terms 'Lexical' for nonNOM/ACC vs. 'Configurational' for NOM/ACC. 42 Notice that the Case Hierarchy stating the precedence of Lexical Case over Configurational Case is not enough to explain the mixed characteristics of Czech Q G E N · Whatever Case Hierarchy would be valid for Czech, it would have to claim that the GEN assigned by Q N is distinct from the GEN assigned by Q G E N · 4 3 A S illustrated in (35b) Q N assigns GEN to the following D/N1™" in both Lexical and Configurational Case structures, while QGEN does so only in the latter case (see (41)). In my analysis I will follow the Case Hierarchy proposed by Babby (1987) and Franks (1995) but I will use another mechanism to explain the mixed properties of Q G E N ·

7.2.

Distinct levels of insertion into derivation

A Case Hierarchy of the form (43) or (44) assumes that distinct syntactic processes can be ordered, in other words, they can apply at distinct levels of derivation. Moreover, if derivations operate on trees built from discrete syntactic elements (lexical entries, morphemes), one can also consider the

298 Ludmila Veselovska presence and/or absence of those elements in trees at various relevant levels of derivation. According to Emonds (1997), lexical entries universally contain two kinds of features: (a) Semantic f-features (only in the open classes: Ν, V, A, P) stored in the Dictionary which for the author represents the interface of syntax with human memory. (b) Grammatical/conceptual F-features (closed class elements of any category) which are those grammaticalized in a given language and they are stored in what Emonds calls a Syntacticon. Accordingly, the lexical items can be inserted into a derivation on three distinct levels, as in (45). 44 (45)

Three levels of insertion (Emonds 1997) a. The elements containing semantic features /(open class elements) must be inserted in the derivation from the beginning of the computation (at D-Structure or via Merge). b. The closed class elements, which contain only grammatical features F interpretable at LF, can be inserted into a derivation late, just prior to Spell Out. c. The grammatical formatives containing only uninterpretable grammatical features F are inserted into a structure at PF. These include (i) contextual F-features (ii) lack of content F-features (iii) the F-features of some empty category being spelled out in non-canonical position (Alternatively Realized features).

(45) can be summarized as (46). (46)

The feature content of a lexical element decides the level of its insertion into derivations.

I assume that apart from Case assignment various other syntactic processes may apply cyclically, in a specific order. The distinct characteristics of structures containing quantifiers may then follow from the distinct feature specifications of the Q elements which force their distinct levels of insertion into a cyclic derivation. Let us consider the feature content of the quantifiers first.

Agreement patterns of Czech group nouns and quantifiers 299

7.2.1.

Feature specification of quantifiers

Jackendoff (1977) discusses similarities and distinctions among the English group nouns, semi-numerals, and cardinals, and he claims that the nature of their distinct behavior is based on the presence of the specific features [+Num] with semi-numerals and [+Num], [+Card] with cardinals. According to the author, the English structure of [-Num][-Card] group nouns requires (a) the article, and (b) the o/-formative, as in (47) (see Jackendoff 1977: 130): (47)

Group Nouns:

a bunch of weeks

The presence of [+Num], which may be subject to dialect variation, causes Numeral of-Deletion and results in the pattern of semi-numerals exemplified in (48). (48)

Seminumerals: a hundred (*of) weeks

Apart from the Numeral of-Deletion, the additional presence of the [+Card] feature requires Cardinal α-Deletion resulting in the pattern of the cardinals demonstrated in (49). (49)

Cardinals:

(*a) six (*of) weeks

Thus, for Jackendoff, the distinction between group nouns, semi-numerals, and cardinals is based on a more detailed/rich feature specification of quantifiers/cardinals. In section 4 I demonstrated that Czech quantifiers and cardinals are less rather than more specific with respect to their semantic/morphological feature content than group nouns QN. Therefore I argue that this distinction is to be explained rather by the absence of features with the quantifiers and cardinals. I want to relate this absence of semantic features to the Case Hierarchy (43/44) and to Emonds's levels of lexical insertion (45/46). In section 6.11 demonstrated that the Case and Agreement pattern of QN is Heterogeneous (i.e. it shows no distinction depending on the Case assigned to the complex structure). QN reflects any case assigned to the complex, and it always assigns GEN to the lower DP. This signals that in (37) the QN head is present at all levels at which Case is assigned. As mentioned in section 5.1 Czech QNs have their independent Phi features, including formal Gender and Number, and these are overtly reflected

300 Ludmila Veselovskä

in the morphology of QN premodifiers and in V-Agreement. Given that this is the feature content which makes QN distinct from Q A and QGEN, I propose that (50)

Phi features of QN force its deep insertion in Czech. 45

As a D-Structure head, QN will show a homogeneous Case and agreement pattern, and it will be able to create a separate domain of its own, which is schematically demonstrated in (37). As for the quantifiers and cardinals, the same feature analysis is not appropriate. The semantic content of quantifiers is very restricted and the 'openness' of the class of cardinals is clearly due to arithmetic combination not lexical. Therefore I claim that quantifiers/cardinals are not part of an open class semantic Dictionary. They are, instead, a part of the Syntax of a human language (i.e. a Syntacticon) and as such they can be, following (45), inserted into the derivation after D-Structure. Some economy criteria may in fact force their late insertion.

7.3.

Analysis

of

QGEN

Assuming that QGEN lacks semantic features f, (45b) allows it to be absent at the beginning of the derivation (D-Structure). On the other hand, since QGEN is plausibly interpreted at LF, it is to be inserted before Spell Out, i.e. not after S-Structure. The theoretical option is confirmed by the properties of the Case and Agreement Pattern of the QGEN. If the (configurational) GEN case assigned to the following DP is to be a signal of 'headedness' of the Q, then, unlike Q a analyzed in (40), the QGENhas properties of a head in Configurational ('D-Structure') but not in Lexical ('S-Structure') Case patterns. I propose the D-Structure of QGEN in (51) for (a) Lexical and (b) Configurational Case positions.(Cf. (37) for QN and (40) for QA.)

Agreement patterns of Czech group nouns and quantifiers 301

(51) D-Structure of

QGEN

9?

a. b.

INSTR.:

-0with NOM/ACC: - 0 -

takovymi

nadanymi

[those

talented

studentv

takov—

nadan—

those—

talented— students—

students] .INSTR student—

If QGEN is absent at D-Structure, a Lexical Case is assigned in (51a) to the highest head of the complex which is lexical at that level, i.e. to lexical head of the following DP, to N°. On the other hand, Configurational Case is not assigned yet and therefore in the Configurational Case positions the lower DP/NP remains unmarked for Case in (51b). (52) shows the S-Structure of QGEN in the (a) Lexical and (b) Configurational Case position.46 The QGEN inserted at S-Structure does not assign any Case to the following DP if this DP is already case marked, as in (51a)/(52a). On the other hand, in (51b)/(52b) Q G E N assigns GEN to the head of the following DP which has not been case marked yet. Being the highest syntactically active head of the complex, QGEN also receives a configurational NOM/ACC. 47

302 Ludmila Veselovskä

(52)

S-Structure of

QGEN

D™

QGEN

Nmax

SP(D) SP(N)

N' N°

a. b.

INSTR.: s mnoha takovymi nadanymi studentv with many [those talented students] .INSTR NOM/ACC: mnoho takovych nadanych studentü many.NOM/ACC [those talented students] .GEN

The absence of QGEN in the D-Structure can also explain the puzzling 'discontinuity' of the Heterogeneous Discontinuous Pattern of QGEN mentioned in (42aii), namely the fact that DEM/POSS and (rare) A"1"" modifiers preceding QGEN in (52b) pattern show GEN case.48 This suggests that the head relevant for NP-agreement in (33a) is to be the syntactic head of the complex in the D-Structure. Because QGEN is absent on that level, it is not able to create its own agreement domain and the Phi features of the D-Structure head, including its Case features, appear on DEM/POSS. On the other hand, the relevant head for the V-Agreement in (33b) is the S-Structure head, because the V-Agreement with the [QGEN+D/N"18"] structures does not reflect the features of Ν but it is default, as mentioned in (42ai). 8.

Elements with dual lexical specification

I have argued that the distinct properties of three kinds of Czech quantifiers can be explained in terms of distinct levels of their insertion into a cyclic derivation. I demonstrated that a number of co-occurring properties, which otherwise seem to be unrelated and accidental, can be derived from this distinction. In the final section I want to support my analysis illustrating some examples with 'ambiguous' behavior. I will demonstrate that a level of

Agreement patterns of Czech group nouns and quantifiers

303

insertion of a specific lexical item is not always fixed and can be subject to variation. I will show, however, that the correlation between the whole complex of semantic, morphological, and syntactic properties is valid and once the level of insertion is chosen, the structure shows the full range of expected properties. 49

8.1.

"pär"

'couple'

One of the meanings of the Czech word pär 'a couple/pair' is a couple of something (i.e. a set of two). The word is a masculine Noun, i.e. as such its contains a feature of Gender. According to (45) and (50) it is to be inserted into derivation at D-Structure either as a full Noun (NQ) or as a Group Noun (QN). The same word, on the other hand, can be interpreted as some 'unspecified low amount' of elements - i.e. it can be a kind of existential (indefinite) quantifier Q G E N · I argued in the preceding sections that Q G E N is absent at the beginning of the derivation and it is inserted later in syntax, which results in number of distinct properties when compared with N.

8.1.1.

Relativization and pragmatic reference

As discussed in section 6.1 both NQ and QN are lexical heads of complex projections and enter into V-Agreement (see (36c)). As for the distinctions between common Nouns (NQ) and Group Nouns (QN), they are discussed in section 2. The following examples (53) show that pär 'a couple/pair', when it enters V-agreement (i.e. when it is NQ or QN), can be a head of a relative clause and in (54) can serve as pragmatic antecedent of a pronoun. In other words, it behaves as NQ. (The following N™"1 is feminine plural to provide a recognizable contrast with masculine singular pär 'a pair'.)

304 Ludmila

(53)

Veselovskä

[Nq] Prisel pär devcat, arrived-3MS pair-MS.NOM girls-FP.GEN a. -ktere byly predstaveny... - who-FP.NOM were-FP introduced-FP b. -ktery byl predstaven... - which-MS.NOM was-SM introduced-MS 'There arrived a pair of girls j who were introduced...' ι I which was introduced...'J

(54)

[Nq] ... α ja jsemho /je predstavil. ... and I AUX 3SM/3PF-ACC introduced ... 'and I introduced it/them.'

Compare then (53)/(54) with the contrasting (55)/(56), which demonstrate the same patterns with pär 'a couple of with indefinite quantifier interpretation. The QGEN characteristics are signaled by a default V-Agreement. As QGEN. it cannot be a head of a relative clause and it cannot serve as pragmatic antecedent of a pronoun. In (55b) the relative pronoun cannot agree with neither the MS of the QGEN (if those are -MS) nor with the default V-agreement. (55)

[QGEN]

Prislo pär devcat, arrived-3NS couple-MS.NOM girls-FP.GEN a. - ktere byly predstaveny... - who-FP.NOM were-FP introduced-FP b. - *ktery/-e byl /-o predstaven /-o... - *which-M/NS.NOM was-M/NS introduced-M/NS 'There arrived a couple of girls I who were introduced...' 1 1 * which was introduced...' J

Agreement patterns of Czech group nouns and quantifiers

(56)

305

[QGEN]

... α ja jsem * ho /je predstavil. ...and I AUX * 3M(N)S/3PF-ACC introduced ... and I introduced *it/them.'

8.1.2.

Countability / Premodification

In sections 3 and 5 I also discuss the distinction between N ( Q ) and QGEN with respect to their ability to be counted and to be premodified by agreeing DEM/POSS/Adj. The following example (57) demonstrates that pär 'a pair' as NQ (signaled by a V-Agreement) can appear as plural and can be premodified by agreeing elements. (57)

[Nq] a.

b.

Prisel ten jeden kräsny pär devcat. came-3SM [the one nice pairsl-SM.NOM girls-PF.GEN 'There arrived the nice pair of (=a set of two) girls.' Prisly ty dva kräsne parv devcat. came-3SM [the two nice pairsl-PM.NOM girls-PF.GEN 'There arrived the two nice pair of (=two sets of two) girls.'

Contrary to (57), the following (58) shows that the same word pär 'a couple' as QGEN (signaled by a default V-Agreement) cannot be counted and cannot have agreeing adjectival premodifiers (with the exception of the DEM in discontinuous -GEN). (58)

[QGEN]

Prislo (tech /*jeden/*kräsny(ch)) pär devcat. came-3SN the-PF.GEN/*one /*nice-PF couple-? girls-PF.GEN 'There arrived a couple of (=some small amount of) girls.'

8.1.3.

Differing case and agreement patterns

The examples (55), (57), and (58) demonstrate the Case and Agreement Pattern of the word pär 'a couple/pair' in NOM/ACC positions. Notice that

306 Ludmila Veselovskä in NOM/ACC positions it is always followed by a GEN DP and the distinction is signaled mainly by distinct V-agreement. The examples (59)/(67) demonstrate the Lexical Case Pattern, where the distinction between NQ and QGEN is more significant. Notice that in (59) the word pär 'a pair' is inflected for instrumental (including the consonantal ending typical for nominal instrumental) and it is followed by GEN Ν™"χ. (57) and (59) thus show a typical nominal pattern described for Group Nouns as the Homogeneous Double Agreement pattern. I give examples in both singular and plural to show again that as NQ, the word pär 'a pair' is countable and can have adjectival premodifiers. (59)

[Nq] a.

s tim jednim krdsnym pärem mladych devcat with [the one nice pairl-SM.INS [young girls]-PF.GEN 'with the one nice pair of (=set of two) young girls.'

b.

s temi dvema krasnvmi pärv mladych devcat with [the two nice pairl-PM.INS [young girls]-PF.GEN 'with the two nice pairs of (=two sets of two) young girls.'

Contrary to (59), where the word pär 'a pair' as NQ is marked for Lexical Case and assigns GEN, in (60) the same word appears in its frozen form (not inflected at all). There is no G E N following QGEN and the instrumental case is reflected by the N™* and its premodifying elements, including the initial demonstrative. (58) and (60) show the typical Heterogeneous Double Agreement Pattern of QGEN illustrated earlier in (41a/b) and described in (42). (60)

[QGEN]

s temi pär (*-em) mladymi devcaty with the-PF.INS couple (*-INS) [young girls]-PFJNS 'with a couple (=a few) of young girls.' The contrast between the inflection of the word pär 'couple' in (59) and (60) (and the similar contrast between elements demonstrated below) argues against any analysis which would take the default Oblique Case agreement of QGENS for true agreement with the Phi features of the Noun. (59) demonstrates the ability of the lexical element pär 'a pair' to appear in INSTR, and it shows the morphology of its INSTR. (60) thus clearly

Agreement patterns of Czech group nouns and quantifiers 307 exemplifies that the element ( Q G E N ) does not reflect either Case or NP Agreement. The examples above show that a specific lexical entry, once recognized/used as N Q or as QGEN, will follow the pattern typical for its feature specification. According to the analysis presented in this paper the lexical entry for par 'a couple/pair' can be inserted in either D-Structure (as N, i.e. an element containing semantic features: at least those of Gender and independent Number) or S-structure (as QGEN, i-e. having no semantic features). The feature specification determines the level and location of insertion, and the distinct interpretation and syntactic properties follow from the syntactic position and level of insertion into the derivation.

8.2.

Seminumerals sto 'hundred', tisic 'thousand', milion 'million'

While the lexical entry for pär 'a couple/pair' exemplified in 8.1 has distinct semantic interpretations reflected by morphological and syntactic properties, the lexical entry for the Czech cardinal sto 'a hundred' does not have two clearly distinguishable meanings which could motivate its optional feature specification in the Lexicon. (But see footnote 44.) The cardinal sto 'hundred' has a neuter Gender.50 The neuter V-agreement in (61) signals that the cardinal sto 'hundred' is a syntactic head of the structure (i.e. it is not QA). The discontinuous agreement pattern with a demonstrative in GEN makes it moreover distinct from QN- (61) demonstrates the by now familiar characteristics of a discontinuous pattern of QGEN·

(61)

[QGEN]

Prislo /*-i tech /*to sto came-3SN /*P the-PM.GEN /*NOM hundred-SM.NOM chlapcu boys-PM.GEN 'There arrived the hundred of boys.' Compare, however, two available forms which can occur in the Lexical Case pattern of sto 'hundred' which are demonstrated in (62) and (63). In (62) the pattern is still the Heterogeneous Double agreement pattern of QGEN:

308 Ludmila Veselovskä

(a) the following nominal complex is in INSTR, (b) the demonstrative agrees with the Noun (i.e. PF.INSTR), (c) the cardinal sto 'hundred' does not inflect but it appears with a frozen morphology, and

(d) it does not allow premodification apart from a DEM. (62)

[QGEN]

s

temi

(*iedne mi) sto

with the-PF.INS (*one) mladymi

(*-em)

hundred-SN/ (*INS)

devcaty

young girls-PF.INS 'with the hundred of young girls.' Contrary to (62), (63) shows the Homogeneous Double Agreement pattern of NQ: (a) the following nominal complex is in GEN, (b) the initial demonstrative agrees with the quantifier, (c) the cardinal sto 'hundred' caries a standard nominal (consonantal) INSTR case morphology, and (d) it allows a premodifying A™". (63) [Nq] s

tim

(jednim) stem

mladych

with [the one hundred]-SN.INS [young 'with the (one) hundred of young girls.'

devcat

girls]-PF.GEN

The pattern of NQ is in fact forced whenever the cardinal is counted, because real countability is a property only of NQ, not of Q(GEN/A)· 5 1 Dual characteristics are typical for other Czech high cardinals as well, above all for the cardinals tisic 'thousand' and milion 'million'. The Czech Numerals/Cardinals are partitioned among all agreement patterns illustrated in this paper. I give a tentative summary of their characteristics in (64).52

Agreement patterns of Czech group nouns and quantifiers 309

(64)

Table 6. Specification of higher Cardinals (Czech) Cardinal Dve/tri/ötyri '2/3/4' Simple Cardinals Pet, sest... devatenact '5,6...-19' Sto, tisic, milion '100/1,000/1,000,000' Miliarda 'billion' Compound a) Dvacet jedna 'twenty-one' [=QGEN QA] b) jednadvacet. 'one-and-twenty' [=QA + QGEN] Cardinals Tri tisice ' 3 0 0 0 ' [ = Q A + NQ/QN]

9.

Characteristics QA QGEN QGEN or NQ NQ QA or QGEN QGEN QGEN or N 0

Summary

In this paper I first demonstrated in sections 2-5 a number of morphological and syntactic properties of the structures containing three distinct kinds of quantifying elements in Czech: group nouns Q N , existential quantifiers QGEN, and universal quantifiers QA· I compared these quantifiers with Nouns (NQ), showing the gradual loss of 'nominal' properties which is illustrated in a summarizing Table 7 in (65). (65)

Table 7: Summary of the properties of Ν and Q The observed property

1 1 2 2 3 4 5 3 6 7 4

8 9 10 11 5 12 13 14

Independent relative clause Pragmatic antecedent to pronoun Q and Ν™3* have separate sets of DEM/POSS/ A™" agreeing premodifiers The following DP can be independently quantified Q has adverbial prequantifiers (nearly, exactly, etc.) The following DP can be omitted The elliptically omitted DP in NOM/ACC must be represented by a GEN clitic Semantic Gender Formal Gender (non-interpreted) Interpreted Singular/Plural Formal Singular/Plural (non-interpreted) Q's features trigger V agreement D/N™" following the Q in NOM/ACC is in GEN D/N"1" following the Q in Lexical Case is in GEN

NO

+ + +

QN

QGEN

QA

-

-

-

-

-

-

-

-

+

-

-

-

-

+

+

±53

+

-

-

+

-

+

+

-

-

-

agr. i4 agr. agr. agr.

+ + +

+

-

-

-

+

%

-

+ +

+ +

+

-

+

+

-

-

-

310 Ludmila Veselovskä I proposed that the distinctions (1-5) in (65) follow from the distinct categorial natures, which allow distinct subcategorizations for a lexical Ν and functional Q stated in (16) as (16)

a. b. c.

D, [_N/*Q] N/Q, [ D] N/*Q, [ Q]

Other distinctions result from distinct feature contents of the Qs, which are also reflected in the morphology of Qs as they are summarized in (8)-(ll) in Table 7. The properties listed as (12)-(14) in Table 7 have been taken as criteria of headedness, and based on comparison with lexical Nouns I have proposed that the distinctions (3)-(7) in Case and Agreement Patterns result from: a. Distinct positions in the structure, i.e. head vs. SPEC position was used to explain the distinction between QN/GEN and QA. I argued that QN/GENS are inserted into the head position above N™"*, while QA occurs in the SPEC position as illustrated in (40). The SPEC position of QA explains why QA does not show any properties of a syntactic head and why it is similar to adjectival premodifiers. 56 b. Distinct levels of insertion into a cyclic derivation, i.e. D-Structure vs. S-Structure insertion of lexical vs. grammatical elements as proposed in Emonds (1997). I claimed that Q N are pre-transformational lexical elements as long as they contain features of [Number] and [Gender], On the other hand QA and QGENS are grammatical elements and are therefore inserted into a derivation later, prior to Spell Out. The relevant structures are given in (37) for QN and in (51)-(52) for QGEN· The structures assume that distinct syntactic processes apply in distinct levels of derivation as proposed for the Case hierarchy in Babby (1987) and Franks (1995) and cited here in (43)/(44).

In the final section I supported my analysis by examples of some lexical entries with dual specifications, i.e. those which can be inserted at DStructure or as late as the pre-Spell Out level. I demonstrated that distinct syntactic behaviors are always accompanied by relevant morphology and distinct interpretations of the elements.

Agreement patterns of Czech group nouns and quantifiers 311

Notes 1. This paper was presented in May 1999 in the workshop on Semi-lexical Heads at Tilburg University (Netherlands). I am grateful to the audience for inspiring questions and suggestions. The remaining mistakes are mine 2. Babby (1987) discusses case and agreement structures with quantifiers (cardinals) for Russian. The author uses examples and tests similar to those given here and I use some of his terms. There are some distinctions between Czech and Russian, however, which I will not discuss here. I will not argue against Babby's analysis. A detailed analysis of quantified structures can be found also in Franks (1995: chapters 4 and S) concerning mainly Russian and Serbo-Croatian. Franks analysis is closer to the one presented here, though it also differs in many basic aspects. As for the methodology, the analyses mentioned above do not crucially compare the quantifier structures with nominal structures, which is the way chosen in this paper. For more about the distinctions in interpretation see also Pesetsky (1982) and Franks (1995). 3. All these structures belong to the kind of pseudo-partitive constructions described in e.g. Jackendoff (1977: chapter 5), where the author uses the names: 'Group Nouns' for most elements which I will label QN, 'Numerals/Articles' for QGEN. and 'Existential quantifiers' for some of my QA. Jackendoff concentrates on linear order between the English elements occurring in the field of quantifiers and numerals and works with a form of a bar notation not accepted here. I divide the quantifiers/cardinals mainly according to the overt morphology, and the labels used in this work are therefore chosen to suggest the agreement patterns described in the following sections. 4. Traditional label for some of these nouns is 'collective nouns'. For detailed semantic characteristics see Landman (1996). 5. Priruöni mlnvnice destiny (1995: 303) gives other examples. The authors group these expressions among numerals which express a non-specific amount. 6. E.g. nekolik 'several' or tolik 'so many', nemnoho 'not many', pramalo 'very little', and comparative/superlative forms of the Q G E N quantifiers, especially of the 'many/much' and 'few/little' group. With these the comparative and superlative are formed by a suppletion using another stem vice/nej-vice 'more/most' and mini/nej-meni 'less/least'. In Prirucni mlttvnice öeltiny (1995) they are grouped among 'non-counting' numerals and in Mluvnice destiny (1986) among adverbials of quantity. 7. By intersective I mean that these Qs refer to a specific set within some other independendy quantified potential set. E.g. 'many/few boys arrived' states that a set of 'many/few' boys arrived from a set of 'boys'. 8. By 'non-intersective' I mean that the quantifiers do not count over some other (independently quantified) set expressed by the following nominal complex. Instead, they state the completeness of the following set. E.g. 'all/both boys arrived' states that the complete (=all/both) set of 'boys'. In Prirucni mluvnice

312 Ludmila Veselovskä

cestiny (1995) and Mluvnice cestiny (1986) these quantifiers are grouped among pronouns expressing 'extreme quantity' and labeled as 'totalisators'. 9. This paper does not argue in favor of any specific form of the Czech GEN complement. Based on Veselovskä (1998) I assume it is a DP which receives GEN Case from its N-sister, as illustrated in (1). If it proves better to label postnominal configurational GENs in Czech PPs (with Ρ empty), nothing would change in my analysis. 10. I will not say anything here about the interpretative rules which allow the verb in structures like 'drank a bottle of wine* to be interpreted with the complement of 'bottle', i.e. 'wine'. See also footnote 3. 11. By SPEC I mean every X™ which is a daughter to Y " " and a sister to Y, i.e. I will ignore any distinction between thematic and non-thematic adjectives. Although the two may carry specific readings the distinction is not overtly signaled by their morphology or syntax in Czech. 12. Many expressions in (3)/(4) have two possible interpretations and only the one which has a general quantificational meaning (and is glossed in English with the equivalent containing an indefinite article) shows the properties of QN. 13. For more details about Phi features see section 5 below. 14. Unlike English, Czech tolerates both demonstrative and possessive to occur in one complex nominal phrase, as illustrated in (11). 15. I argue here for the abstract syntactic distinction between NQ and QN. Many specific lexical entries, however, function as both NQ and QN. Apart from standard duality (for some examples see section 8), the distinction is also obscured by the fact that any QN can be (in marked contexts) used as NQ. Therefore the elements which can appear with overt premodifiers (i.e. are used as NQ) can also be relativised and referred back to by pronouns. The correlation between the presence of premodifiers and the ability of the element to have a separate relative clause is illustrated below. In (i) 'a plenty/deal-F of is not premodified by any agreeing adjectival and the relative clause formed with the 'plenty' is strongly deviant. On the other hand, in (ii) 'a plenty/deal-F of has a premodifying agreeing adjectival and the relative clause sounds much better. (i) Prosäkla tam hromada oleje, ktery /*-a se roztekl /*-a ... soaked-F there plenty-F oil-M which-M/*F REFL spread-M/*F ... Ά deal of oil soaked there, which spread around...' (ii) Prosäkla tam velkd hromada oleje, ktery/-a se roztekU-a soaked-F there great-F plenty-F oil-M which-F/M REFL spread-M/F ... Ά great deal of oil soaked there, which spread around....' 16. The position of DEM/POSS is fully felicitous following the QN. In marked and emphatic contexts the DEM/POSS can also precede the QN. 17. In emphatic and marked structures where DEM/POSS can precede the Q G E N quantifier as in (i). (i) Tech tvych mnoho prdtel mi nevadi. the your many friends meDAT not-bother. Ί don't mind those many friends of yours'

Agreement patterns of Czech group nouns and quantifiers 313

18. The positioning of Cardinals with respect to DEM/POSS exemplified in (13)/(14b) results in distinct interpretations in Czech: When DEM/POSS follow the cardinal the interpretation is pseudo-partitive (intersective in the sense of footnote 6) and the scope of DEM/POSS does not coincide with the cardinal, (i) below means that we have in mind some specific group of (more than four/five) boys and we recall only four/five of them. (i) dtyri ti chlapci pet tech chlapcu [four the boys]-MP.NOM five-NS.NOM [the boys]-MP.GEN 'four of the boys' 'five of the boys' On the other hand, when DEM/POSS precede the cardinal, the structure is non-partitive (non-intersective in the sense of footnote 6): The scope of DEM/POSS coincides with the cardinal, (ii) means that we have in mind just four/five boys and these are those boys. (ii) ti ctyri chlapci tech pet chlapcu [the four boys]-MP.NOM the-MP.GEN five-NS.NOM boys-MP.GEN 'the four boys' 'the five boys' As can be seen from (13)/(14b) and from (i) and (ii) above, the position of the cardinal does not change the distinct Case and Agreement patterns of the whole structures, which are discussed in detail in section 6. 19. In Veselovskä (1998) I propose that Czech adjectival agreement is a signal of a syntactic coindexation between Ν and its functional projection D. The absence of a separate D projection above Q can thus also explain the absence of agreeing elements related to Q, which is mentioned in section 3. 20. For similar structures see above all Giusti (1992) who proposes the functional category Q located above DP. QP and some other functional projections of Ν are used also in e.g. Ritter (1993). Some modification of (15b) may indeed be needed due to the position of cardinals which was mentioned in (13b)/(14b). Namely the position of Q with respect to D. In (15b)/(16a) I propose the obligatory ordering D-Q, but some additional 'lower' position must be available for Cardinals (and marginally for all quantifiers) to allow the distinct interpretation of examples mentioned in footnote 13. I will not discuss this 'lower' position here. For some related English data see Jackendoff (1977). 21. I will not discuss here the obligatoriness of N-complements based on criteria developed in e.g. Grimshaw (1991) although they may be relevant for the distinction between a standard Noun and NQ. 22. Clitics in Czech appear in the second position in the sentence, following the first phrasal constituent (see Toman 1993; Veselovskä 1994; Riet and Veselovskä 1999). Postnominal GEN pronouns are highly restricted in Czech (see Veselovskä 1998) but the clitic GEN extracted from a DP with elliptically omitted NP is a frequent phenomena. 23. The GEN clitic is usually only optional in Lexical (Oblique) Case contexts with both Q n and QGEN· The judgements are, however, murky, and therefore I will not give any examples here.

314 Ludmila Veselovskä

24. The ability of QN to form an autonomous plural contrasts with the other quantifying elements discussed in this paper. In spite of the fact that QN group nouns express 'a group of* inherently, the formal ability to create plural is often retained as well. As anonymous reviewer suggests, the plural is often related to non-quantificational collective reading. I propose this to be true especially when Q N or QGEN is 'nominalized', i.e. used as N Q (see e.g. section 8.1.2 example (59)). 25. As in English, some Czech mass nouns can be counted with distinct reading of 'many different kinds of or 'many portions of as in (i) and (ii). The presence of number on Q does not contribute to the readings. (i) Privezli spoustu/-y vin. brought-3PM plenty/-ies wine-NP.GEN 'They brought plenty of wines.' (ii) Vypil spoustu/ spousty /pet piv. drank-3SM plenty-FS/plenty-FP/five beers-NP.GEN 'He drank five beers.' 26. According to Havränek and Jedliöka (1981: 217) the 'numeral' morphology of Czech cardinals was originally uniform and followed the feminine pattern. In modern Czech the signs of the feminine pattern are identifiable with the cardinals 3 and 4 only. The authors state that with the higher cardinals only the complementarity of -zero and ^i endings survived. Following the same logic, the complementarity between and ^a exemplified for [-DEF] quantifiers in (27-29) is similar to the case paradigm of the nominal neuter pattern. In both cases, however, the same complementarity (-zeroAi and Aa) can be formulated also as e.g. a Number distinction of the productive Feminine (i) and Neuter (ii) paradigms. (i) kost - Icost-i bone-SF.NOM/ACC bone-SF.GEN/DAT/LOC = PF.NOM/ACC (ii) mist-o - mist-a city-SN.NOM/ACC city-SN.GEN = PN.NOM/ACC 27. Pronominal means a kind of morphology typical of definite pronouns: ten/ta/to/ti/ty/ta 'the-3SM/SF/SN/PM/PF/PN'. With the exception of nominative, the paradigms are very similar to the paradigms of one class of adjectives. For complete paradigms see e.g. Havränek and Jedliöka (1981:207/208,217). 28. Babby (1987: 100-101, notes 13/14) argues that the numerals 2/3/4 were adjectives in Old Russian, while the numerals 5 and 10 were nouns (feminine, singular), and in his paper he gives examples that illustrate their adjectival/nominal properties and the gradual loss of their adjectival/nominal character. The frozen morphology is also discussed in Franks (1995, section 5.2.5) for some Russian quantifiers/cardinals. In Czech, contrary to Russian the frozen morphology is characteristic for all Q G E N ' S · 29. The domain of NP Agreement in Czech covers (apart from complex adjectival postmodifiers and predicative adjectives) only adjectival premodifiers. It does not include the N-complement position to the right of Ν (stolu 'table.GEN' in

Agreement patterns of Czech group nouns and quantifiers 315

(i)), and the modifiers of the premodifiers (strasni 'terribly-ADV' in (i)). This agreement concerns the combination of features of Gender, Animacy, Number, and Case. (i) takov-ά stra Sne dlouh-ά noha stolu such-FS.NOM terribly-ADV long-FS.NOM legg-FS.NOM table-GEN 'such a terribly long leg of the table' Franks (1995) states the field of Slavic NP Agreement is in terms of sisterhood: Modifiers are sisters of the nominal structure they modify and agreement is a percolation of features targeting 'modifier'. In does not apparently make a difference for the author whether the status of the modifier is a head of a maximal projection, though most of them clearly do head maximal projections. 30. Apart from the Subject-verb Agreement , a kind of Object -verb Agreement can also be seen in Czech in pseudo-Perfect structures where the predicative (passive) adjective reflects the Case/Number/Gender of the direct object. (i) Spoustu strdnek mil napsanou uz vcera plenty-SF.ACC page-PF.GEN had-3S written-SF.ACC already yesterday Ά large amount of long pages he had written yesterday already.' 31. I will not follow Babby's (1987) proposal that Phi features of the Noun are distributed separately from Case features. I will also not use here the term 'percolation', especially when it concerns sharing of features rightwards, i.e. downwards. In this paper agreement is a defined as in (33a), i.e. only as a result of SPEC-head relation . If it were labeled as 'percolation', this 'percolation' would go upwards, to the left, and would target only maximal projections within the relevant domain . 32. In (35a) the quantified subject appears after the verb. Position, however, is irrelevant for Czech morphology contrary to e.g. Russian, where some authors propose distinct structures for preverbal and postverbal quantified subjects. 33. Recall, however, that the distinction between NQ and QN is often obscured by dual characteristics of individual elements mentioned in footnotes 11 and 14. The properties of QN are often derived referring to the analogy with the other Qs. 34. Moreover, if transitive verbs subcategorize for + Ν (or + D), a QP must be able to satisfy this subcategorization. To propose nominal characteristics for Q is a way to state this fact, as well as claiming that Q is a nominal functional head . As such it contains some [+N] (or [+D]) feature which is able to satisfy the subcategorization. The QP is an extended projection of Ν in the same sense as a DP is. 35. Another argument in favor of the SPEC analysis could be the ability of QA to be extracted into a sentence initial focus position leaving the remaining structure behind. However, given that all Qs can be extracted in such a way and that GEN clitic is separated from its Q N / G E N AS well, I am not making any conclusions based on the extraction abilities of the element. 36. A sentence can contain several structures containing Q elements of the same or of distinct type. For example, as illustrated below, the subject, direct object

316 Ludmila Veselovskä

37.

38. 39.

40.

41. 42.

43.

and also the ACC object of preposition can contain Q G E N quantifiers and all of them show the identical Q G E N patterns described in (41). Also, any ordering of the three constituents is grammatical in Czech, although if both subject and object are animate, the order S-0 is unmarked. (i) Mnoho divek poslalo mnoho chlapcu pro mnoho kamarädek. many girlsGEN sent-3SN many boysGEN for many friendsGEN· 'Many girls sent many boys for many friends.' Apart from dative, locative, and instrumental, the genitive with verbs is also a Lexical Case in Czech (contrary to the postnominal genitives which appear in this paper). I am not using any verbal genitival structures here, however, since they cannot overtly illustrate the distinctions between the agreement patterns. The terminology for the patterns is adopted from Franks (1995), The discontinuous agreement of Q G E N cardinals is illustrated in (13bii) and footnote 13. The term 'discontinuous agreement' is used in Babby (1987) for similar structures in Russian. In Russian, however, the discontinuous structures occur in more complex form. Compare my example (13b) with Babby's (1987: 125) (59). Lexical Case is assigned by either a verb and/or a (possibly empty) preposition that subcategorizes for a specific Lexical Case. Apart from the statement that these non-NOM/ACC cases are related to the subcategorial frame of the lexical element, i.e. they are part of the lexical entry, I will not further discuss the nature of Lexical Case. C-command is assumed by Babby (1987) to be the relevant notion for cases assigned under government and so it is here, as stated in (33c). Recall that ACC is, apart from direct object case, also the case assigned after a number of Czech prepositions. I illustrated in (41) that ACC after prepositions shows the same agreement and case patterns as ACC after verbs. The above stipulation therefore means that I assume that both verbs and prepositions in Czech are either specified for Lexical (non-ACC/NOM) case or they are not. Subsequently, the Configurational Case assigned 'on the basis of overall structural environment' surfaces as ACC in Czech after both verbs and prepositions. What I call GEN here is the 'postnominal genitive', i.e. the case which appears with a nominal phrase in the complement position of a Noun or Quantifier. I assume that the two are the same, both distinct from genitives appearing with nominal phrases selected by a verb. For a similar analysis see Babby (1987: 117) where the author defines the Russian postnominal GEN as Structural/Configurational Case. All his arguments apply to Czech as well; therefore I follow Babby's proposal and not the claim made in Franks (1995), who takes the discontinuous pattern as a signal of the [+Oblique] nature of a GEN assigned by a quantifier. Franks, however, does not propose the ±Oblique characteristics of a specific Case as a cross-language universal. Slavic postnominal Genitive is presented in his work as either + or -Oblique. In Veselovskä (1998) I argue in detail that postnominal GEN in Czech has all the

Agreement patterns of Czech group nouns and quantifiers 317

properties of a Configurational Case in structures containing common Nouns, and nothing signals that it should be distinct from NQ. 44. In the following text I use the traditional terminology 'D-Structure' for (45a), 'S-Structure' for (45b) and 'PF' for (45c). I assume that these levels of derivation refer to cyclic extension of subsequent projections via Merge or movement. 45. Recall however, that neither the Gender nor Number of Q N have true semantic content. The inanimate group nouns cannot show any male vs. female distinctions. As for the Number of QN, there is no distinction in interpretation of a formal plural and formal singular of group nouns. Because they are not semantically relevant, I do not assume that (50) applies to all lexical items in Czech. I can imagine some pronouns, which also reflect Phi features, qualify for (45b/c). Moreover, (50) makes NQ and Q N identical and in sections 2,3,4 I demonstrated a number of distinctions between NQ and Q N . I tentatively explained the distinctions by a distinct subcategorization (16). Possibly more could be said about this, which I leave to future research. 46. Notice that (51) is not a full equivalent of (37) as long as the maximal projection of the complex is not unambiguously QP. In fact, at D-Structure the complex is more likely D T At S-Structure, however, the complex must be syntactically defined as a maximal projection of Q as in (52). 47. Within the proposed hypothesis the QGEN in (52a) remains unmarked for Case. Recall my discussion about the character of the morphology of Q G E N in section 5.2. (27). The 'lack' of Case with Q G E N need not violate the Case Filter if the Case Filter applies only to DPs. 48. See especially (13bii) where the QGEN cardinal follows DEM/POSS. Due to its morphology the cardinal is the only element in the structure which qualifies as a possible NOM-subject, although it is located in the middle of the GEN field (preceded by GEN DEM/POSS and GEN D/Nmax). 49. The dual characteristics of some elements including seminumerals discussed in the following section 8.2 apparently violates (46/50). However, the lexical entries with dual characteristics are always formally neuter or masculine, though also feminine expressions exist in Czech which have a very similar meaning. They never show any dual characteristics, but are always only N Q or Q n , i.e. deep inserted. Therefore it seems that (50), which refers to Phi features, allows [+MASC] to be ignored but never [+FEM]. Assuming that [+FEM] is the marked Gender in Czech, the ambiguity seems less unexpected and (50) can be kept. 50. sto 'hundred' is also the only Czech cardinal which retains its dual form, i.e. it has a special 'dual' morphology when it follows 'two'. See dve ste 'two hundred-DUAL' vs. tri/ctyri st-a 'three/four hundred-PL' 51. Though they are at the first sight synonyms, the interpretations of (62) and (63) are not fully identical. In (63), the presence of a premodifier forces the form of NQ with the interpretation of 'a set of hundred girls', i.e. with a consistive read-

318 Ludmila Veselovskä

52.

53.

54.

55. 56.

ing in the terminology of Franks (1995), while with QGEN IN (63) the reading is more partitive, i.e. 'a hundred of individual girls'. With compound cardinals in Czech the dual characteristics follows from their composition: in Czech the ordering is optional and in a compound of [QGEN + QA] form, any of its parts can be analyzed as a head. The following DP contains another Q element. This element, however, does not add an independent quantification. The structures comprise a QA quantifier (totalisator) and a cardinal of QA or QGEN type (e.g. both two brothers, all five sons). In section 5.3 I show that QA have a rich morphology reflecting the Gender and Number , but this morphology copies the Phi (and Case) features of the Noun, i.e. they do not belong to the QA. V-agreement is 3SN ('default'), and so is better stated as 'N cannot agree with the Verb' in the presence of QGEN· A challenging alternative in the framework chosen here could be proposed for QA if they were analyzed as 'grammatical formatives containing only uninterpretable grammatical features F'. As such, according to (45c) QA would be inserted into the structure after Spell Out, at PF. An element inserted into derivation at PF plausibly cannot project and cannot be analyzed as a syntactic head at any lower (pre-Spell Out) level of derivation. The lack of head properties demonstrated for QA in (39) is formalized in (40) by locating QA in a position in SPEC. One can hypothesize that a PF inserted element which does not substitute for any existing syntactic node would have to be analyzed as SPEC, simply due to the fact that it does not project and does not show any syntactic properties of a head. Such an analysis could be supported by a non-intersective interpretation of QAs, which are not interpreted as Qs counting over a set but as 'totalisators' (see footnote 6), i.e. as some default unmarked Q. This may also explain why there are no distinctions with respect to extractability between QGEN and Q A .

On the other hand, one may wonder how a PF-inserted element can license ellipsis and why the cardinals in the lower position (following DEM/POSS) are non-intersective, too. Also because of framework-internal problems (e.g. what mechanism would force such a PF insertion) I propose the SPEC analysis here and do not argue in favor of a PF level of insertion for QA.

Agreement patterns of Czech group nouns and quantifiers 319

References Abney, Steven 1987 The English Noun Phrase in Its Sentential Aspect. PhD dissertation, MIT. Babby, Leonard 1987 Case, Prequantifiers, and Discontinuous Agreement in Russian. Natural Language and Linguistic Theory 5: 91-138. Chomsky, Noam 1981 Lectures on Government and Binding. Dordrecht: Foris. 1986 Barriers. Cambridge, Mass.: MIT Press Emonds, Joseph 1997 Lexicon & Grammar: the English Syntacticon. Manuscript, University of Durham. Franks, Steven 1995 Parameters of Slavic Morphosyntax. New York and Oxford: Oxford University Press. Giusti, Giuliana 1992 La sintassi dei sintagmi nominali quantificati. Tesi di Laurea, University of Venice. Grimshaw, Jane 1991 Argument Structure. Cambridge, Mass.: MIT Press Havränek, Bohuslav and JedliCka, Alois 1981 Ceskä mluvnice. [Czech Grammar]. Praha: Stätni pedagogickd nakladatelstvi. Jackendoff, Ray 1977 X-bar Syntax: A Study of Phrase Structure. Cambridge, Mass.: MIT Press. Landman, Fred 1996 Plurality. In S. Lapin (ed.) Handbook of Contemporary Semantics. London: Blackwell. Mluvnice öestiny 2 1986 [Czech Grammar 2] Czech Academy of Science, Academia Praha. Pesetsky, David 1982 Paths and Categories. PhD dissertation, ΜΓΓ. Prirucni mluvnice destiny 1995 [Czech Grammar Handbook]. Philosophical Faculty Masaryk University, Brno. Nakladatelstvi Lidov6 noviny. Ritter, Elisabeth 1993 Where's Gender? Linguistic Inquiry 47: 795-803. Selkirk, Elisabeth 1977 Some Remarks on Noun Phrase Structure. In: P. Culicover et al. (eds.), Formal Syntax. New York: Academic Press.

320 Ludmila Veselovskä

Toman, Jan 1993

A Discussion of the Second Position in Czech. Talk given at the Wackernagel workshop, Geneva, June 1993. Veselovskä, Ludmila 1994 Phrasal Movement and X-Morphology:Word Order Parallels in Czech and English Nominal and Verbal Projections. PhD dissertation, Palacky University, Olomouc. 1998 Possessive Movement in the Czech Nominal Phrase. Journal of Slavic Linguistics 1998.2. Indiana: Bloomington. Vos, Riet and Veselovskä, Ludmila 1999 Clitic Questionnaire. In: Henk van Riemsdijk (ed.), Clitics in the Languages of Europe. The Hague: Mouton de Gruyter.

Part III Semi-lexicality in the Verbal Domain

On the (semi)lexical status of light verbs Miriam Butt and Wilhelm Geuder

1.

A brief introduction to light verbs

Light verbs have proven to be an interesting challenge for theories of syntax and semantics because they have properties that make it difficult to class them either with function words or with proper lexical verbs. In this paper1, we make an attempt to arrive at a more precise understanding of light verbs by combining syntactic and semantic considerations, and showing that light verbs have to be considered as a class in their own right. Our claim is that a particular combination of syntactic and semantic properties can be identified that characterizes light verbs. While we principally consider both V+N and V+V light verb constructions, this paper predominantly centers on the analysis of the latter. To this end, large parts of the paper consist of a case study of the properties of V+V light verb constructions in Urdu.2 In order to clarify and delimit the subject matter, we will first introduce the concept of a light verb in this section, and review some of the major approaches to light verbs that have been proposed in the literature. We then go on to our own proposal.

1.1.

Characteristics

Jespersen (1965, Volume VI: 117) is generally credited with first coining the term of light verb for English V+NP constructions such as have a rest, give a shout. Since then, the term light verb has been extended to cover V+V constructions in a variety of languages, such as Japanese, Romance, and Urdu. An example from Urdu is given in (1), with the main verb 'eat' and the light verb 'take'.

324 Miriam Butt and Wilhelm Geuder

(1)

yaasiin=ne kek khaa lii-yaa Yassin.M=Erg cake eat take-Perf.M.Sg 'Yassin ate the cake (completely, for the benefit of himself).'

A light verb is generally part of a complex predicate. Based on an in-depth examination of Urdu V+V complex predicates in comparison with work done on Japanese iwrw-constructions and Romance causatives and restructuring verbs, Butt (1995) proposes the characteristics in (2) for the crosslinguistic identification of complex predicates. (2)

Defining Characteristics of Complex Predicates: • The argument structure is complex (two or more semantic heads contribute to it). • The grammatical functional structure is that of a simple predicate (there is only a single subject and no embedding). • Complex predicates may be formed either morphologically (lexically) or syntactically.

While these syntactic characteristics accurately describe a number of complex predicates, the semantics of complex predicates is not as well understood. In most cases, the systematic semantic contribution of the light verb to the complex predicate is extremely difficult to identify.

1.2.

Previous work on light verbs

It seems that the difficulty of giving a notional characterization for the function of light verbs has led researchers to center, in their analyses, on features that are predominantly syntactic in nature. One approach has been to treat the light verb as a functional element which simply serves as a licenser of predication. Cattell's (1984) survey of V+NP phenomena in English can be seen as falling into this category, as well as Grimshaw and Mester's (1988) Argument Transfer analysis of N+V (suru 'do') complex predicates in Japanese. This type of analysis can only be considered acceptable if the light verbs in question really do not make any semantic contribution other than licensing the predication of the nominal. However, this is clearly not the case for English and has been shown not to be the case for Japanese (Isoda 1991), where suru 'do' at the very least is responsible for contributing a selectional restriction with respect to agentivity (also see Matsumoto 1996 for

On the (semi)lexical status of light verbs 325

some discussion of further problems with the Argument Transfer account). As such, while the analysis of light verbs as functional elements is simple in the sense that it avoids the difficulties of addressing semantic issues, it is unsatisfactory because there are clearly some issues to be accounted for. Another approach is to identify light verbs as a type of auxiliary, though even in this category the approaches range from seeing light verbs as essentially redundant elements (e.g. Hacker 1958) to placing auxiliaries on a par with French tense auxiliaries (Abeille, Godard and Sag 1998), to analyzing them as vector verbs which perform an explicating or aspectual function (Hook 1974,1991). On the other extreme, Huang (1992), for example, analyzes Chinese ba and de constructions as an instance of control. In this type of approach light verbs are treated on a par with main verbs that allow for control/raising. A more complex approach, which begins to account for the syntactic and semantic properties of light verbs more adequately, is to see them as contributing to the argument structure of a predicate which they are syntactically and semantically dependent on. Rosen (1989), for example, posits three differing types of argument mergers for Romance restructuring verbs and causatives: light (empty), partial, and complete merger. Some approaches invest light verbs with almost the same predicational power as full verbs: Alsina (1996), Mohanan (1994), or Butt (1995) posit a notion of Predicate Composition or Argument Fusion in which the light verb and the main verb are seen as both significantly contributing to the complex predication. These argument structure approaches manage to approximate the idea that both the main and the light verb are jointly contributing to the predicational power of the complex predicate. However, even these more complex approaches leave unexplained the wide variation in usage and the subtle shades of meaning that accompany one and the same argument structure composition.

1.3.

Our proposal

We see light verb constructions such as those in (1) as a type of complex predicate which consists of a main (or full) verb in combination with a lexically defective verb. Our first point, to be addressed in section 2, will be to show that light verbs of the V+V type in Urdu constitute a syntactic class that is distinct from auxiliaries. We analyze light verb constructions as doubly headed structures; the full verb and the light verb form a unit, and

326 Miriam Butt and Wilhelm Geuder

the light verb in many respects patterns with fully lexical elements, not with functional heads. This distinguishes light verbs syntactically from auxiliaries, in that auxiliaries typically involve a VP-embedding construction and do not form part of the lexical core of the clause. These syntactic facts will later be correlated to the particular semantic function that light verbs perform. As a first step towards a more accurate semantic analysis, section 3 argues that there is no reason to subsume light verbs under aspectual markers, contrary to claims in the literature (e.g., Hook 1991, 1993; Hopper and Traugott 1993). Furthermore, a historical investigation into the origin of light verbs (Butt and Lahiri 1998) has revealed that light verbs have actually been attested over thousands of years in much the same form as today. As such, we cast doubt on the popular idea that the semantic "lightness" of light verbs must have arisen from progressive semantic bleaching that resulted in their grammaticalization. In this, light verbs again prove to be different from auxiliaries. We therefore develop an alternative view on the lexical semantic properties of light verbs in sections 4 and 5. In section 4, we argue that light verbs, though being lexically defective, are not completely empty elements (as assumed by Grimshaw and Mester 1988, for example). A light verb conveys a special type of lexical meaning which consists in a modulation of the event description provided by the full verb that it is in a construction with. Apart from these compositional issues, we point out the importance of the fact that all light verbs in Urdu are paired with formally identical full verb uses. The relation between the light verb use of a verbal item and its use as a full verb is identified as a case of lexical polysemy (not grammaticalization). On the other hand, we show that in spite of the well-motivated conceptual connections between full and light verb meanings, a marked difference in semantic type appears with the light verb variant. Crucially, these fundamental semantic properties are shared by both V+V and V+N light verb constructions. Light verb constructions do take on different properties depending on the different syntactic frames, but these differences are shown to be predictable from the constructions and their diachronic origins; they do not entail a deeper difference in the lexicalsemantic status of the light verbs. Hence, a category of light verbs can be delineated independently on lexical-semantic grounds. In section 5, we briefly consider the semantic types of complex predicate constructions in another language in which the use of complex predicates is close to being obligatory - the case in point is the Australian language Wagiman. Here, we find more semantic types of complex predicate

On the (semi)lexical status of light verbs 327 constructions than in Urdu, and this case helps to establish the point that the semantic notion of light verbs we are envisaging is in fact independent of limitations on the available constructions in single languages. We thus believe that the criteria we provide will allow the identification of light verbs crosslinguistically. In sum, while not all light verbs function exactly alike, they can be characterized by a unique collection of semantic properties, which are consequently mapped, in V+V constructions, onto a corresponding syntactic class with distinctive syntactic properties. Light verbs in V+V constructions can thus be considered semi-lexical heads on both the syntactic and the semantic dimensions. 2.

The syntactic status of light verbs in Urdu

This section and the next draws its data exclusively from Urdu. The issues under discussion are relatively well understood for Urdu and thus are useful for making the larger point about light verbs in general. In Urdu, light verbs can be distinguished very clearly both from main verbs and auxiliaries in terms of syntactic properties. The next few sections first discuss examples of Urdu V+V light verbs in more detail, then go on to show how they differ from main verbs and auxiliaries with respect to word order, topicalization, case marking and reduplication.

2.1.

Urdu V+ V complex predicates

Examples of typical Urdu V+V complex predicate constructions are shown in (3). The form of the first verb in the sequence is identical to the stem form and never carries any inflection. This is the main verb. The second verb carries tense and aspect marking and is inflected according to the standard paradigms governing simple verb inflection. This is the light verb.3 (3)

a.

b.

naadyaa=ne xat (H^QJ (άϋ-γαα)ω Nadya.F=Erg letter.M.Nom write give-Perf.M.Sg 'Nadya wrote a letter (completely).' naadyaa (αά)ω {bet-ϊϊ)ω Nadya.F.Nom. come sit-Perf.F.Sg 'Nadya has arrived.'

328 Miriam Butt and Wilhelm Geuder

Light verbs form a separate prosodic phrase, as also indicated in (3), and are separable from the main verb, as is shown in subsequent sections. The light verbs are also always form-identical with a main verb, but are restricted to a finite subset (about 24) of the main verbs (about 200). Some of the most commonly used light verbs in Urdu/Hindi (Hook 1974) are shown in Table (4). (4)

2.2.

Common light verbs le 'take' daal 'put' aa 'come' par 'fall' bet" 'sit' kar 'do' maar 'hit' nikal 'emerge'

de 'give' jaa 'go' ut* 'rise' mar 'die'

Syntactic differences between light verbs and main verbs

Given that light verb uses are restricted to a subset of the main verbs, this already entails that the syntactic distribution of light and main verbs, as a class, will differ. When both uses exist, it is still to be noted that two differing syntactic analyses of the same surface V+V string may be possible, depending on whether the second verb is interpreted as a main verb, as in example (5b), or as a light verb, as in (5a). (5)

a.

b.

naadyaa=ne [xat lik? dii-yaa] Nadya.F=Erg letter.M.Nom write give-Perf.M.Sg 'Nadya wrote a letter (completely).' naadyaa=ne xat [lik? (kar)] dii-yaa Nadya.F=Erg letter.M.Nom write having give.Perf.M.Sg 'Having written a letter, Nadya gave (it).'

The analysis in (5b) is a control structure in which the embedded verb is the head of an adverbial clause. So, light verbs must be clearly differentiated in terms of phrase structure from constructions involving two main verbs. In addition, the light verb construction proves asymmetrical in that the main verb may be topicalized independently of the light verb, as shown in (6b), though the light verb may not be topicalized away from the main verb, as shown in (6c).

On the (semi)lexical status of light verbs 329

(6)

a.

baccaa so ga-yaa child.M.Nom sleep go-Perf.M.Sg 'The child has gone to sleep.' b. so to baccaa ga-yaa sleep Top child.M.Nom go-Perf.M.Sg 'The child has gone to sleep.' c. *ga-yaa to baccaa so go-Perf.M.Sg Top child.M.Nom sleep 'The child has gone to sleep.'

While we show in the following that light verbs have certain properties in common with main verbs, it must be kept in mind that they are distinct from them by the criteria just shown.

2.3.

Light verbs vs. auxiliaries

One popular alternative analysis with regard to light verbs has been to regard them as a type of auxiliary. However, auxiliaries and light verbs display clear behavioral differences. Unlike auxiliaries, light verbs always span the entire verbal paradigm. That is, light verbs are not restricted to appear with just one tense or aspectual form. Light verbs also never display a defective paradigm, which is again unlike what one might expect of an auxiliary. Furthermore, auxiliaries and light verbs show distinct syntactic behaviors with regard to case marking, word order, reduplication and topicalization.

2.3.1.

An introduction to Urdu/Hindi auxiliaries

In order to be able to make the comparison between auxiliaries and light verbs in Urdu more immediately accessible to the reader not familiar with this language, Table (7) shows the forms that are clearly auxiliaries in Urdu.4 (7) Urdu auxiliaries Form Meaning Inflection Pres/Fut/Impf/Perf ho-naa to be th -aa/ii/e be (Orig. stand) Past rak11 -naa Progressive (Orig. stay) Perf* * (Fut/Impf only with differing morphology on the main verb)

Defective Cells Past All Others Pres/Past/Prog *

330 Miriam Butt and Wilhelm Geuder

Some explanation of these forms is in order. Synchronically, the be auxiliaries are regarded as belonging to the same paradigm, whereby th - functions as the past form. However, as is evident from differences in agreement properties (the ho auxiliaries inflect according to tense, person and number, while th - inflects according to gender and number), these auxiliaries really represent a case of two paradigms in complementary distribution. The fact that the auxiliaries inflect differently can be traced directly to their historical origins: the ho auxiliary is the descendant of the verb 'be', while the th is derived from the past participle form of the Sanskrit verb sthä 'stand'. The progressive rah is form-identical with the full verb 'stay'. However, it is not a light verb because its syntactic distribution differs from that of light verbs and it has a defective paradigm, unlike light verbs. The morphological inflection on the auxiliary rah is identical to perfective and imperfective morphology. These bits of morphology have been reinterpreted to denote progressive action in the case of the perfective morphology and repeated action in the case of the imperfective morphology.5

2.3.2.

Case

One difference between light verbs and auxiliaries is that light verbs in Urdu/Hindi have an influence on the case marking of the subject that goes beyond the usual type of effect observed with split-ergativity (see Butt 1995). As can be seen in (8), the verb 'write' usually takes an ergative subject when formed with the perfect tense, regardless of the auxiliary used.6 Light verbs like par 'fall', on the other hand, have the ability to require that the subject be nominative, even in the perfect. (8)

a.

b.

us=ne/*vo xat Pron.3.Sg=Erg / =Nom letter.M.Nom 'He wrote a letter.' us=ne/*vo xat Pron.3.Sg=Erg/=Nom letter.M.Nom (he / faa) be.Pres.3.Sg / be.Past.Sg.M 'He wrote (has/had written) a letter.'

lit -aa write-Perf.M.Sg lit -aa write-Perf.M.S

On the (semi)lexical status of light verbs 331

(9)

a.

b.

2.3.3.

*us=ne/vo xat lilt* par-aa Pron.3.Sg=Erg / =Nom letter.M.Nom write fall-Perf.M.Sg 'He fell to writing a letter.' us=ne/*vo xat lik? lii-yaa Pron.3.Sg=Erg / =Nom letter.M.Nom write take-Perf. M.Sg 'He wrote a letter (completely).'

Word order

A consideration of ordering in the Urdu verbal complex furthermore shows that auxiliaries and light verbs do not compete for the same positional slot. In contrast to the general freedom in word order that Urdu exhibits with respect to NPs and obliques in the clause, the word order inside the Urdu verbal complex is very rigid and can be characterized as in (10). (10)

Main verb (light verb) (Progressive) (Be Auxiliary)

As (11) shows, light verbs and auxiliaries are not interchangeable. (11)

2.3.4.

a.

baccaa so jaa rah-aa he child.M.Nom sleep go Prog-M.Sg be.Pres.3.Sg 'The child is going to sleep.' b. *baccaa so rah ga-yaa he child.M.Nom sleep Prog go-Perf.M.Sg be.Pres.3.Sg 'The child is going to sleep.'

Reduplication

In complex predicates the light verb, the main verb, or both may be reduplicated (see Fitzpatrick-Cole 1994, 1996 for a detailed study on the sister language Bengali). Example (12) illustrates reduplication of the light verb.

332 Miriam Butt and Wilhelm Geuder

(12)

a.

b.

vo so jaa-ti Pron.3.Sg.Nom sleep go-Impf.F.Sg 'She to used to go to sleep.' vo so jaa-tii Pron.3.Sg.Nom sleep go-Impf.F.Sg th-ii be.Past-Sg.F

th-ii be.Past-Sg.F vaa-tii go.Redup-Impf.F.Sg

'She used to keep going to sleep (at inopportune moments).' An auxiliary may not be similarly reduplicated, (13) a. vo so rah-ii Pron.3.Sg.Nom sleep Prog-F.Sg 'She was sleeping.' b. *vo so rah-ii Pron.3.Sg. Nom sleep Prog-F.Sg 'She was sleeping.' (14)

2.3.5.

as (13) and (14) show. th-ii be.Past-Sg. F vah-ii th-ii Prog-Redup be.Past-Sg.F

vo so-tii th-ii Pron.3.Sg.Nom sleep-Impf.F.Sg be.Past-Sg.F 'She used to sleep.' b. *vo so-tii th-ii s-ii Pron.3.Sg.Nom sleep-Impf.F.Sg be.Past-Sg.F be.Redup 'She used to sleep.' a.

Topicalization

Further evidence for a clear difference between auxiliaries and light verbs comes from topicalization. As already shown in (6), a main verb may be topicalized away from the light verb. However, it may not be topicalized away from a cluster of auxiliaries, as (15) illustrates: (15)

a

b

so to baccaa ga-yaa sleep Top child.M.Nom.go-Perf. M.Sg 'The child has gone to sleep.' *so to baccaa rah-aa he sleep Top child.M.Nom Prog-M.Sg be.Pres.3.Sg 'The child is sleeping.'

On the (semi)lexical status of light verbs 333

2.3.6.

Summary

The above data point towards the conclusion that light verbs must be acknowledged as a separate class of items; in particular, they do not fall into a class with auxiliaries. There is a parallelism in the behavior of light and full verbs with respect to case marking of the subject and reduplication which suggests that light verbs constitute a subtype of the lexical category V. Further, following argumentation by Alsina (1996), we consider light verbs to be a type of co-head (co-predicator). The structure in (16) illustrates this idea for the purposes of this paper.

(16) IP SpecIP TOPIC

I' AspP

SpecAspP

1° Asp' Asp0 (Progressive)

VP SpecVP

V' NP

V' γ° (main)

γ0 (light)

The notion of co-predication carries with it the idea that both elements of the complex predicate (i.e., both the light and the main verb) jointly determine the predicative power of the complex predicate. Main verbs have a complete argument structure and introduce an event description. They are predicational elements in their own right. Light verbs have an incomplete argument structure that can only be licensed in conjunction with a main verb. A solid amount of evidence has been accumulated for this view in previous work on argument structure in Urdu (Mohanan 1994; Butt 1995), and which we substantiate from a more semantic angle in section 4.

334 Miriam Butt and Wilhelm Geuder

3.

The issue of perfectivity and aspectogenesis

The presence of light verbs in Urdu/Hindi has sometimes been viewed as due to a process of aspectogenesis (e.g., Hook 1991). Under this view, the light verb is taken to mark perfectivity (in the sense of the Slavicists). Since we have spent part of the last section arguing that light verbs need to be differentiated from auxiliaries and hence should not be implicated in tense/aspect phenomena (or if, then only indirectly), we need to establish that light verbs in fact are not an instance of aspectogenesis. And indeed, the supposed connection with perfectivity can be shown to be weak, if not non-existent. The idea that light verbs are aspectual elements arises from the fact that among other things, Urdu light verbs signal the inception or completion of an event (e.g., Hook 1974; Singh 1994; Chakraborty 1992; Butt 1995). However, even if their overall meaning is often found quite elusive, clear examples can be found that demonstrate that they can also signal volitionality, forcefulness, benefaction, and a host of pragmatic information that cannot be characterized systematically (cf. section 4). It is not clear how this other type of information can be reconciled cleanly with the idea of aspectogenesis.

3.1.

The independence of the tense/aspect system

The mixed periphrastic and inflectional system of Urdu/Hindi is quite rich already. In addition to the forms in Table (17), further constructions such as reduplication of the verb are available in order to express repeated actions or habituality.

(17) Pres Urdu

Past

Fut

maaraa

maaregaa

Impf Pres/Past maartaa +Aux (be)

Perf Pres/Past maaraa +Aux (be)

Prog Pres/Past maar raahaa +Aux (be)

maar- 'hit' - 3.Sg.M

Now, if light verbs arose as part of aspectogenesis, one would expect that they form a clear subclass of the tense/aspect system. However, this does not appear to be the case. Instead, light verbs themselves get marked for [±perfective] and take on different meanings depending on this (we do not

On the (semi)lexical status of light verbs 335

deal with this complication in the present paper, but see Butt 1997). So, rather than being part of the aspectual paradigm, light verbs interact with it. In addition, while the tense/aspect system of Urdu is the result of rather massive changes from an originally purely inflectional Sanskrit system, traces of light verb constructions can be found as far back as Sanskrit and Middle Indo-Aryan (Pali) (see Butt 1999). These historical facts again do not coincide with a notion of aspectogenesis by which light verbs arose in order to mark perfectivity.

3.2.

Perfectivity

In fact, while light verbs do denote the inception and completion of an event (among other things), the proposed connection to perfectivity can also be cast into a doubtful light through examples in which light verbs occur in non-perfective readings. Consider the definition of perfectivity on which Hook (1991) bases his argumentation: Perfectivity indicates the view of the situation as a single whole, it denotes a complete situation with a beginning, a middle and an end. Perfective verbs can be used to indicate the beginning (ingressive) or the end (completive) of a situation. (Comrie 1976)

While this definition is rather loose, it is still strong enough to allow us to detect counterexamples to the thesis that light verbs in Urdu are markers for perfectivity. Consider the examples in (18), involving when - clauses: (18)

a.

mariiam imel lik? rah-ii Γ-ii Miriam.F.Nom e-mail.F.Nom write Prog-F.Sg be.Past-F.Sg jab viilii kamre=ke andar aa-yaa when Willi.M.Nom room.M=Obl in come-Perf.M.Sg 'Miriam was writing an E-mail when Willi came into the room.'

b.

mariam imel lilf maar rah-ii f-ii Miriam e-mail.F.Nom write hit Prog-F.Sg be.Past-F.Sg jab viilii kamre=ke andar aa-yaa when Willi.M.Nom room.M=Obl in come-Perf.M.Sg 'Miriam was dashing off an E-mail when Willi came into the room.'

336 Miriam Butt and Wilhelm Geuder

Example (18a) shows a construction without a light verb. The head clause, containing a progressive marker, describes an extended situation of writing which functions as the time frame for the perfective wAe/i-clause that marks Willi's coming in; the latter must be the situation that is "viewed as a whole", because it is embedded into another ongoing action. Example (18b) shows the same two clauses, but with the light verb 'hit' added in the head clause. If the light verb were a perfective aspect marker, its occurrence in (18b) would be paradoxical. It is the when-clause that introduces the event that is viewed as a whole in this construction, but this is not the one that contains the light verb. Moreover, the light verb 'hit' not only occurs in the wrong clause, it is even embedded by a progressive, although the progressive is clearly an imperfective construction. Finally, the addition of the light verb changes the content of the event description in the head clause. The perfective variant of 'to write' would be expected to express a complete view on the situation of writing, as the definition above suggests; however, the combination 'write-hit' amounts to a new situation, namely that of sending off something. This is indeed a bounded situation, but this is due to the new lexical content that has been added. As already indicated, this predicate 'write-hit' is then put in the progressive. The moral is that we have to be careful with the notion of "aspect". In light of the above data, we conclude that what is involved with these light verbs is not aspect in the temporal sense, in the sense of locating an event relative to the reference time of the clause (i.e., for instance, "viewing it as a whole"). Rather, what is involved is some type of lexcially specified content which interacts with the semantics of the main verb. The interaction of the lexical content of the light verb with the event semantics of the main verb is what gives rise to all such effects as the specification of inception/completion, volitionality, force, benefaction, etc. This is discussed further in the next section.

3.3.

Conclusion

We have seen that light verbs do not enter into an aspectual paradigm, and so it is also very doubtful that one can speak of "grammaticalization" in this case. What is involved in light verb constructions is some type of lexcially specified content which interacts with the semantics of the main verb.

On the (semi)lexical status of light verbs 337

4. 4.1.

On the semantics of light verb constructions General remarks

The agenda for the semantic analysis of light verb constructions can be broken down into the following three basic questions: 1. What is the interpretation of the light verb construction as a whole? 2. What is the semantic contribution of the light verb itself, and what is the mode of composition between light verb and full verb? 3. How is the interpretation of the light verb connected to the lexical meaning of the corresponding full verb use? In order to be able to view the lexical semantic issues in sufficient detail, we will conduct a detailed case study of just one light verb: give (Urdu de) and then integrate our findings for this light verb into a more general typology of light verb constructions. The Ürdu verb de 'give' may be used in a variety of constructions. In addition to occurring in the V+V complex predicate as presented in the previous sections (an example is repeated here in (22)), the verb de can also be used as a full verb as in (19), as expressing a permissive as in (20), and as a light verb in an N+V complex predicate as in (21). (19)

naadyaa=ne yaasiin=ko xat dii-yaa Nadya.F=Erg Yassin.M=Dat letter.M.Sg.Nom give-Perf.M.Sg 'Nadya gave Yassin a/the letter.'

(20)

naadyaa=ne yaasiin=ko xat lilt -ne Nadya.F=Erg Yassin.M=Dat letter.M.Sg.Nom write-Inf.Obl dii-yaa give-Perf.M.Sg 'Nadya let Yassin write a/the letter.'

(21)

a.

niinaa=ne kahaani=par ctyaan Nina.F=Erg story.F.Sg=on attention.M.Sg.Nom dii-yaa give-Perf.M.Sg 'Nina paid attention to the story.' (from Mohanan 1994: 209)

338 Miriam Butt and Wilhelm Geuder

b.

(22)

naadyaa=ne kamre=ko jhaamu Nadya.F=Erg room.M.Sg.Obl=Dat broom.M.Sg.Nom dii-yaa give-Perf.M.Sg 'Nadya swept the room.'

naadyaa=ne xat lilf dii-yaa Nadya.F=Erg letter.M.Sg.Nom write give-Perf. M.Sg 'Nadya wrote a letter (completely).'

Note that while the form of the verb 'give' is the same in all of these constructions, the overt morphosyntactic forms of the other parts of the predicate differ. The affix -ne on the permissive has been identified by Butt (1995) as a frozen oblique form of a morpheme that marks verbal nouns. That is, the closest English literal rendition of (20) would be something like: "Nadya gave the writing of the letter to Yassin." Non-literally, this translates into a permissive.7 Similarly, ctyaan 'attention' in (21) is very clearly a noun (see Mohanan 1994 for a detailed discussion of the syntax and semantics of N+V complex predicates in Hindi/Urdu). Prima facie one would therefore expect to be able to assimilate the Urdu permissive and the N+V complex predicates under one type of analysis, and not be surprised if the V+V type of complex predicate required another type of analysis altogether. Ultimately, however, we believe that a semantic characterization of light verb constructions does not fall in with the categorial division of N+V vs. V+V in every respect. Rather, it turns out that certain N+V constructions (but not others) display enough similarities to the V+V complex predicates to allow us to group them together in one semantic class (syntactic differences notwithstanding). Showing this, among other things, is the objective of the following digression on light verbs in English.

4.2.

The light verb 'give' in English

Unlike Urdu, English is more restricted in that it only exhibits complex predicate type constructions in a V+N constellation. Nevertheless, English is of interest for the discussion at hand because we would like to show that the lexical-semantic part of the analysis that is required is largely the same for these V+N and for the Urdu V+V constructions. The English V+N construction also provides us with a way of easing into the more detailed

On the (semi)lexical status of light verbs 339

discussion of Urdu via a language and a phenomenon that is presumably already familiar to the reader. In the literature, we find at least one detailed typological overview of different variants of 'give'-type verbs, viz. Newman (1996). Concerning light verbs, however, Newman simply follows the line of Hook (1991) in classing the light verb 'give' in V+V complex predicate constructions crosslinguistically under aspectual markers. In contrast, the light verb give as it occurs in English examples such as He gave the car a wash is given a completely different analysis. As we have argued at length in earlier sections, Urdu light verbs could be clearly distinguished in terms of their syntax. English is different in this respect: The light verb give is found in the same syntactic frames as the corresponding full verb use. This is to say, light verbs occur in the typical ditransitive construction (cf. the examples (24)-(28) below), but for both full and light verb there are also monotransitive variants, like (23a) and (23b), respectively: (23)

a. b.

to give heat / light / milk to give a scream

What will be said about the ditransitive case can mostly be transferred to this type as well, but for the sake of brevity we would like to leave this variant largely out of consideration. So let us consider the following list of English ditransitive examples, ordered in terms of increasing "abstractness" or "lightness" of the verb meaning (the first example in (25) and (27) is also discussed in Newman 1996): (24)

give him the ball give the dog a bone give the customer a receipt

(25)

Tom gave the children their inheritance money before he died The king gave the settlers land

(26)

give advice give someone the right to do something give someone information

(27)

give someone emotional support give someone one's regards

340 Miriam Butt and Wilhelm Geuder (28)

a. b.

give someone a kiss / a push / a punch / a nudge / a hug give the car a wash, give the soup a stir

Our aim in the following in-depth discussion of these examples is to show that these different uses, including the light verb use, constitute a gradual extension of a prototypical concrete meaning of give. In other words, we argue that we are dealing with a dense meaning spectrum that represents an instance of lexical polysemy. The light verb use of give is characterized by the appearance of an event-denoting nominalization in the role of the direct object argument, which we will denote as the THEME slot for the purposes of this paper. In general, the different lexical properties of all variants of give seem to covary with the sort of entity occupying the THEME slot (a state of affairs that is typical of verbal polysemy). Example (24) illustrates a maximally concrete meaning of give. To describe it, we can start out with a lexical decomposition as follows: (29)

DO (x, [x CAUSE CHANGE ])

The relation DO (taken from Dowty 1979) encodes the agentive-intentional role of χ in bringing about the event of giving. Inadvertent causation, which would still fall under the CAUSE relation, is excluded for literal give; therefore, DO is needed in addition. Instead of the usual result function BEC (become), we posit a relation CHANGE which explicitly specifies the initial state as well. This allows one to capture the meaning component of give according to which the agent transfers an object out of his own sphere of possession. Ultimately this is the source of the sense extensions of give that convey only the notion of emission (as in the monotransitive examples in (23)), or only exertion of force by the agent, cf. the Urdu example (46) farther below. The relation HAVE in this representation is a vague predicate that is specified in different ways according to the different uses of the verb. The prototypical variant of give as in (24) involves a change of location accompanied by a change in possession. We can render this by using two more specific relations: AT (as a purely local relation of being contained in a space defined by its second argument) and POSS ("possession": embodying a special relation of control over an object). So for (24), both of the following inferences from (29) are valid: (30)

DO (x, [x CAUSE CHANGE ]) DO (x, [x CAUSE CHANGE ])

On the (semi)lexical status of light verbs 341 The referent of the THEME NP, the ζ argument in the formula above, decides on the kind of resultant state that can be asserted in relation to give. With a theme argument that denotes a more abstract entity or something unable to move in space, the notion of a transfer in space is found to be absent from the verb meaning. For instance, giving the money in the first example of (25) might consist of a bank transfer without any concrete object being moved. In the second example in (25), only the transfer of a right of possession to the land is asserted. Although the land is an entity that is localized in space, it is not moved. Consequently, (25) shows uses that instantiate HAVE only by POSS, i.e., the first clause of (30) - and it can be seen that this is correlated to a particular sort of entity occupying the THEME slot. In (26), repeated below, we find cases that neither imply transfer in space nor transfer of possession: (26)

give advice give someone the right to do something give someone information

Entities like information and rights do not behave like property and cannot be disposed of in the same way. Still, these cases are related to the former group in that giving still entails an increase in the recipient's possibilities to act on things; thus, this state of affairs shares certain abstract properties with the notion of control inherent in POSS. The lexical representations for (26) would have to replace POSS by another, quite vague, two-place relation. The point is that there is still some resultant state, and it can in most cases still be described by the verb have (i.e., to have the right to do something, to have information). Also, the sense of transfer is still clearly present. In relation to this variant, (27) then displays a further extension, (i.e., weakening) of the meaning of give in that these examples only describe a beneficial effect for the recipient, but drop the notion of control over these effects altogether. (27)

Kim gave Lee emotional support give someone one's regards

Another aspect of semantic weakening that occurs with these examples is that it is hard to say what kind of resultant state could come about by these

342 Miriam Butt and Wilhelm Geuder actions. Turning finally to (28), the light verb use can now be seen as another, even further extension of the previous sense: (28)

a. b.

give someone a kiss / a push / a punch / a nudge / a hug give the car a wash, give the soup a stir

(28) seems to describe merely the exertion of some effect on the "recipient". The (28a)-examples are somewhat intermediary between (27) and the (28b)-cases: in (28a) there is at least a visible movement that can be construed as the "transfer"; in (28b) there is no such concrete directed movement. Here, the mere application of an action to an entity is all that is left of the transaction meaning of give. While there is all this variation with respect to the kinds of results and THEMES, a number of meaning features appear to remain constant. The first such feature is agentivity: The subject (in the ditransitive construction) always denotes a sentient being that performs an intentional act. There is not much variation in the choice of the recipient either: As long as some entailment of control is present as part of a resultant state, the recipient will have to be a sentient being; only when the result has been reduced to just the exertion of an effect (as with light verb uses) an extension to inanimate concrete objects becomes possible (e.g. the car in (28b)). A further trait that is present in all variants is a vague sense of emission. This commonality is noted in Newman (1996) who uses it to explain the unacceptability of sentences like */ gave the ball a good catch as opposed to I gave it a good throw. We can further observe that the THEME argument of the concrete, prototypical use of give does not possess the property of being mapped onto the progressing event, i.e., this argument role is not "incremental" (cf. Dowty 1991). This means that the boundedness of the event does not vary along with the boundedness of the THEME argument, rather, the verb meaning always enforces a specific reading on the amount of matter that is transferred, even if the NP that realizes the THEME consists of a mass noun. So for example, in the sentence I gave chocolate to the child, we do not get an unbounded event of giving, going on as long as there is chocolate. Instead, there is one single event of giving which is construed as involving a specific amount of chocolate. For the light verb uses with their event nominals in object position, we can make the related observation that they always involve count nouns and that the resulting interpretation is a bounded event of pushing, hugging etc. (cf. the examples in (28) above). This constant boundedness of the events described by give could be

On the (semi)lexical status of light verbs 343

explained as resulting from the fact that the THEME argument is always forced into the reading of a specific quantity. If an event-denoting expression is put in this place, it would be plausible that it will then be required to be a bounded event, as the bounded-unbounded distinction has been argued to be the verbal counterpart to the nominal count-mass distinction (Krifka 1992). It may be that the event nominalizations that are used here would not allow for a mass noun interpretation anyway, but it is certainly not a coincidence that we end up with a construction that only introduces bounded events. We return to the issue of the interpretation of events that are thus given in section 4.6 in connection with issues of semantic representation. To summarize this somewhat lengthy discussion, the following picture has emerged: The array of different uses of give in English yields the impression that the semantic features present in light verb constructions can be understood as continual extensions from the prototypical concrete meaning of give, with the extension of the meaning being correlated with the increasing abstractness of the referents filling the THEME argument slot.8 This would seem to foster an analysis that derives light verbs exclusively by means of mechanisms of lexical sense extension (e.g. like metaphor or generalization). This impression, however, is due to the fact that both lexical and light verbs occurred in the very same grammatical construction.

4.3.

The light verb 'give' in V+Vconstructions

As has been argued in section 2, light verbs can be identified as a distinct syntactic class in Urdu. Therefore, the existence of light and full verb uses of the same verb would not seem to be a matter of a purely lexical-semantic meaning extension as it does in the English V+N constructions. In this and the following subsections, we examine whether the constructional difference really entails a fundamental semantic difference between these two types of light verbs. While Urdu complex predicates do not give rise to a continuous spectrum of readings that are more or less concrete, there is still a whole range of different light verb meanings.9 Consider the following collection of data, taken and adapted from Hook (1974).

344 Miriam Butt and Wilhelm Geuder (31)

us-ne prezident-ko xat Pron.3.Sg=Erg president=Dat letter.M.Sg.Nom lit dii-yaa write give-Perf.M.Sg 'He/She wrote a letter to the president.'

(32)

us=ne bHikhaarii=ko do pese de dii-ye Pron.3.Sg=Erg beggar=Dat two paisa.Pl give give-Perf.Pl 'He gave the beggar two paisa.' (paisa is a type of money)

(33)

laksman-ne surpanld1aa=ko bhag-aa dii-yaa Laxman.M=Erg Shurpanka=Acc run-Caus give-Perf.M.Sg 'Laxman drove off Shurpankha.'

(34)

us-ne vusse-me aa kar Pron.3.Sg=Erg anger.Obl=in come having kitaab zaamiin-par patak dii book.F.Sg.Nom floor.F=on throw give.Perf.F.Sg 'In a fit of rage he dashed the book to the floor.'

(35)

us=ne gilaas banaa dii-yaa Pron.3.Sg=Erg glass.M.Sg.Nom make give-Perf.M.Sg 'He made a/the glass.'

(36)

us=ne gilaas tor Pron.3.Sg=Erg glass.M.Sg.Nom break 'He broke a/the glass.'

(37)

us=ne bhuul=se gilaas Pron.3.Sg=Erg forget=Inst glass.M.Sg.Nom tor dii-yaa break give-Perf.M.Sg 'He broke a/the glass by accident.'

(38)

vo ghoiaa laa-yaa Pron.3.Sg.Nom horse.M.Sg.Nom bring-Perf.M.Sg aur hamaare=upar caf-aa dii-yaa and Pron.l.Pl=on climb-Caus give-Perf.M.Sg 'He brought his horse right on top of us.'

dii-yaa give-Perf.M.Sg

On the (semi)lexical status of light verbs 345 (39)

baap=ne apne bete=ko father.M.Sg=Erg self.Obl son.M.Sg.Obl ghar=se nikaal kar baahar kar dii-yaa house.M.Sg=Inst remove having out do give-Perf.M.Sg 'The father turned his son out of the house.'

(40)

tum=ne in baccö=ko rul-aa kar Pron.2.Sg=Erg Dem.Obl.Pl child.Pl.Obl cry-Caus having chor dii-yaa, ab xus ho leave give-Perf.M.Sg, now happy be.Pres.2.Sg 'You've left these children in tears, are you satisfied now?'

(41)

bhuul=se mufe apnaa forget=Inst Pronl.Sg.Dat self.M.Sg.Nom sahii naam bataa dii-yaa true name.M.Sg.Nom tell give-Perf.M.Sg 'He inadvertently told me his real name.'

(42)

kisii=ne batuaa someone=Erg wallet.M.Sg.Nom 'Someone lost a/the wallet.'

(43)

tum=ne kuuraa us=ke Pron.2.Sg=Erg garbage.M.Sg.Nom Pron.3.Sg=Gen.Obl sir=par phek dii-yaa head=on throw give-Perf.M.Sg 'You threw the garbage on his head.'

id1ο dii-yaa lose give-Perf.M.Sg

This array of examples clearly shows that Urdu allows for many more combinations of event descriptions with give than the corresponding English light verb. As in English, however, all uses entail that the event described by the main verb is bounded. Consider (44), for example, which can only be interpreted under a scenario in which one singing event took place and was completed, to be followed by another singing event. Under no circumstances could the second half of the sentence refer to the same stretch of a singing event as the first half of the conjunction.

346 Miriam Butt and Wilhelm Geuder

(44)

naadyaa=ne gaanaa gaa dii-yaa Nadya.F.Sg=Erg song.M.Sg.Nom sing give-Perf.M.Sg aur phir us-ne aur bhii gaa-yaa and then Pron.3.Sg=Erg too sing-Perf.M.Sg 'Nadya sang a song (completely) and then sang some more.'

This often repeated observation that light verbs appear to give rise to completive readings is what has prompted the conception that Urdu light verbs are aspectual markers. However, another telling observation about the examples in (31)-(43) is that they all imply some sort of agentivity - though sometimes in a weaker sense than that found with English. With the Urdu light verb 'give', agentivity can be reduced to something like "responsibility". This includes unintentional causation, as shown by (37) and (41), as well as the case of (42), where the effect of adding the light verb 'give' (de) is to ascribe the agent some ultimate responsibility for the loss (otherwise not an intentional act). Concerning Newman's (1996) proposal that 'give' also requires an "emission" schema, we can add that we are not aware of any examples that involve combinations as in the unacceptable English sentence give the ball a catch adduced by Newman. So an overall directedness of the interaction appears to apply to Urdu as well. The meaning features of boundedness and agentivity can thus be found to make up a constant core of the light verb's meaning. However, other typical features of the verb give seem to be defeasible. Some examples imply the presence of a recipient ((31), (32), (41)) while others are merely concerned with the exertion of an unspecific effect on a target ((33), (34), (38), (39), (40), (43)). Still others fail to imply even this ((35), (42)). With respect to this feature, the interpretation depends on the kind of event described by the accompanying full verb. If the meaning of the full verb suggests the presence of a recipient or at least a particular (possibly beneficial) effect, then the light verb 'give' appears to enhance this meaning component. If, however, such a trait is absent from the full verb, the light verb does not enforce it on the construction. Something similar can be said about the recurring intuition that give-light verbs also express a quality of "forcefulness". This can indeed be claimed for some of the examples ((33), (34), (36) through (39), and (43)), but not for the others. So again, we would assume that the meaning of the full verb decides on whether this trait is "activated" in a light verb construction. This calls for a treatment of the light verb's lexical meaning in terms of an underspecified lexical entry, the details of which we have to leave for future research. The important thing is, however, to note that we are really dealing with default semantic

On the (semi)lexical status of light verbs 347

features that reside in the light verb; although the "activation" of such meaning components may be prompted by the main verb, the features at issue do not originate from the main verb. Consider the minimal pair in (45) and (46), in which the use of 'give' is interpreted as expressing more force than the use of le 'take' (which appears to be the least meaningful and therefore the most unmarked of the "completive" light verbs; see Butt 1995 for some discussion). Thus, the combination of light verb and main verb interacts to produce or activate differing components of meaning. In the example (46) below, we can see an instantiation of the "emission" schema that is a generalized extension of the meaning feature that, in giving, an agent produces something out of his own sphere of control. (45)

us=ne dusman=ko paanii-me Pron.3.Sg=Erg enemy.M.Sg=Acc water=in dibaa lii-yaa depress take-Perf.M.Sg 'He/She drowned the enemy in the water.'

(46)

us=ne duSman=ko paanii= me Pron.3.Sg=Erg enemy.M.Sg=Acc water=in dibaa dii-yaa depress give-Perf.M.Sg 'He/She drowned the enemy in the water (forcefully).'

In comparison with English give, the Urdu V+V construction with 'give' can be seen as exhibiting meaning features that are similar in a very general sense. The difference lies in the observation that while 'give' in the V+V constructions retains as a core meaning such components as agentivity and completion, its other typical features, such as "transfer" or a beneficial component, are defeasible features - their existence has to be inferred in accordance with the specific context, or otherwise cancelled. So Urdu light verbs have a meaning that is very flexible, but the meanings are of the same overall type as the meaning that has been found for the English V+N construction.

348 Miriam Butt and Wilhelm Geuder

4.4.

The argument structure ofN+ V and V+ V constructions in comparison

In English the light verb must be assumed to contribute to the marking of the subject as an agent, because the agent is the grammatical subject of the VP headed by give, and give in turn expresses a notion of agentivity. Something similar seems to hold for the Urdu V+V constructions: as we saw in section 2.3.2, light verbs such as 'give' have effects on the case marking of the subject. In particular, 'give' selects an ergative case and as the ergative in Urdu appears to be connected with volitionality/conscious control (e.g., Mohanan 1994; Butt and King 1991), this confirms the existence of a selectional relation with regard to agentivity in this construction as well. With regard to the role of indirect objects, a difference between the two constructions is apparent. If we take the English give construction at face value, the "recipient of the giving" (speaking metaphorically) can be identified with the patient of the event denoted by the nominal, as in (47), where the recipient of give can only be the car, which in turn can only be the patient of the event a wash: (47)

He gave the car

a wash

REC(give) THEME(give) PATffiNT(wash)

The V+V constructions, however, clearly cannot be analyzed along these lines. Consider the Urdu example (48). (48)

tum=ne (us=ke-liye) skuutar Pron.2.Sg.=Erg (Pron.3.Sg=for) scooter.M.Sg.Nom cto dii-yaa wash give-Perf.M.Sg 'You washed the scooter (for him).'

The attempt at a schematic interpretation in parallel to (47) above would result in something like (49): (49)

(to-him)

scooter

REC(wash) PATffiNT(wash) REC(give)?

wash

give

On the (semi)lexical status of light verbs 349

Here, the beneficiary is not linked to the patient of 'wash'; it is a recipient/beneficiary that can be optionally added as an adjunct argument. The direct object 'scooter' can only be construed as the patient of washing. In other words, the direct object position of the complex predicate construction (48) cannot be thematically linked to the light verb 'give' (it cannot be understood as bearing a theme relation to 'give', whatever its exact meaning). The absence of a beneficiary argument contributed by the verb 'give' is precisely what allows the introduction of an optional beneficiary adjunct. This stands in marked contrast to the English V+N or Urdu N+V constructions.

4.5.

A difference in structural pedigree

If there is no thematic relation between the light 'give' and the direct object in the Urdu V+V construction, we are forced to the conclusion that the THEME argument slot of 'give' has been lost in this construction. This is an astonishing finding, for in spite of all variation that we have observed before, the THEME was never found missing. Note, moreover, that the absence of the theme slot is a distinctive feature of just the V+V construction, for, even in Urdu, it still appears in the N+V and permissive complex predicates. In the permissive (20), repeated below, the infinitive 'write' can be analyzed as a theme argument of 'give' (due to its nominal inflection). In the Urdu N+V complex predicate construction (21b), also repeated below, it is the nominal ('broom' in this example) that can be analyzed as functioning as an argument, just as in English. (20)

naadyaa=ne yaasiin=ko xat Nadya.F=Erg Yassin.M.Sg=Dat letter.M.Sg.Nom Zi** -ne dii-yaa write-Inf.Obl give-Perf.M.Sg 'Nadya let Yassin write a/the letter.'

(21)

b.

naadyaa-ne kamre=ko jhaamu Nadya.F=Erg room.M.Sg.Obl=Dat broom.M.Sg.Nom dii-yaa give-Perf.M.Sg 'Nadya swept the room.'

350 Miriam Butt and Wilhelm Geuder

The reason behind these different argument structures can be traced back to different diachronic sources, and to different synchronic connections that can be drawn between the complex predicates and other constructions in the language. The permissive use is most probably connected to, or derived from, constructions with an embedded infinitive such as the one shown in (50). (50)

naadyaa-ne

us=ko

jaa-ne

delf-aa

Nadya.F=Erg Pron.3.Sg=Acc go-Inf.Obl see-Perf.M.Sg 'Nadya saw him/her leaving.' The N+V complex predicate is connected to a typical object-verb construction and presumably arose from such a syntactic constellation. These two Urdu complex predicates and the English V+N construction can thus presumably be traced to constructions in which the main predicate is embedded in the argument frame of the verb that is now analyzed as light. However, this is not the case for the Urdu V+V complex predicate. Both synchronically and diachronically, the V+V complex predicate is connected to an adverbial modificatory clause, not an embedding construction. Consider the modern Urdu example in (51), which in principle has two readings. One is the complex predicate reading discussed in this paper. The other is a sequential reading in which the embedded adverbial clause 'having V-ed' modifies the main event. (51)

naadyaa

aa

ga-yii

Nadya.F.Nom come go-Perf.F.Sg 'Nadya has arrived.' 'Having come, Nadya went.' This adverbial construction can be traced back through the ages to an indeclinable absolutive construction formed with -tvä or -ya. Consider the ambiguity present in the in the Old Indo-Aryan (=1200-600 BCE) Sanskrit example in (52), and the similarity to modern day complex predicates in the Sanskrit and Pali (=600 BCE - 1000 CE) examples in (53) and (54), respectively. (52)

indram

ärabhya

cara

Indra-Acc grasp-Gd go-Imp.2Sg 'Having taken hold of Indra, move!' 'Keep yourself to Indra!'

(Sanskrit, Tikkanen 1987: 7)

On the (semi)lexical status of light verbs 351

(53)

yathä mandükä utplutyotplutya as frog.Pl.Nom up-spring-Gd-up-spring-Gd gacchanti go.3.Pl.Pres.Ind 'Just like frogs move by jumping, jumping ...' ['Just like frogs jumping...'] (Mähäbhäsya, Tikkanen 1987: 21)

(54)

assamapadam änetvä aggim katvä adäsi hermitage.Acc lead.Gd fire.Acc make.Gd give.Impf.3.Sg '... brought her to his hermitage and made a fire for her' ['having brought (her) to the hermitage, made a fire (for her)'] (Pali, Jätaka Tales, Sri Lanka, Hendriksen 1944: 134)

The interpretation of these gerundive constructions in the ancestor languages of Urdu/Hindi was tricky (see Tikkanen 1987 for a detailed discussion): The modificatory -tvä/-ya clause could indicate an event preceding or following the main clause, but it could also indicate an event which temporally overlapped that of the main clause. There is furthermore some indication from control and the placement and interpretation of negation that these -tvä/-ya clauses together with the matrix verb formed a single domain of predication: a property that is characteristic of complex predicates. The stark difference between N+V and V+V constructions in terms of argument structure properties and diachronic origins would appear to preclude an analysis which sought to capture both N+V and V+V constructions within the same type of explanation. On the other hand, it is hard to see what an alternative explanation could be for the rise of light give in V+V constructions. It must be noted that the constructions in (52)-(54) do not yet bridge the gap between adverbial constructions with -tvä/-ya and complex predicates involving give as a light verb: The problem is precisely that light give suddenly lacks the theme argument position, whereas the verb go in (53) is a saturated predicate. No meaning variants of give have so far been found that lacked a theme argument, and moreover, the Urdu light verb variant was found to express semantic features not qualitatively different from the ones in the English V+N construction. So the absence of the theme slot is intriguing. It is tempting to speculate that the open argument slot for give might originally have been saturated by an implicit argument, for instance by an empty anaphor in a construction with a resumptive function. If such an account could be given, the event argument

352 Miriam Butt and Wilhelm Geuder

would again enter into the theme slot of the verb give, though not being represented by a syntactic constituent. Example (54) might be amenable to such an analysis; however, we are not able to really substantiate any such speculations. To conclude, V+V light verb constructions have been shown to be markedly different from N+V ones in terms of argument identification and also in terms of their history. Still, there is a remarkable lack of semantic alternatives, since the diachrony of the V+V construction does not indicate any alternative path of lexical development from the full verb to the light verb semantics of give. The construal of the event as a theme argument of give continues to be the only way to connect full and light verb usage that we know of. This state of affairs will be found to carry over in some sense to the synchronic semantic analysis of the construction that we turn to in the next section. There, we show that the semantic analysis of light verb constructions of both kinds forces us to assume a similar kind of lexical meaning in spite of the different constructions.

4.6.

The transfer metaphor and the semantics of V+N/V+ V constructions

It is now time to come back to the representation of light verb meanings. The discussion in section 4.2 already hinted at the idea that light verb meanings could be connected to the concrete, prototypical meaning of give by way of a very vague and abstract "transfer" schema as the trait that they still share with the full verb meaning. In fact, works in the cognitivesemantic framework explicitly propose such an abstract "transfer schema" as a semantic analysis of the English light verb construction (Newman 1996; cf. also Norvig and Lakoff s 1987 paper on the semantically related case of the light verb take). The main point about such schemata seems to be that the event, denoted by the nominalization in object position, can be construed as a transferred entity in a metaphorical sense. This is descriptively appealing - as far as the English V+N construction is concerned - but we do not think that invoking the notion of metaphor constitutes a complete analysis in and of itself. As an illustrative example of the type of analysis which involves increasing abstractness and a metaphorical use of the notion of transfer in order to arrive at a complex predicate reading, consider a simple event schema of "transfer" (cf. the Langacker-style representation used in Newman 1996). Example (55) shows the transfer schema in which a theme

On the (semi)lexical status of light verbs 353

is transferred from an agent to a recipient. There is exactly one event of giving which contains all of the elements in the diagram. (55)

[Agent]

Ο

[Recipient]



Q

[Theme:] [Concrete / Abstract Entity] This schema could also be invoked if the object of give is replaced by an abstract entity; the shared schematic structure in both the domain of spatial transfer and in the domain of abstract transmission is what is perceived as the metaphorical link. This idea of "metaphorical transfer" might work quite well for examples such as giving someone one's regards. However, the cases in (28), repeated here in (56), are not exactly of the same kind: (56)

a. b. c.

to give someone a kiss to give someone a hug to give the soup a stir

If these, too, were taken as instances of metaphorical transfer, they would be subsumed under the same schema: (57)

[Agent]

Π

[Recipient]



Π

[Theme: Event] However, what kind of conceptualization could the metaphor be said to contribute here? The examples in (56) seem to mean little more than just 'to

354 Miriam Butt and Wilhelm Geuder

kiss/hug someone', 'to stir the soup', etc. So, obviously, the schema must be interpreted as saying that the kissing (etc.) itself is metaphorically characterized as an act of giving (something). But this points to a fundamental difference in the interpretation of the pictorial schemata (55) and (57). The problem lies in the referential distinctiveness of the verb give and the event nominalization in its object position: The full verb give, taking an object that denotes a thing, asserts that this thing is transferred, and that this transfer constitutes an event. Such an analysis can be carried through as long as the object denotes a distinctive (possibly abstract) entity. This requirement causes problems. For example (56a) to fit into this schema, we would have to assert that there is an entity "kiss" which exists independently besides the event of its transferal. It is hard to see whether this view makes sense, and which criteria could be used to decide the question: Is "a kiss" one of the things that are involved in an event of kissing, or is a kiss just tantamount to such an event? Perhaps "a kiss" could be conceived of as a more narrow part of the event that in turn would be denoted by the verb kiss (the culmination part, as it were). While the issue may be difficult to resolve with respect to give a kiss alone, there is no such uncertainty with example (56c). "A stir" cannot be thought of as an entity that exists independently of the event of "stirring", nor could it be a subpart of this event. They are just the same. Clearly, however, one criterion for achieving a light verb reading is that it must occur in conjunction with a true event description. In the case of V+N complex predicates the event fills the object position. As soon as we assume that an event occupies an argument slot, an important issue with regard to event interpretation arises. The point is that the notion of a concrete transfer in space is a relation between individuals, which itself constitutes an event. On the other hand, while the "giving of an event" can still be represented as a relation between three entities, this representation does not introduce an event of its own. This point can be made clearer if we consider the following parallel cases: Events can entail movement, but obviously they cannot themselves be moved, not even in a concrete, spatial sense. Whenever there is movement in connection with an event, this appears as a component of the event description itself. So, the example They danced into the garden could be conceived of as an event of dancing that moves into the garden. But it still behaves as one and the same event; the movement of the dancing event does not yield a second distinct event with the first as one of its arguments. Likewise, note that the location of a thing generally constitutes a state, but the location of an event does not constitute an additional state on top of it: Localization again appears as just

On the (semi)lexical status of light verbs 355 one of the constituent features of the event. If even the spatial movement of an event does not give rise to a new event of its own, then, a fortiori, a metaphorical transfer of an event does not either. The pictorial schema in (57), then, does not seem very helpful, for it cannot represent the fact that the whole transmission schema and the Theme that occurs inside it must be taken to denote the very same thing. As a more perspicuous formulation, we propose (58), using χ as a variable restricted to the sort of "things/individuals" and e as a variable ranging over "events". The detached first variable represents the position of the referential argument, which is missing in the second case: (58)

a.

give (full verb):

R(e) (x^xj.x^)

b.

give (light verb): R (xj,xj,e)

To put it very simply, the event variable occurs in different "roles" here. In notional terms, (58a) represents an event concept, which can be used to refer to an event, but (58b) represents a pure relation between individuals and events. It is of a type that looks rather similar to thematic role predicates as they occur in event semantics. The view that emerges here thus seems to be somewhat similar to the stance taken in Higginbotham (1989), who adumbrates an analysis that light verbs express a thematic role, e.g. like agent-ofle). However, the notion of thematic role is not really applicable. There are at least two clear differences: First, the meaning of 'give' cannot be reduced to a single one of the known thematic roles, such as agentivity; it comprises a whole bundle of other features of the same semantic type (like, possibly, application of force or a beneficiary effect etc). So "thematic role" could at best be understood in the sense of Dowty's (1991) proto-roles, i.e., as entailments from the verb meaning that concern specifically one participant, but maybe also a whole number of them. The second crucial difference between light verb entries of the type (58b) and thematic roles is that thematic roles serve to index arguments by bundling into a single label semantic features that are already inherent in the verb meaning. Light verbs, on the other hand, can in principle introduce new meaning features, although they can perhaps also be redundant or serve a purpose that we have earlier referred to as the mere "enhancement" of meaning features. From this description, it rather seems that light verbs are most similar to modifiers. The easiest way to implement this idea would be to render the light verb meaning in the format of the neo-Davidsonian representation for manner adverbs (cf. Parsons 1990), i.e., as a predicate of events that is

356 Miriam Butt and Wilhelm Geuder

interpreted in conjunction with the main event description.10 The contents of such modifiers can still be called schematic features of events, i.e., force transmission, benefaction etc. Light verbs are thus invested with a status that is quite different from the contents of the corresponding full verb in that the modifiers are inherently specified for a type of construal that makes them unable to denote classes of events on their own (as we put it above, they do not express an event concept). Schematically, let us postulate an event predicate GlVE-TYPE(e) to cover the intended modifier meaning and apply it to a concrete example in (59): (59)

a.

John gave the car a wash wash(e)(John, the car) & GlVE-TYPE(e)

- where the contents of GiVE-TYPE(e) might be expanded as follows: b.

GlVE-TYPE(e) or:

= e has beneficial effects on THEME(e); = e involves the force transmission pattern AGENT(e)-THEME(e)

etc. As has already been pointed out, light verbs show an enormous flexibility of meaning. The precise semantic interpretation is determined through contextual factors, which are adduced via some mechanism of non-monotonous inference. It seems that such an analysis runs into a compositional problem with respect to English, since the "modifier" is projected as the head of a VP and the "modified" is its complement. So our claim must be put more modestly - the lexical features of the light verb in English are comparable to the type of lexical contents found in modifiers, while the construction introduces additional complexity.11 We want to leave this question for further research, since the same problem will not arise for Urdu. Let us just point out that the mismatch between a verbal complementation construction and an interpretation that identifies the two event variables is really an independent problem: A very similar issue appears in the interpretation of socalled cognate object constructions, as pointed out by Mittwoch (1998) (though with the reverse emphasis).12 For Urdu, the complex predicate construction appears to be a quite straightforward implementation of a semantic conjunction. So we claim that despite the overt syntactic difference in the construction of the complex predicates (V+N vs. V+V, complementation vs. a sequence of verbs), the

On the (semi)lexical status of light verbs 357 semantics of the constructions are to be rendered quite similarly. Let us go back to the examples (42) and (43): (42)

kisii=ne batuaa someone=Erg wallet.M.Sg.Nom 'Someone lost a/the wallet.'

to dii-yaa lose give-Perf.M.Sg

(43)

tum-ne kuuraa us=ke Pron.2.Sg=Erg garbage.M.Sg.Nom Pron.3.Sg=Gen.Obl sir=par phek dii-yaa head=on throw give-Perf.M.Sg 'You threw the garbage on his/her head.'

Similar representations as above can be given for these cases. Example (43) for instance would yield: (60)

a.

throw(e)(you, the garbage) & GOAL(e, on his/her head) &

b.

GIVE-TYPE(e) GlVE-TYPE(e) = e involves the force emission/transmission pattern AGENT(e)-THEME(e)

It can be seen that the modification expressed by light verbs can be very weak, and at times it seems almost redundant. In principle, this would not seem problematic, since redundancy does not block the use of modifiers in other cases either (for instance, there is no less redundancy in adverbial cases like He rushed quickly to the door). What this points to, however, is that light verbs may receive a special function, namely as a device for classifying events (rather than elaborating them). The extent to which this is the case depends on the degree to which light verbs are entrenched in the grammatical system of a language. In section 5, we shall resume this topic.

4.7.

Conclusion

The connection between light and full verb meaning does consist of a kind of schematic generalization of meaning features that are present in the full verb use as well, i.e., there is indeed a continuous extension from full to light verb meanings, as evidenced by the case of the V+N light verb construction in English. At the same time, even in English, the semantic properties of the construction require an analysis which involves identifica-

358 Miriam Butt and Wilhelm Geuder

tion of the events expressed by V and Ν. This results in a type of element that is crucially different from the usual denotation of verbs: Light verbs make no independent reference to a class of events. Representations such as in (55) and (57) that exclusively rely on image schemata obscure this crucial difference: In one case, the circles and arrows constitute an event, and in the other they do not. This is to say, the (metaphorical) "transfer" and the "transferred event" collapse into one and the same thing, and the transfer schema boils down to an elaboration of the action which is, metaphorically, "transferred". We instead propose a semantics for light verb constructions that treats light verbs as a kind of modifying elements: They contribute additional semantic features to the event description. This mirrors the intuition in the syntactic literature that the light verb can only be licensed in combination with a main verb. While it is true that V+N and V+V constructions behave differently, the differences do not point to conceptual-semantic differences between these two occurrences of light verbs; the differences that can be observed are predictable from the constructions in which these lexical elements occur. 5.

Light verbs and complex predicates: Extending the picture

In this section, we move on towards a general characterization of light verbs. We want to suggest that the findings of the case study on give in the previous section are part of a larger picture, in which light verbs come out as a special class of elements that are distinct from lexical verbs and auxiliaries or aspectual categories. The discussion of V+N vs. V+V structures has touched on the issue of differences in the grammatical system of different languages. The status of light verbs is clearly influenced by the degree to which light verbs are made obligatory in the grammar of a language. While in English the employment of light verbs appears to have more the status of a primarily stylistic device, the use of light verbs is very entrenched in Urdu. It is maybe for this reason that they take on weaker and more general meanings there. In this section, we want to adduce the case of the Australian language Wagiman as described in Wilson (1999), and thus set our ideas on the interpretation of light verbs in a larger perspective. Wilson conveniently (for us) provides a solid identification of a syntactic class of light verbs on a par with the work done in the first half of this paper. He furthermore combines this syntactic identification of light verbs with a detailed lexical-

On the (semi)lexical status of light verbs 359

semantic analysis of the relations within complex predicates. One point of special interest for us is that, in Wagiman, complex predicates are even more pervasive than in Urdu. The use of full verbs as the only predicator in a sentence is quite restricted. There seems to be a ban on using a single full verb as the inflected predicate of the clause; one must either use a second, possibly light, verb that carries the inflection, or the verb must be affixed with a special derivational element in order to be allowed as the single inflected item. The main point of this section will be to situate the semantic analysis proposed in the previous section within a typology of the semantic fünctions of complex predicates. In this, we build on the semantic analysis offered in Wilson (1999) for the Wagiman cases. However, it must be emphasized that what we present here is actually a reanalysis of Wilson's data. Wilson organizes his treatment in terms of the different types of "full verbs" (i.e., those in first position, not those in the second slot in which the light verbs appear). Furthermore, Wilson assumes that all complex predicates can essentially be analyzed by way of a functor-argument relationship. We have not taken this route in the previous section, and we also believe that it is not the optimal treatment for many of the Wagiman cases. From Wilson's data, we extract the following three main semantic types of complex predicates, which we illustrate and analyze in the next few sections. The analysis we have proposed for the English and Urdu cases so far falls under Type C, event modification. However, as the typology extracted from Wagiman shows, this is not the only type that can be associated with complex predication (which is even true for Urdu with respect to other types of complex predicates that have not been explicitly addressed in this paper). (61)

5.1.

Types of Complex Predication A: Event summation B: Event augmentation C: Event modification / Classification

Type A: Event summation

A main characteristic of the event summation type is that there is no asymmetry apparent from the semantics of the construction, and both verbs seem to be used in their full and literal meaning. A Wagiman example is given in (62). This type is absent in Urdu (and English).

360 Miriam Butt and Wilhelm Geuder

(62)

bing-nga ga-ya be glowing-Asp 3.Sg-go.Pres 'it's going along glowing'

(Wilson 1999: 77)

We call this type event summation because the construction connects two independent event descriptions. The result is a complex event that unites two single independent events to yield a collective entity - so note that this sum of events does not constitute a complex event of the type that is normally the result of lexical decomposition. Eckardt (1998) proposes that the proper interpretation of conjunctions of event descriptions is by way of summation, not by the logical (i.e., Boolean) and. The construction in (62) seems to constitute a similar case. Eckardt (1998) proposes the formalization in (63a) with Θ* as the summation operator (connecting properties of events). This can be rewritten as in (63b) by using the related operator Φ, connecting individual events. (63)

a.

[Xej. glowing(ej)] Θ* [Xe2. going-along(e2)]

b.

Xe* [3e 1 (glowing (e^) & 3e 2 (going-along(e2)) & e* = e j © e 2 ]

5.2.

Type B: Event augmentation

Event augmentation is actually the well-known case in which a second predicate adds to the first to yield a complex event in the sense of lexical decomposition theory. That is, collocations of predicates that take a causative, inchoative, or resultative meaning fall under this heading. Examples from Wagiman are provided below. (64)

dewh ngaha-ny-ma nibulin be white.pfv l.Sg.become.ppfv-ds eye(=face) 'my face went white (in fear)' (Wilson 1999: 74)

(65)

dajup nge-ge-ng gahan ngal-door? be-shut.pfv 2.Sg-put-ppfv that-fem door 'did you shut that door?' (Wilson 1999: 78)

Sample Analysis of (64):13

On the (semi)lexical status of light verbs

(66)

a. b.

361

dewh ngaha-ny-ma nibulin bewhite.pfv l.Sg.become.ppfv-ds eye(=face) XP.BEC(e)(P) [Xs.white(s)(the face)]

A decisive feature of the augmentation construction is the existence of a functor-argument relation between the two verbs, which can change the sort of event expressed by the lexical verb, e.g. from a state to a transformation event in (64)-(66).

5.3.

Type C: Event modification

The third type is marked by the fact that two event descriptions are tied together that are not independent. Consider the following example: (67)

liri-ma ga-di-n lamarra swim-asp 3.Sg-come-pres dog 'the dog is coming swimming'

(Wilson 1999: 101)

In (67), light verb and lexical verb can be seen to stand in a modifying relationship. To express this, we can proceed in the same way as we did when proposing a translation of 'give' in terms of GIVE-TYPE(e). So the format of the representation will be as in (68). (68)

swim(e)(the dog) & TO-CENTER (e)

This is admittedly a very simplified notation for paths involved in an event. Parsons (1990) analyzes all directional PP's in this format. The modifier use of come would semantically be on a par with a directional PP, i.e., a modifier that specifies the path parameter of a dynamic event. The point is that the representation cannot simply consist of a conjunction of 'swim' with the main verb reading of 'come'; we need a special link that enables a modificatory relationship. The light verb 'come' - and this shows the difference to type A - specifies one of the semantic dimensions of the event concept that is provided by the main verb, namely the path of movement. Thus we have two properties of the same coherent event, in distinction to type A which involved a collection of two events as parts of a sum event (correspondingly, a logical and appears in (68), in distinction to the summation operator of (63)). What is more, we know that swim as a manner of movement verb entails the existence of a path, while going along

362 Miriam Butt and Wilhelm Geuder and glowing in (62) happened independently without making entailments on each other. In section 4, we analyzed 'give' as an instance of a semantically weak modifier. A logical extension of this idea would be a modifier that is semantically void and, ultimately, just serves as a device to classify event types. This case is more or less hypothetic; i.e., we do not know whether it exists in a pure form. However, some candidates can be found for such a limiting case of modification: (69)

(70)

wal yaha-ny lagiyi grow-pfv 3.Sg.become-ppfv body 'her body has grown'

(Wilson 1999: 80)

gartgart-da ga-ya yimbanma laugh-asp 3sg-go.pres always 'he is always laughing'

(Wilson 1999: 97)

Of course, it has to be decided on a case by case basis whether the light verb is making any contribution at all, and the information on the interpretation of the Wagiman sentences that is made available in Wilson (1999) is somewhat limited for these purposes. It seems to us that the verb 'go' in (70) is functioning more as a general marker for an activity and is in that sense just like a classifier for the verb 'laugh'. This phenomenon of "empty" modifiers seems to be absent in Urdu. This is probably due to the fact that the use of complex predicates in Urdu is not compulsory, so the need to use light verbs without a meaning contribution does not arise.

5.4.

Implications for a typology of light verbs

Wagiman is the case of a grammatical system in which only a limited set of verbs are allowed to appear in the inflected slot. The other verbs have to enter into a complex predicate construction with an inflected verb. This system gives rise to a particularly broad range of complex predicate uses, and it therefore contributes to our understanding of what characterizes a light verb. The first point that emerges from an inspection of the data presented above concerns the question of whether light verbs have to be viewed as constituting a syntactic class that is distinct from fully lexical verbs. This would be denied for the case of English, but the Urdu V+V construction

On the (semi)lexical status of light verbs 363

admitted for such an interpretation. In Wagiman, we now see the case of a V+V construction in which full verbs are licensed in the same slot as "light" verbs are; compare the meaningless variant of 'go' in (70) ('to go laughing') and the lexically intact variant in (62) ('to go along glowing'). In view of the range of constructions found in Wagiman, we conclude that the continuous meaning extension from full to light verb is possible in V+V constructions, too. So the difference between light and full verb uses is once again reduced to a pure lexical-semantic gradation. This is in line with our thesis that V+V complex predicate constructions do not imply a lexicalsemantic status for light verbs that is different from those in V+N constructions. The comparison between English, Urdu, and Wagiman thus leads to the general conclusion that the existence of syntactic distinctions between lexical and semi-lexical (i.e., light) verbs can vary across languages, whereas we hypothesize that the distinction between light verbs and auxiliaries is an invariable consequence of the semi-lexical status of light verbs. This leads us to a second point: The typology of the semantics of complex predicate constructions suggests the possibility of coming up with a more precise notion of light verb than has been possible before. Type A definitely cannot be called as "light" as Types Β and C, because it involves independent event concepts. Type Β is markedly different from C in that it augments the event in terms of aktionsart. For example, there is a minimal contrast to be noted between Wagiman complex predicates such as (65) and certain superficially similar Urdu V+V constructions. Consider (65) from Wagiman (repeated here): (65)

dajup nge-ge-ng gahan ngal-door? be-shut.pfv 2.Sg-put-ppfv that-fem door 'did you shut that door?' (Wilson 1999: 78)

The first verb is the stative predicate 'to be shut', augmented by ge ('put'). So the complex predicate construction serves to add the accomplishment part. However, examples such as Urdu sulaa diyaa 'sleep-CAUS put' are not to be analyzed as event augmentation because here the verb 'sleep+CAUSATTVE' already denotes an accomplishment, so the light verb does not change the aktionsart. For another example, note that there is no need to call the Wagiman Type-B verb 'become' in (66) ('be-white become') light, because, actually, in this augmenting function it constitutes a semantically stronger variant as compared to the merely modifying use in (69) ('grow become').

364 Miriam Butt and Wilhelm Geuder

Therefore, it would make sense to reserve the term "light verb" for the modifying type C alone (and hence employ a definition of light verbs that is based on their semantics). This more restrictive definition of light verbs would yield a coherent set of properties of light verbs, over and above their modifying semantics alone. Their specific semantic function ties in with the fact that their "lightness" is intuitively asserted on the grounds of a comparison with some full verb variant - and indeed, light verbs with a modifier meaning can only arise via semantic generalization of some full verb meaning, for this would seem to be the only way to extract an event schema. Moreover, light verbs can now be characterized as consistently not suggesting a functor-argument relationship. And finally, their modifier status would entail that they should generally not be able to project arguments of their own; instead, they modulate the contents of the semantic roles of the full verb's arguments.14 (In contrast, causatives clearly introduce new arguments).

6.

Conclusion

The first part of this paper has presented syntactic and morphological evidence from Urdu which showed that light verbs cannot be analyzed on a par with auxiliaries. We instead proposed that light verbs are syntactically verbs (of category V), but are only licensed in conjunction with a main verb. Through a comparison of Urdu with English V+N constructions and V+V complex predicates of the Australian language Wagiman, we showed that while the overt syntactic frame that light verbs occur in may differ markedly from language to language and even within languages, a unified semantic notion of light verbs appears that is characterized by their function to modulate an event description in the way of a modifier (or perhaps even classifier). Some of the relevant differences between main verbs, light verbs and auxiliaries can thus be presented as in Table (71). Light verbs do not introduce an event description of their own, but instead hook into an existing event description and modify it. Auxiliaries, on the other hand, do not bear on argument structure, and they do not introduce an event description, or modify it, but situate a given event with respect to temporal or modal parameters.

On the (semi)lexical status of light verbs 365

Main verb Light verb Auxiliary

Full a-structure Incomplete a-structure No a-structure

event description modifies event situates event (e.g. Reichenbachian E,R,S)

Category V V AUXorl

Notes 1. An earlier version of this paper was presented at the workshop on semi-lexical heads in Tilburg, May 1999. Since then, it has also profited from discussions with Cristiano Broccias, Daniel Heitz, Shin-Sook Kim, Tara Mohanan, Peter Pause, the audience of the Konstanz SFB Colloquium, and from the comments of an anonymous reviewer. Elena Bashir pointed out numerous errors in the transcriptions, which we were able to correct, thanks to her help. The work was written with support from the SFB 471 "Variation und Entwicklung im Lexikon", funded by the Deutsche Forschungsgemeinschaft (DFG). Miriam Butt would also like to thank Ranu and Venkat Raman for providing the friendly environment that made an (almost) timely revision of this paper possible even during an extended research stay in Calcutta. 2.

The South Asian languages Urdu and Hindi are closely related. Both are among the 16 official languages of India and are spoken primarily in the north of India. Urdu is the official language of Pakistan. The data presented in this paper are drawn primarily from the dialect of Urdu spoken in Lahore, Pakistan, as well as from examples cited in the literature on both Urdu and Hindi. 3. Example (3b) might appear to contradict a claim we put forth in section 3 of this paper, namely that light verbs are different from aspectual auxiliaries and do not have a temporal semantics. However, the problem only lies in the translation. (3b) is really in line with our analysis because what the sentence essentially describes is the resultant state of 'coming', i.e., 'being here*. This is a lexical content, not a temporal assertion. For the English verb 'arrive', however, a perfect construction is the only way of approximating this resultative meaning (which could be conveyed by an adjectival passive in certain other cases). 4. Modals such as 'can', 'want', etc. behave much like main verbs in Urdu from a syntactic point of view and are thus kept out of the discussion here. 5. The tense/aspect system of Urdu/Hindi is rather complex as it also includes the possibility of reduplication for repeated action. We do not pursue a complete understanding of the Urdu/Hindi tense/aspect system within the scope of this paper.

366 Miriam Butt and Wilhelm Geuder

6. The progressive rah cannot be used in this test as per definition the sentence then ceases to be in the perfect. 7. As an anonymous reviewer pointed out to us, English V+N constructions also exhibit some "permissive" uses of give, e.g. He gave them a look into the room. 8. In general, the above sketch of meaning features is not very unlike the features that Newman (1996) identifies as constitutive of the meaning of give. He proposes to distinguish four different semantic "dimensions": (1) a spatio-temporal dimension (movement of a thing along a path), (2) a dimension of force-transmission (distinguishing an active party, i.e., the giver, from an inactive party that is acted on, the recipient), (3) a dimension of control (a change in possession, or at least disposal, occurs), and (4) a dimension of "human interest" (something with a positive value is added to the recipient's sphere of "control", e.g. sphere of possession). However, we would like to emphasize the fact emerging from our discussion that such semantic dimensions do not exist side by side but reveal mutual dependencies when it comes to lexical change. 9. Note that in this section we confine our discussion to the V+V constructions as the N+V complex predicates with 'give' for our purposes do not differ significantly enough from the English V+N constructions to merit a separate discussion (but see Mohanan (1994) for a detailed discussion of the distinct morphosyntactic properties of Hindi N+V complex predicates). 10. For detailed arguments as to why the event variable of adverbial modifiers is not to be equated with a verb's referential argument, in line with (58), see Geuder (2000). 11. Most authors concerned with the V+N English construction focus on the properties of the complement in light verb constructions that are supposed to show its "non-referentiality", and it is from this that the claim is then derived that light verbs do not denote events of their own. The problem is that the relevant effects, such as a ban on definites, are shared with, e.g., presentational constructions. Mulders (1992) advances a number of arguments to show that the evidence against referentiality of the nominalization in object position is not that strong. What we have been doing here is rather to focus, in a sense, on the "non-referentiality" of the light verb itself. 12. Take one of Mittwoch's (1998) examples of a cognate object construction from Hebrew: (i) Hu nixsal kisalon gamur He failed failure complete 'He failed a complete failure' = He failed completely This is another, even more dramatic case where verb and object are not referentially distinct but collapse into one event description. Note also that the modifier (gamur - 'complete') appears on the object, not on the V; so here, too, it is the object position that seems to host the "main predicate". Furthermore, in cognate object constructions, the same uncertainty appears with respect to the decision on the referential distinctness of V and object that we noted for the case of give a kiss·, consider e.g.:

On the (semi)lexical status of light verbs 367

(ii) He smiled an enigmatic smile Here, we have to ask whether a smile is just an event of smiling or something distinct from the process, e.g. the (content of the) expression on the face. 13. In the formalization (66b), (e) is the referential argument of BEC, and Ρ is a type of state that the event takes as a (complement) argument. Hence, BEC is to be read here as a predicate of events, which denotes the class of "transformation" events. Cf. Parsons (1990) for the issue of an event-semantic reformulation of the BEC operator in Dowty (1979). 14. For the case of light verb constructions that appear with a benefactive PP such as the Urdu example (48), the light verb must be regarded not as the verb that projects this argument but as a device which marks the appearance of an optional argument. Likewise, the subject of English light verb constructions must be seen as being associated with the VP as a whole, not just with give. Clearly, this is an issue for further research.

References Abeilld, Anne, Daniele Godard and Ivan A. Sag 1998 Two Kinds of Composition in French Complex Predicates. In: Erhard Hinrichs, Andreas Kathol and Tsuneko Nakazawa (eds.), Complex Predicates in Nonderivational Syntax. Syntax and Semantics Volume 30. San Diego: Academic Press. Alsina, Alex The Role of Argument Structure in Grammar. Stanford: CSLI 1996 Publications. Butt, Miriam 1995 The Structure of Complex Predicates. Stanford: CSLI Publications. 1997 Aspectual Complex Predicates, Passives and Disposition/Ability. Paper presented at LAGB 97, Edinburgh, http://www.ling.unikonstanz.de/pages/home/butt/ 1999 Differing Paths of (Non)Grammaticalization. Paper presented at the Workshop on New Reflections on Grammaticalization, Potsdam. http://www.ling.uni-konstanz.de/pages/home/butt/ Butt, Miriam and Tracy Holloway King 1991 Semantic Case in Urdu. In: Lisa Dobrin, Lynn Nichols and Rosa M. Rodriguez (eds.), Papers from the 27th Regional Meeting of the Chicago Linguistic Society, 31-45. Butt, Miriam and Aditi Lahiri 1998 The Status of Light Verbs in Historical Change. Paper presented at DIGS 5, York, http://www.ling.uni-konstanz.de/pages/home/butt/

368 Miriam Butt and Wilhelm Geuder

Cattell, Ray 1984

Composite Predicates in English. Syntax and Semantics Volume 17. Sydney: Academic Press Australia. Chakraborty, Jayshree 1992 Perfectivity and the Resultative State in Hindi. South Asian Language Review Volume Π (1): 55-67. Comrie, Bernard 1976 Aspect. Cambridge: Cambridge University Press. Dowty, David 1979 Word Meaning and Montague Grammar. Dordrecht: Reidel 1991 Thematic Proto-Roles and Argument Selection. Language 67: 547619 Eckardt, Regine 1998 Adverbs, Events, and Other Things. Tübingen: Max Niemeyer Verlag. Fitzpatrick-Cole, Jennifer 1994 The Prosodic Domain Hierarchy in Reduplication. PhD thesis, Stanford University. 1996 Reduplication meets the phonological phrase in Bengali. The Linguistic Review 13: 305-356. Geuder, Wilhelm 2000 Oriented Adverbs. Issues in the Lexical Semantics of Event Adverbs.Doctoral Dissertation, Univ. of Tübingen Grimshaw, Jane, and Armin Mester 1988 Light verbs and 77ieia-marking. Linguistic Inquiry 19 (2): 205-232. Hacker, Paul 1958 Zur Funktion einiger Hilfsverben im modernen Hindi. Mainz: Akademie der Wissenschaften und der Literatur. Higginbotham, James 1989 Elucidations of Meaning. Linguistics and Philosophy 12: 465-517 Hook, Peter E. 1974 The Compound Verb in Hindi. Center for South and Southeast Asian Studies: The University of Michigan. 1991 The Emergence of Perfective Aspect in Indo-Aryan Languages. In: Elizabeth C. Traugott and Bernd Heine (ed.), Approaches to Grammaticalization, Volume II. Amsterdam: John Benjamins. 1993 Aspectogenesis and the Compound Verb in Indo-Aryan. In: Manindra K. Verma (ed.), Complex Predicates in South Asian Languages. Delhi: Manohar Publishers. Hopper, Paul J. and. Elizabeth C Traugott 1993 Grammaticalization. Cambridge: Cambridge University Press.

On the (semi)lexical status of light verbs 369

Huang, C.-T. James 1992 Complex Predicates in Control. In: R. Larson, S. Iatridou and U. Lahiri (eds.), Control and Grammar, 109-146. Dordrecht: Kluwer Academic Publishers. Isoda, Michio 1991 The light verb construction in Japanese. In: L. Dobrin, L. Nichols, and R.M. Rodriguez (eds.), Papers from the 27th Regional Meeting of the Chicago Linguistics Society, 261—275. Jespersen, Otto 1965 A Modern English Grammar on Historical Principles, Part VI, Morphology. London: George Allen and Unwin Ltd. Krifka, Manfred 1992 Thematic relations as links between nominal reference and temporal constitution. In: Ivan Sag and Anna Szabolcsi (eds.), Lexical Matters, 29-53. Stanford, CA: CSLI Publications Matsumoto, Yo 1996 Complex Predicates in Japanese: A Syntactic and Semantic Study of the Notion 'Word'. Studies in Japanese Linguistics, Volume 7. Stanford, CA: CSLI Publications. Meißner, Η. Konrad 1964 Simplex und Verbalkompositum in Tulsi Das' Ramayana: cale jana - cali ana - uthi dhana. PhD thesis, Phillip-Universität zu Marburg. Mittwoch, Anita 1998 Cognate Objects as Reflections of Davidsonian Event Arguments. In: Susan Rothstein (ed.), Events and Grammar, 309-332. Dordrecht: Kluwer Mohanan, Tara 1994 Argument Structure in Hindi. Stanford: CSLI Publications. Mulders, Ren6 1992 On the aspectual nature of syntactic complementation. Den Haag: Holland Academic Graphics (HIL Dissertations) Newman, John 1996 Give: A Cognitive Linguistic Study. Berlin: Mouton de Gruyter. Norvig, Peter and George Lakoff 1987 Taking: A Study in Lexical Network Theory. In: Jon Aske et al. (eds.), Berkeley Linguistics Society. Proceedings of the 13th Annual Meeting, 185-206 Parsons, Terence 1990 Events in the Semantics of English. Cambridge MA: MIT Press Rosen, Sara 1989 Argument Structure and Complex Predicates. PhD thesis, Brandeis University.

370 Miriam Butt and Wilhelm Geuder

Singh, Mona 1994

Perfectivity, Definiteness, and Specificity: A Classification of Verbal Predicates Hindi. Phd Thesis, University of Texas, Austin. Tikkanen, Bertil 1987 The Sanskrit Gerund: A Synchronic, Diachronic and Typological Analysis. Studia Orientalia. Helsinki: Finnish Oriental Society. Wilson, Stephen 1999 Coverbs and complex predicates in Wagiman. Stanford, CA: CSLI Publications.

"Semi-lexical" motion verbs in Romance and Germanic Anna Cardinaletti and Giuliana Giusti

1.

Introduction

The distinction between lexical and functional elements is a central one in current grammatical theory.1 This distinction is often based on the assumption that the two different types of categories constitute mirror patterns with respect to a wide range of semantic, morphological, and syntactic properties (cf. Abney 1987). However, it is sometimes not easy to decide whether an element belongs to either of the two classes. This may lead one to assume a third, half-way class of "semi-lexical heads" which share some properties of lexical categories and others of functional ones. In this paper, we discuss one instance of these cases, namely motion verbs such as "go" and "come" which enter a particular "inflected construction" across unrelated languages such as some Southern Italian dialects, American English and Swedish. We will see that in this construction, motion verbs share many properties with functional verbs such as auxiliaries, but they maintain their semantic content and few other lexical properties. They could thus provide a very good candidate for the class of semilexical heads. It will turn out, however, that these semi-lexical heads cannot be identified by a fix set of properties which are shared by a coherent class of syntactic entities. We will observe cross-linguistic variation in the functional vs. lexical properties displayed by motion verbs. This disfavors the assumption of a "third type" category in addition to the other two opposite types. In the same way, it would not help to group different semi-lexical motion verbs into different categories, since this would multiply in no principled way the number of intermediate categories between lexical and functional categories. Our general claim is opposite to such a line of reasoning and it is set in (1):

372 Anna Cardinaletti and Giuliana Giusti

(1)

Semi-lexical motion verbs are lexical categories merged as functional heads.

(1) captures the fact that all the functional properties that can be claimed for motion verbs in the cases under consideration are actually lack or suppression of their canonical lexical properties. We will take this as evidence for the functional usage of such verbs.2 The paper is organized as follows. In section 2, we give the basic data and briefly dismiss three potential analyses of the inflected construction. In section 3, we focus on the functional usage of motion verbs in the inflected construction. This usage blocks a number of lexical properties that motion verbs usually display. In the inflected construction, they appear in a monoclausal structure forcing clitic climbing, "single event" interpretation, and a fix order of the two verbs. In some languages, they disallow the arguments and adjuncts typical of lexically used motion verbs. We account for these facts by proposing that the motion verbs under consideration are merged as functional heads in the extended projection (in Grimshaw's 1991 sense) of the lexical verb which follows them. A number of parallelisms with auxiliary verbs will confirm this hypothesis. In section 4, we see that an analysis of these verbs as functional categories tout court cannot capture the fact that these verbs, differently from auxiliary usages of motion verbs, maintain their motion semantic content in the three languages under consideration, in American English they require an animate subject, and in Swedish they may have locative arguments and adjuncts. We will therefore propose that these verbs are lexical elements that are merged as functional elements. In section 5, we set our analysis in the recent minimalist framework of Chomsky (1995, 1998). We assume a bare phrase structure building procedure which merges the relevant items from the lexicon in a strict cycle. By focussing on the merging points of motion verbs in the three languages, we find what independent properties of the languages under consideration give rise to the linguistic variation found in the lexical/functional behavior displayed by these verbs. 2.

The data and three potential analyses

Let us look at some data in Marsalese, a Western Sicilian dialect. In (2), two different constructions appear to be possible with the verb "go": The former, exemplified in (2a), is parallel to the infinitival construction also

"Semi-lexical" motion verbs in Romance and Germanic 373 found in Italian (cf.: Vado a prendere il pane). The latter, exemplified in (2b), displays two inflected verbs: (2)

a. b.

Vaju a pigghiari go-Is to fetch-lNF Vaju a pigghiu go-Is to fetch-Is Ί go to fetch bread.'

u pani. the bread u pani. the bread

("infinitival construction") ("inflected construction")

Constructions similar to (2b) are also found in other Southern Italian dialects, such as Eastern Sicilian, Southern Apulian and Southern Calabrian (cf. Rohlfs 1969: §710, §761). The inflected construction is not limited to the Romance area, but is also found in Germanic languages such as American English (3b,c) (cf. Carden and Pesetsky 1977; Jaeggli and Hyams 1993) and Swedish (4b) (cf. Wiklund 1996): (3)

a. b. c.

I go to buy bread. I go and buy bread. / go buy bread.

(4)

a.

Jag gär I

b.

för att göra

mig

en grogg.

go-PRES for to make-lNF myself a grogg

Jag gär

och

gör

mig

en grogg.

I go-PRES and make-PRES myself a grogg Ί go to make myself a grogg.'

Before we proceed with our investigation, we must dismiss three potential analyses of the inflected construction which might come to mind. The inflected construction is not a coordination despite the apparently coordinative morphemes and in American English, och in Swedish and a in Marsalese diachronically derived from Latin AC (cf. Rohlfs (1969: §761), and section 4.2 below). Abundant syntactic evidence against a coordination analysis has been provided by Faraci (1970), Carden and Pesetsky (1977), Jaeggli and Hyams (1993) for American English, by Wiklund (1996) for Swedish, and by Cardinaletti and Giusti (1998) for Marsalese. We refer the interested reader to those works. The inflected construction is not parallel to the finite construction found in some Southern Italian dialects in which the second finite verb is introduced by the particle ku, mu/mi (Rohlfs 1969: §717) or u (Francesco

374 Anna Cardinaletti and Giuliana Giusti

Giardinazzo, p.c.)· As shown by Calabrese (1993) for Salentino, the finite construction contains a full clausal complement to a wide class of lexical verbs, while, as we will show extensively below, the inflected construction is monoclausal. If the inflected construction is different from the finite construction with ku/mu/mi/u, it is expected that the two can coexist in one and the same language. This is indeed the case of Milazzese (spoken in the town of Milazzo, North-Eastern Sicily), where both the inflected construction and the finite construction exist in addition to the infinitival construction (cf. Ruggeri 1999): (5)

a. b.

pigghiari

Vaju

a

go-Is

to fetch-lNF

Vaju

a

pigghiu

u pani.

("infinitival construction")

the bread u pani.

("inflected construction")

go-lS to fetch-IS the bread c.

Vaju

mi pigghiu

u pani.

("finite construction")

go-lS MI fetch-IS the bread Ί go to fetch bread.' A detailed analysis of the finite construction and of the differences with the inflected construction is beyond the scope of this paper. Finally, the lack of object sharing between the two verbs excludes that the inflected construction is parallel to serial verb constructions (cf. Baker 1989; Lee 1992; Collins 1997 for recent discussion). This conclusion has been pointed out for American English by Baker (1989: 519, fn.3), Jaeggli and Hyams (1993: 322, fn.7) and Pollock (1994: 303, fh.19). The very same conclusion, we claim, holds for Marsalese and Swedish, since the inflected construction of these languages also lacks object sharing.

3.

Functional behavior

In this section, we show that motion verbs in the inflected construction share many properties with auxiliaries. A general property that unifies auxiliaries and motion verbs is that both can occur either as lexical verbs or as functional verbs. In the latter case, auxiliaries lose all of their selectional and semantic properties and just retain their morphological properties. They can thus be considered as prototypical functional verbs. Auxiliaries differ in this respect from motion verbs which also retain some of their lexical properties, as will be claimed in section 4. Therefore, functional behavior can be characterized as lack of (all or part of) lexical behavior.

"Semi-lexical" motion verbs in Romance and Germanic 375

3.1.

Closed classes

The motion verbs which can enter the inflected construction belong to a closed class:4 (6)

Marsalese a. iri 'go', viniri 'come', passari 'come by', mannari 'send' b. *acchianari 'go up', *scinniri 'go down', *trasiri 'go into', *curriri 'run',...

(7)

American English (Shopen 1971; Carden and Pesetsky 1977) a. go, come, run b. *walk, *fly, *rush,...

(8)

Swedish (Wiklund 1996) gä 'go', komma 'come', springa 'run',...

Belonging to a closed class is a typical property of functional categories which is shared by the motion verbs entering the inflected construction. On the other hand, all motion verbs, including those in (6b)-(7b), can enter the infinitival construction, displaying a lexical behavior in this respect. As for which specific verbs belong to this closed class, they appear to be the "weaker" motion verbs in the sense of Ritter and Rosen (1996), i.e., the semantically most basic ones. The two "weakest" verbs "come" and "go" enter the inflected construction in all languages for all speakers, while "less weak" verbs such as "run" and "come by" display variation across languages and among speakers.

3.2.

Fix order

In the languages under consideration, auxiliaries precede lexical verbs. If motion verbs are functional elements, we expect that in the hierarchical configuration, they precede the lexical verbs they are associated with. This is indeed what we find. The order of the two verbs is fix: the motion verb obligatorily precedes the other verb:5

376 Anna Cardinaletti and Giuliana Giusti

(9)

a.

Vaju a pigghiu u pani. go-Is to fetch-IS the bread b. *Pigghiu u pani a vaju. fetch-Is the bread to go-Is

(10)

a. I go and buy bread. b. #/ buy bread and go. c. / go buy bread. d. */ buy bread go.

The same observation has been made by Wiklund (1996: 36) for all verbs entering the inflected construction in Swedish. She provides examples with the verb "sit": (11)

a.

Denboken

satt

Lars och

thatbook-DEF sit-PAST Lars and

b. *Denboken

löste

Lars och

thatbook-DEF read-PAST Lars and

laste. read-PAST

satt. sit-PAST

'Lars was sitting and reading a book.' Since, as we said, auxiliaries precede lexical verbs in the languages under consideration, the word order in (9)-(ll) is expected under the hypothesis that the motion verb is merged as a functional head.

3.3.

No arguments

The infinitival and the inflected construction are different with respect to the possibility of argument insertion. While the former is always compatible with the presence of an argument of the motion verb, the latter is never so in Marsalese and American English. In Marsalese, a directional complement can be present only with motion verbs in the infinitival construction, as shown in (12a), where it immediately follows the motion verbs itself, but it is not possible in the inflected construction in (12b):6

"Semi-lexical" motion verbs in Romance and Germanic 377

(12)

a.

b.

Va (agghiri a casa) a mangiari (*agghiri a casa). go-3s (towards to home) to eat-lNF (*towards to home) 'He goes towards home to eat.' Va (*agghiri a casa) a mangia (*agghiri a casa). go-3s (*towards to home) to eat-3s (*towards to home)

The contrast in (12) follows from the assumption that a verb merged as a functional head cannot project its arguments. A further case is provided by a specific property of the lexical verb iri. In (13), iri occurs with a semantically void clitic cluster. The cluster is formed by a reflexive and the locative ni 'from here' and provides no semantic meaning. It is obligatory if no other complement of the verb is merged in the clause, as in (13a), and optional in the presence of another argument, as in (13b): (13)

a.

b.

Minni vaju. REFLcl-LOCcl go-ls Ί am going.' (Minni) vaju a casa. (reflql-loCcl) go-Is home Ί am going home.'

If the motion verb combines with an infinitival clause, as in (14a), the clitic cluster is optional. In the inflected construction (14b), the cluster is impossible: (14)

a.

b.

(Minni) vaju a mangiari. ( r e f l c l - l o c c l ) go-Is toeat-iNF Ί am going to eat.' (*Minni) vaju amangiu. ( * r e f l c l - l o c c l ) go-Is to eat-Is Ί am going to eat.'

In (14a), a mangiari is the complement clause selected by the lexical motion verb. The presence of the complement clause makes minni optional as does the directional complement a casa in (13b). In (14b), the sequence a mangiu does not constitute a clausal complement to a lexical instance of the motion verb. The only possible analysis of the inflected construction is a monoclausal one (as will be argued for in more detail in section 3.7), in

378 Anna Cardinaletti and Giuliana Giusti

which the motion verb is merged as a functional head of the extended projection of the second verb (mangio, in this case). If the motion verb is merged as a functional element, the clitic cluster which usually combines with the lexical verb cannot be merged. The same observation holds with other verbs that have a double usage as lexical verbs and as auxiliaries. The lexical verb stari optionally takes the same semantically void clitic cluster, as in (15a), but the cluster becomes ungrammatical when stari is used as an auxiliary of the progressive aspect, (15b):7 (15)

a.

(Minni) staju a casa. (refLcl-loCql) stay-Is at home

b.

(*Minni)

Ί am staying at home.'

staju

mangiannu a casa.

(*REFLCL-LOCCL) stay-Is eat-GER

at home

Ί am eating at home.' Aviri combines with the clitic ci in its lexical usage, as in (16a). When aviri takes a clausal complement as in (16b), the clitic is optional parallel to the case of iri in (14a). Aviri cannot be combined with ci in its auxiliary usage, as in (16c): (16)

a.

Ci

b.

there CL have a sister Ί have a sister.' (Ci) haju raaggiustari (therecjJ

c.

haju 'na soro.

have-Is

to fix-lNF

'stamachina. this car

Ί have to fix the car.' (*Ci) haju mangiatu a casa. (*thereCL) have-ls

eaten

at home

Ί have eaten at home.' In all cases seen so far, an analysis of the inflected construction in terms of a functional usage of motion verbs is straightforward. The same line of reasoning holds for American English. A directional complement of the motion verb can only occur with an infinitival complement, but it is impossible in both inflected constructions: (17)

I go all the way there to eat.

"Semi-lexical" motion verbs in Romance and Germanic 379 (18)

a. #/ go all the way there and eat./*/ go and eat all the way there. b. *I go all the way there eat. / *I go eat all the way there.

As we will see in section 4.4 below, this does not hold for Swedish, where it is possible to merge a locative argument of the motion verb in the inflected construction. The fact that language variation is found with respect to what lexical properties are present or absent suggests that the correct way to look at the items under consideration is not to label them as either lexical, functional, semi-lexical or even semi-functional, depending on whether they are more lexical or more functional, but to view them as lexical items which can dispense with their selectional requirements and thereby be merged in the extended projection of another verb. In section 5, an attempt is made to understand the language variation observed.

3.4.

No adjuncts

Contrary to what happens in the infinitival construction (19a), in the Marsalese inflected construction the motion verb cannot combine with adjuncts such as c'a machina 'by car'. In (19b) the adjunct could only be construed with the lexical verb mangiari, which produces a semantically anomalous sentence:8 (19)

a.

Peppe va a mangiari c'a machina. Peppe go-3s toeat-lNF by car 'Peppe goes to eat by car.' b. * Peppe va a mangia c'a machina. Peppe go-3s toeat-3s by car

As for American English, an adjunct is possible both in the infinitival and in the "V and V" construction, but it is not allowed in the "V V" construction: (20)

a. They go to eat by car. b. They go and eat by car. c. *They go eat by car.

In American English, only the motion verb in the "V V" construction loses the lexical property of being modified by an adjunct. Here, we observe syn-

380 Anna Cardinaletti and Giuliana Giusti tactic variation in one and the same language as to the behavior of the motion verb in slightly different constructions. In Swedish, adjuncts behave like complements in that they can appear with the motion verb (cf. section 4.4 below).

3.5.

Morphological restrictions

Marsalese and American English are parallel in that the motion verb in the inflected construction is possible with interesting (although not completely understood) morphological restrictions. In Marsalese, only the "unmarked" indicative present and imperative forms are allowed. First and second plural persons are excluded, as shown in (21d,e) and (22b). The past indicative (23a), the imperfect indicative (23b) and the past subjunctive (23c) are all ungrammatical:9 Vaju a pigghiu u pani. go-lS to fetch-Is the bread b. Vai a pigghi u pani. go-2s tofetch-2s the bread c. Va a pigghia u pani. go-3s tofetch-3s the bread d. *Emu a pigghiamu u pani. go-lPL to fetch-lPL the bread e. *Iti a pigghiati u pani. go-2PL to fetch-2PL the bread f. Vannu a pigghianu u pani. go-3PL to fetch-3PL the bread Ί / you / etc. go to fetch bread.' a.

(22)

a.

Va pigghia u pani! go-MP-2s buy-lMP-2s the bread 'Go to fetch bread!' b. *lti pigghiati u pani! go-!MP-2PL buy-lMP-2PL the bread

"Semi-lexical" motion verbs in Romance and Germanic 381 (23)

a.

*Ii

a pigghiai

go-PAST-ls

b. *Ia

a pigghiava

go-lMPERF-lS

c.

*Si tinn' if

u pani.

to fetch-PAST-lS the bread

u pani.

to fetch-lMPERF-lS the bread

issi

aaccattassi

REFL CL -LOC CL go-SUBJ-2s to buy-suBJ

sta butia,

spinnissi

this shop,

spend-SUBJ-2s less

upani

ne

the bread in

chiu picca.

The descriptive generalization for the Marsalese verb iri in (21)-(23) is that the forms that contain the root να-, but not the forms that contain the root e-/i- can instantiate the inflected construction. The existence of two allomorphs is not specific of the pattern of the verb iri. Two allomorphs are also found overtly for the verb viniri 'come', which has the allomorph venfor the 1st, 2nd, 3rd singular and 3rd plural persons of the present indicative and for the 2nd singular imperative and the allomorph vin- for the 1st and 2nd plural of the present indicative and for all persons of other tenses and moods. The fact that the other verb entering the inflected construction, passari 'come by', does not display an overt differentiation in its morphological pattern does not necessarily mean that this verb does not have two homophonous allomorphs. In the inflected construction, viniri and passari undergo the same morphological restrictions as iri. Allomorphy cannot be taken as the ultimate cause of the inflectional restrictions on the inflected construction. It shows that whatever the reason to single out the three singular persons and the 3rd plural of the present indicative and the 2nd singular of the imperative may be, this holds independently of the analysis of the inflected construction. Allomorphy must thus be the result of some general property of the verbal inflectional system which interacts with syntactic principles. Notice that, in some obvious sense, the forms which are built with the allomorph va- or ven- are less marked than the forms built with e-/i- or vin-: present tense is less marked than past, indicative is less marked than subjunctive, singular is less marked than plural, 3rd person is less marked than 1st and 2nd person. We can say that the functional usage of motion verbs in Marsalese is not only limited to semantically less marked verbs (namely, "weaker" verbs in the sense of Ritter and Rosen 1996, cf. section 3.1 above), but it is also limited to the less marked allomorphs of such verbs. A tentative generalization is in (24):

382 Anna Cardinaletti and Giuliana Giusti

(24)

The inflected construction is possible with the less marked forms of a verbal paradigm.

The restrictions on the American English constructions are different, due to the different morphological systems of the two languages. However, they are rather reminiscent of the Marsalese facts in that the only allowed form is the unmarked "base" form (cf. Shopen 1971; Carden and Pesetsky 1977): (25)

a. b. c. d. e. f. g. h.

John managed to go visit Harry every week. Go visit Harry tomorrow! John will go visit Harry tomorrow. 1/You /We/ They go visit Harry every Thursday. *John goes visit Harry every afternoon. *John went visit Harry tomorrow. *John has gone visit Harry already. *John is going see Harry tomorrow.

(26)

a. #John went and visited Harry. b. *John went and visit Harry. c. John didn 't go and visit Harry. d. Did John go and visit Harry ?

Contrary to Marsalese and American English, in Swedish all tenses are allowed. In (27), for instance, the lexical verb has the same past form as the motion verb. To make sure that (27a) is not a coordination, in (27b) we provide an instance of extraction, which gives perfectly acceptable results: (27)

a.

Han gick he

b.

pa affären och köpte

go-PAST to shop-the and buy-PAST

bröd. bread

'He went to the shop to buy bread.' Vadi gick han pä affären och köpte U? what go-PAST he

to shop-the and buy-PAST?

'What did he go to buy in that shop?' But a very general property of the Swedish verb morphology is that there is no person specification. All tenses are unmarked for this feature. The morphological restrictions in Marsalese and American English and the great freedom allowed to the inflected construction in Swedish, comply with the generalization in (24). The differences can be reduced to language

"Semi-lexical" motion verbs in Romance and Germanic 383 specific properties of verbal morphology and how it interacts with the inflected construction in each language (see also section 5 below).

3.6.

Invariant forms in Marsalese

The less marked specification of motion verbs in the Marsalese inflected construction can be proved empirically. Only in this construction, the motion verb iri can display an invariant form, homophonous to the 3rd person singular. Compare the infinitival construction in (28) with the inflected construction in (29): (28)

a. b. c.

a. b. c.

(Eu) (I) (Tu) (you) (Iddi/Idde) (they)

vaju / *va go-lS / *go vai/*va go-2s / *go vannu / *va go-3PL / *go

a pigghiari u pani. to fetch-INF the bread a pigghiari upani. to fetch-INF the bread a pigghiari u pani. to fetch-INF the bread

(Eu) (I) (Tu) (you) (Iddi/Idde) (they)

vaju/va go-IS / go vai/va go-2s / go vannu/va go-3PL / go

apigghiu to fetch-IS apigghi to fetch-2s a pigghianu to fetch-3PL

u pani. the bread u pani. the bread u pani. the bread

The invariant form va can be regarded as a reduction of vaju, vai and vannu, namely a reduction of the inflected forms built with the allomorph va- of the verb iri. This explains why the invariant form is not possible in the 1st and 2nd person plural of the indicative present and with any tenses other than indicative present, which are built with a different allomorph of the verb iri, namely e-/i-: (30)

a.

*(Niatri) (we) b. *(Viatri) (you) *(Eu) c. (I)

va go va go va go

a pigghiamu to fetch-lPL a pigghiati to fetch-2PL a pigghiai to fetch-PAST-lS

u pani. the bread u pani. the bread u pani. the bread

384 Anna Cardinaletti and Giuliana Giusti

A language internal comparison shows that the choice of an invariant form is also available with other functional verbs, such as the auxiliaries aviri and stari, with similar restrictions: the invariant form is allowed with the three singular persons of the present indicative and incompatible with the plural persons and with tenses other than present indicative. We provide some examples of these restrictions in (31) and (32): (31)

a.

c.

d. (32)

(Eu) unci haju/ha (I) not there CL have-ls / have Ί have never been there.' (Tu) un ci hai / ha

statu mai. been never

statu been emu / *ha statu have-1 PL / have been avia/*ha

(you)

not therecL have-2s / have

(Niatri)

un ci

(we)

not there CL

mai. never mai. never statu mat.

(Eu)

unci

(I)

not there CL have-lMPERF-lS / have been never

(Eu)

ci

staju/sta

there CL stay-Is / stay (D Ί am going there.'

stai/sta

ennu. go-GER

(Tu)

ci

(you)

thereCL stay-2s / stay

go-GER

c.

(Niatri)

ci

ennu.

(we)

thereCL stay-1 PL / stay

d.

(Eu)

ci

(I)

there CL stay-lMPERF-lS / stay go-GER

stamu/*sta

ennu.

go-GER

stava/*sta

ennu.

In their lexical usage, none of these verbs allows invariant forms, as is apparent in (33): (33)

a.

b.

c.

(Eu) ci vaju/*va. (I) there CL go-1S / go Ί go there.' (Eu) ci haju / *ha 'na soro. (I) there CL have-ls / have a sister Ί have a sister.' (Eu) ci staju/*sta. (I) there CL stay-lS / stay Ί stay there.'

"Semi-lexical" motion verbs in Romance and Germanic 385 A full account of the optionally reduced forms displayed by auxiliaries in Marsalese is beyond the scope of this article. We can speculate that auxiliaries are merged into the relevant functional heads directly (cf. Cinque 1999a). As a consequence, they cannot check the features which are realized lower in the structure. With the additional assumption that auxiliaries may be merged higher than the person head, this can capture the lack of person features in (31)-(32). As we will see in section 5.1 below, a similar analysis can be proposed for the motion verbs in the inflected construction: motion verbs are always higher than the person head, since the inflectional features are checked by the lexical verb.

3.7.

For a monoclausal analysis

A functional element is merged in the structure as part of an extended projection and it does not project its own. In our case, we expect the motion verb to be merged in the extended projection of the verb which follows it, giving rise to a monoclausal structure. This is supported by a wide range of syntactic properties.

3.7.1.

Unique person, tense, and mood specifications

A defining property of the inflected construction in the three languages under consideration is expressed in (34) and exemplified in (35)-(39): (34)

The two verbs in the inflected construction must share inflectional features.

(35)

a.

*Ia go-lMPERF-1S

b.

*Vaju go-PRES-lS

(36)

a.

*li go-PAST-lS

b.

*Vaju go-PRES-lS

(37)

a. b.

a pigghiu

u pani.

to fetch-PRES-lS the bread

apigghiava

upani.

to fetch-lMPERF-lS the bread

a pigghiu

u pani.

to fetch-PRES-lS the bread

apigghiaiu

pani.

to fetch-PAST-lS the bread

*I went buy bread. */ went and buy bread.

386 Anna Cardinaletti and Giuliana Giusti (38)

a. */ go bought bread. b. *I go and bought bread.

(39)

a.

*Vi we

b. *Vi We

gick

och köper

go-PAST and buy-PRES

gar

ochköpte

go-PRES and buy-PAST

bröd. bread bröd. bread

In Marsalese (35), we observe that an imperfect form of either verb cannot combine with a present form of the other verb. In (36), we observe the same case for past forms. In American English (37), a past form of either verb cannot combine with a base form of the other verb in the "V V" construction. In (38), we observe the same case for the "V and V" construction. Swedish (39) shows the same phenomenon. If we assume, as is currently done, that only one tense, mood and person specification can be assigned to the extended projection of the verb, (34) strongly suggests a monoclausal analysis for the inflected construction.

3.7.2.

Single event interpretation

Another defining property of the inflected construction in the three languages under consideration is expressed in (40): (40)

The two verbs in the inflected construction refer to a single event.

Shopen (1971: 257-258) notices that in American English the inflected construction does not have the same meaning as the infinitival construction. In the inflected construction, the motion verb and the predicate are interpreted as building a single event. Thus, while the sentence in (41a) refers to two different events and is true even if the purchase does not take place, (41b) is a contradiction. The event of going and that of purchasing must coincide, hence the continuation which negates the event of purchasing makes the sentence ungrammatical (sentences from Shopen 1971: 258): (41)

a.

They go to buy vegetables every day, but there never are any vegetables. b. *They go buy vegetables every day, but there never are any vegetables.

"Semi-lexical" motion verbs in Romance and Germanic 387 This state of affairs is expected if the two verbs belong to one and the same extended projection. Since a similar analysis is also valid for the American English "V and V" construction, the fact that (42) patterns with (41b) is not surprising: (42) *They go and buy vegetables every day, but there never are any vegetables. The same observation obtains in Marsalese. The negation of the event of purchasing is possible with the infinitival construction, but impossible with the inflected construction: (43)

a.

Vaju aaccattari acicoria gnignornu, ma go-IS to fetch-lNF the chicory every day but unn'a trovu mai. not itcL find-IS never Ί go to buy chicory every day, but I can never find any.' b. *Vaju a accattu a cicoria gnignornu, ma go-IS to fetch-Is the chicory every day but unn 'a trovu mai. not itcL find-IS never

Swedish gives the same result: The inflected constructions in (44) involve a single event interpretation. The second sentence negating the event of buying, thus, produces a very marginal acceptability: (44)

a.

b.

77

De gar och köper grönsaker varje dag, they go-PRES andbuy-PRES vegetables everyday, men det finns aldrig nigra. but there are never any 'They go to buy vegetables every day, but there aren't ever any.' "Hon gick ochköpte grönsaker, she go-PAST andbuy-PAST vegetables, men affären var stängd. but store-the was closed 'She went to buy vegetables, but the shop was closed.'

We may wonder why the sentences in (44) are not judged with a star. This is because there is the marginal possibility to interpret them as coordina-

388 Anna Cardinaletti and Giuliana Giusti tions. If the complement of the lexical verb is extracted, forcing the inflected construction interpretation, we obtain a deviant sentence: (45) *Vad\ gär

de

och köper t; varje dag

what go-PRES they and buy-PRES every day

men det finns aldrig nägra? but there are

never

any?

The single event interpretation can straightforwardly be captured by monoclausality.

3.7.3.

Clitic climbing in Marsalese

In the Marsalese inflected construction, clitic climbing is obligatory. In other words, clitic pronouns obligatorily appear on the motion verb: (46)

a.

*Vaju

a pigghiulu.

go-Is

b.

U

to fetch-lS-it CL

vaju

a pigghiu.

it CL go-IS to fetch-Is Ί go and fetch it.'

In this respect, the motion verb in the inflected construction patterns with the two functional verbs aviri and stari discussed above, (47)-(48), and differs from the motion verb in the infinitival construction in (49) and from modal verbs (50), where clitic climbing is preferred but not strictly obligatory (hence, the question mark on (49a) and (50a)): (47)

a.

*Haju have-Is

b.

L'haju

pigghiätulu. fetch-PAST.PART-itCL

pigghiatu.

itcL have-IS fetch-PAST.PART Ί have fetched it.'

"Semi-lexical" motion verbs in Romance and Germanic 389

(48)

a.

*Staju pigghiannulu. stay-IS fetch-GER-itcL b. U staju pigghiannu. itCL stay-IS fetch-GER 'I'm fetching it.'

(49)

a. b.

(50)

a. b.

?Vaju a pigghiallu. go-IS to fetch-INF-itcL U vaju α pigghiäri. 0 s to ΐ* external argument > internal arguments

398 Anna Cardinaletti and Giuliana Giusti

5.1.

The structure of the inflected construction in Marsalese

Let us start with Marsalese. In this language, the lexical verb moves at least as high in the functional structure as in Italian. This is established by looking at the relative position of verbs and adverbs, using the adverb hierarchy argued for by Cinque (1999a): ora 'now' > forse 'maybe' > di solito 'usually' > spesso 'often' >giä 'already' > sempre/mai 'always /never': 1 8 (66)

a.

b.

Peppe (*sempre) pigghia {sempre) u pani Peppe (*always) fetch-3s (always) the bread ne 'sta butia. in this shop 'Peppe always buys bread in this shop.' Peppe (*giä) pigghia (giä) upani Peppe (*already) fetch-3s (already) the bread ne 'sta butia. in this shop 'Peppe already buys bread in this shop.' Peppe (ora) pigghia (*ora) upani ne 'sta butia. Peppe (now) fetch-3s (*now) the bread in this shop 'Peppe now buys bread in this shop.'

The sequence "motion verb + a + lexical verb" in (67) appears below ora and above giä just like the simple lexical verb in (66): (67)

a.

b.

c.

Peppe (*sempre) va (*sempre) a(*sempre) pigghia Peppe (*always) go-3s (*always) to (*always) fetch-3s (sempre) u pani ne 'sta butia. (always) the bread in this shop Peppe (*giä) να (*giä) α (*giä) Peppe (*already) go-3s (*already) to (*already) pigghia (giä) u pani ne 'sta butia. fetch 3S (already) the bread in this shop Peppe (ora) va (*ora) a (*ora) pigghia (*ora) Peppe (now) go-3s (*now) to (*now) fetch-3s (*now) u pani ne 'sta butia. the bread in this shop

We propose that in the inflected construction (67), first the verb pigghia checks its features in the functional position above giä, then the connecting

"Semi-lexical" motion verbs in Romance and Germanic 399 element a is merged: the resulting node is an extended projection of the lower lexical verb. Successively the motion verb va is merged, and the resulting node extends the extended projection of the lower verb. After this, the structure building procedure continues, for example, to merge a different class of adverbs, such as ora in (67c), or a clitic pronoun, as in (68a) (cf. section 3.7.3 above): (68)

a.

U

vaju

a pigghiu. t0

b. c.

itCL go-Is fetch-Is Ί go and fetch it.' *Vaju a u pigghiu. go-IS to itcL fetch-Is U pigghiu. it C L fetch-Is Ί take it.'

The ungrammaticality of (68b) shows that the clitic pronoun is not merged immediately higher than the lexical verb, in contrast with the simple case in (68c), but after the motion verb is merged, (68a). In the imperative mood, the lexical verb moves higher than in the indicative mood, as is apparent from the position of the hierarchically high adverb ora and of the clitic pronoun. Compare (69a) with (67c) and (69b) with (68c): (69)

a.

b.

Pigghia ora u pani! fetch-lMP-2s now the bread 'Take bread now!' Pigghialu! fetch-IMP-2s-itCL 'Take it!'

The same is true for the sequence "motion verb + lexical verb" in an imperative inflected construction, which precedes both the hierarchically high adverb ora and a clitic pronoun: 19 (70)

a. b.

Va pigghia ora u pani! go-lMP-2s fetch-lMP-2s now the bread Va pigghialu! go-imp-2s fetch-imp-2s-it c i

400 Anna Cardinaletti and Giuliana Giusti The evidence concerning indicative and imperative verbs, thus, points to the same conclusion: the checking of verb-related features such as tense, agreement and mood precedes the merging of the motion verb. Being merged in such a high head, the motion verb cannot interact in any way with the thematic structure of the lexical verb. As we have seen in the first part of the paper, in Marsalese the motion verb cannot project any arguments and adjuncts, nor can it assign a secondary θ-role to the external argument. As a second consequence, being merged in a head higher than the checking domain of the lexical verb, the motion verb cannot check its features in the canonical way (by moving to a designated functional head). It is thus not surprising that the motion verb cannot have all the features of an inflected verb and is restricted to some persons and moods (the 1st, 2nd, 3rd singular and 3rd plural present indicative and 2nd singular imperative, cf. section 3.5). Two questions now arise: how can these features be checked, and why can the motion verb have only these features? We propose that these features are checked in a parasitic way, by copying the features of the inflected lexical verb onto the motion verb. The copying procedure guarantees that the motion verb has the same features as the lexical verb (cf. section 3.7.1). The restriction of the inflected construction to the verbal forms mentioned above is not surprising if these features, being unmarked (cf. section 3.5), do not need to be checked in the canonical way, but it suffices for them to be copied from an inflected verbal form. Remember that Marsalese motion verbs may display an invariant form in the unmarked persons and moods (cf. section 3.6). We conclude that no feature copying has taken place in this case, or, alternatively, that the copying procedure has an optional morphological manifestation. Notice finally that the motion verb in the inflected construction is a different element from Cinque's (1999b) andative verb, which is very low in the structure in Marsalese as well. In (71), we provide Cinque's hierarchy, exemplified in Italian (72) and in Marsalese (73): (71)

... > Causative > ... Andative > ... V

(72)

a.

Ce lo fecero andare a prendere subito. to-usCL itcL made-3PL go-lNF to fetch-lN immediately 'They made us go to fetch it immediately.'

"Semi-lexical" motion verbs in Romance and Germanic 401

b.

(73)

a.

b.

siamo andate/i afar flrmare t0 them^L are gone FEM/MASC make-INF sign-LNF a Gianni. by Gianni 'We went to make Gianni sign them.' U fazzu iri apigghiari aPaola. it CL make-IS go-lNF to fetch-lNF by Paola Ί make go Paola fetch it.' "U vaju afari pigghiari a Paola. it C L go-Is tomake-lNF fetch-lNF by Paola

In (72a) and (73a) the andative verb appears lower than the causative verb. In (72b) and (73b) it appears higher than the causative, yielding very marginal results. Contrary to the andative verb in (71), the motion verb in the Marsalese inflected construction appears higher than the causative verb, as expressed in the hierarchy (74) and exemplified in (75): (74)

Inflected construction: ... > Motion verb > ... Causative > ... V

(75)

U picciriddu the child

u ya it CL go-3s

a fa lavaria su matri. to make-3s wash-lNF by his mother

'The child goes to make it be washed by his mother' If the structure building procedure obeys the universal hierarchy proposed by Cinque (1999a, 1999b), we must consider the motion verb and the andative verb as two different elements given their occurrence at different points in the structure. In such a framework, it is predicted that the two can co-occur in the order expressed in (76), as is indeed the case in (77): (76)

... > Motion Verb > ... Causative > ... Andative > ... V

(77)

U va(ju) it C L go(-ls)

a fazzu iri apigghiari a Paola. to make-Is go-lNF to fetch-lNF by Paola

Ί go to make Paola go to fetch it.' The co-occurrence of the motion verb and the andative verb reinforces Cinque's hypothesis of a universal hierarchy on the one hand. On the other

402 Anna Cardinaletti and Giuliana Giusti hand, it also reinforces our hypothesis that some languages have the possibility to merge a motion verb (belonging to a closed class) after merging the lexical verb in its spell-out position. In other words, the motion verb in the inflected construction is a language particular device which adds up to the universal hierarchy present in all languages. The natural question arises as to how the child can acquire this construction. We take the presence of uninflected auxiliaries and/or the presence of uninflected forms of the motion verb as evidence for the possibility to merge a verbal head which does not need to check its features in the canonical way, but can check features via copying. As a matter of fact, all dialects which display the inflected construction also display invariant forms of the motion verb. On the other hand, Italian lacks both invariant forms and the inflected construction. The same is true of a Calabrian dialect spoken in the town of Bovalino Marina (in the province of Reggio Calabria), which displays neither invariant forms nor the inflected construction (Francesco Giardinazzo, p.c.). We take this not to be accidental, but depending on the different possible checking procedures in the two types of languages. Further comparative research is needed to establish whether this correlation holds universally and whether the presence of uninflected form is not only a necessary but also a sufficient condition for the presence of the inflected construction.

5.2.

The structure of inflected constructions in American English

In American English, we have observed two constructions: "V V" and "V and V". In both constructions, the lexical verb must remain very low and does not check all its features prior to the merging of the motion verb. The combinations of modals and auxiliaries with the motion verb in (78)-(79) show that, like lexical verbs and differently from auxiliaries and modals, the sequence "motion verb (and)" cannot precede auxiliaries, may follow modal verbs and the infinitival marker to, and can co-occur with dosupport: (78)

a. *They go have eaten. b. *They go be eating. c. He should go eat. d. He tried to go eat. e. Did they go borrow the money right away?

(Shopen 1971)

"Semi-lexical" motion verbs in Romance and Germanic 403 (79)

a. *They go and have eaten. b. *They go and be eating. c. He should go and eat. d. He tried to go and eat. e. Did they go and borrow the money right away ?

According to the classical analysis by Pollock (1989), in English the verb moves very little or does not move at all, procrastinating the checking procedure after Spell Out. As for the inflected construction, we must assume that after the motion verb is merged (regardless of the presence of the connecting element and, which we treat parallel to the Marsalese a), it is impossible for the lexical verb to further move to a functional head at LF, due to minimality considerations. This captures the fact that in American English, only the unmarked bare form, which need not be checked, enters the inflected constructions. The motion verb does not undergo a similar structural restriction and could in principle check its features at LF. However, being in the extended projection of the lexical verb, it cannot display different features with respect to the lexical verb. This is the reason why also the motion verb must appear in its base form.21 Since in American English, the motion verb is merged very low, it may retain some of its lexical properties, among which the possibility to assign a secondary θ-role to the subject in the "V V" construction, as observed in section 4.3. This is not the case in the "V and V" construction, where the presence of the functional head realized by and clearly shows that the motion verb is merged higher than in "V V". The English-internal difference can thus be due to the different merging point of the motion verb. Partial support for this claim comes from the distribution of the motion verb with respect to negation and sentential adverbs. While in the "V V" construction, the motion verb must follow negation and adverbs (80a,b), (81a,b), in the "V and V" construction, it can marginally precede them (sentences (80c,d), (81c,d)):22 (80)

a. Do not go visit Harry. b. *Go not visit Harry. c. Do not go and visit Harry. d. IGo and not visit Harry.

(Carden and Pesetsky 1977: 90)

404 Anna Cardinaletti and Giuliana Giusti

(81)

a.

I seldom / often go talk to my advisor. (Jaeggli and Hyams 1993: 319) b. *I go seldom / often talk to my advisor. c. / seldom / often go and talk to my advisor. d. ?/ go seldom / often and talk to my advisor.

Furthermore, while the motion verb in the "V V" construction must follow the causative verb make, it can marginally precede it in the "V and V" construction: (82)

a. They always make me go buy bread. b. *They always go make me buy bread.

(83)

a. They always make me go and buy bread. b. IThey always go and make me buy bread.

As seen above, the different placement of the motion verb in the two American English constructions correlates with their different capacity of imposing selectional restrictions on the external argument. In both cases, however, the motion verb is merged too high to be able to select internal arguments. As for the capacity of projecting adjuncts, this is preserved in the "V and V" construction, but it is unexpectedly not found with the "V V" construction. We repeat the relevant examples for convenience: (84)

a. They go to eat by car. b. They go and eat by car. c. *They go eat by car.

We suspect that the impossibility of (84c) correlates with the fact that the adjunct selected by the motion verb must follow the lexical verb, as shown in the grammatical (84a) and (84b). In antisymmetric terms, this word order is obtained by moving a projection of the lexical verb across the adjunct. While this is possible for the bigger XPs to eat in (84a) and and eat (84b), it is impossible for the bare verb eat in (84c). 5.3.

The structure of the inflected construction in Swedish

In Swedish, tensed verbs occur very low in the structure. They follow all adverbs and negation:

"Semi-lexical" motion verbs in Romance and Germanic 405

(85)

a. b.

Jag vet att Kalle I know that Kalle Jag vet att Kalle I know that Kalle Ί know that Kalle does not

förmodligen sjunger. probably sing-PRES inte sjunger. not sing-PRES sing.'

(85) shows that Swedish tensed verbs do not move to the high Infi positions to check their features before Spell Out. Since Swedish does not have any person features in its verbal inflection, nor does it have any counterpart to English do-support (cf. Vikner 1997), we assume that verb movement to the high Infi positions also does not take place after Spell Out. Let's see the consequences of this assumption for the inflected construction. After the lexical verb is merged, the connecting element och and successively the motion verb are merged. At this point, no further movement is required for either verb. The two verbs have the same tense/mood features. As in Marsalese, the features of the lexical verb are copied onto the motion 23

verb. The bottom-up procedure continues to merge adverbs such as förmodligen, which are universally merged rather early, according to Cinque's hierarchy. This is represented in the embedded clause in (86): (86)

Jag vet att Kalle förmodligen sitter och I know that Kalle probably sit-PRES and Ί know that Kalle probably sits and sings.'

sjunger. sing-PRES

In main clauses, the bare phrase structure building procedure continues, as expected, and moves the higher verb (namely the motion verb) to the verbsecond position, as represented in (87) (sentences from Josefsson 1991: 142): (87)

a.

Kalle sitter Kalle

säkerligen ochfiskar

sit-PRES probably

abborre.

and catch-PRES perch

'Kalle is probably sitting and catching perches.' b. * Kalle sitter ochfiskar säkerligen abborre. Kalle

sit-PRES

and catch-PRES probably

perch

(87a) confirms that merging of the motion verb in the inflected construction in Swedish is very early, much earlier than the application of the "Verbsecond rule". This excludes (87b), where first the lexical verb is moved to "second position", then the motion verb is merged.

406 Anna Cardinaletti and Giuliana Giusti

Very similar observations hold for imperative verbs. While in simple sentences, the lexical verb moves to a high position preceding negation, as in (88), in the inflected construction it is the imperative motion verb alone which appears before negation, (89)·?* (88)

Rök inte här! smoke-lMP not here 'Don't smoke here!'

(89)

a.

Korn inte och rök här! come-lMP not and smoke-lMP here 'Don't come and smoke here!' b. *Kom och rök inte! come-lMP and smoke-lMP not

In order to understand the fact that in Swedish, the motion verb can retain most of its lexical properties, we propose that it is merged in the structure earlier than its American English counterpart. Notably, it can extend the Larsonian VP-shell projecting a lexical VP in the specifier of which a locative argument is merged. The possibility of extraction displayed in (90) simply makes sure that we are dealing with an inflected construction and not with a coordination: (90)

a.

Jag vet

inte

vad

I

not

what they go-PRES on restaurant

know

de

aker

pä restaurang

och äter. and eat-PRES

b.

Ί don't now what they go to the restaurant to eat.' Jag vet inte vad de akte pa affären I know not what they go-PAST on shop och köpte. and buy-PAST

Ί don't know what they went to the shop to buy.' In the embedded construction (90), the motion verb precedes the locative complement. This shows that the motion verb projects its complement and then is moved to a higher level. The reasons of this movement are irrelevant to our point here and must be the same that apply when the motion verb is a lexical verb. In this respect, (90) is parallel to (91):

"Semi-lexical" motion verbs in Romance and Germanic 407 (91)

a.

b.

Jag vet inte vem som aker pa re staurang. I know not who that go-PRES on restaurant Ί do not know who goes to the restaurant.' Jag vet inte vem som äkte päaffären. I know not who that go-PAST on shop Ί don't know who went to the shop.'

Being so low in the structure, we would expect the motion verb to be able to assign a secondary theta-role as in the American English "V V" construction. This is however not the case, as observed in section 4.3. Since the property to assign a secondary theta role is certainly something to be specified in the lexicon, we assume that this is a matter of lexical variation and that this idiosyncratic property of the English motion verb is not present in Swedish.

6.

Conclusions

In this paper, we have analyzed the behavior of a class of elements, namely motion verbs in the inflected construction, which might seem to provide good candidates for semi-lexical items. We arrived at the conclusion that it is not desirable to assume a third-type category which displays some functional and some lexical properties. There is no empirical generalization as to what subset of properties these should be, and language variation is found in this area. A promising way of looking at the fact that some lexical categories lack some of their typical lexical properties is to propose that this is due to their merging into a non-lexical (functional) projection. Only less marked verbs have the property of having a functional usage. In the course of the paper, we have discussed the four generalizations in (92): (92)

a. b. c. d.

The inflected construction is possible with the less marked forms of a verbal paradigm. The two verbs in the inflected construction must share inflectional features. The two verbs in the inflected construction refer to a single event. The motion verb immediately precedes the lexical verb which has reached its spell-out position in the derivation.

408 Anna Cardinaletti and Giuliana Giusti

To account for these generalizations, we have formulated the hypothesis in (93): (93)

The motion verb is merged in the derivation immediately higher than the spell-out position of the lexical verb.

We have proposed that first, the lexical verb (or other functional verbs present in the structure) reaches its spell-out position in the clause, then the connecting element (if present) and successively the motion verb are merged immediately higher than the spell-out position of the lexical verb. After this operation, the structure building procedure continues regularly as in any other sentence. In this way, we capture the cross-linguistic variation found in the inflected construction in the location of the motion verb. The discussion so far led us to formulate the parametrized property in (94): (94)

The motion verb loses some but not all its lexical properties. (Language variation is found as to which properties are lost and which are retained.)

We have proposed that (93) and (94) are related. The higher up in the structure the motion verb is merged, the less lexical properties of the motion verb can be realized in the structure. Being merged in a head higher than the checking domain of the lexical verb, the motion verb cannot check its features in the canonical way (by moving to a designated functional head). We have proposed a copying procedure, which copies the features of the inflected verb onto the motion verb. The copying procedure can account for a wide range of restrictions observed in the inflected construction: (95)

a. b. c.

the impossibility of the construction in many languages the restriction on tense, mood, and person the occurrence of uninflected forms

The restrictions above can be reduced to the limited capacity of some verbs to check their features parasitically on the lexical verb. We can speculate that only less marked forms (with less features) can do so. Notably, the three languages under consideration either have verbal morphology unmarked for person, as in Swedish and American English, or have special verbal forms with this characteristic, as in Marsalese. In languages which

"Semi-lexical" motion verbs in Romance and Germanic 409

do not have the inflected construction, such as Italian, invariant forms do not exist. For the sake of the presentation, we have assumed that functional categories share a number of defining properties. But this matter is far from being settled. As a point of fact, functional properties are a bundle of descriptive facts that constitute more of a tendency than a diagnostics. Our approach to motion verbs in the inflected construction may shed some light on the treatment of functional elements as elements deprived of some (or all) of their lexical properties.

Notes 1. We would like to thank Paola Benincä, Judy Bernstein, Guglielmo Cinque, Francesco Giardinazzo, David Pesetsky, Giuseppe Rallo, Lori Repetti, Michal Starke, Tarald Taraldsen and Anna-Lena Wiklund for having discussed different sets of the data with us. Most of the Marsalese data have been collected in a field work by Giuliana Giusti in summer 1998. All the Marsalese data have received the final approbation of Giuseppe Rallo, who is a native speaker of this language. Special thanks also go to Anna-Lena Wiklund for her extreme availability at any moment with the Swedish data and for very helpful comments. Previous versions of this work have been presented at the Quarta giornata italo-americana di dialettologia held in Padua on June 9, 1998, in classes at the University of Geneva in February 1999 (by Anna Cardinaletti), and at the Incontro di Grammatica Generativa held in Siena in February 1999. We would like to thank the audiences of these events for comments and criticism. We regret that for fortuitous reasons we could not deliver this paper at the Conference on semi-lexical heads in Tilburg for which it was prepared. We would also like to thank an anonymous reviewer whose punctual comments have helped us improve the final version. All the usual dismissals of the case apply here. Although the whole paper is a joint enterprise, for all academic purposes Anna Cardinaletti takes responsibility for sections 4-6, Giuliana Giusti takes responsibility for sections 1-3. 2. The proposal in (1) is a weak version of a more general proposal that does away with the lexical/functional distinction and takes all functional elements as lexical categories merged in the extended projection of a lexical item. In this way, functional elements lose some or all of their lexical properties while lexical categories remain the only categories in the lexicon. 3. An anonymous reviewer observes that the presence of the connecting element is further evidence against an analysis of (2)-(4) in terms of serial verbs. For tense and aspect specifications, cf. sections 3.5 and 3.7.1 below.

410 Anna Cardinaletti and Giuliana Giusti

4. In American English, the "V and V" construction is also found with aspectual verbs such as hurry up, try, be sure, etc., the "V V" construction is also found with hurry, try, stay, sit (cf. Shopen 1971; Carden and Pesetsky 1977). In Swedish, the inflected construction is found with many other verbs, such as the following, from Wiklund (1996): (i) Aspectual verbs: fortsätta 'continue', börja 'begin', sluta 'stop', ... Control verbs: se till 'make sure', glömma 'forget', pröva 'try',... Locative verbs: sitta 'sit', stä 'stand', ligga 'lie',... ta 'take' Verbs denoting channel for a speech act: skriva 'write', ringa 'phone', ... 5. The # diacritic in (10b), as well as in the rest of the paper, signals that the sentence is grammatical under the irrelevant interpretation as a coordination. 6. In (12), the complex preposition agghiri a, which can only be directional, clearly distinguishes between the two constructions. Any other locative preposition, such as a, da, in, etc., is ambiguous between a directional and a Stative interpretation. When these prepositions are found in the inflected construction, they can only follow the lexical verb and must be interpreted as introducing a Stative complement associated with the lexical verb. Cf. a casa in (i), which can only refer to the location where the event of eating takes place: (i) Va (*acasa) amangia {a casa). go-3s (*home) to eat-3s (at home) 'He goes to eat at home.' The same directional/stative ambiguity holds for the locative clitic ci. When it appears in the inflected construction, as in (ii), it can only refer to the locative complement of the lexical verb mangiari 'eat': (ii) Ci va amangia. there CL go-3s toeat-3s 'He goes to eat there.' 7. Notice that lexical stari, contrary to lexical iri, always requires a locative complement. This is why with stari, the clitic cluster is never obligatory (contrary to what we have seen with iri in (13a)), but always optional, as in (15a). 8. In (19a), the adjunct c'a machina follows the clausal complement a mangiari, as is always the case in the co-occurrence of adjuncts and complements: (i) Peppe va a casa c'a machina. Peppe go-3s to home by car 'Peppe goes home by car.' The clausal complement a mangiari in (19a) thus behaves like a PP complement, while the lexical verb a mangia in the inflected construction (19b) does not. 9. The present subjunctive and the future indicative are not different from the present indicative and cannot be tested. 10. This is also true of the occurrence of go in (ia), the "unexpected-event" construction discussed in Carden and Pesetsky (1977: 89), whose paraphrase is (ib):

"Semi-lexical" motion verbs in Romance and Germanic 411

(i)

a. He went and hit me. b. He up and hit me. Italian has a similar pseudo-coordination construction withprendere 'take': (ii) Prese e mi diede un colpo. took and to-me C L gave a punch 'He went and hit me.' 11. Here, we agree with Pollock's (1994: 304) criticism of Jaeggli and Hyams's (1993) treatment of American English come and go as aspectual auxiliaries. 12. This is confirmed by some related dialects; cf., e.g., Calabrese in (i), in which the connecting element of the inflected construction is the same as the coordinative conjunction (from Rohlfs 1969: §759): (i) Sutta a late finestra vegnu e staju. under the your window come-ls and stay-Is Ί come and stand under your window' 13. Some variation is found among speakers. According to David Pesetsky (p.c.), the "V and V" construction has an agentivity restriction as well. 14. The same holds for the other verbs which enter the inflected construction (cf. fn. 4). 15. We have no claim as to the order of these operations. 16. We abstract away from the presence of the connecting element which is irrelevant to our discussion here. We also disregard language variation as to whether the connecting element is present or not. 17. The discussion leads us to establish a further parallelism between motion verbs and auxiliaries. In the inflected construction, the motion verb must combine with a specific form of the lexical verb, namely an inflected form, and its merging point depends on the spell-out point of the lexical verb. Being the lexical verb inflected, the motion verb is also inflected and merged after the lexical verb has finished its overt checking. In this framework, we expect that auxiliaries, which combine with specific verbal forms (i.e., "have" combines with a past participle and "be" combines with the progressive form), are merged at different points, depending on the different morphological specifications of the lexical verb. A hint to this effect is provided by the different distribution of e.g. floating quantifiers with the Marsalese auxiliaries aviri and stari. Compare (54a) above, repeated here as (ia), with (ib): (i) a. Ci hannu (*tutti) statu (tutti). there CL have-3PL

b.

(*all)

stay-PAST.PART

(all)

'Everybody has been there.' Ci stannu (tutti) ennu (tutti). there C L stay-3PL (all) go-GER (all)

'Everybody is going there.' The different distribution of the floating quantifier suggests that the gerundive lexical verb ennu in (ib) is lower than the past participle statu in (ia), which implies that auxiliary stari in (ib) is merged lower than auxiliary aviri in (ia).

412 Anna Cardinaletti and Giuliana Giusti

Much like the motion verb in the inflected construction, the different merging point of the auxiliary is related to the different morphological specifications on the lexical verb. 18. The adverbs forse 'maybe', di solito 'usually* and spesso 'often' do not exist in Marsalese. 19. Notice that in (70) the connecting element a is missing. This is another signal that an imperative verb is higher than an indicative verb. 20. (75) and (77) raise a general question as to the nature of the second inflected verb in the inflected construction, which is not necessarily the lexical verb. In both cases, the second inflected verb is the causative, which in (75) embeds the lexical verb lavari 'wash' and in (77) embeds the andative verb iri which, in turn, embeds the lexical verb pigghiari 'fetch'. In Cinque's (1999b) analysis, both the causative and the andative are functional verbs merged in the extended projection of the lexical verb. Differently from the motion verb in the inflected construction, they are merged before the lower verb checks tense and person features. As a consequence, they retain the possibility to check tense and person features and to reach the spell-out merging point of the verb. The motion verb in the inflected construction is inserted after this point is reached. 21. This analysis is different from Pollock's (1994) in that we do not assume that the lexical verb incorporates into the motion verb. However, the motivation of the restriction to the base form reduces in both Pollock's analysis and ours to the impossibility for one of the two verbs or both to move to a higher functional head for checking. In Pollock's analysis, it is the motion verb which is blocked from checking its features due to the fact that it is part of an incorporated head. In our analysis, it is the lower verb which should check its features, and the impossibility to do so is reduced to minimality considerations. 22. Cf. also the perfect sentence in (i): (i) Go and not bother Harry. The construction try and allows a similar placement of the aspectual verb (from Carden and Pesetsky 1977: 90): (ii) ITry and not get lost. 23. Although feature copying is present in both Wiklund's (1998) analysis and ours, the two proposals are very different. Wiklund takes both the motion verb and the verb following it to be lexical verbs. According to Wiklund, the verb following the motion verb projects a reduced extended projection, which is a complement to the motion verb. The copying procedure goes the opposite direction with respect to our analysis: features of the motion verb are copied onto the verb following it. 24. The inflected construction thus provides definitive evidence that the movement of the inflected verb to the verb-second position (ending up in second position in e.g. declarative clauses and in first position in e.g. imperative clauses) is not triggered by feature checking. If it were, (87b) and (89b) should be grammatical.

"Semi-lexical" motion verbs in Romance and Germanic 413

References Abney, Steven P. 1987 The English Noun Phrase in its sentential Aspect. Ph.D. dissertation, MIT. Baker, Mark C. 1989 Object Sharing and Projection in Serial Verb Constructions. Linguistic Inquiry 20.4: 513-553. Calabrese, Andrea 1993 The sentential complementation of Salentino: A study of a language without infinitival clauses. In: Adriana Belletti (ed.), Syntactic Theory and the Dialects of Italy, 28-98. Torino: Rosenberg and Sellier. Carden, Guy and David Pesetsky 1977 Double-Verb Constructions, Markedness, and a Fake Co-ordination. Papers from the Thirteenth Regional Meeting of the Chicago Linguistic Society, 82-92. Cardinaletti, Anna and Giuliana Giusti 1998 Motion verbs as functional heads. GenGenP 6.1: 50-60 (revised version to appear in Christina Tortora (ed.), The Syntax of Italian Dialects. New York: Oxford University Press). Chomsky, Noam 1995 The Minimalist Program. Cambridge, Mass.: The MIT Press. 1998 Minimalist Inquiries: The Framework. MIT Working Papers in Linguistics 15. Cinque, Guglielmo 1999a Adverbs and Functional Heads. New York-Oxford: Oxford University Press. 1999b The interaction of passive, causative, and "restructuring" in Romance. To appear in Christina Tortora (ed.), The Syntax of Italian Dialects. New York: Oxford University Press. Collins, Chris 1997 Agreement Sharing in Serial Verb Constructions. Linguistic Inquiry 28.3: 461-497. Faraci, Robert 1970 And as a verb complementizer. I s ' NELS. Grimshaw, Jane 1991 Extended Projections. Ms., Brandeis University. Jaeggli, Osvaldo A. and Nina M. Hyams 1993 On the independence and interdependence of syntactic and morphological properties: English aspectual come and go. Natural Language and Linguistic Theory 11: 313-346.

414 Anna Cardinaletti and Giuliana Giusti

Josefsson, Gunlög 1991 Pseudocoordination - A VP + VP coordination. Working Papers in Scandinavian Syntax 47: 130-156. Larson, Richard 1988 On the Double Object Construction. Linguistic Inquiry 19: 335391. Lee, Sook-Hee 1992 The Syntax and Semantics of Serial Verb Constructions. Ph.D. dissertation, University of Washington, Seattle. Pollock, Jean-Yves 1989 Verb Movement, Universal grammar and the Structure of IP. Linguistic Inquiry 20: 365-424. 1994 Checking Theory and Bare Verbs. In: Guglielmo Cinque, Jan Köster, Jean-Yves Pollock, Luigi Rizzi and Raffaella Zanuttini (eds.), Paths Towards Universal Grammar. Studies in Honor of Richard S. Kayne, 293-310. Washington, D.C.: Georgetown University Press. Ritter, Elizabeth and Sara Th. Rosen 1996 Strong and weak predicates: Reducing the lexical burden. Linguistic Analysis 26.1-2: 29-62. Rohlfs, Gerhard 1969 Grammatica storica della lingua italiana e dei suoi dialetti, vol. 3: Sintassi e formazione delle parole. Torino: Einaudi. Ruggeri, Manuela R. 1999 Untitled term paper, University of Venice. Shopen, Timothy 1971 Caught in the act. Papers from the Seventh Regional Meeting of the Chicago Linguistic Society, 254-263. Vikner, Sten 1997 V°-to-I° movement and inflection for person in all tenses. In: Liliane Haegeman (ed.), The New Comparative Syntax, 189-213. London/New York: Longman. Wiklund, Anna-Lena 1996 Pseudocoordination is subordination. Working Papers in Scandinavian Syntax 58: 29-54. 1998 Morphosyntactic Parasites as Underspecified Heads: On the Double Supine Construction. Licentiatexamen Thesis, University of Umeä.

Underspecification in serial verb constructions Tjerk Hagemeijer

1.

Introduction

Serial verb constructions (henceforth SVCs) occur in several Asian (e.g. Mandarin), West-African (e.g. Yoruba) and Creole languages (e.g. Haitian). Despite the crosslinguistic differences due to language specific properties, at surface structure many similarities can be pointed out between serializing languages (Lord 1993). Benefactive constructions expressed by a (reanalyzed) verb meaning 'to give' or directional constructions involving a manner of motion and a directed motion verb are examples of constructions that occur in most serializing languages. In the present paper1, I will draw mainly on evidence from SVCs in Säo-Tomense (henceforth ST), a Portuguese-based Creole language spoken on the island of Säo Tome in the Gulf of Guinea, in order to show that in these constructions verbs are often defective with regard to their semantic and/or categorial selection properties. More specifically, it will be argued that verbs in the first position (henceforth Vis) are to some extent semantically unspecified, although categorially well specified. Here I draw a parallel between Vis and light verbs. I will argue that inside the verbal complex predicate the verb in second position (henceforth V2s) contributes semantic features lacking on VI. V2s, on the other hand, exhibit an unstable categorial label in some cases. Syntactic tests show that several lexical items with a verbal appearance do not carry a verbal categorial label. More interestingly, in some cases verbs in the V2 slot display hybrid categorial behavior between V and P. I will present some evidence supporting the claim that these lexical items are underspecified for the [V] feature in the lexicon. In the next section, a brief definition of SVCs in ST will be given; section 3 gives an outline of SVCs under a complex verbal predicate approach; in section 4, I will bring forth some arguments that support the semantic

416 Tjerk Hagemeijer

defectivity or semi-lexicality of verbs in the first position; section 5 then takes a closer look at the verbs in second position, i.e., reanalyzed and mixed lexical items; section 6, finally, gives an insight in the grammaticalization paths mainly based upon the evidence from section 4 and 5. 2.

Serial Verb Constructions

The next definition of SVCs will be used to set SVCs apart from other constructions in ST. (a) (b) (c) (d) (e) (f) (g)

two (or more) verbs behaving like a single event; one overt subject; one tense marker on V1; one negation marker on V1; only one or a repeated aspect marker; no subordinate or coordinate conjunctions; no discourse pauses.

In (l)-(4), some types of SVCs that occur in ST are presented. (1) (2) (3)

(4)

Bisu vwa subli. bird fly go.up 'The bird flew upwards.' Ome tufu djelu pe djibela d'e. man put money put pocket of-3SG 'The man put the money in his pocket.' Zon toma mantchin kota po Zon take machete cut tree 'Zon cut the tree with the machete.' Tlabado d'e ku po mata. worker give-3SG with stick kill 'The worker beat him to death with a stick.'

(directional SVC) (locative SVC) (instrumental SVC)

(resultative SVC)

Serializing languages mainly seem to differ among each other with regard to the tense and aspect criteria and the directionality of scope as a result of basic word order settings. Serializing SOV languages, like Ijo, are rare and although they govern to the left the order between events is iconic, i.e. the order between the verb phrases is not the inverted one of serializing SVO languages (Muysken 1987).

Underspecification in serial verb constructions 417 In a rigid SVO language like ST, the tense, mood, aspect (TMA) and negation nodes are always on VI with scope over V2. Aspect marker ka (generally used for habitual actions), however, can be placed on both verbs in a SVC in order to emphasize an iterative or durative reading, but crucially none of the other lexicalized functional nodes can be repeated.2 Yet, not always does there exist a semantically clear-cut distinction between only one or a repeated aspect marker, which might ultimately be seen as a tendency towards the weakening of V2's verbal status (cf. section 5). Nevertheless, the double aspect marking criterion is helpful in the sense that it precludes all apparently related constructions from being SVCs. Compare example (5) to (6)-(9).3 (5)

a.

b.

Zon ka dese ba poson. (SVC) Zon ASP go.down go city 'Zon uses to go down to the city of S. Tome.' Zon ka dese ka ba poson. Zon ASP go.down ASP go city 'Zon always goes down to the city of S. Tome.' fla. talk

(aspectual verbs)

(6)

Inen komesa *(ka) 3PL start (ASP) 'They started to talk.'

(7)

Ε dese Iwa *(ka) glita. 3SG go-down street ASP scream 'He went down the street screaming.'

(overlapping events)

(8)

Zwana manda Zon (*ka) tlaba. Zwana order John (ASP) work 'Zwana orders Zon to work.'

(causative verbs)

(9)

Zon ka tende Zwana (ka) fla. Zon ASP hear Zwana (ASP) speak 'Zon hears Zwana speaking.'

(perception verbs)

From the examples it follows that in the non-serializing constructions in ST the aspect marker is either obligatory or prohibited on the second verb. In constructions with perception verbs, aspect marking on V2 is optional, but crucially not interacting semantically with aspect marking on VI, since it does not induce the required iterativity effects.

418 Τ jerk Hagemeijer 3.

SVCs as verbal complex predicates

A great many different analyses for SVCs have been proposed, most of which have been analyses of coordination or subordination (either adjunction or complementation). Yet, there remains little doubt that crosslinguistically, or even language internally, SVCs cannot be simply subsumed under one and the same analysis. Perhaps the main factor of convergence consists in the tight and specific relation between both verbs. This section will therefore focus essentially on the complex predicate effects exhibited by these constructions, since the required close interaction between verbs forms the basis for the results I will arrive at later. In the first place, it has frequently been noted that the properties displayed by SVCs do not immediately support a coordination analysis. Ready argument extraction out of both conjuncts in (10) and (11) seems to enter in conflict with Ross' (1967) Coordinate Structure Constraint: (10)

[Ke kwa\i ku Zon koye tj pe kwali. what thing REL Zon pick put basket 'What did Zon put in the basket.'

(11)

[Andji]i ku Zon koye gweva pe tj. 'Where did Zon put the guavas.'

Ross' Constraint applies specifically to symmetric coordinations. If anything, SVCs are asymetric coordination structures. The strict order between both verb phrases and the c-command relations illustrate the tight organization these constructions obey. Veenstra (1997) claims that only an adjunct analysis is able to account for the different syntactic properties exhibited by SVCs. Crucially, c-command asymetries, adjunct extraction (cf. 12) and reflexivization (cf. 13) do not hold under a complementation or an analysis where the first verb phrase is adjoined to the second. (12)

a.

[De ke modo\\ ku bo toma kwa se tj pe Of which way KU 2SG take thing DEM put n'ala%? in-over.there 'How did you take that object and put it over there?'

Underspecification in serial verb constructions 419 b.

c.

(13)

Ν tom'e ku do-dosu mon. 1SG take-3SG with REDUPL-two hand Ί took it with both hands.' "N pe'e n'ala sosegadu achi. 1SG put-3SG in-there calmly like.this Ί put it over there in a gentle way.'

Zorii toma mina} PROj pya kala d'e plopi/»j/*k. Zon take child see faceof-3SG self. 'Zon recognized himself (his own face) through the child('s face).'

In example (12) it becomes clear that wh-extraction out of the second conjunct is worse, since (12c) is only a marginally accepted answer to (12a). The comitative take-serial in (13) has two possible binders c-commanding the reflexive anaphor d'e plopi (note that pronouns lack gender specification). Yet, there is an intervening element that assigns the preferred reading to the matrix subject and not the theoretically closest binder of the reflexive. Like most authors, I assume there is a PRO establishing a control relation between the matrix subject and the second verb phrase. Munn (1993) provides another important argument for an adjunction analysis, since it is claimed that coordination structures behave similar to parasitic gap structures. Therefore it is argued by this author that coordinations are best treated as adjunction structures. Additional evidence for adjunction structure comes from adverb placement. In ST, adverbs can be stacked between the two verb phrases in both the constructions and presumably delimit the right edge of the first verb phrase (cf. 14).4 These adverbs do not interfere with the observed free extraction patterns, in (15). (14)

Zon tomafaka {nifogon/ tres ola} va mpon. Zon take knife {in kitchen/three hour} cut bread 'Zon cut the bread with the knife in the kitchen/at three o'clock.'

(15)

Sa mporii se ku Zon tomafaka{nifogon/tres ola} va tj. 'It was the bread Zon cut with the knife in the kitchen/at three o'clock.'

Apparently, island violations still seem to contradict the extraction patterns out of the second verb phrase on either a coordinate or an adjunct account. However, as both Dechaine (1993) and Williams (1994) observe, in the specific case of asymmetric coordination the second verb phrase is trans-

420 Tjerk Hagemeijer

parent to extraction. Therefore, it seems that a asymmetrically coordinated adjunct is basically on the right track in order to explain many of the facts about SVCs, and especifically take-serials. Additional evidence for the specific status of SVCs comes from negation. Negation in ST consist of two disjoined particles, na...fa, one of them occurring in preverbal position and one globally in clause final position. As is shown, the negation particles have the whole construction in their scope (on the pretended single event reading). (16)

Zon na tomafaka (*fa) va mpon Zon NEG1 take (NEG2) cut bread 'Zon didn't cut the bread with a knife.'

(ku-e) fa. (with-3SG) NEG2

Although fa is able to cross for example complementizers introducing finite clauses without negating the subordinated clause, as in (17), this item will never cross clear-cut cases of overt and covert conjunction (cf. 18). Moreover, (19) shows that adjuncts are always in the syntactic scope of na...fa. (17)

Non na sebe kuma kwa se pasa fa. 1PL NEG1 know how thing DEM happen NEG2 'We don't know how that thing happened.'

(18)

Inenna kumefa, na bebe fa, dansa montchi. 3PL NEG1 eat NEG2 NEG1 drink NEG2 dance much 'They didn't eat, they didn't drink, (but) danced a lot.'

(19)

Inen na kota po ku-inen mantchin se n'obo fa. 3PL NEG1 cut tree with-3PL machete DEM in-jungle NEG2 'They didn't cut trees in the jungle with the machetes.'

Contrary to what Collins (1997) claims for Ewe and Campbell (1996) for Akan (and its dialects), ST cannot be considered an object-drop language. Examples with object-drop inside a verb of two and three places, in (20) and (21) respectively, and an instance of topicalization, in (22), crucially show that take-serials are distinct in not allowing an overt cliticized pronoun on V2 in order to recover the preceding object (cf. 23). (20)

Ch' inen kuji kalu, non ka {zug'e/*zuga} buta. If-3PL cook kalu 1PL ASP throw-3SG/throw throw 'If they cook kalu (typical plate), we will throw it away.'

Underspecification

in serial verb constructions

(21)

Ch' inen ka fe Icume, non ka da mina *{ele). If-3PL ASP make food, 1PL ASP give child 3SG 'If they prepare food, we will give it to the child.'

(22)

Inen livlu se, Zon {tchil'inen/*tchila} se 3PL book DEM Zon take/take-3PL without 'These books, Zon took them without paying.'

(23)

Kaso mode bisu {mata/*mat'e}. dog bite bird {kill/kill-3SG} 'The dog bit the bird to death.'

421

poga. paying.

Predicate cleft of either VI or V2 in SVCs is in most cases accepted. If V2 lost all or part of its verbal features, however, clefting becomes ungrammatical or highly marginal. It is important to note that, whenever instances of predicate or verb phrase clefting (cf. 24) or fronting (cf. (25)-(26)) occur, the fronted V is always recovered in the original sentence.5 We take this structure preserving constraint as another strong piece of evidence for the tight relation between both verbs in SVCs. (24)

Sa [va mpon] ku sun Pedu toma faka *(va). be cut bread REL mister Pedu take knife cut 'It was cutting the bread that mr. Pedu did with the knife.'

(25)

[Tlega san], ome ligi mina *{tlega). Hand.over lady, man lift.up child hand.over 'To the lady, the man gave handed over the child.'

(26)

[Bi fesa], inen migu kole *(bi). come party, 3PL friend run come 'From the party, the friends came running.'

A final argument for the specific relation between the two verb phrases comes from instrumental take-serials. If a discourse pause occurs, the pronoun on the clause final preposition has to agree with the sandwiched object (mantchin), identifying an instance of covert coordination (cf. 28); if not, this agreement is crucially lacking and the serial reading gets activated (cf. 27).

422 Tjerk Hagemeijer

(27)

Ome torn' inert mantchin se kota po (lcu-e/*ku-inen). Ome take-3PL machete DEM cut tree {with-3SG/with-3PL} 'The men cut the trees with the machetes.'

(28)

Tlabado torn' inen mantchin se, kota po (ku-ineti/*ku-e). 'The workers take the machetes and cut the trees with them.'

Implicit is the fact that verb compounding or at least verb adjacency at some level of the derivation are a wanted consequence in order to support any claim for complex predicates. Lefebvre (1991) argues that the lexical conceptual structures of both verbs are conflated in the lexicon (but crucially not in the syntax). Veenstra's (1996), however, postulates a deep structure adjacency requirement based on data from tonal sandhi in Saramaccan. Furthermore, Collins (1997) claims that in SVCs the second verb incorporates into the first one at LF, his proposal respects adjacency at some early level in the derivation, which in turn is similar to Larson's (1991) shelled proposal for SVCs. Yoruba's nominalizing double predicate cleft and Igbo's strict V-V compounding, rather than serialization, should be seen as intermediate stages between SVCs and (Bantu) incorporation (Baker 1991). 4.

Verbs in the first position as semi-lexical heads

In this section I will provide some evidence for the semi-lexical or defective semantic status of VI. The central claim will be that VI inherits (part of) its semantic structure from V2, the semantic head of the construction. In other words, VI is unspecified for certain semantic features, but crucially not for categorial ones, since it receives all the lexicalized functional nodes associated to verbs. It should be noted though that, according to the type of SVC, semantic feature un(der)specification can come in different ways. In the verbal domain, other instances of semi-lexicality can arguably be detected in Dutch and German V-to-V raising constructions (Van Riemsdijk 1998), with restructuring verbs latu sensu (Emonds; Haider; Cardinaletti and Giusti, this volume), or with light verbs (in the Grimshaw and Mester 1988 sense) forming a verb class distinct from both auxiliaries and main verbs (Butt and Geuder, this volume).

Underspecification in serial verb constructions 423 4.1.

Against the auxiliary status of VI

TMA in serializing (Creole) languages is generally expressed by means of preverbal functionalized lexical elements. These lexical items are often stored as "particles" without any further assumptions. A quick survey of ST, and presumably also other serializing languages, tells us that these particles do not constitute a homogeneous class when submitted to several syntactic tests. Like regular verbs, past tense marker tava 'had' (approx.) for nonstative predicates is able to occur as a copula or in constructions of verb ellipsis. Habitual aspect marker ka, however, does not occur in either of these concepts, while progressive aspect marker s(a)ka only accepts the copula environment. Further research is required in order to test if the designation lexical-functional element is sufficient to cover all these items as being distinct from purely (covert) functional, lexical, and semi-lexical categories. Verbs in the first position of SVCs cannot be treated as auxiliary verbs. Perhaps the most relevant test to demonstrate this is constituted by question-answer environments.6 It follows from the contrast between the SVCs in (29) and the auxiliaries latu sensu (TMA, aspectual verbs, motion verbs, etc.) in (30)-(32) that only in the former case both verbs are able to occur as an answer in a question-answer environment.7 (29)

a.

b.

(30)

a.

b. (31)

a.

b.

Zon kole hi ke? Zon run come home 'Did Zon run home?' Efan, e kole/bi/kole bi. 'Yes, he did.' Zon ka/ska kole? Zon ASP/ASP run 'Does Zon run' / 'Is Zon running.' *E ka/ska. Zon tava kole? Zon Τ run 'Zon had run?' Efan, e tava/*kole. 'Yes, he had.'

(SVC)

(Aspect markers)

424 Tjerk Hagemeijer

(32)

a.

b.

4.2.

Zon bi kume pichi? Zon come eat fish 'Did Zon come to eat fish?' Efan, e bi/*kume. 'Yes, he did.'

(motion auxiliary)

Light verbs

In the spirit of the Grimshaw and Mester (1988), light verbs are distinct from Chomsky's (1995) hypothesis, in which a small ν stands for null elements or (abstract) affixes attracting features. Grimshaw and Mester study the properties of Japanese suru, a full lexical item, which is claimed to be a thematically incomplete verb: it carries inflection and assigns accusative Case to a non-argument position. Through a process of Argument Transfer, light verbs become theta-marking and form a complex item with the item they Case-mark. According to these authors, light verbs show lexical variation. On their account, English do, for example, may be seen as an intransitive counterpart of Japanese suru. The following example was taken from Grimshaw and Mester (1988: 212): (33)

John-wa murabito-ni [[ookami-ga kuru-to]-o John-Top villager-to wolf-NOM come-COMP-GEN keikoku]-o shita. warning-ACC suru 'John warned the villagers that the wolf was coming.'

Keikoku is a noun meaning 'warning' and composes with shita (past form g of suru). The process of Transfer goes as follows. (34)

a. b. c.

keikoku (Agent, Goal, Theme) suru ( ) keikoku (Theme) + suru (Agent, Goal)

In a more radical version suru absorbs all the arguments of the noun. In other words, keikoku always loses argument positions to suru. Grimshaw and Mester (1988: 229) state that "[t]he Verb suru illustrates one kind of Light Verb: its argument structure is so highly underspecified that it is incapable of Θ-assignment of any kind. Other Light verbs, like saseru [causative form of suru], have a more fully specified argument structure:

Underspecification in serial verb constructions

425

incomplete, but with some arguments specified.". It is also presumed that this analysis extends to lexicalized expressions like idiom chunks. Hence, it readily follows that light verbs show different degrees of defectivity. In example (35) from Saramaccan, the light status of άά 'to give' results from a definiteness restriction on the noun it combines with. Veenstra (1996) claims that the NP in this construction has to be indefinite, otherwise the reading would be one of regular transfer of the knife. Syntactically, da behaves as a normal double object verb. (35)

Mi dä hen wan fäka. 1SG give 3SG DET knife Ί stabbed him.'

(Saramaccan; Veenstra 1996: 163)

In these crosslinguistically common verb-noun complexes, the lexically weakened verb leans on the semantic information of the noun. Instances of light Vis inside SVCs can readily be detected in ST. In some cases, these verbs form a unit with their internal argument without respecting the typical selection properties associated to them. Manner of motion verbs like kole 'to run' do not usually take direct internal arguments, but in (36) it combines exceptionally with the noun we 'eye(s)'. Yet, the ungrammatical examples in (37) show that the contribution of we transitivizes the complex predicate (cf. 38) rather than affecting the properties of one of the verbs only. (36)

Ε kole we dese. 3SG run eye go.down. 'He looked down.'

(37) *E kole we./*We dese./*E dese we. (38)

Ε kole dese. 'He ran down( wards).'

Hence, the data above do not immediately support a traditional small clause configuration, but rather support an analysis like the one proposed by Neeleman (1994), who treats verb-particle and verb-resultative constructions as complex predicates. In the spirit of his hypothesis, the prominent relation is between the verb and the non-verbal predicate (in English, Dutch, etc.) rather than between the internal argument of the verb and the non-verbal predicate. Hence, the verb and the non-verbal predicate form

426 Tjerk Hagemeijer

complex heads. If we extend this hypothesis to serializing languages, these are arguably distinct from non-serializing languages in Qallowing verbal f

predicates instead of the non-verbal ones referred above. Both verbs m SVCs would therefore have to be in adjacent position in deep structure or before spell-out, which corresponds to (39). To derive the surface structure, the first verb needs to move out of the complex V to a higher shell. VP

(39) DP

V' DP

V

we kole

subli

Verbal idiom chunks, as in (40), can also be subsumed under this analysis. (40)

Mwala bila we subli. Woman turn eye go.up 'The woman cursed.'

Butt and Geuder (this volume) treat a restricted class of verbs in Urdu as semi-lexical heads, i.e., light verbs forming complex predicates with another verbal head. It is argued that these verbs are lexically defective (but crucially still lexical) exhibiting a different behavior from both auxiliaries and main verbs. Since Urdu's complex predicates resemble SVCs in many respects, this approach to light verbs seems readily extendable to the latter constructions, since they propose a theory which accounts for the different properties of verbs within verbal complex predicates. Semantic defectivity may involve an incomplete theta-grid or simply absence of specific semantic features. The underlying idea is that verbs may not in all circumstances be full verbs in the different modules of grammar (e.g. restructuring environments), in a similar way nouns may not be fully specified in certain environments (compare verbal complex predicates to container noun constructions, etc.).

Underspecification in serial verb constructions ATI 4.3.

Take-serials

The properties of the take-verb in so-called take-serials differ from language to language. In some languages, like Mandarin (from Den Dikken and Sybesma 1998: 3) and Kwawu (Campbell 1996: 93), take-verbs are grammaticalized items. (41)

Zhang San ba beizi da-sui-le. Zhang San take cup hit-break-PRF 'Zhang san broke the cup.'

(Mandarin)

(42)

Me-de nwoma no maa Kofi.10 I-de book that gave Kofi Ί gave the book to Kofi.'

(Kwawu)

Despite the somewhat different details of their analyses, these authors converge on the fact that these take-verbs are "dummies" assigning accusative Case but no thematic role. Since it is documented that Mandarin's ba (and arguably also Kwawu's de) was once a verb, this means it lost all of its verbal properties. Yet, it seems natural to assume that they passed through different stages before full grammaticalization was achieved. Den Dikken and Sybesma (1998) and Den Dikken (1998) build on evidence from Mandarin and Fon to support their claim that these take-verbs are the overt representation of Chomsky's little v. Fon is indeed an interesting case where a take-verb (s6) exhibits a different behavior in and outside SVCs. Inside SVCs, this verb can select for abstract complements; outside of these constructions, this option is simply not available. Here, a totally distinct take-verb is required. Lefebvre (1991: 71) thus argues that the abstract complement fläse in (43) is not directly selected for by the take-verb. In other words, it can be deduced that it belongs to the argument structure of V2. (43)

Jacques Cartier {s6/ze} fläse yi Quebec. Jacques Cartier take French go Quebec 'Jacques Cartier brought French to Quebec.'

(Fon)

It is also shown by Lefebvre that in some cases where a Goal and a Theme argument occur Fon exhibits phenomena of clitic climbing. According to her, this is possible if the two verbs that form a SVC have a common thematic grid.

428 Tjerk Hagemeijer (44)

Koku so äso nä e. Koku take crab give her 'Koku gave her a crab.'

(Fon; Lefebvre: 1991:72)

(45)

Kdku. so e äsö ηά (e). Koku take her crab give (her) 'Koku gave her a crab.'

Therefore, the case of take-serials in Fon presents us with a double argument supporting the idea that VI is semi-lexical: the "shared" argument belongs to V2 rather than VI and clitic climbing, a typical property of Italian restructuring predicates (Rizzi 1982), is possible. It has also been noticed that take-verbs contribute an external argument to the clause. The following examples show that there exists a relevant difference between a SVC and its non-serial counterpart with regard to volitionality. In Mandarin, there is no restriction on the semantics of the external argument in the foz-constraction, which forms an additional argument for its high degree of defectivity, while take-serials in ST (or Fon) exhibit a clear volitionality constraint. The contrast is between (46) and (47)-(48). (46)

Feng ba shu gua-dao-le. (Mandarin; Sybesma 1992: 121) Wind take tree blow-fall-LE 'The wind has blown over the trees.'

(47)

{Zon/ventu} kebla lodoma. {Zon/wind} break bottle 'Zon broke the bottle.'

(on purpose or accidentally)

Zon/*ventu toma lodoma Zon/wind take bottle 'Zon broke the bottle.'

(on purpose)

(48)

kebla. break.

Several authors (Baker 1989; Campbell 1996; Collins 1997, etc.) claim that internal argument sharing by both verbs is an obligatory property of SVCs. In the spirit of Den Dikken and Sybesma (1998) and Campbell (1996), Mandarin's ba, Fon's so, and Kwawu's de do not select for an internal argument, i.e, this argument presumably originates from inside the second verb phrase.

Underspecification in serial verb constructions

429

In ST, take-verbs (and Vis in SVCs in general) still exhibit strong lexical behavior. Outside and inside SVCs, take-verbs select for direct objects or double object constructions with a Source and a Theme argument. In Hagemeijer (2000) it is therefore argued that Theme/Instrumental arguments are base-generated in this language and that the relation between the shared argument is mediated by an operator-variable chain, which follows essentially from evidence of preposition stranding and parasitic gaps. Moreover, I argue that instruments in take-serials should be considered Themes to VI but Instruments to the complex predicate. Yet, there are a few arguments suggesting that the first verb phrase is less a constituent than the second. In the first place, the second verb phrase - but crucially not the first - can be fronted. (49)

Va mpon, Zon toma faka να.11 Take bread Zon take knife cut 'Cut the bread, Zon did with a knife.'

(50) *Toma faka, Zon toma va mpon. Take knife Zon take cut bread In (51), ellipsis of the second verb phrase does not license the first verb phrase with its instrumental meaning, although the instrumental meaning comes about if V2 is repeated and followed by a regular instrumental PP (ku faka 'with a knife'). (51)

Zon toma mantchin kota po, magi Maya {*(kota) ku Zon take machete cut tree but Maya {(cut) with faka/Homa faka). knife/take knife} 'Zon cut the tree with a machete, but Maya (did so) with a knife.'

There is also a sharp contrast between "frontings" of the first and the second verb phrase in those constructions without instruments. (52) *Toma mpon, Zon toma va. take bread Zon take cut (53)

Va mpon, Zon toma va. 'It was cutting the bread Zon did.'

430 Tjerk Hagemeijer

Although take-verbs in ST are far more lexical than its counterparts in other serializing languages, it followed from the ungrammatical sentences above that the first verb phrase is considered less a constituent than the second. Hence, in the instrumental construction, the second verb phrase contributes to a change of the semantic information of the first verb phrase, which (partially) reanalyzes as the Instrument of the complex predicate. More generally, the semantic properties of the second verb phrase rather than those of the first derive these constructions.

4.4.

Evidence from motion SVCs

Another argument for the specific relation between the first and the second verb phrase comes from motion SVCs. These require a combination of two verbs sharing a [Motion] feature. In typical motion construction, VI is a manner of motion or directed motion verb and V2 a directed motion verb. The following table presents some combinations between intransitive motion verbs. Table 1. Motion SVCs: combinations based on intransitive verbs. Figure

Figure

12

V2: Dir. Motion+Path (+Deixis)

VI: Manner of Motion + kole, landa, nda, vwa, ... (run, swim, walk, fly)

be, ba, bi, subli, dese (go, go, come, go up, go down)

VI: Directed Motion + subli, dese (go up, go down)

V2: Dir. Motion+Path (+Deixis) be, ba, bi (go, go, come)

Ground ((Ρ) NP)

Ground ((P) NP)

Non-directional contexts with ba in V2 are ungrammatical under a serial reading, although they can be acceptable as instances of covert coordination with an intervening discourse pause (cf. 55). Therefore we are dealing with closed-class items in both slots. (54)

Zon kole ba he. Zon run go home 'Zon ran home.'

(55) *Zon {kume/fla} ba... (but: Zon kume, bake...) Zon {eat/speak} go

Underspecification in serial verb constructions 431 The altered semantic properties of VI become clear once we find the prototypical verb of directed motion ba in the V2 slot, since ba (to go) is crucially found in complementary distribution with be (to go). This is illustrated in Table 2. Table 2. Complementary distribution of ba and be 'to go'.

Be

Ba Ε ba [pp n'ala], 'he went over there.' Ε ba [ΝΡ&«]. 'He went home'

Ε be [pp d'ai]. 'He went from here.' Ε be [ppku bo\. 'He went with you.'

Ε ba [NP omali]. 'He went to the sea.' Ε be [pp d'omali]. 'He went by sea.' Ε ba [ppwe karu]. 'He went to the front Ε be [pp ni we karu]. 'He went in the seat of the car.'

front seat of the car.'

From Table 2 it follows that ba is used for contexts of motion with a specified Goal, while in the case of be the Goal is unspecified. In others word, a feature [a telic] is responsible for this distribution. I assume that ba is an unaccusative and be an unergative verb. Let us also assume that the pair ba/be is exceptional in the sense that there is overt morphology marked in the verb, whereas in all other cases the unaccusative/unergative distinction is not overtly marked. Additional language internal evidence for the unaccusative/unergative distinction comes from verb reduplication patterns in ST. Unlike unergatives (cf. (56)-(57)), unaccusatives (cf. (58)-(59)) cannot be reduplicated, which follows from their semantic nature. (56)

Zon landa-landa, so chiga kanwa. Zon swim-swim then arrive canoe 'Zon kept swimming and then made it to the canoe.'

(57)

Zon be-be, so chiga losa. Zon go-go then arrive plantation 'Zon kept going and then made it to the plantation.'

(58) *Zon mole-mole ku dolo muntu. Zon die-die with pain much (59) *Zon ba-ba poson. Zon go-go town. Furthermore, ba conflates the features [Directed Motion], [Deixis] and [Path], while be crucially has a negative value for the latter of these three

432 Tjerk Hagemeijer

features. If we assume that intransitive motion verbs are underspecified for [Path] in the lexicon, the derivation determines whether ba or be surfaces. This prediction is borne in the following examples, where adjacency determines whether ba or be is selected. (60)

Ε

ba [np ke

3SG go (61)

home

Zon] [pp ku- inen

Zon

mina se\.

with-3PL child DEM

Ε be [pp ku inen mina se] [np ke Zon].

Both: 'He went to John's place with those children.' [Path] corresponds to [Telic], a feature that is activated compositionally by the operation of merge with the adjacent constituent. Hence, the reduced tree structure in (44) accounts for feature matching between both verb phrases: after composition, the second verb phrase assigns a telic or nontelic role to the intervening aspectual node, where the value is picked up by the higher verb phrase. This approach predicts that in directional SVCs unaccusatives go strictly with unaccusatives and unergatives strictly with unergatives.13 I argue that a telic feature spreads from the second verb phrase to the telicity unspecified first verb phrase by means of regular percolation, as represented in (62). (62)

V'

AspP' [atel] Asp A

VP [atel] V'

Θ

V

The next two examples illustrate the difference between telic and non-telic events.

Underspecification in serial verb constructions 433

(63)

Inert ladlon se [_teiic kole [be]]. 3PL thief DEM run go 'The thieves ran away.'

(unerg. motion SVC)

(64)

Inert ladlon se [+teiiC kole [ba karu]]. 3PL thief DEM run go car 'The thieves ran to the car.'

(unacc. motion SVC)

Locative constructions constitute another subtype of complex motion predicates. I argue that locative V2 imposes a semantic constraint on VI, since it requires that a transitive verb with a dynamic or motion feature is in the VI slot. Then, as expected, a clear contrast arises in those cases where a directional and a non-directional interpretation can be obtained. Maurer (1999: 3-4)14 gives the following pair for ST. (65)

Ε soya kanwa pe matu. (directional) 3SG pull canoe put bush 'He pulled the canoe towards the bushes.'

(66)

Ε saya kanwa η matu. (non-directional) 3SG pull canoe in bush 'He pulled the canoe inside the bushes.' (i.e., the place where he does the pulling)

Given the constraints on VI, the verbs that may occur in the VI position belong to a restricted class. In section 5, it will be shown that locative pe is halfway being reanalyzed as a preposition. I will conclude this subsection with a short note on verb reduplication, a strategy we first presented in examples (56)-(57). Verb reduplication is a property of many (non-stative) predicates in ST, allowing predicates to obtain an iterative, continuative or focus reading. An example of these strategies can be seen in (67)-(69) respectively. (67)

Bendepanu posa-posa ni tudu floli ku-e kontla. Butterfly sit.down-sit.down in all flower REL-3SG find 'The butterfly sits down on all the flowers she finds on her way.'

(68)

Ν dumu-dumu andjin ante ku fomi da mu. 1SG pound-pound andjin until that hunger give 1SG Ί kept pounding andjin until hunger struck me.'

434 Tjerk Hagemeijer (69)

Ome pobli bi-bi kume. man poor come-come eat 'The poor man came exclusively for the purpose of eating.'

In SVCs, however, I could not find any instance of VI (or V2) reduplication. (70) *Zwana (saka) kole-kole be/ba ke. Zwana (ASP) run-run {go/go home} (71) *Zon (saka) saya-saya kanwa pe matu. Zon (ASP) pull-pull canoe put bush (72) *Mosu (saka) toma-toma inert faka se boy (ASP) take-take 3PL knife DEM

va mpon. cut bread

I assume this happens because the second verb phrase functions as the result or culmination point of the preceding event and therefore blocks reduplication. Verb reduplication is even ungrammatical in those constructions with repeated aspect markers which normally trigger the iterative/durative reading. In these cases it is likely that reduplication is precluded because the complex predicate remains resultative in nature in all the subintervals of the global aspect. Hence, reduplication may be considered an additional piece of evidence for the way the second verb phrase constrains the higher one. Still, it is not clear to me why the reduplicative focus reading is blocked.

4.5.

The semi-lexical configuration

In conclusion, Vis in SVCs belong to a restricted (or semi-open) class of verbs and are underspecified for different types of semantic information, although they maintain a great deal of their lexical properties. The following reduced tree structure represents the relation between both Vs in verbal complex predicates.15

Underspecification in serial verb constructions 435

(73)

TP T° (F/L?)

AspP

Asp°(F)

VP

V° (SL) A

AspP

Asp°(F) VP s-features V°(L) While comparing SVCs and direct partitive constructions in the nominal domain, which Van Riemsdijk treats as semi-lexical heads, it turns out that several parallels can be drawn. N2 in the latter constructions is defective for taking functional heads of Determiner/Quantifier type, whereas V2 is defective for tense, mood and negation. Moreover, both SVCs and partitive constructions are able to take adjuncts (adverbial and adjectival ones respectively). The indirect partitive construction studied by the same author then resembles covert coordination of verb phrases, where V2 is able to be marked for those nodes that are missing in SVCs proper. In the latter constructions, we are not dealing with semi-lexical heads. 5.

Verbs in the second position

In this section, a closer look will be taken at the properties of V2 in SVCs. This verb is defective because it is always in the scope of the lexicalized functional nodes on VI. For several reasons pointed out later on, the weakened verbal status of V2 leads to reanalysis. First I will present two cases of reanalysis and then study two cases of verbs that are only halfway a new categorial label.

5.1.

"Reanalyzed"

verbs

16

Instances of lexical items in the V2 slot with a phonetically verbal shape 17 which do not show any verbal behavior are readily identifiable.

436 Tjerk Hagemeijer

Pasa'to (sur)pass' Besides other functions (directional serials, main verbs), pasa occurs as a comparative or degree marker in respectively (74) and (75M76). (74)

Ope bo sa tamen pasa ope mu. foot 2SG be big surpass foot POS.1SG 'Your feet are bigger than mine.'

(75)

Kaso se sa bluku pasa. dog DEM be mean surpass 'That dog is extremely mean.'

(76)

Ome ba liba pasa. Man go up surpass 'The man went up very high.'

The non-verbal behavior of pasa in these examples follows from the impossibility of predicate cleft and aspect marking. (77) *Sa pasa ku kaso se sa bluku pasa. (78)

Ope bo sa tamen (*ka) pasa ope mu.

In contexts (75)-(76), pasa lost the capacity of taking an internal or external argument and modifies verbs, adjectives, prepositions or adverbs, which is the typical distribution of degree-words. In the view of Grimshaw (1991), degree elements are category neutral functional elements receiving its specifications by inheritance from the lexical head. In that particular analysis, degree elements are unspecified for the feature that distinguishes Adv from Adj. Since Degree-like pasa is a closed-class item modifying Adv, Adj, Ρ and V, I will adopt this category neutrality hypothesis, an idea that can be traced back to Lefebvre and Massam's (1988) claim made for the noun phrase domain in Haitian Creole. Comparative pasa in (74), on the other hand, still shows clear signs of transitivity and, given its non-verbal behavior, could then either be a preposition or a complementizer. Although pronouns are able to cliticize phonologically on prepositions and complementizers, the former option is preferred, since the argument of pasa is not part of an elliptic structure like in English and other languages. The contrast reads between (79) and (80).

Underspecification in serial verb constructions 437 (79)

Ζon sa tamen pasa mu (*sa). Zon be tall surpass lSG(be) 'Zon is taller than I.'

(80)

John is taller than I (am). Da 'to give'

This lexical item in the V2 slot behaves in all respects like a normal Case marking preposition. It introduces different types of thematic roles, like Source (cf. 81), Experiencer (cf. 82) Benefactive (cf. 83) and Goal (cf. 84). (81)

Inen mina se tava ka kole da koblo. 3PL child DEM Τ ASP run give snake 'Those children were running away from the snake.'

(82)

Fogon ka sa kentchi da non. kitchen ASP be hot give 1PL 'The kitchen is becoming hot for us.'

(83)

Zon tava ka tlaba da sun Glomo. Zon TNS ASP work give mister Glomo 'Zon used to work for mister Glomo.'

(84)

...so n'ga zuga vunvu se da nglentuke. ...then 1SG-ASP throw bee DEM give inside house '...then I will throw these bees inside the house.'

Like true prepositions, da in the environments above cannot be clefted nor receive any preverbal markers. In addition, unlike SVCs, subjects can be dropped in non pro-drop environments (cf. 85b) and null verb phrases are not allowed (cf. 85c), which is the typical distribution of prepositions. (85)

a.

Ε tlaba da sun Glomo? 3SG work give mister Glomo 'Did he work for Mr. Glomo?'

438 Tjerk Hagemeijer b.

c.

(*E) da sun Glomo? (3SG) give mister Glomo 'For Mr. Glomo?' Efan, e tlaba {*da/*d-e}. yes 3SG work {give/give-3SG} 'He did.'

subject-drop

null VP

This type of reanalysis is crosslinguistically very common in serializing languages. Mandarin for example has a gei (to give) as a normal main verb and a gei in serial-like constructions. Zhang (1990) argues that the former is a verb, since it receives aspect marking and can be stranded, while the latter 18

gei behaves like a regular preposition. This is roughly the same distribution we find in other Asian, West-African and Creole languages. Saramaccan is claimed to have at least three different uses of da ('to give'): preposition, verb and prepositional complementizer (Veenstra 1996). In ST, however, there is at least one syntactic test that provides us with a piece of evidence that da presumably did not start out right away as a fullfledged preposition. The relevant contrast is between (86), on the one hand, and (87)-(88) on the other. (86)

[Ke nge], ku Zon tlaba {datj*d'e}? which person KU Zon work {give/give-3SG} 'Who did Zon work for?'

(87)

[Andji]i ku mina-ome be {n'e, /*ni}? where KU child-man go {in-3G/in} 'In where did the boy go?'

(88)

[Ke kwa\ ku piskado bili vwado {ku-e[/*ku}? what thing KU fisherman open flying.fish {with-3G/with} 'What did the fisherman open the flying fish with?'

These instances of wh-movement into the specifier position of CP show that regular prepositions like ni (or its contracting variant η) and ku in respectively (87) and (88) are always stranded with an invariable spelledout trace with the phonetic shape of the third person singular pronoun.19 Da in (86) crucially lacks this spelled-out trace. We take this to be evidence of its former verbal status, since verbs also show up without this trace.

Underspecification in serial verb constructions 439 5.2.

Mixed items

At his point we will discuss several lexical items in the V2 slot that do not show a uniform behavior with respect to their categorial label. The discussion centers on the properties of V2 in locative and directional serials. It will be claimed that V2s in the former two constructions display simultaneously verbal and prepositional behavior.

5.2.1.

Locative serials

Tests concerning (repeated) aspect marking, verb cleft, subject-drop and null verb phrases show the following results when applied to pe. (89)

(90)

(91)

Zon ka bloka awa ka pe lata. Zon ASP pour water ASP put tin 'Zon always pours the water in the tin(s).' 11

Sa pe Be put a.

b.

(92)

a.

b.

ku Zon bloka that Zon pour

awa water

pe lata. put tin

aspect marking

verb cleft

Andji ku manad-e deta kabesa {pe/*pe-e}? Where that sister of-3SG put head {put/put-3SG} 'Where did his sister put her head?' Pe (liba) meza. subject-drop Put (on) table O n top of the table.' Ε bloka awa pe lata? 3SG throw water put tin 'Did he throw the water in the tin.' Ε blok'e *(pe). 3SG throw-3SG (put) 'He did.'

null VP

From the examples it follows that pe simultaneously shows verbal and prepositional features. On the one hand, evidence for the former behavior are context (89) and (92), where pe, like verbs, receives respectively aspect marking and, unlike prepositions, occurs together with VI in answers to global interrogative clauses. On the other hand, there is a parallel between pe and true prepositions because, similar to what happens to this latter cat-

440 Tjerk Hagemeijer

egory, predicate clefting is considered ungrammatical (cf. 90) and in answers to partial interrogatives pe occurs without the expected pronoun (cf. 91), similarly to da in (85b). Once more it needs to be pointed out that ST cannot drop subjects in these environments. Additional evidence for the hybrid status of pe comes from external argument sharing. If pe still is a verb, we expect the subject of the second verb phrase to be a PRO controlled by the matrix subject. If pe would behave like a prepositions, however, PRO is excluded. In (93), PRO is ccommanded by the matrix subject and the (human) Theme argument. Both are able to bind the reflexive. (93)

Zorii saya Mayaj pe nglentu Ice d'e plopiy}. Zon pull Maya put inside house of-3SG self 'Zon pulled Maya inside his/her own home.'

Thus, it seems that the verbal interpretation of pe triggers a PRO reading, while the prepositional interpretation precludes that same PRO reading. Partial categorial reanalysis is presumably eroding the PRO argument. With repeated aspectual markers, a verbhood test for V2, the prediction is that the Theme cannot bind the reflexive anymore. This, however, is not borne out. (94)

Zoni ka saya Maya) ka pe nglentu ke d'e plopiyj. 'Zon keeps on pulling Maya inside his/her own home.'

Since it is not directly relevant for the present purposes, I will not pursue whether we are dealing with a case of semantics preceding syntax or rather with a PRO argument that is able to take an arbitrary (but specified) referent.

5.2.2.

Directional serials

The same syntactic tests applied to pe in the previous section are now applied to ba. Note that the results also apply to other verbs of directed motion. (95)

Zon ka subli ka ba losa. Zon ASP go.up ASP go plantation. 'Zon always goes up to the plantation.'

aspect marking

Underspecification in serial verb constructions 441 (96)

Sa be ku Zon subli ba losa. verb cleft be go that Zon go.up go plantation. 'It was going up to the plantation that Zon did.' (approx.)

(97)

a.

Andji love subli be? Where that-3SG go.up go 'Where did he go up to?' b. *(E) ba losa. subject-drop (3SG) go plantation 'He went to the plantation.' (not: 'To the plantation.')

(98)

a.

b.

Ε subli ba losa? 3SG go.up go plantation 'Did he go up to the plantation?' Ε subli (be). 'He did (go up).'

null VP

From the grammaticality of verb cleft and the impossibility of licensing null subjects in question-answer contexts, it follows that ba is quite distinct from pe in the sense that it still exhibits strong verbal behavior. Nevertheless, below it will be claimed that directed motion verbs have a prepositional counterpart in cleft constructions. Comitative SVCs constitute some evidence for the fact that verbs of directed motion in ST are slightly affected from a categorial point of view. Comitatives typically make use of directed motion verbs (essentially 'to come' and 'to go') in the V2 slot. Although we do not predict a "hybrid" PRO interpretation along the same lines for what was observed with respect to pe, in (99) coreference of PRO with a Theme is somewhat marginally acceptable, despite the preference for the matrix subject control. (99)

5.3.

Mwalai toma omej bi ku-e ke d'e plopiyy,*k. Woman take man come with-3SG home of-3SG self 'The woman brought the man to her own place.'

Evidence for splitting features

There is some interesting evidence which seems to support the claim for the hybrid nature of the two items studied in the subsection above. I would like

442 Tjerk Hagemeijer

to argue that cleft constructions allow the prepositional features of pe to be split off of the verbal one at surface structure. (100) Sa be

[?ppefogo]ku ome ka tufu mon [Aspp [ASP (?) ka [yp [v pe]]]]. put fire that man ASP put hand ASP put

(101) Sa [ppnifogo] ku ome ka tufu djelu [Aspp Usp (?) ka [yp [v pe]\]]. both: 'It is in the fire that the man always puts his hands.' In the former section it was shown that pe is hybrid as to its categorial specification. Furthermore, clefted pe in (100) can be easily replaced by true preposition ni (cf. 101), which follows naturally if we claim that pe djibela is an instance of PP cleft. I assume that the clefted element in this construction has lost its dynamic or motion feature, which remains exclusively on the clause final item. That the latter item displays verbal behavior follows readily from the presence of aspect marker ka. Just like the case of ba, ni can be claimed to be the [Path] feature that was split off of V. Now let us try to extend this splitting hypothesis to motion verbs like ba. I will argue again that ?P cleft, in (102), show that it is possible to split the verbal or motion feature off of the inherently preposition-like path feature. (102) Sa [?Pbalosa] ku Zonka subli [ A S P [ A S P k a [ V P [ V · [ V ^ E ] ] ] ] ] . be go plantation that Zon ASP go.up ASP go 'It is upwards to the plantation Zon always goes.' P

Despite the fact that ba in SVCs resembles a preposition, I argued it still exhibits strong verbal behavior. However, earlier it was proposed that verbs of directed motion are [otPath] in the lexicon. At surface structure, a positive value rewrites as a incorporated Ρ feature, which crucially is lacking for be. (103)

ba V

be Ρ

V

Unlike the case of pe in (100)-(101), we do not dispose of an alternative construction for (103) with a true preposition. Yet, another verb of directed motion in (104) provides us with a piece of evidence which allows us to argue that the clefted ?P is reanalyzed as a PP.

Underspecification in serial verb constructions 443

(104) Sa [npsubli/liba oke] ku munken ka nda [Aspp [ASP ka [ v p J « W / ] ] ] . be go.up/up hill REL munken ASP walk ASP go.up 'It is up the hill that the munken (dove species) always walks.' The example above shows in fact that a motion verb in a clefted position can be readily replaced by a regular preposition (liba 'up'), similarly to what we claimed for pe. But how to deal now with the items left behind, i.e. the mirror image of (102)? (105)

Sa [v be] ku

Zon

ka

subli [ Asp p [A SP ka [ypba

losa]]].

be go that Zon ASP go.up ASP plantation. 'It is really going up to the plantation that Zon always does.' In both (102) and (105), V2 can be preceded by an aspect marker, which warrants the verbal features of the item in this position. I argued that be is verbal in nature and can be split off of ba losa. In this environment, clause final ba (losa) needs to be a V, because the V feature has to be preserved inside the complex predicate to guarantee that the first verb phrase can match its features with the lower one. Earlier on it was argued that da had grammaticalized. Therefore the prediction that it behaves prepositionally in both environments should be borne out in (106). (106)

Sa [pp da

sum] Glomo

ku

Zon

ka

tlaba (*ka) [p

Be give mr. Glomo that Zon ASP work (ASP) 'It is for mr. Glomo that Zon uses to work.'

da].

give

As we saw in a section 5.1, da lacks verbal features. Therefore the prediction that it cannot be preceded by aspect marker ka is borne out, since we expect da to behave as a preposition in both the clefted and the clause final position. If items in the V2 slot go through a complete process of reanalysis (cf. da), the complex predicate effects cease to exist. In consequence of predicate restructuring, VI will have to abandon its defective semantic status. Given the facts above, it follows that SVCs are potentially problematic for several claims which have been made concerning feature specification. If an item shows simultaneously prepositional and verbal behavior at surface structure, we have to find a mechanism to explain this distribution. On the classic account, Ps are [-N, -V] (Emonds 1985; Van Riemsdijk 1998)20 and Vs [-N,+V]. Therefore, the Ν feature unifies both categories.

444 Tjerk Hagemeijer

Hence, I claim that the V feature is unspecified, i.e. [aV]. If we then assume that a lexical insertion process in the course of the derivation determines the exact value of a , items like pe and ba can be stored in the lexicon without any further assumptions. The main verb counterparts of mixed items do not require a separate storage in the lexicon, since they can be derived by the very same mechanism. For a grammaticalized item like da, no additional claims are necessary, because it behaves prepositionally in any relevant context when found in the V2 slot of SVCs. Hence on a traditional account its features would be [-N,-V] in that specific position and [-N,+V] whenever it patterns as a main verb. In alternative, it might deserve some consideration to posit an [aV] feature for this and other items too.

6.

Grammaticalization paths

Table 3 centers on the shift from V to Ρ and summarizes several of the properties of the verbs we have been dealing with. Table 3. From verbs towards prepositions.

V, VI negation tense/mood markers aspect marker internal argument verb clefting pied piping VP/PP cleft + copy in situ stranding + trace null VP/PP in question-answer subject-drop in question-answer

V2-+P ba,be, etc.

pe

da

+

-

-

+

-

-

+

pasa

Ρ

-

-

-

-

-

-

-

-

-

+

+

+

-

-

-

+

+ +

-/+

+ +

+

+

??

+ +

+ +

+

-

+

+

-

-

-

-

-

-

-

+

n.a.

+/-

-

+

n.a.

+

-

+

+

-

+

-

-

Table 3 gives us the insight that each item exhibits distinct properties. Unlike in other languages, in ST any verb in the VI slot is still a full-fledged lexical verb which receives all the lexicalized functional nodes (TMA) and Case and, if transitive, theta-marks an internal argument. Verbs like pe and ba in the V2 slot are always preceded by verbs with a motion/(a)telic fea-

Underspecification in serial verb constructions 445

ture. Therefore, VI belongs to a semi-open class of verbs matching some of its feature(s) with the second verb phrase. Verbs in the second position cannot receive TMA or negation markers. It was shown that pasa and da cannot be treated as verbs. Locative pe, however, exhibits hybrid behavior, while verbs of directed motion were arguably closer to full-fledged verbs. I assume that the following strongly interrelated factors can be identified in the unidirectional process by which verbs in the V2 slot of SVCs start losing their verbal behavior: (a) Scope: all the (lexicalized) functional heads (TMA and negation) concentrate on VI and have simultaneously scope over V2; (b) Frequency: high utilization frequency of an item supports its grammaticalization (Hopper and Traugott 1993); (c) Tightness: the tighter the relation between two verbs, the higher the grammaticalization probabilities of V; (d) Availability: only a restricted number of verbs is selected for functional positions; (e) Superstate pressure: Portuguese uses non-verbal lexical items in the relevant environments; (f) Identification: reduced verbal morphology facilitates a categorial shift (of V2), since both V and Ρ are relational, assigning their internal Θrole directly. Therefore the status of V2 is of crucial importance to what is going on in SVCs. If V2 becomes fully recategorized, VI cannot maintain the unspecified semantic features it exhibited inside the verbal complex predicate due to predicate restructuring effects, i.e VP-VP structures become, for example, VP-PP structures. It is suggestive to conclude that there are several diachronic paths for a semi-lexical head: they may fully functionalize, which is arguably what happened in many cases to TMA markers, which in Creoles are often claimed to derive from verbs; they may receive full specifications again, i.e. lexicalize, when V2 becomes, for example, relabeled as a preposition; finally, they may remain semantically underspecified up to different degrees, like Mandarin's ba, Fon's so, etc. In the light of the latter scenario, a tendency towards full functionalization may be expected over time.

446 Tjerk Hagemeijer

(107) VI: a. lexical —» semi-lexical —> lexical-functional (TMA, Mandarin's ba, Kwawu's de) b. lexical —> semi-lexical —> semi-lexical (restructuring predicates, Urdu's light verbs, SVCs) c. lexical —> semi-lexical —> lexical (reanalyzed da in V2: VsL+VL becomes VL+P) (108) V2: lexical—» (lexical-)functional (pasa constructions) I will not address the complex issue concerning the status of prepositions, i.e whether they form (semi-)lexical and/or functional classes of items. This applies mainly to (107c) and (108). 7.

Concluding remarks

The main conclusion one draws from the facts described above is that SVCs as verbal complex predicates have significant effects on both verbs in these constructions. Unlike covert coordination structures, SVCs are combinations of lexically restricted classes of syntactic and semantic heads fusing two argument structures into a single one. The semantic head (V2) imposes s-selection restrictions on the syntactic head, which we argued to be semantically unspecified up to some degree. The syntactic head (VI), in turn, is able to weaken the verbal status of V2, which can ultimately become recategorized. Notes 1. I would like to thank Ines Duarte for her valuable comments, my informants, especially Jerönimo Pontes, for helping me out at any time, and an anonymous reviewer for several accurate comments. The source of this paper was within the context of the Subprogram of Science and Technology of PRAXIS XXI, Lisbon, Portugal. 2. Saramaccan, a Surinam Creole, shows a similar pattern as ST with regard to aspect marking in SVCs, but Haitian Creole, for example, does not allow aspect marking on the second verb. A crosslinguistic survey of SVCs arguably suggests that repeated tense marking is not a property of SVCs, but rather a symptom of covert coordination. Collins (1997) forcefully argues that in Ewe constructions with repeated tense markers are instances of covert coordination. In the spirit of this author, serializing languages are distinct from non-serializing

Underspecification in serial verb constructions 447

languages because, in the former, tense is able to license multiple verbal heads (see also Campbell 1996 for the same line of reasoning). 3. ST does not yet have an official orthography (Ferraz 1979: 56-7 for a brief proposal). I will not distinguish between low and middle vowels, since they are minimally contrastive; affricates are realized as dj and tch\ |J] is represented by ch before vowels and as s before consonants; nasals are represented as m (before bilabials) or η (other contexts); as for homorganics, I take the articulation point to be determining (ex. mpon 'bread'; nge 'person'); vowel drop resulting from contractions is represented by means of an apostrophe (ex. da+e [de] > d'e 'give him/her'); semivocalization as a consequence of contraction of two words or compounds are represented by means of a hyphen (ex. ku+e [kwe] > ku-e 'with him/her'); nearly all verbs receive their accent on the final syllable and other items usually on the last but one. 4. Campbell (1996) uses the adverb placement test in defense of a distinction between accusative and non-accusative SVCs, the latter being a typical ECM configuration. Crucially, in Kwawu, an Akan dialect, only the accusative (adjunction) type of SVCs allows for adverb stacking. 5. Prepositions undergo the same constraint, but the item that remains in situ always receives a spelled-out trace: (i) N' kanwa, piskado che η 'e. in-canoe fisherman leave in-3SG 'In the canoe, the fisherman took off.' In Hagemeijer (2000), some evidence is presented for the fact that in ST topicalizations (cf. (ii)) are base-generated, while it can be argued that cleft constructions (cf. (iii)) involve instances of real movement. (ii) Zon, bo kunda kuma/*ku- e mata ploko. Zon, 2SG think that/that-3SG kill pig 'Zon, you thought he killed the pig.' (iii) Sa Zon; ku bo kunda ku/*kuma tj mata ploko. 'It was Zon you thought that/that killed the pig.' Like most or all verbs of mental activity, kunda 'to think' takes complementizer kuma. Whenever (wh-)arguments are moved out of the embedded clause, the extracted item moves through SPEC,CP where it agrees with the head kuma, which undergoes a phonetic change. Old Irish, Ewe or French are examples of languages with similar phenomena. 6. Note that the possibility of predicate cleft of VI and V2 (separately) distinguishes SVCs from most but not all constructions with a tendency to monoclausal behavior. 7. In some locative constructions with pe 'to put', both verbs obligatory co-occur in answers: (i) a. Sun tufa djelu pe djibela? gentleman put money put pocket 'Did the gentleman put money in his pocket?'

448 Tjerk Hagemeijer

b.

Efan, e tufü *(pe). 'Yes, he did.' 8. Grimshaw and Mester, p. 212. 9. Larson (1991) claims that serializing languages allow for verbal secondary predicates. 10. Campbell mentions that de is sometimes glossed as 'to take'. 11. It should be noticed that verbs like να 'to cut' do not take obligatory instrumental adjuncts nor do they license double object constructions, i.e. instruments like faka 'knife' are not part of the (explicit) argument structure of these verbs. 12. Table adapted from Winford (1993) for English based Caribbean Creoles. 13. Collins (1997: 468, footnote 12) briefly refers that manner of motion verbs may become unaccusative when telic. Similarly, in languages like Dutch and Italian, telicity determines the selection of perfective auxiliaries in motion predicates. Arguably, perfective auxiliaries are also semi-lexical in nature. 14. The original glosses and translations are in Spanish. 15. It could very well be the case that tense marker tava heads an independent verb phrase, since exceptionally some adverbs like semple 'always' may occur in between the tense phrase and the aspectual phrase (cf. also section 4.1). 16. The quotation marks mean to illustrate that it is unclear whether the items in this section were reanalyzed or not from a diachronic point of view. 17. It is important to note that all V2s in SVCs also show up as regular main verbs. 18. In South East Asian languages, grammaticalized V2s are usually referred to as co-verbs. 19. At first hand it might look as if we are dealing with a resumptive strategy, but from the next examples taken from Hagemeijer (2000) we learn that number agreement (gender is not marked on pronouns) with the antecedent forms the characterizing property of this strategy. For all other cases, I claim that preposition stranding is involved. (i) [Inen nge se [CT ku bo fla {ku-e/*ku-inen}]] [{se konse /*kons'inen}]. 3PL person DEM REL 2SG talk {to-3SG/to-3PL} {without know/know3PL} 'The persons you talked to without knowing.' (ii) [Inen mwala se [cpku η tende [NPome [cpkufla {n'inen/*n'e}]}]] 3PL women DEM REL 1SG hear man REL talk {in-3PL/in-3SG} [{je kons'inen/*konse}] {without know-3PL/know} *'The women that I heard the man that talked about them without knowing them.' I claim that (i) is the stranding context. Not only is agreement lacking between the relativized noun and the spelled-out trace on the preposition, but also the traditional adjunction analysis predicts that the (for the case spelled-out) variable stranded after operator movement licenses parasitic gaps (Chomsky 1986; Cinque 1990, etc.). In the situation in (ii) an extra CP was embedded to

Underspecification in serial verb constructions 449

create a complex NP island. Since we are dealing with a strong island, whmovement is blocked and the only way to rescue the derivation is by inserting a resumptive pronoun. This pronoun agrees in number with the relativized noun and crucially does not license a parasitic gap as in (i). 20. Grimshaw (1991), however, claims that Ps are [+N,-V] forming extended projections with DP and NP.

References Baker, Mark 1988 1989

Incorporation. Chicago: University of Chicago Press. Object Sharing and Projection and Verb Serialization. Linguistic Inquiry 20: 513-553. 1991 On the Relation of Serialization to Verb Extensions. In: C. Lefebvre (ed.), Serial Verbs: grammatical, comparative and cognitive approaches, 79-102 Amsterdam/Philadelphia: John Benjamins. Butt, Miriam and Geuder, Wilhelm 2001 On the Lexical Status of Light Verbs. Paper presented at the workshop on Semi-Lexical Heads. Tilburg University. Campbell, Richard 1996 Serial Verbs and Shared Arguments. The Linguistic Review 13: 83118. Chomsky, Naom 1986 Barriers. Cambridge, MA: MIT Press. 1995 The Minimalist Program. Cambridge, MA: MIT Press. Cinque, Guglielmo 1990 Types of A '-Dependencies. Cambridge, MA: MIT Press. Collins, Chris 1997 Argument Sharing in Serial Verb Constructions. Linguistic Inquiry 28: 461-497. Dechaine, Rose-Marie 1993 Predicates across Categories: towards category-neutral syntax. Ph.D. Dissertation, MIT. Den Dikken, Marcel 1998 Take Serials Seriously, Please. Paper submitted to the celebration of Chomsky's 70th anniversary http://mitpress.mit.edu/chomskydisk/dikken.html. Den Dikken, Marcel and Sybesma, Rint 1998 Take Serials Light up the Middle. Manuscript.

450 Tjerk Hagemeijer

Emonds, Joseph 1985 A Unified Theory of Syntactic Categories. Dordrecht: Foris. 2001 The Flat Structure Expression of Semi-Lexical Heads. Paper presented at the workshop on Semi-Lexical Heads. Tilburg University. Ferraz, Luiz Ivens 1979 The Creole of Säo Tomi. Witwatersrand University Press: Johannesburg. Grimshaw, Jane 1991 Extended Projection. Unpublished manuscript. Brandeis University. Grimshaw, Jane and Mester, A. 1988 Light Verbs and Θ-Marking. Linguistic Inquiry 19: 205-232. Hagemeijer. Tjerk 2000 Serial Verb Constructions in Säo-Tomense. Unpublished manuscript, Universidade de Lisboa. Haider, Hubert 2001 Heads & Selection. Talk given at the workshop on Semi-Lexical Heads. Tilburg University. Hopper, Paul and Traugott, Elizabeth 1993 Grammaticalization. Cambridge University Press: Cambridge. Larson, Richard 1991 Some Issues in Verb Serialization. In: C. Lefebvre (ed.), Serial Verbs: grammatical, comparative and cognitive approaches, 185211. Amsterdam/Philadelphia: John Benjamins. Lefebvre, Claire 1991 Take Serial Verb Constructions in Fon. In: C. Lefebvre (ed.), Serial Verbs: grammatical, comparative and cognitive approaches. Amsterdam/Philadelphia: John Benjamins. Lefebvre, Claire and Massam, Diane 1988 Haitian Creole Syntax: a Case for DET as Head. Journal of Pidgin and Creole Languages 3: 213-243. Levin, Beth and Rappaport, Malka 1995 Unaccusativity. Cambridge/London: MIT Press. Lord, Carol 1993 Historical Change in Serial Verb Constructions. Amsterdam/ Philadelphia: John Benjamins. Maurer, Philippe 1999 El Verbo Locativo poner en Santomense, Principense y Angolar. In: K. Zimmermann (ed.), Lenguas Criollas de Base Lexical Espaiiola y Portuguesa. Frankfurt/Madrid: Vervuert/Iberoamericana.

Underspecification in serial verb constructions 451

Munn, Alan 1993

Topics in the Syntax and semantics of Coordinate Structures. Ph.D. Dissertation, University of Maryland. Muysken, Pieter 1987 Parameters for Serial Verbs. In: V. Manfiredi (ed.), Niger-Congo Semantics and Syntax 1, 65-76. Boston: Boston University-African Studies Center. Riemsdijk, Henk van 1998 Categorial Feature Magnetism: the endocentricity and distribution of projections. Journal of Comparative Germanic Linguistics 2: 148. Rotterdam: Kluwer. Rizzi, Luigi 1982 A Restructuring Rule. In: Issues in Italian Syntax. Dordrecht: Foris. Ross, John Robert 1967 Constraints on Variables in Syntax. Ph.D. Dissertation. Indiana: University Linguistics Club. Sybesma, Rint 1992 Causatives and Accomplishments: The Case of Chinese 'ba\ Ph.D. Dissertation. Leiden: HIL. Veenstra, Tonjes 1996 Serial Verbs in Saramaccan. The Hague: HDL. 1997 X'-Configurations and the Status of Shared Arguments in Serial Verb Constructions. In: R.-M. D6chaine and V. Manfiredi (eds.), Object Positions in Benue-Kwa. The Hague: HIL. Williams, Edwin 1994 Thematic Structure in Syntax. Cambridge/London: ΜΓΓ Press. Winford, Donald 1993 Directional Serial Verb Constructions in Caribbean English Creoles. In: F. Byrne and J. Holm (eds.), Atlantic Meets Pacific, 183-203. Amsterdam/Philadelphia: J. Benjamins. Zhang, Shi 1990 Correlations between the Double Object Construction and Preposition Stranding. Linguistic Inquiry 21: 312-316.

Part IV Semi-lexicality of Adpositional and Adposition-like Elements

Semi-lexical heads in a semantically charged syntax Kristin M. Eide and Tor A. Afarli

Bowers (1993) proposes that the subject-predicate relation is formed by a predication operator (op.), i.e. a function from a property to a propositional function or predicate. This operator heads an independent functional predication projection, as shown for instance in (la), and it is also lexicalized in certain cases, for instance as in (lb). (1)

a. b.

...make [PrP John [Pr' [Pr op.] [AP crazy]}] ... regard [PrP John [Pr' [Pr op./as] [AP crazy]]]

In this paper1 we will first show, mainly using data from Norwegian, that several types of element may lexicalize the predication operator (section 1). Next, we investigate how meaning is constructed - i.e. how a given projection is determined as to its syntactico-semantic content - in a situation where a semantically uniform operator like the predication operator is variously lexicalized by different visible elements, each with its own specific amount of inherent content (section 2). We then go on to argue that a given visible element is often multifunctional in that it may potentially lexicalize different types of functional operator (section 3). Last, we propose that syntactic representations should be construed as structured objects essentially consisting of functional operators that are made visible by various types of head by insertion and movement. Thus, the visible heads are semilexical in virtue of supporting an underlying functional operator. In effect, this amounts to suggesting a program for a semantically charged syntax (section 4).

1.

Lexicalization of the predication operator

The Norwegian counterpart to as in structures like (lb) is som. Thus, we find structures like (2), where we assume that som is the lexicalization of the predication operator.2

456 Kristin Μ. Eide and Tor A. Äfarli (2)

a.

b.

...anse [Jon [som gal]] ...regard Jon as crazy '...regard Jon as crazy.' ...anse [Jon [som forbryter]] ...regard Jon as criminal '...regard Jon as a criminal.'

The bracketed part of the strings in (2) has the representation shown in (3), where XP is the property phrase which is turned into a predicate by the predication operator lexicalized by som. That is, the claim is that the bare adjective/noun in small clauses like (2a,b) is a property phrase that is turned into a predicate by the predication operator, cf. Bowers (1993: 595-597, 647-653); Eide and Äfarli (1999: 156-159). PrP

(3) DP Jon

Pr' Pr som

XP gal/forbryter

In other words, the constituents under consideration here are small clauses, expressing propositions constituted by a predicate (Pr1) and a predication subject (DP), which is indeed in accordance with the semantic intuition that we have regarding these constituents. Additional motivation for the claim that we are here dealing with small clauses is the fact that these constituents may contain an expletive or expletive-like subject, as shown in (4).3 (4)

a. b.

...anse [det [som altfor kaldt for skigaing i dag]]. ...regard it as much-too cold for skiing to day ...anse [det [som uheldig at Jon vil komme]]. ...regard it as unfortunate that Jon will come

Following Eide (1998a, 1998b), Eide and Äfarli (1999) we furthermore assume that the copula may lexicalize the predication operator in environments where a verbal head is required, which amounts to saying that the particle som is a non-verbal counterpart to the copula. The feature that separates the copula from som is exactly the verbal feature [+V] (and thus the capability of supporting tense or other verbal morphology); in all other

Semi-lexical heads in a semantically charged syntax 457

respects the copula-expression and the jom-expression seem to pattern together: regarding case, agreement and type of complement selected, see the discussion in Eide and Äfarli (1999: 164-170). Generally, the structural and semantic similarities between copula-expressions and som-expressions are indicated by the fact that a small clause headed by the particle som can as a rule be paraphrased as a full clause with the particle replaced by the copula; compare the examples in (5) with those in (4). (5)

a. b.

...mene at [det [er ...think that it is ...mene at [det [er ...think that it is

altfor kaldt for skigaing much-too cold for skiing uheldig at Jon vil unfortunate that Jon will

i dag]]. to day komme]]. come

Thus, we conclude that in copula constructions the copula lexicalizes the predication operator; i.e. copula constructions have the structure (3) with som replaced by the copula. However, it seems that the predication operator can have still other lexicalizations. Specifically, we propose that til 'to' in resultative small clauses like (6a) and for 'for' in small clauses like (6b) are lexicalizations of the predication operator. As such, we refer to til and for as prepositional predication particles (see also Eide 1998b: 71-72; Eide and Äfarli 1999: 170). (6)

a.

b.

...gj0re [Jon [til forbryter]]. ...make Jon to criminal '...make Jon into a criminal.' ...ta [Jon Ifor keiner]]. ...take Jon for waiter '...take Jon for being a waiter.'

The til-type exemplified in (6a) is especially interesting for present purposes. Indication that theft'Z-expressionin (6a) is indeed a small clause predicate is the fact that this expression (like small clause predicates headed by som) licenses an expletive-like subject, indicating that the bracketed part in (6a) is a clausal expression, cf. (7).4 (7)

...gj0re [det [til noe skittent at jeg sa dette]]. ...make it to something dirty that I said this '...make it into something dirty that I said this.'

458 Kristin Μ. Eide and Tor A. Afarli

Notice also that the noun complement in (6a) is bare, indicating that the complement is the property phrase of a small clause (like e.g. in (2b)), not the referential complement of a preposition. In the latter case, being a referential argument, the noun phrase is as a rule not bare: (8)

Vi snakket til en forbryter /forbryteren / *forbryter. we talked to a criminal / criminal-the / criminal

Like som, the prepositional predication particle til seems to have a verbal counterpart, namely bli 'become'. This is illustrated by the following Swedish sequence from Selma Lagerlöfs famous novel about Nils Holgersson: (9)

Pojken künde rakt inte forma sig att tro, att han hade blivit förvandlad till tomte [...] orn jag väntar ettpar ögonblick, sä blir jag nog människa igen. 'The boy could not get himself to believe that he had been turned into a goblin [...] if I wait just a couple of moments, then I will surely become a human being again.'

We propose that what distinguishes til from som is an inchoative/directional feature (ID feature). Similarly, we assume that the same ID feature distinguishes bli 'become' from vcere 'be'. Thus, som and vcere are [-ID], whereas til and bli are [+ID]. Verbs like gj0re 'make' or forvandle 'turn' that select inchoative non-verbal complements select the riZ-type small clause, cf. (6a), (7), (9), whereas verbs like anse 'regard' that select non-inchoative non-verbal complements select the iom-type small clause, cf. (2), (4). 2.

The content of the Pr-projection

The underlying predication operator contributes what might be called "predicative content" to the Pr-projection it heads. This accounts for the propositional interpretation of the PrP, as well as observable effects like the possibility of having an expletive subject in its specifier position and a bare noun in its complement position. However, the content of the projection as a whole is also partly constituted by the inherent content of the element that lexicalizes the operator. In cases where the operator is lexicalized by som or the copula, the semantic contribution of the lexical element seems to be

Semi-lexical heads in a semantically charged syntax 459

quite small, and the element is little more than a structural marker of the underlying semantic operator.5 However, this is not so in cases where til lexicalizes the operator. In such cases, the operator and the lexical element each contributes significantly to the apprehended meaning of the projection. As suggested in the previous section, the particle til, which otherwise clearly functions as a preposition, carries an inchoative/directional meaning, and when this element is inserted into the head position of a PrP, the result is the amalgam [pred.op. & inchoative/directional], which accounts for the resultative reading of the small clause PrP. This means that the apprehended meaning of the projection is determined by two separate, but interacting components: the meaning of the underlying predication operator and the specific inherent meaning of the lexical element that makes the operator visible. However, the apprehended meaning of a projection is not exhausted by the content of the operator and the content of the visible element. We also claim that the complement of a head typically plays a crucial role in determining the reading of the head itself, and its projection (see Pustejovsky 1995 for a closely related idea). This point of view implies a dynamic interpretation of the principle of compositionality (Frege's Principle), i.e. the principle that the meaning of the whole is a function of the meaning of the parts and their mode of combination, cf. (10). The plasticity introduced by the idea that the complement of the head influences on the semantic reading of the head itself, implies a notion of a dynamic semantic enrichment of the basic compositional architecture, cf. (11). (10)

Compositionality: "[T]he parts" referred to in the statement of Frege's Principle must be the syntactic constituents of the expression in question. Moreover, the meanings of those constituents must enter into the meaning of the whole expression in a fixed way, determined once and for all by the semantic rule corresponding to the syntactic rule by which those constituents were joined. (Dowty, Wall and Peters 1981: 9)

(11)

Dynamic enrichment: The meanings of the parts enter into the meaning of the whole expression such that the meaning of a given head may be affected by the meaning of the constituent with which it combines.

460 Kristin Μ. Eide and Tor Α. Äfarli

To show how this works, we will use the notion of dynamic enrichment to explain the different readings associated with the copula, namely the pure predicational reading, the equative reading, the existential/spatial reading, and the temporal reading, see (12). (12)

a. b. c. d.

Clark Kent is a man. Clark Kent is Superman. Clark Kent is outside. Superman is tomorrow.

(pure predicational reading) (equative reading) (existential/spatial reading) (temporal reading)

In fact, the verb is frequently called the copula in (12a) only, but our claim is that (12) contains four occurrences of the same verb, the different readings of the verb being due to the semantic nature of the complement.6 In (12a) the complement of the copula is a man, a phrase that denotes a property. The predication operator, lexicalized by the copular verb, turns this phrase into a predicate. The relation is depicted in (13). In (12b) the complement of the copula is a phrase denoting a referent, and the only likely copular relation between this referent and the subject referent is the relation of identity, depicted in (14). In (12c) the complement of the copula is a phrase denoting a place, and the verb gets its spatial reading via interaction with the spatial meaning of the complement, depicted in (15). (13)

copular relation: pure predication copula is

(14)

copular relation: identity copula is

(15)

property a man

referent Superman

copular relation: spatial copula is

place outside

Thus, the apprehended meaning of the copula in a given sentence is determined not only by the meaning of the underlying predication operator and

Semi-lexical heads in a semantically charged syntax 461 any inherent meaning the copula might have. It is also enriched in crucial ways as a result of the interaction of the copula with its complement.7 Now, turn to (12d), repeated here as (16). (16)

Superman is tomorrow.

(temporal reading)

In this case, dynamic enrichment yields a temporal reading of the copula, as indicated. However, this example is particularly interesting because it shows the importance of encyclopedic knowledge in the determination of the semantics of a given string. In (16), the complement of the copula is a phrase denoting a point in time. Therefore, the phrase consisting of the copula and its temporal complement yields the reading takes place tomorrow. But in combining this relation with the subject Superman, our knowledge of the world kicks in and tells us that the word Superman in this case cannot possible denote the referent Superman. A referent couldn't possibly be something that takes place tomorrow - events take place, not objects or persons. As a consequence, the straightforward referent interpretation is very unlikely and is normally rejected. The only way we can make sense of this sentence is to assume that Superman refers to an event, for instance the event of showing the film about Superman. This particular example can serve as a reminder that language is never used in a vacuum, but for communicative purposes in a given discourse, against a vast background of encyclopedic knowledge (knowledge of the world). This knowledge narrows down the list of possible meanings relevant in a given context, and turns communicative use of language into more than a fairly educated guess.

3.

Multifunctionality and support of operators

In section 1, we saw that the predication operator can occur in various disguises in Norwegian: the particle som, the copular verbs vcere, bli, and the prepositional predication particles til, for. Given the basic predicational content of the projection headed by these elements, this variation suggests - as we have in fact been assuming - that the syntactico-semantic substance resides, not primarily in the supporting element, but in the abstract operator made visible by the element. The assumption that the content of a functional projection is not first and foremost derived from the visible element filling the head position, is

462 Kristin Μ. Eide and Tor Α. Äfarli further corroborated by the examples in (17), showing structures containing Old Norse relative clauses (from Iversen 1972: 153). (17)

a. b. c. d.

kringla heimsins su er mannfolkit byggir. 'the world that men live in.' pau helgu ord en ί bokinni varu. 'those holy words that were written in the book.' gera Ms par sent eigi hafi jyrr vent. 'build a house where there had been no house before.' peir allir, at flau tidindi heyrdu. 'everybody that heard these news.'

As these examples show, the complementizer position (C-position) of relative clauses in Old Norse is possibly made visible by different complementizers. Although there are differences as to frequency and the geographical and historical distribution of these complementizers (Äfarli 1995b), more than one type often occur in the same text. Moreover, there is no reason to assume that the C-projections of these clauses have different content, the complementizer ostensibly functioning as nothing more than a marker of the underlying head of the C-position - which in all these cases might well be a higher predication operator, cf. the suggestion in Chomsky (1982: note 11, 92-93) that the relation between a relative clause and its antecedent is predicational. Now, if it is really the case that the visible element marking a functional head position should be viewed primarily as a positional marker, we should in fact expect that the very same element should possibly be able to make different underlying semantic elements or operators visible. This is so since the sparse inherent content that a given visible functional element often has, is presumably compatible with having different syntactico-semantic roles. This expectation seems to be fulfilled. Thus, som may be used e.g. as a question complementizer or comparative particle in addition to its use as a predication particle, see (18): (18)

a.

b.

Vi sp0r hva som har skjedd. (question complementizer) we ask what som has happened 'We ask what has happened.' Han snakker sä fort som henne. (comparative particle) he speaks as fast som her 'He speaks as fast as she does.'

Semi-lexical heads in a semantically charged syntax 463 c.

Vi betrakter ham som sjef. we consider him som boss 'We consider him the boss.'

(predication particle)

The syntactico-semantic function or content of som does not seem to be the same in these instances, as indicated in the parentheses, and also by the translations (in particular, som does not seem to lexicalize a predication operator in (18a,b)). Likewise, vcere 'be' may be used as an ordinary auxiliary verb in addition to its use as a copular verb, and til 'to' and for 'for' may function as ordinary prepositions, in which case they do not lexicalize a predication operator. That is, the very same element appears to be capable of being "recycled" in different syntactico-semantic roles, possibly to instantiate different operators. In other words, overt forms are in principle multifunctional. Assuming that syntactic elements generally make functional operators visible by "supporting" them,8 and furthermore assuming that the content of any functional projection is at least partly constituted by the content of the operator and partly by the content of the supporting element or marker, it follows that the supporting element always underdetermines the content of the projection it heads.9 Summarizing, the fact that not only the same functional form may mark different functional projections, but also that different functional forms may mark the same functional projection, indicates that the head of a given functional projection is an abstract syntactico-semantic item, and that insertion of an overt element in that position is not what gives the position its syntactico-semantic identity. Rather, the overt element marks, supports, or makes visible a position that has already got a syntactico-semantic idenS i tity. 10 4.

The operator structure as the syntactico-semantic backbone of the sentence

Up to now we have been mainly concerned with the predication operator and its lexicalization, although we have suggested that there exist other functional operators, as well. In this section, we would like to pursue the general idea that the syntactic functional projections of a clause are the projections of underlying operators, which are in turn supported by visible items. Specifically, we propose that a syntactic string is the derivative visible expression of a rudimentary I(nternalized)-semantic representation

464 Kristin Μ. Eide and Tor A. Afarli constituted by structurally ordered operator tokens. This rudimentary operator structure can be said to constitute the basic underlying logical form of the sentence. However, before we discuss the syntactic instantiation of the underlying operator structure in more detail, we would like to sketch how it relates to the over-all semantics expressed by the clause. In fact, it has been proposed that language has no semantics at all, i.e. the claim is that the meaning of clauses is only more or less indirectly related to our general cognitive representations of meaning - our general conceptual structure (Fodor 1998: 9; Jackendoff 1983: 95). Such a view is also expressed in the following quotation from Fauconnier (1994: xx-xxi): "Sentences bring together, in one linguistically homogeneous form, heterogeneous and incomplete information as to the cognitive construction to be performed within a context for the purpose of constructing meaning. Meaning ensues when such operations are performed, but is not itself directly assignable to sentences." We accept the idea that sentences in some sense function as triggers for elaborate meaning construction or "backstage cognition", to use a phrase employed by Fauconnier (1994: xvii). However, we go against the idea that natural language has no semantics. If that were the case, it seems to us that it would be impossible to get some particular meaning-related "backstage cognition" started at all. The sentence must have some amount of semantics, however rudimentary, in order to trigger some particular meaning construction, i.e. a given sentence does not trigger any thought; it triggers a corresponding thought. In fact, we read the quotation from Fauconnier given above as a statement to this effect. Our proposal is that the back-bone of the "information" that sentences "bring together" is the operator structure alluded to previously. (We tend to think of this operator structure as a "constricted" and "rigidified" Language of Thought structure brought to linguistic use). Moreover, we assume that this rudimentary operator structure is related to the syntactic structure of the sentence in a homomorphic fashion. In that respect we adopt as our general point of view the framework of selective Grammar Semantics developed in Bouchard (1995), in particular the principle of "Full Identification" which says that every (morpho-)syntactic formative of a sentence must have a corresponding element in the semantic representation, and that every formative of a semantic representation must be identified by a (morpho-)syntactic element in the sentence, which is associated with that representation (Bouchard 1995: 22). In other words, the grammar semantics directly expressed by a clause must be distinguished from the semantics that results from the clause. Notice that Bouchard's principle of "Full Identification" is

Semi-lexical heads in a semantically charged syntax 465

open to interpretation. On a strong interpretation, it disallows any syntactico-semantic representation that is not homomorphically correlated with phonetically realized items. Here, however, we opt for a weaker interpretation, i.e. one that dictates that any phonetically realized item must have a syntactico-semantic correlate, but not necessarily the other way round. Now, on the assumption that syntactic structure expresses the operator structure in a homomorphic fashion, our method will simply be as stated in (19): (19)

To find the I-semantic representation of a sentence, identify the syntactic elements and relations and find their semantic correlates.

Notice that we are committed to the view that any visible syntactic element/relation has a corresponding semantic element/relation, but, as just mentioned, we still leave room for semantic elements/relations that are not directly expressed in overt syntax." Let's see how this approach works. The consensus on the basic syntactic structure of clauses in the last fifteen years or so is roughly that they consist of at least two functional projections - a C-projection and an I- or T-projection - on top of the basic lexical VP, which expresses the basic proposition. There are several variations on this theme, with various proposals regarding the number and order of functional categories, the most recent Chomsky an analysis assuming a so-called light verb projection on top of the VP (Chomsky 1995: 315-316). However, for reasons that will be mentioned below we will assume the simple CP-TP-VP structure depicted in (20) for a clause employing a transitive verb. We propose that the C-projection and the T-projection have the operator correlates shown in (21), which, to use a metaphor, may be seen as semantic seeds that the syntactic structure grows out of. (20)

CP

c

TP τ

VP

su

V' V

DO

466 Kristin Μ. Eide and Tor A. Afarli

(21)

a.

b.

Semantic correlate of C: basic force operator (interrogative, imperative, declarative), cf. Stenius (1967), Kitahara (1997: 9), Elvsaas (1998). Semantic correlate of T: tense operator (e.g. Pollock 1989)

What about the lexical V-projection? Given our earlier adoption of Bowers (1993) regarding the analysis of secondary predication (i.e. non-finite predication), it seems natural to adopt Bowers' analysis of the nexus of the full clause, as well. Bowers argues that the verb phrase is non-unitary, with the basic VP being the complement of a Pr-projection. The main verb obligatorily lexicalizes the predication operator by raising from V to Pr in a (partial) representation like (22).

However, contrary to Bowers' non-unitary analysis of the verb phrase and in keeping with our analysis of the copula, we have argued elsewhere (Eide and Äfarli 1999: 171-179) that the predication operator is directly lexicalized by the main verb, so that the V-projection may be said to be a joint projection of the predication operator and the main verb. Thus, we claim that the structure of the verb phrase is not as in (22), but rather as in (23).12

V

DO

Semi-lexical heads in a semantically charged syntax 467

In other words, our analysis amounts to a chunking of the attributive content of the verb and the predication operator.13 On the basis of the above reasoning, we hypothesize that the syntactic structure shown in (20) corresponds to, i.e. is the visible counterpart to, the I-semantic structure shown in (24): (24) ^ ^ ^ ^ ctense op.>



Consider how the functional operator tokens in (24) are made syntactically visible. We assume that each operator has a designated slot for an overt element that makes the operator visible, so one way of making for instance the predication operator visible is for som or a verb to fill that slot. Technically, we adopt the device proposed in Rizzi and Roberts (1996: 106) whereby incorporation is construed as substitution into a subcategorized slot of the head. Thus, "where an incorporation trigger X° has the feature [+Y° ], this means that the slot for is base-generated within triggering substitution of Y0...". In the predication operator case, e.g. som or a verb is substituted into the subcategorized slot of the predication operator. Next, consider the tense operator. The tense operator may be made visible by the insertion of an overt particle into the visibility slot. This is the strategy typically employed in Creole languages, where specific particles mark tense, see e.g. Muysken (1981). A sample of such tense particles in some Creole languages is shown in (25): (25)

Hong Kong Macanese Haitian

ja te

Jamaican ben Negerhollands ha

In other languages, tense is marked by a tense affix on the verb, which we take as indication that the verb has been raised to the tense operator. Thus, in such languages the tense operator is made visible by movement (analyzed as a complex operation involving insertion as a subpart). In sum, there are two main mechanisms that language uses to make a functional operator visible. Either an element is directly inserted from the lexicon or an element is inserted from some other position in the structure (by move-

468 Kristin Μ. Eide and Tor Α. Äfarli

ment). In other words, there are two main means of "supporting" a functional operator. One interesting idea implied by this, is that movement is semantically driven, not feature driven as is assumed in current Minimalist syntax. That is, verb movement is "semantically" motivated in order to make functional operators visible (see Roberts and Roussou 1997 for a related idea). Thus, we want to reinterpret the feature driven movement (Last Resort, feature checking) assumed in Minimalist syntax as movement triggered by the need to fill the visibility slot of semantic operators. Roberts and Roussou (1997) note several problems with the mechanism of feature checking, among others that it introduces features into the derivation whose sole purpose is to be deleted, and that it requires the presence of the same feature twice. However, if movement/insertion is triggered by a general requirement (possibly subject to parametric variation) that the visibility slot of semantic operators is filled, these problems are eliminated. In fact, we see the Minimalist reinterpretation of movement as Attraction (Chomsky 1995: 297) as a step in this direction. This reinterpretation shifts the triggering factor from the lower, moved element to the upper element that the moved element is checked against, which in our analysis corresponds to a functional operator.14

5.

Conclusion

We have shown that the predication operator (Bowers 1993) is lexicalized by various overt heads in Norwegian (the predication particle som, copular verbs, prepositional predication particles, main verbs), each overt head constituting a semi-lexical head in the sense that it lexicalizes an underlying functional operator. Moreover, observing that the elements that may lexicalize the predication operator often have syntactico-semantic functions besides making the predication operator visible, we concluded not only that the content of a functional projection is primarily constituted by an abstract underlying element or operator, but also that the visible elements are possibly, even typically, multifunctional, often vastly underdetermining the functional projection they head. Generalizing that idea, we hypothesized that the syntactic structure is really the homomorphic instantiation of an underlying semantic structure consisting of an ordered array of operators such as a predication operator, a tense operator, and a basic force operator, to mention the minimum of operator types that we assume are present in finite main clauses in Norwegian. Thus, we in effect propose the outlines of

Semi-lexical heads in a semantically charged syntax 469

a program for a semantically charged syntax with semantic operators rather than features (as in current Minimalist syntax) as the driving force for insertion/movement.

Notes 1. Earlier versions of most of the material contained in this paper were presented at the Conference on (Preferably) Non-Lexical Semantics, Paris, May 1998, at the 23rd Annual Penn Linguistics Colloquium, Philadelphia, February 1999, at the Workshop on Semi-Lexical Heads, Tilburg, May 1999, at the SKY Symposium, Helsinki, September 1999, and at seminars at the University of Bergen and at the NTNU, Trondheim. We want to thank the audiences at these occasions for useful comments. Also, we want to thank our anonymous reviewer for useful suggestions. 2. Som is obligatory in small clause complements of evaluative verbs like anse 'regard', betrakte 'consider', oppfatte 'perceive', bed0mme 'evaluate', vurdere 'assess', regne 'reckon' (and also in small clause complements of a few other verbs). Otherwise it is obligatory with nominal predicates (unless, depending on the particular verb involved, a prepositional predication particle or zero lexicalization is used) and prohibited with adjectival predicates, see Eide and Äfarli (1999) for discussion. 3. Det in (4b) is probably not an expletive subject, but a cataphoric subject referring to the that-clause. Nevertheless, the example clearly indicates the clause-like structure of the som-expression. 4. A corresponding argument can be made for (6b). 5. There are nevertheless some subtleties: som yields a more hypothetical reading than does νcere 'be'. Thus, ...se Jon som sp0kelse '...see Jon as a ghost' differs from ...se Jon vcere sp0kelse '...see Jon be a ghost' in that the former is potentially hypothetical while the latter is not. 6. As for the two readings in (12a,b) Rapoport (1987: 142) says: "Assuming the existence of two verbs be, rather than two (or more) uses of one verb, leaves unaccounted for the fact that the two verbs are phonetically identical. While this could well be an accident of English, most of the languages I have examined have only one verb for the two uses. If there were indeed two verbs, the fact that in (almost) every language the two verbs are phonetically identical would be an extremely odd, and inexplicably universal, coincidence." 7. Notice that the configurational complement of the predication operator lexicalized by is, is a property-denoting phrase in (12a) only. Thus, (12b-d) do not seem to instantiate the scheme in (3), where the complement denotes a property that is input to the predication operator (recall that the predication operator is a function from a property to a propositional function). The problem is that (12b-d) seemingly do not contain a property element that can be input to the predication operator. A similar kind of problem regarding the

470 Kristin Μ. Eide and Tor Α. Äfarli

8.

9.

10.

11.

12.

13.

instantiation of the predication operator by main verbs (cf. section 4), is discussed in Eide and Afarli (1999: 177). There we in essence propose as a possibility that in cases where a property element cannot be identified by a constituent in the structure, a function is imposed on the lexical element, whose output is a property. In present terms, that function is imposed as part of the meaning of the copular verb by dynamic enrichment. For example, in (14) the underlying predication operator is a function from a property to a propostional function. However, Superman denotes an entity. The content imposed on the copular verb can be construed as a function from an entity to a property, which, when applied to the entity-denoting complement, yields a property as output. That (abstract) property in turn is input to the predication operator, yielding a propositional function, i.e. a predicate, instantiated by is Superman. Much in the same way as English do is thought to support an underlying tense element in so-called do-support, see e.g. Pollock (1989), Chomsky (1995: 139140). Consider the alternative. If the syntactico-semantic function or content is determined by inherent properties of the element actually filling the head position, the multifunctionality witnessed in e.g. (18) is only possible if the overt element is polysemous, so that there are at least three different som. However, it seems to us that polysemy is not a solution, but rather a problem to be solved, and we would like to propose that our analysis suggests a fruitful approach to that problem. According to that analysis, polysemy is an epiphenomenon, derived from the fact that the same form may be used to mark different syntactico-semantic head positions, the content in each case being to a con-siderable degree determined by the underlying semantic element or operator. Interestingly, we have become aware that a similar conception is found in the Distributed Morphology approach. Thus, in Harley and Noyer (1999: 7) we read: 'Theories endorsing Separationism are attractive because (a) they allow similar syntactico-semantic forms to be realized in quite different ways phonologically, and (b) they permit polyfunctionality of phonological expressions: a single piece [...] might correspond to a set of distinct and unrelated syntacticosemantic functions." See also Marantz (1997), Halle and Marantz (1993). We assume that the number and types of functional projections that languages employ may vary from language to language, and even from clause type to clause type within the same language, see Afarli (1995a), Thräinsson (1996) for motivation. The projection of a predication operator lexicalized by a (main) verb could be labeled either V or Pr. We have chosen V since intuitively the verbal content of the verb (Theta-properties etc.) is at least as important for the nature of the projection as is the predication operator. This chunking is otherwise motivated on grounds of processing efficiency, see Jackendoff (1983: 125), Bouchard (1995: 95). Notice also that our proposal amounts to a restoration of the Port Royal idea that a given verb is constituted

Semi-lexical heads in a semantically charged syntax 471

by a concealed copula (expressing the predication operation) and an attribute (expressing the conceptual content of the verb), cf. Buroker (1994: 14-22) and Arnauld and Nicole ([1662] 1996: 78-82). 14. The arguments we have suggested for semantically driven movement are of a conceptual nature. Empirically, very little seems to distinguish semantically driven movement from Minimalist-type feature driven movement. However, we have argued elsewhere (Afarli and Eide forthcoming) that certain EPP facts are better explained on a semantic approach to movement/insertion as opposed to a feature-based approach.

References Afarli, Tor A. 1995a Seeds and Functional Projections. In: Patxi Goenaga (ed.), De Grammatica Generativa, 139-150. Donostia: Euskal Herriko Unibertsitatea. 1995b A Unified Analysis of Relative Clause Formation in Old Norse. In: Inger Moen, Hanne Gram Simonsen and Helge L0drup (ed.), Papers from The XVth Scandinavian Conference of Linguistics, 533-543. Oslo: Department of Linguistics, University of Oslo. Afarli, Tor A. and Kristin M. Eide forthcoming The Subject Requirement and Predication. Nordic Journal of Linguistics. Arnauld, Antoine and Pierre Nicole 1996 Logic or the Art of Thinking. Cambridge: Cambridge University Press. [1662] Bouchard, Denis 1995 The Semantics of Syntax. Chicago: University of Chicago Press. Bowers, John. 1993 The Syntax of Predication. Linguistic Inquiry 24: 591-656. Buroker, Jill Vance 1994 Judgement and Predication in the Port-Royal Logic. In: J. Elmar Kremer (ed.), The Great Arnauld and Some of His Philosophical Correspondents, 3-27. Toronto: University of Toronto Press. Chomsky, Noam 1982 Some Concepts and Consequences of the Theory of Government and Binding. Cambridge, Mass.: MIT Press. 1995 The Minimalist Program. Cambridge, Mass.: MIT Press. Dowty, David R„ Robert E. Wall and Stanley Peters 1981 Introduction to Montague Semantics. Dordrecht: Reidel.

472 Kristin Μ. Eide and Tor Α. Äfarli

Eide, Kristin M. 1998a

Som-Predicatives: Exploring the Predication Operator. In: Timo Haukioja (ed.), Papers from the 16th Scandinavian Conference of Linguistics, 64-74. Turku: The Department of Finnish and General Linguistics of the University of Turku. 1998b Lexical vs. Structural Meaning: A Case Study. Talk at The 7th Postgraduate Conference, University of Manchester. Eide, Kristin M. and Tor A. Äfarli 1999 The Syntactic Disguises of the Predication Operator. Studia Linguistica 53: 155-181. Elvsaas, Cecilie 0rjasster 1998 Talehandlingers syntaks. Thesis, INL, NTNU, Trondheim. Fauconnier, Gilles 1994 Mental Spaces. Cambridge: Cambridge University Press. Fodor, Jerry 1998 Concepts. Oxford: Clarendon Press. Halle, Morris and Alec Marantz 1993 Distributed Morphology and the Pieces of Inflection. In: Kenneth Hale and Samuel Jay Keyser (ed.), The View from Building 20. Essays in Linguistics in Honor of Sylvain Bromberg er, 111-176. Cambridge, Mass.: MIT Press. Harley, Heidi and Rolf Noyer 1999 Distributed Morphology. GLOT International 4, issue 4. Iversen, Ragnvald 1972 Norr0n grammatikk. Oslo: Tano/Aschehoug. Jackendoff, Ray 1983 Semantics and Cognition. Cambridge, Mass.: MIT Press. Kitahara, Hisatsugu 1997 Elementary Operations and Optimal Derivations. Cambridge, Mass.: MIT Press. Marantz, Alec 1997 No Escape from Syntax: Don't Try Morphological Analysis in the Privacy of Your Own Lexicon. University of Pennsylvania Working Papers in Linguistics 4: 201-225. Muysken, Pieter 1981 Creole tense/mood/aspect systems: the unmarked case? In: Pieter Muysken (ed.), Generative Studies on Creole Languages, 181-199. Dordrect: Foris. Pollock, Jean-Yves 1989 Verb Movement, Universal Grammar, and the Structure of IP. Linguistic Inquiry 20: 365-424.

Semi-lexical heads in a semantically charged syntax 473

Pustejovsky, James 1995 The Generative Lexicon. Cambridge, Mass.: MIT Press. Rapoport, Tova Rebecca 1987 Copular, Nominal, and Small Clauses: A Study of Israeli Hebrew. PhD dissertation, MIT. Rizzi, Luigi and Ian Roberts 1996 Complex Inversion in French. In: Adriana Belletti and Luigi Rizzi (ed.), Parameters and Functional Heads, 91-116. New York: Oxford University Press. Roberts, Ian and Anna Roussou 1997 Interface Interpretation. Talk at GLOW 20, Rabat. Stenius, Erik 1967 Mood and Language-game. Synthese 17: 254-274. Thräinsson, Höskuldur 1996 On the (Non-)Universality of Functional Categories. In: Werner Abraham, Samuel David Epstein, Höskuldur Thräinsson and C. Jan-Wouter Zwart (ed.), Minimal Ideas, 253-281. Amsterdam: Benjamins.

As for as /for, they are semi-lexical heads Joan Rafel

1.

Introduction

The type of construction that is examined in this work1 appears embedded in examples like the ones in (1). (1)

a. b.

They regard [John as smart]. They took [John for a fool].

Apparently these constructions are composed of a DP, a particle, and an XP. The particle is as and for in (la) and (lb), respectively, whereas the categorial value of X is adjectival (A) in (la) and nominal (N) in (lb). In these constructions, the XP describes a property or an entity that holds for the DP. This means that the AP smart in (la) and the DP a fool in (lb) are predicated of the DP John} In fact, the predicative relationship between these two phrases can be independently established within a copular sentence. This is illustrated in (2). (2)

a. b.

John is smart. John is a fool.

The predicative relationship that is found in the embedded constructions in (1) suggests that at some point of the derivation the subject (DP) and the head of the predicate (X) set up a local Spec-Head agreement relationship, presumably within a functional projection. Spanish, for instance, morphologically manifests the agreement in gender and number that holds between these two elements when the syntactic category of X is A. Thus, in (3) the DP tus hermanas 'your sisters' triggers the agreement in gender (feminine) and number (plural) on the A tontas 'fools'.3

476 JoanRafel (3)

Tomaron a tus hermanas por {*tonto / took.they to-ACC your sisters FEM-PL for fool-MASC-SG *tonta / * tontos / tontas.} fool-FEM-SG fool-MASC-PL fool-FEM-PL 'They took your sisters for fools.'

On the other hand, it seems that the DP, the particle, and the XP form a single complex constituent. For instance, the fact that the sentences in (1) do not presuppose the meaning of the sentences in (4) indicates that, in the former cases, the DP John does not have the same semantic relationship with the verb as it would have were this DP the semantic object of that verb. (4)

a. #They regard John. b. #They took John.

This is another clue that suggests that the DP John is an argument of the embedded complex constituent in the sentences in (1). The organization of the article is as follows. In section 2,1 present what I call the standard analysis of these constructions, and point out some nontrivial problems that derive from this approach. In section 3,1 introduce the analysis of this type of construction in terms of what I call a Complex Small Clause. In section 4, I provide arguments that indicate that the particles as and for behave like semi-lexical heads in these structures. This idea is reinforced in section 5 where crosslinguistic evidence is used. Finally, I draw some conclusions in section 6.

2.

The standard analysis

The standard, and at first sight the most simple, analysis that has been proposed for the constructions under discussion is the one depicted in (5) (see, for instance, Hantson 1989 and Starke 1995). (5)

[Cp Yc C ... [«,(*/>) DP X' ]]]

If we apply this structure to the constructions that appear embedded in the examples in (1), we obtain the representations in (6).

As for as / for, they are semi-lexical heads 4ΊΊ (6)

a. b.

They regard Johny [Cp tj [ c as... They took Johrii [CP TI[C for...

[SCUAP) [SCKPF)

tj smart]]], tj a fool]]].

First of all, the structure in (5) states that the X', smart in (6a) and a fool in (6b), and the DP, John in both cases, constitute a Small Clause (SCI). If we adopt the structural model of SCI defended by the Small Clause Theory (since Stowell 1981, 1983), which here is reproduced in (7), then we would say that in (6) the DP John is an argument that is base-generated in the specifier of the phrase projected by the head of the SCl-predicate. This position is Spec, AP in (6a) and, presumably, Spec, NP in (6b). In this position, the DP John is assigned a theta-role by the SCl-predicate. Later, at some point of the derivation, the DP and the head of the predicate will move up to a functional projection, say FP, in order to establish a local Spec-Head agreement. In the structure in (5), the functional projection FP would be found where the dots appear, that is, between the lexical XP-shell andC. (7)

Small Clause structure IFP IF [XP DP LT X ]]]]

where X = Ν, P, A, V

According to the analysis in (5), this SCI, in turn, is selected by the particle as / for, which heads the construction that is subcategorized for by the verb to regard and to take in the examples in (6). As represented in the structure in (5), these particles are treated as prepositional complementizers (see Emonds 1985; Hantson 1989; Starke 1995). On the one hand, they are claimed to be 'prepositional' (P) because they are homophonous with Ps, (8), and because they can be left stranded in those languages that normally allow P-stranding, as English, (9). (8)

a. b.

As his friend, my opinions are probably biased. This is for me.

(9)

a. b.

What do you regard him as? Who do you take me for?

On the other hand, they are called 'complementizers' (C) since both can be found introducing a full sentence. This is illustrated in (10).5 (10)

a. b.

I did as I was told. For me to go is impossible.

478 JoanRafel An advantage that is often pointed out by the proponents of the C status of the particles as and for in (6) is that, by adopting this hypothesis, we avoid having to say that in (6a), for instance, "P selects AP, an otherwise unattested fact" (Starke 1995: 245). But notice that this argument immediately collapses when we say that these particles do select a (small) clause in these constructions. Semantically, this means that Ρ would not select a property, namely an A, but a subject-predicate relationship. And, from a syntactic viewpoint, Ρ would not subcategorize for a bare AP-shell, but for the functional domain introducing that AP-shell (perhaps the FP-node). Nonetheless, this new perspective turns out to reinforce the idea of treating as and for as prepositional Cs, because now these particles may be considered the highest head of a clausal domain, that is, the clausal domain of the SCI. I will come back to this point in section 4.1 below. Finally, in the representations in (6) it is assumed that the subject of the SCI, John, moves to the matrix clause to check structural Case. This movement is already carried out at Syntax, and it goes from Spec, XP, to Spec, CP and, finally, to a position within the matrix clause, presumably the specifier of an agreement projection.6 In the examples in (11), the DP checks accusative Case, whereas in (12) it checks nominative. (11)

a. b.

They regard him as smart. They took him for a fool.

(12)

a. b.

He was regarded as smart. He was taken for a fool.

In the remainder of this section I show that the analysis in (5) does not account for many nontrivial properties that characterize the constructions under investigation.

2.1.

The relationship between the SCl-subject and the prepositional C

A first question that immediately comes up by adopting the analysis in (5) is why the prepositional C, in particular for, cannot assign Case to the subject of the SCI that it precedes in these constructions. This is surprising, bearing in mind that in English the prepositional C for does generally assign Case to the subject of the infinitival clause that it may introduce. The puzzling contrast, then, is that between the example in (13) and the one in (14).

As for as / for, they are semi-lexical heads 479

(13)

It is impossible |for him to go\.

(14) * They took [for him a fool]. A related issue is that, in those contexts in which for introduces an infinitival clause, this prepositional C cannot license the trace that the subject of the infinitival clause leaves behind when this subject moves to a higher position in the structure. An example is given in (15). (15) *Who\ would you prefer [cp t, [c for [n> tj to leave first?]]] Again the prepositional Cfor would surprisingly behave in a different way in the constructions under discussion if their analysis were as depicted in (5). As shown in (16), the C for would license the trace here, as opposed to what happens in the sentence in (15). (16)

2.2.

Who\ did they take [c? tj [cfor

[fp tj a fool?]]]

The movement of the subject creates a mixed [A, A', A] chain

In the analysis in (5), the movement of the DP-subject from the embedded position where it is base-generated to the matrix clause would create a mixed (sandwiched) [A, A', A] chain. This is so because that DP would move from Spec, XP, an argumental position (A); to Spec, CP, a nonargumental position (A'); and, finally, to a position within the matrix clause, say Spec, AgroP in examples like the ones in (17) and Spec, IP in (18), again π an argumental position (A). (17)

a. b.

They regard him as smart. They took him for a fool.

(18)

a. b.

He was taken for a fool. g He passes for a soldier.

Obviously, we would expect this operation to yield an ungrammatical sentence, contrary to the facts.9

480 JoanRafel 2.3.

The subject cannot check Case at Logical Form

In languages like Spanish, the subject of a sentence can check Case either overtly or covertly, that is, at Syntax or at Logical Form (LF). As a result, the subject can show up in the sentence in Spec, IP or, presumably, in the position where it is base-generated, namely in Spec, VP. An example is given in (19).10 (19)

Ayer (Juan) corrio (Juan). yesterday Juan ran.he Juan 'Juan ran yesterday.'

Now a nontrivial question that arises from the analysis in (5) is why the DP that is base-generated in a position following the particle por 'for' in the construction in (20) cannot ever remain in that position in such languages and check nominative Case at LF. The impossibility of this option is demonstrated by the ungrammaticality of (21). (20)

Pocos estudiantes pasaron por jugadores few students passed.they for players 'Few students passed for basketball players.'

de baloncesto. of basketball

(21) * Pasaron por pocos estudiantes jugadores de baloncesto.

2.4.

Verbs like 'to take' do not subcategorize for a full CP

Finally, the standard analysis assumes that verbs like to take subcategorize for an ordinary CP-projection. This move is difficult to reconcile with the fact that verbs of this type do not normally subcategorize for CPs containing a full sentence, instead of a SCI: (22)

a. *They took that John is a fool. b. *They took for John to be a fool.

The ungrammaticality of these sentences shows that the ontological category of "proposition", which is typically associated to a CP-projection at Syntax, cannot be combined with the meaning that a verb like to take has.11

As for as / for, they are semi-lexical heads 481

3.

The Complex Small Clause-analysis

The analysis that I would like to propose in this article for the constructions that appear embedded in (1), repeated here as (23), is depicted in (24). (23)

a. b.

They regard [John as smart]. They took [John for a fool].

482 JoanRafel In this structure, X heads a SCI, the subject of which is the argumental PRO that appears in Spec, XP. In (23a) the categorial value of X is adjectival (A) (smart), whereas in (23b) it is nominal (N) (fool).12 This SCI is presumably introduced by a functional domain, which in the structure in (24) is represented by the FP-node. The subindex χ indicates that the category that bears it is an extended projection of the lexical head X.13 As can be observed, the particle as / for in the structure in (24) heads the highest extended projection of X. Finally, the subject of the construction is the lexical DP that is merged with C', namely John.14 As indicated, I call this structure a Complex Small Clause (CSC1). I use this term to make reference to a construction that behaves like an ordinary SCI but the predicate of which is a predication. In the structure in (24), the subject of the construction, John, is merged with a nonverbal head, as or for. Hence the idea that we are dealing with a SCI here. But, on the other hand, this nonverbal head is the highest head of a clause, the clause that the predicative relationship between PRO and X' forms. I call this latter predicative relationship the internal predication of a CSC1.15 In the remainder of this section, I discuss some properties of the constructions under examination that derive from the CSCl-structure in (24).

3.1.

PRO and X'

The CSCl-structure in (24) tells us that an argumental PRO is directly merged with X'. The result of this operation is a lexical XP-shell in which X' assigns a theta-role to the null argument PRO. In the example in (25), for instance, the categorial value of X is A. (25)

They regard John as PRO smart.

Nonetheless, we also have similar constructions (that will not be discussed here) in which the categorial value of X is V. An example is provided in (26). (26)

I regard John as PRO having too much property.

This latter possibility is important since it stands as an argument in favor of the idea that there is a subject between the particle and the embedded (verbal) predicate. Now this subject cannot be the lexical DP John in (25) or

As for as / for, they are semi-lexical heads 483 (26). Otherwise all the problems attributed to the standard analysis in section 2 above would arise once again here. At some point of the derivation, PRO and the head of the lexical XPshell, namely X, will have to license their predication relationship. Presumably, this licensing is satisfied within a functional projection FP by means of a Spec-Head agreement relationship between PRO and X. As represented in (24), this functional projection FP would be found in a position higher than the lexical XP-shell but, importantly, lower than C. Examples like the one cited in (27) in Spanish indicate that there is "space" between the lexical XP-shell and C. In other words, (27) shows that the particle, in this case por 'for', is not a sort of prefix attached to the predicate X'. This example is from J. Eslava Galan, En busca del unicornio, (1987), Spain. (27)

aquellas pocas cosas que ellos tenian pormuy necesarias... those few things that they had for very necessary 'those few things that they held as being very necessary...'

Here the intensifier muy 'muy' shows up between the particle por 'for' and the A necesarias 'necessary'. As a result of the Spec-Head configuration between PRO and X within the functional projection FP, PRO will trigger the gender and number agreement on X if the categorial value of that X is A. The Spanish sentence in (28) illustrates this effect. (28)

Tomaron a tus hermanas\ por PRO, {*tonto / took.they to your sisters-FEM-PL for fool-MASC-SG *tonta / Hontos / tontas.} fool-FEM-SG fool-MASC-PL fool-FEM-PL

In this example, it can be observed that X, which here is the A tontas 'fools', must obligatorily agree in gender (feminine) and number (plural) with the subject PRO, which, as we will see shortly below, is controlled by the lexical DP that is base-generated in Spec, CP, namely tus hermanas 'your sisters' in (28). The relationship that we find between PRO and X in the internal predication in the CSCl-structure, then, would be just like the relationship that is established between PRO and X in constructions of the following kind in Spanish:

484 JoanRafel (29)

[Por PROi {*tonto / *tonta / *tontos for fool-MASC-SG fool-FEM-SG fool-MASC-PL /tontas,}] tus hermanas; perdieron todo lo fool-FEM-PL your sisters- FEM-PL lost.they all what que habian ganado. that had.they won 'Your sisters lost everything they had won because of their foolishness.'

In this construction the subject PRO also triggers the gender (feminine) and number (plural) agreement on the predicate tontas 'fools'. But differently from the construction that we are examining in this work, here this PRO is controlled by an argument of the matrix clause, that is, the lexical subject tus hermanas 'your sisters'.16 As far as Case is concerned, it seems reasonable to suppose that the mechanism that serves to license the PRO that is found in sentences like the one in (29) or the PRO that typically appears functioning as the subject of an adjunct SCI, as in (30), also licenses the PRO that is contained in the 17 internal predication of the CSCls in (31). (30)

I found the glasss [PRO; empty].

(31)

a. b.

Tomaron a [Juani por PRO; tonto]. They took [.John, for PROi a fool].

So, in all three constructions, PRO presumably checks some kind of null Case by default. The analysis in (24), on the other hand, predicts that PRO is the argument that must bind the anaphor that may appear in the complement position of the CSCl-internal predication. Consider the example in (32). (32)

Mariaj tomo a [Juanj \por [Sci (DP) PROj un apasionado Maria took, she to Juan for a fanatic {desimismoi / *de simismas.} ]]] of himself of herself 'Maria took Juan for a fanatic {of himself / *of herself.}'

In this sentence, we can see that the anaphor cannot be bound by the subject of the matrix clause. According to the CSCl-analysis in (24), the lexical DP Juan would not be a candidate to bind the anaphor either, because this

As for as / for, they are semi-lexical heads 485

argument would not bind the anaphor locally, that is, within the domain of the internal predication of the CSC1. Thus the only candidate that would be able to do so would be the subject of the SCI, namely PRO. Summarizing, the CSCl-structure in (24) sustains that the particle as / for in English, or por 'for' in Spanish, introduces a predicative relationship. The domain that these particles precede in these constructions, then, would clearly differ from the domain introduced by a Ρ in ordinary SCls, (33), whereas it would resemble the domain introduced by a copula, (34). (33)

John, is [sei (PP) ti [p on [DP the table]]].

(34)

Yoursistersi are

[SCKAP)

tj

[A·

happy]].

On the one hand, the Ρ on that shows up in (33) selects the DP the table. Now the basic difference between this projection (DP) and the projection headed by X in a CSC1 (XP) lies in that, in the former case, the projection does not contain any (null) subject. The reason for this is simple: in this type of construction the DP the table is not predicated of any subject. For instance, this DP cannot be predicated of a PRO controlled by the DP John. We know this because the sentence in (35) is ungrammatical. (35) *John is the table. Conversely, according to the analysis that is defended in this work, the projection headed by X in a CSC1 does contain a subject. This subject is a PRO and is base-generated in Spec, XP. This straightforwardly accounts for the fact that the internal predicate, namely X', can be combined independently with a lexical subject in a copulative sentence. Compare (33) and (35) with (36)-(37). (36)

They regard John as their best friend.

(37)

John is their best friend.

3.2.

DP and C

Differently from the standard approach, I claim that the particle that appears in these constructions is not the head of an ordinary CP-constituent, but the head of a CSC1 (cf. subsection 2.4 above). In other words, I claim that C is the head of the highest extended projection of X, and the head of

486 JoanRafel the projection that hosts the subject of the construction. As represented in (24), the subject of the construction is the DP that is base-generated in Spec, CP. This lexical DP would be licensed in this position either by Θmarking, that is, it would receive a theta-role from the C'-constituent, or by predication, that is, it would saturate the X'-predicate (see Rothstein (1995) 18 and references cited there). Like the lexical subject in any other type of argumental SCI (see (38)), this DP checks structural Case outside its clausal domain. In the examples in (39), the Case checked is the accusative assigned by the matrix verb, whereas in (40) it is the nominative Case that is assigned by the matrix IPhead. (38)

a. b.

They consider him smart. He was considered smart.

(39)

They regard him as smart.

(40)

a. b.

He was regarded as smart. He passes for a soldier.

As already pointed out above, examples of this kind demonstrate that the lexical subject of the embedded construction can move from the position where it is base-generated to an argumental position of the matrix sentence. Now the analysis of these constructions in terms of a CSC1 correctly predicts the legitimacy of this operation. This is so since, in the examples in (40), for instance, the lexical subject would move from Spec, CP, which in this construction would be an Α-position, to the subject position of the matrix sentence, again an Α-position. This movement, then, would create a well-formed [A, A] chain (cf. subsection 2.2). As the sentences in (41) show, the subject of this type of construction can also be an anaphor bound by the subject of the matrix sentence in those contexts in which the CSC1 functions as the object of a verb. (41)

a. b.

Johnj regards himself\ as smart. John\ took himself for intelligent.

Again this fact is expected, bearing in mind that this kind of relationship between the matrix subject and the subject of the embedded clause can also be established when the subordinate clause is an ordinary SCI. This is illustrated in (42).

As for as / for, they are semi-lexical heads 487

(42)

John, considers himself intelligent.

Now the subject of a CSC1 must corefer obligatorily with the grammatical subject of the internal predication. This means that, in this construction, the lexical DP that is base-generated in Spec, FP must necessarily control the null subject PRO that is base-generated in Spec, XP. This relationship is represented in (43). (43)

a. b.

Mary regard Peter, as PRO; a fool. Mary took Peter, for PRO] a fool.

As expected, the construction is ruled out if this control relationship is not satisfied. This is what the ungrammatically of the examples in (44) tells us. (44)

a. b.

*Mary·, regard the children as PROj their best friend. *Mary, took the children for PRO, a fool.

The control relationship that is established between the lexical and the null subject in this construction, then, parallels the control relationship that is set up between the lexical subject of the matrix sentence and the null subject of the embedded SCI in constructions like the one in (45) (cf. (43)). (45)

Peter, doesn 't want PROj to be a fool.

The only difference between the control relationship in the CSC1 in (43) and the control relationship in the construction in (45) lies in that only in the former structure is the control relationship established within the same extended projection, that is, within the extended projection initiated by X (fool in (43)). This is so since, according to the analysis proposed here, the particles as/for form part of the extended projection of this lexical head. In contrast, the controller in the construction in (45), namely Peter, is situated in the extended projection headed by the verb want, that is, an extended projection that does not include the null subject PRO. Finally, the CSCl-structure in (24) helps us understand why in languages like Spanish the subject of the CSC1, that is, the lexical DP, can appear at Syntax either in the subject position of the matrix clause if the construction combines with a raising verb, (46a), or immediately preceding the particle, (46b), but never between the particle and the internal predicate, (46c).

488 JoanRafel (46)

a.

Pocos estudiantes pasaron por jugadores de few students passed.they for players of baloncesto. basketball 'Few students passed for basketball players.' b. Pasaron pocos estudiantes por jugadores de baloncesto. c. *Pasaron por pocos estudiantes jugadores de baloncesto.

The answer that the CSCl-structure offers us is that this lexical subject cannot appear between the particle and the SCl-predicate because this argument is base-generated in a position structurally higher than the particle, namely in Spec, CP. In (46a), for instance, the CSCl-subject pocos estudiantes 'few students' shows up in the subject position of the matrix clause, which is the position where that DP overtly moves from Spec, CP. In this position, the DP checks the EPP-feature of the matrix IP and nominative Case. In the example in (46b), on the other hand, the subject of the CSC1 presumably remains in the position where it is base-generated, namely in Spec, CP. In this example, the EPP-feature of the matrix IP would be checked by a null expletive subject proexpi, whereas nominative Case would be checked off covertly by the CSCl-subject. And, finally, the ungrammatical sentence in (46c) simply indicates that the CSCl-subject cannot be lowered to a position following the CSCl-head (cf. subsection 2.3 above). 19

Now, apart from solving all the problems listed in section 2 above, the analysis of this type of construction in terms of a CSC1 allows us to represent the idea that in these constructions a situation, [PRO XP], is predicated of an individual, [DP]. My purpose in the next section is to show that the particle that links the situation and the individual in this type of CSC1 behaves like a semi-lexical head in the sense that it functions as both the highest functional head that introduces the SCl-domain and the lexical element that heads the CSC1. 4. 4.1.

The semi-lexical nature of the Complex Small Clause-head 'As/for' as functional heads

As commented above, one of the reasons that leads us to consider as and for as Cs in the constructions under discussion is that these particles introduce a clausal domain, more specifically a SCI. Now the treatment of as

As for as / for, they are semi-lexical heads 489

and for as Cs entails that these particles must form part of the extended projection of the lexical element X that heads the SCI (see Grimshaw 1991 and Van Riemsdijk 1998). From this perspective, then, we would expect as and for to behave like functional heads. The idea that these particles are functional heads in these constructions accounts for nontrivial syntactic and semantic facts. For instance, it explains why the prepositional C for cannot assign Case to the subject of the SCI that it precedes, (47), in contrast to what occurs when the prepositional C for introduces an infinitival clause, (48). (47) *They took John for him a fool. (48)

It is impossible for him to go.

In order to show that this contrast can be attributed to the functional status of the particle for in both cases, let me first present a general hypothesis that has recently been defended in the literature. This will allow us to draw the relevant parallelisms. To begin with, consider the familiar scopal ambiguity that is observed in a sentence like that in (49). (49)

John left [after Sheila said he should leave].

This sentence has two interpretations. It can mean that 'John left after the time of Sheila's saying that he should leave' or that 'John left after the time which Sheila said he should leave at'. In Larson (1988), it is proposed that this ambiguity stems from the two different positions in which the temporal operator (TO) can be base-generated in this type of construction. On the one hand, that TO can appear modifying the clause headed by the verb to say, in which case the former interpretation is obtained. On the other hand, it may appear modifying the clause headed by the verb to leave, yielding the latter interpretation. It is claimed, furthermore, that in both readings the TO would end up moving from the position where it is base-generated to the specifier of the embedded CP. This is schematically represented in (50). (50)

a. b.

John left [PP [p· after [Cp Opi [c 0 [n> Sheila said [cp he should leave] tj ]]]]] John left [PP [p· after [CP Op; [c 0 [n> Sheila said [cp he should leave tj ]]]]]]

490 JoanRafel More recently, Dubinsky and Williams (1995) have convincingly shown that in English temporal Ps (after, before, while, etc.) recategorize as Cs. In other words, they demonstrate that in this language temporal Ps occupy the head of a CP-projection, instead of the head of a PP-projection situated 20

immediately above a CP-node containing a null C (cf. (50)). According to their proposal, then, the structures in (51) should be modified in the following way: 1 (51)

a. b.

John left [cρ Op; [c after [n> Sheila said [cp he should leave] tj ]]] John left [cp Opi [c after [n> Sheila said [cp he should leave tj ]]]]

Dubinsky and Williams take advantage of the idea that there is a TO in Spec, CP in this type of construction to account for the contrast between (52) and (53). (52)

John left [CP Opi [cafter [n> {*me/PRO} telling him to tj ]]].

(53) It is impossible [cp [cfar [n> {me/ *PRO} to leave]]]. As one can observe, a null subject PRO must follow the prepositional C in the example in (52), whereas in (53) the embedded subject must be a phonologically realized pronoun. Following Manzini's (1992) work, they arrive at the conclusion that a prepositional C can assign Case to the subject of the clause that it introduces only if that prepositional C does not have to establish a Spec-Head agreement relationship with an element in Spec, CP. For a sentence like (52), this means that the temporal prepositional C after cannot assign Case to the subject of the embedded clause because this C must set up a Spec-Head agreement relationship with the TO occupying the specifier of its projection, that is, Spec, CP. This accounts for the fact that the subject of this clause can only be a PRO, since there is no element in the clause that can assign structural Case to this argument. Conversely, the nontemporal prepositional Cfor in the example in (53) can assign structural Case to the subject of the embedded clause because there is no element in Spec, CP which that C must establish a Spec-Head agreement with. Hence, that subject has to be a phonologically realized pronoun.22 Now compare the sentence in (52) with the construction in (54) as analyzed here.

As for as / for, they are semi-lexical heads 491

(54)

They took [CP John [cfor [sa{*him/PRO]

a fool]]].

At this point, the answer to the ungrammatically of They took for him a fool or They took John for him a fool becomes crystal-clear: in this construction, the head of the highest extended projection of the N, namely for, cannot assign Case to the subject to its right because this head must establish a Spec-Head agreement relationship with the subject that is basegenerated in its specifier. In this sense, the particle for in (54) behaves just like the prepositional C after that appears in the sentence in (52) (cf. subsection 2.1 above). Interestingly enough, the contrast between (54) and (55) tells us that this property only holds of prepositional Cs, or more generally of functional Ps, that host an element in their specifier. (55)

Thisbooki is [PP tj [P· for [DP me]]].

In the sentence in (55), the DP the book is the subject of the SCI, whereas the lexical Ρ for is the head of this SCI. This means that the DP the book and the Ρ for must eventually establish a Spec-Head agreement relationship, as usual. Nonetheless, this agreement relationship does not prevent the lexical Ρ for from assigning Case to its complement. 3 Compare (52)/(54) with (55). From a semantic viewpoint, I would like to suggest here, first, that the C that introduces the SCI in the constructions under discussion behaves like a modal marker that modifies the internal predication of the CSC1; and, secondly, that the construction that results from merging a lexical DP with the constituent headed by this modal C, namely C' in the structure in (24), expresses a situation in which the individual that the lexical DP refers to is taken as a participant of a "possible" state of affairs.24 In order to illustrate this idea, let me comment very briefly on the semantic interpretation of a construction headed by the particle for. Consider the example in (56a). According to the CSCl-analysis, the embedded construction in (56a) would be analyzed as represented in (56b). (56)

a. b.

The committee took you for intelligent. The committee took [csci you\ (for [sciPROj intelligent]]].

In the structure in (56b), one can observe that PRO and the A intelligent form a SCI. This SCI describes a situation or, more specifically, a state of affairs. On the other hand, the head of the highest extended projection of

492 JoanRafel the SCl-head is for. This particle projects a CP-projection, which hosts the subject of the CSC1 in its specifier. Now my claim is that, as a modal head, the particle for here indicates that the state of affairs that it introduces, namely [PRO intelligent], is a presumed situation. That is, a situation that holds in a possible world, say W2.25 The syntactic configuration of the CSC1 (see (56b)), then, allows this presumed state of affairs to be predicated of the individual that the CSC1subject refers to. In the example in (56b), this individual happens to be you. Thus, according to this configuration, the speaker who utters the sentence in (56) says something like 'the committee attributed something to you, and this something was a presumed state of affairs in which you were intelligent. The presumptive state of affairs that the speaker has attributed to you by constructing this type of CSC1, or to put it differently, the presumption that the SCI [PRO intelligent] conveys in (56), explains why a sentence like "Don't you think I am intelligent?" sounds appropriate as a response to the statement in (56). This is so since what we are doing by uttering this question is challenging the presupposition that the speaker has constructed. Now the meaning of the CSCl-construction in (56) clearly contrasts with the meaning of a regular SCI. Consider the example in (57). (57)

a. b.

The committee considered you intelligent. The committee considered [sei you intelligent].

In contrast to the sentence in (56), the speaker who utters the sentence in (57) is simply saying that the committee attributed something to you, and this something was the property of being intelligent. So the basic contrast with regard to the CSC1 in (56) is that (57) does not convey a presumptive state of affairs. This immediately explains why the question "Don't you think I am intelligent?" sounds inappropriate as a response to the statement in (57).26

4.2.

'As/for' as lexical heads

In the previous section, we have seen that the head of a CSC1 behaves like a functional head or, in other words, like a modal marker that introduces a clause. Let us see now the other side of the coin, that is, the lexical prop27

erties of this head. The lexical nature of the head in the construction under examination is shown by both syntactic and semantic facts. From a syntactic viewpoint, for

As for as / for, they are semi-lexical heads 493 example, these particles may be left stranded in those languages where a Ρ can be left stranded, as in English. This is shown in (58). (58)

a. b.

What do you regard him as? Who do you take me for?

From a semantic perspective, for instance, this property is demonstrated by the fact that these particles clearly implement the semantic specification of the matrix verb. Consider the following sentences: (59)

a. b. c.

I regard John *(as) my best friend. They took John *ifor) a fool. John passes *(for) a fool.

As one can observe, these examples are grammatical only if the modal particle appears. Otherwise they are ruled out. Now this fact leads us to say that in these constructions the particle has a lexical specification, and that this lexical specification must contribute to the ultimate semantic value of the matrix verb. From a technical point of view, this lexical contribution of the particle to the matrix verb may be achieved at LF by means of a "semantic checking" between the modal particle, namely the CSCl-head, 28

and the matrix verb. A second semantic argument that shows that the head of these constructions also behaves like a lexical category is provided by the fact that light verbs can combine with these structures. Some examples in Spanish are cited in (60). (60)

a.

b.

Dieron a Juan por {muerto /desaparecido / iniitil.} gave.they to Juan for dead disappeared useless 'Juan was taken for {dead / missing / useless.}' Tienen a Juan por {un buen chico / mentiroso.} have.they to Juan for a good boy liar 'Juan is regarded as {a good boy / a liar.}'

The light verbs that appear in these sentences are dar 'to give' in (60a) and tener 'to have' in (60b). This possibility can also be attributed to the lexical properties of the CSCl-head, which in these cases would implement the semantics of the matrix verb.

494 JoanRafel And thirdly, the lexical nature of the head is supported by the versions of the construction that can be found when the head is the particle as. Consider the examples in (61). (61)

a. b.

I regard John as my best friend. I regard John as if he were my best friend.

The embedded constructions that we have in these two sentences look like a CSC1 headed by the modal particle as. But, as the reader might have already noticed, the crucial difference between these two examples is found in that the particle as introduces a SCI in (61a), just like in the examples 29

explored so far in this work, and a complete clause in (61b). Now the embedded construction in (61b) is interesting to us for various reasons. First of all, it clearly shows that the situation described by the predicative relationship introduced by the particle as is a presumptive situation. This is indicated here by the subordinating conjunction ι/that precedes the embedded finite clause. Secondly, it explicitly shows that a participant in this clause is the individual that the lexical DP preceding the particle as refers to. As one can observe, in (61b) the pronoun he and the lexical DP John refer to the same individual. And thirdly, the embedded construction in (61b) illustrates the lexical nature of the particle as in this example. That is, here the particle as introduces a complete clause the highest functional projection of which is the subordinating conjunction if. This suggests that in (61b) the particle as initiates its own extended projection (EP). This is 30 schematically represented in (62). (62)

I regard [ep2 John as [epi if he were my best friend\\.

The properties that the particle as has in this type of sentence, then, would be similar to the properties that the particle like has in examples like the one in (nonstandard) English in (63). (63)

You look like you've just seen a ghost.

In conclusion, the arguments that have been provided in this section indicate that the head of the CSCls that we are examining in this work behaves like what here I have call a 'semi-lexical' head. That is, it does not behave like a pure functional head nor like a pure lexical head. On the one hand, we have seen that it behaves like a functional head in the sense that it is a category that, structurally, stands as the highest extended projection of the

As for as / for, they are semi-lexical heads 495

lexical head X and, semantically, provides the predication that it introduces with modal information. On the other hand, it behaves like a lexical head in the sense that it allows stranding like a lexical P, and contributes to the semantics of the matrix verb. The idea that the particles as and for are semilexical heads in these constructions, then, goes against previous proposals that have been cited in the literature.31 5. 5.1.

Crosslinguistic evidence The Pseudorelative in Romance: From functional to [+lexical]

The so-called Pseudo-Relative (PR) is a construction that is found in the majority of the Romance languages. This construction is formed by a DP, the C que 'that', and a finite verb which agrees in person and number with this DP.32 An example in Spanish is provided in (64). (64)

He visto a Juan que corria. have.I seen to-ACC Juan that ran-IMPERF-3SG Ί saw Juan running.'

Crucially the PR forms a single constituent in one of its possible readings and behaves like a SCI (see Guasti 1988; Rafel 2000a, 2000b). In previous work, I have assigned to the PR an analysis in terms of a CSC1. This analysis is as depicted in (65) (cf. (24)). (65)

[csci (CP) Juani[c que fo proj corria]]]

In this structure, the lexical DP Juan is base-generated in the specifier of the C que, that is, in Spec, CP. From this position, this DP eventually raises to the matrix clause to check accusative Case. That this is so is clearly shown by examples like the one in (66), where the DP appears cliticized onto the main verb by means of the accusative form lo 'him'. (66)

Lo he visto que corria. him-ACC have.I seen that ran-IMPERF-3SG Ί saw him running.'

On the other hand, this lexical DP must obligatorily bind a pro that is situated in the subject position of the finite IP selected by the C que. Note that

496 JoanRafel this pro is the argument that checks the nominative Case assigned by the embedded finite I. I have also provided evidence in favor of the idea that in this construction the C que does not behave like the neuter C que that is found in regular embedded clauses, as in (67). (67)

Maria dijo que Juan corrio. 'Maria said that Juan ran.'

More specifically, I have shown that in the PR the C que behaves like an aspectual head that modifies the internal temporal structure of the imperfective finite verb. Hence the progressive interpretation of this construction 33

(see the translations into English). If this approach turns out to be true, then in the PR the C que would function as a semi-functional head, in the sense that it does not behave like the typical functional C que that is found in regular embedded clauses (see (67)). 5.2.

The resultative construction in Chinese: From lexical to [+functional]

Huang (1982, 1992) observes that resultative constructions in Chinese constitute a structure of internal control. The elements that are involved in this construction are a DP, the aspectual head de, and an adjectival predicate. An example is given in (68). (68)

Zhangsan ba ma qi-de hen lei. Zhangsan ACC horse ride-ASP very tired 'Zhangsan rode the horse and got it very tired.'

According to Huang, the DP ma 'horse' is an object of a complex predicate. This complex predicate is formed by the matrix verb qi 'ride' and the aspectual head de, which here appears incorporated into the verb. Furthermore, he argues that this DP controls a null subject Pro (for PRO or pro) that functions as the subject of the adjectival predicate hen lei 'very tired'. Now, since the construction in (68) behaves in many ways like both the type of construction that we have explored in this article and the PR (see Huang 1982, 1992), it seems plausible, in principle, to hold that, before the

As for as / for, they are semi-lexical heads 497 incorporation of the aspectual head de into the matrix verb, the structure that we have looks like (69) (cf. (24)). (69)

Zhangsan qi [csci(CP) [c· de Zhangsan ride horse ASP

[ap

Pro, hen lei\}}. very tired

Many linguists, Huang included, have treated the aspectual head de in (69) as a C that ends up suffixing to the matrix verb. At this stage, statements of this kind sound quite familiar to us. We have seen that the prepositional Cs as and for are a sort of modal heads that may wind up incorporating into the matrix verb at LF, since they head a CSCI. Something similar may also occur with the C que that appears in the PR, since in this construction que functions as an aspectual marker that also heads a CSCI. Now, for our purposes here, the resultative construction in Chinese provides us with two interesting facts. First, it overtly shows that the CSC1head ends up incorporating into the matrix verb. The second relevant point concerns the nature of the aspectual head de. Huang (1992: 140, fh 1) points out that "[t]he morpheme de is a suffix (or clitic, depending on one's analysis) that developed historically from the full verb de meaning 'obtain (the result of)'." Now, from this we can say that, in the resultative construction in Chinese, the lexical item de undergoes the same process that the C que 'that' undergoes in the PR in Romance, but that the direction of this process is just the other way around. That is, the head de goes (diachronically) from lexical to semi-lexical, in the sense that, in the resultative construction, de does not behave like a regular verb. Summarizing, the observation that semi-functional or semi-lexical categories, namely categories that come from two opposite extremes, functional or lexical, can function as the head of a CSCI serves us to confirm the idea that the particles as and for that appear in the type of CSCI that has been investigated in this article are semi-lexical or semi-functional heads. 6.

Conclusions

In this article we have seen that not only properties and states can be predicated of an individual in a SCl-configuration, but also situations. This occurs when the syntactic configuration is that of a Complex Small Clause. The particle that links the ontological categories 'situation' and 'individual' is the head of the highest extended projection of the lexical head of the clause (situation). From this position, this particle provides the predica-

498 JoanRafel

tion (situation) that it precedes with a modal or aspectual information. On the other hand, this particle does also function as the head of the construction. That is, the specifier of the projection that this particle heads is the position where the subject of the construction, namely the subject of the CSC1, is base-generated. This particle, then, is the element that may be selected by a lexical head, and the element that eventually will have to incorporate into that lexical head, just like the head of any other type of SCI (see Stowell 1991). The behavior of this particle, then, shows us that in the structural configuration of a CSC1 it functions as an [a functional] [a lexical] head, that is, as a 'semi-lexical' or 'semi-functional' head.

Notes 1. I would like to thank an anonymous reviewer and the audience at the Workshop on Semi-lexical Heads for very helpful comments. 2. In this type of construction, proper names can also appear as predicates. This is shown in (i). (i) They took John for Michael. As predicates, the proper names that can appear in this context cannot be taken as referential. They would rather denote the physical or nonphysical properties that characterize the individual called Michael in the example (i). In other words, the proper name is taken as intentional. This can also be seen in copular sentences: (ii) Mary Clarissa is Agatha Cristie. For different philosophical argumentations regarding the semantics of proper names see Ludlow (1997, part IV). 3. In contrast to English, the phrase tonto 'fool' is adjectival in this construction in Spanish. Note that in this language tonto can be modified by the degree phrase muy 'very', as shown in (i) (see also section 3.1 below). (i) Tomaron a Juan por muy tonto. 'They took Juan for a big fool.' For an interesting discussion of these constructions in Spanish, see Suner (1990). 4. I say "presumably" since, as far as I know, we still lack substantial evidence regarding the position in which the subject of a SCI is base-generated when the predicate of that SCI is a DP. 5. The idea that these particles are Cs is in tune with the fact that, in some languages, SCls can be introduced by a C. In Irish, for instance, a SCI can be introduced by the prepositional C gan, which is a C that expresses sentential negation (see Chung and McCloskey 1987). This is illustrated in (i), from Chung and McCloskey (1987: 186).

As for as / for, they are semi-lexical heads 499

(i) Ba mhinic [cp [c· gan [Sci e sa teach.]]] was often NEG him in-the house 'He was often not in the house.' 6. I ignore the functional projection FP here just for ease of exposition. 7. In the literature, it has been assumed that Α-positions are those bearing grammatical relations. That is, potential theta-role positions like Spec, VP, Spec, XP in a Small Clause, and the specifier of agreement projections, for instance, Spec, AgrsP -or Spec, IP- and Spec, AgroP. On the other hand, A'-positions are operator positions, typically Spec, CP. 8. In this example, the construction under discussion combines with the raising verb to pass. 9. The only sort of "mixed" chains that are allowed in the Minimalist Program (Chomsky 1995) are (αϊ a j . where αϊ occupies an Α-bar position and α^ an Α-position and a t and a„ are interpreted at LF as an operator and a variable, respectively. 10. The pre- or post-verbal position that the subject may occupy in the sentence in Spanish and in other Romance languages has been related to discourse functions. On the one hand, it has been claimed that the subjects that appear in a pre-verbal position, presumably in Spec, IP, make reference to old information (unless they bear a heavy stress), whereas the subjects that show up postverbally, presumably in Spec, VP, indicate new information (see Contreras 1976; Vallduvi 1990; Cinque 1993; Picallo 1999). Recall, furthermore, that in Spanish finite verbs move from the head of the VPprojection to the head of IP at Syntax. 11. The same argument can also be extended to verbs like to regard and to pass. 12. Here I ignore the presence of the D a since, as already pointed out above, we still lack evidence regarding the position in which the subject of a SCI is basegenerated when the predicate of that SCI is a DP. 13. For the notion of extended projection, see Grimshaw (1991) and Van Riemsdijk (1998). 14. This analysis should be easy to translate into your favorite SCl-model without altering the basic idea that is defended here, that is, the CSCl-structure in (24). 15. For a detailed exposition on the notion of the structure of a Complex Small Clause and an exemplification of the constructions that respond to this structure, see Rafel (2000b). 16. In examples of this type, the reference of PRO can also be contextually established. That is, a sentence like that in (i) is also grammatical. (i) Ρor PRO tontas, Juan perdio todo lo que habia ganado. for fool-FEM-PLJuan lost.he all what that had.he won 'Juan lost everything he had won because of their foolishness.' As one can observe, here PRO must refer to a feminine plural N, which does not appear in the sentence.

500

JoanRafel

17. It is important to remark at this point that the predicate of the SCI, namely X', will not need Case even if it is nominal. The reason is simple: this phrase is not an argument, but a predicate that expresses a property that holds of its subject. 18. In Rothstein (1995: 511), for instance, it is claimed that "NPs that are not predicational are licensed by internal θ-marking or as subjects of predicates." 19. The question why the prepositional C for cannot assign Case to the SCl-subject is answered in section 4.1 shortly below. 20. They observe, for instance, that the C that cannot follow the Ρ after in Modern English, (i), as opposed to Old English, (ii). These examples are from Dubinsky and Williams. The example in (ii) is from the fifth century (1464). (i) They came after (*that) they read their Bibles. (ii) It is solde rythe well aftyr that the wole was. 21. Presumably this modification would not apply to those languages that admit the C that following a temporal P, like Spanish, (i), and Dutch, (ii). (i) Despues de que Juan llegara. after of that Juan arrived 'After Juan arrived.' (ii) Na dat Jan arriveerde. 22. As suggested in the previous footnote, the TO would be occupying the specifier of a CP headed by an ordinary C in languages like Spanish and Dutch. This means that, in these languages, the temporal Ρ would precede a CP, instead of a DP. So, in this context, the question whether the Ρ assigns Case reduces to a more general question, that is, whether or not CPs need Case. 23. In this sense, the lexical Ρ in (55) would behave just like a transitive verb. 24. Here I understand the term 'state of affairs' as a static situation that holds throughout some stretch of time. 25. The presumptive nature that a clause has when it is introduced by the particle for is already mentioned in Jespersen (1927): "It should be noted that in nearly all sentences the combination of for and an infinitive denotes some vague possibility or something imagined." (taken from Gee 1975: 301). The only difference is that in (56) for combines with a SCI, instead of an infinitive, or infinitival clause. 26. Of course, the opinion that the sentence in (57) expresses may not be shared by the speaker who utters this sentence. This is shown by the fact that the example in (i) does not express a contradiction. (i) The committee considered you intelligent, but you aren 't. The relevant point here is that, as opposed to the example in (56), in (57) the speaker does not construct a presumed situation that leads the hearer to infer that the speaker questions the truth value of the embedded clause in the actual world, W l . 27. The double nature of these particles is perfectly in tune with the ambiguous properties that modal verbs usually have. That is, they may behave like either full verbs or mere auxiliaries.

As for as / for, they are semi-lexical heads 501

28. Although the semantic value of the combination take-for is similar to the semantic specification of a simple verb like to consider, there are arguments that show that these two verbal forms are not interchangeable. For example, the SCl-predicate that the particle for introduces can be a proper noun (see footnote 1), (i). As (ii) shows, this possibility is not available when the SCI is selected by the verb to consider. (i) They took John for Richard. (ü)*They considered John Richard. And, secondly, there are prepositional predicates that cannot appear in the SCI introduced by the particle por 'for' in Spanish, (iii), whereas these prepositional predicates are admitted by the SCI subcategorized for by the verb considerar 'to consider', (iv). (iii) *Lo tomaron por [de muy mal gusto / de mala educacion.} it took, they for of very bad taste / of bad manner (iv) Lo considero {de muy mal gusto / de mala educacion.} it consider.I of very bad taste / of bad manner Ί consider it {bad taste / bad manners.}' 29. By the term 'complete clause', I mean a full sentence, namely a domain headed by the lexical head V, that possesses its own temporal specification. 30. This option is not possible when the particle is for / por, in English, (i), or in Spanish, (ii). (i) *They took John for if he were intelligent. (ii)*Tomaron a Juan por si (el) fuera inteligente. This could indicate that the particle for / por cannot be used as a pure lexical head, according to the position that we have adopted here. The different behavior of as and for / por in this context deserves a careful examination on its own, so here I leave this point as a simple observation. 31. In Bowers (1993), for instance, it is claimed that as is simply "a direct lexical realization of Pr" (p. 597), that is, the head of the Predicate Phrase (PrP) that introduces any type of SCI. Similarly, in Haegeman (1994), it is held that as is a lexical realization of agreement (Agr). 32. There are many arguments that conclusively demonstrate that this construction is not a relative clause (see the translations into English). I refer the reader to the references cited in the text. 33. In this sense, the C que 'that' behaves like the Ρ aan 'at' that appears in the Dutch construction in (i): (i) Jan is het kind aan het uitschelden. Jan is the kid at the insult-INF 'Jan is insulting the kid.' This is not so surprising if we bear in mind that in some Romance languages the C que 'that' can have an adverbial value. This is the case of Catalan: (ii) He arribat que Γ autobus ja havia marxat. have.I arrived that the.bus already had.it gone Ί arrive {at the time at which / when} the bus had already left.'

502 JoanRafel References Bowers, John 1993 The syntax of predication. Linguistic Inquiry 24: 591-656. Chomsky, Noam 1995 The Minimalist Program. Cambridge, Mass.: MIT Press. Chung, Sandra and James McCloskey 1987 Government, barriers, and small clauses in Modern Irish. Linguistic Inquiry 18: 173-237. Cinque, Guglielmo 1993 A null theory of phrase and compound stress. Linguistic Inquiry 24: 239-297. Contreras, Heles 1976 A Theory of Word Order with Special Reference to Spanish. Amsterdam: North Holland. Dubinsky, Stanley and Kemp Williams 1995 Recategorization of prepositions as complementizers: The case of temporal prepositions in English. Linguistic Inquiry 26: 125-136. Emonds, Joseph 1985 A Unified Theory of Syntactic Categories. Dordrecht: Foris. Gee, James 1975 Perception, intentionality, and naked infinitives: A Study in Linguistics and Philosophy. Ph.D. dissertation, Standford University. Grimshaw, Jane 1991 Extended projection. Manuscript., Brandeis University. Guasti, Maria Teresa 1988 La pseudorelative et les phenomenes d'accord [The Pseudo-Relative and the agreement phenomena]. Rivista di Grammatica Generativa 13: 35-58. Haegeman, Liliane 1994 Introduction to Government and Binding Theory. Oxford: Blackwell. Hantson, Andre 1989 The complementizer 'as'. In: Dany Jaspers (ed.), Sentential Complementation and the Lexicon, 207-219. Dordrecht: Foris. Huang, C.-T. James 1982 Logical relations in Chinese and the theory of Grammar. Ph.D. dissertation, MIT. 1992 Complex predicates in Control. In: Richard K. Larson, Sabine Iatridou, Utpal Lahiri, and James Higginbotham (eds.), Control and Grammar, 109-147. Dordrecht: Kluwer Academic Publishers.

As for as / for, they are semi-lexical heads 503

Jespersen, Otto 1927 A Modern English Grammar, 7 vols. London: George Allen and Unwin. Larson, Richard K. 1988 Extraction and multiple selection in PP. MIT Working Papers in Linguistics 9: 119-135. Ludlow, Peter 1997 Readings in the Philosophy of Language. Cambridge, Mass.: MIT Press. Manzini, Maria Rita 1992 Locality: A Theory and Some of its Empirical Consequences. Cambridge, Mass.: The MIT Press. Picallo, Maria Carme 1999 Rasgos sintäcticos y posiciones argumentales en catalän. Report de recerca GGT, Universität Autönoma de Barcelona. Rafel, Joan 2000a From C to P: Evidence from Romance. Probus 12: 139-163. 2000b Complex Small Clauses. Ph.D. dissertation, Universität Autönoma de Barcelona. Riemsdijk, Henk van 1998 Categorial feature magnetism: The endocentricity and distribution of projections. Journal of Comparative Germanic Linguistics 2: 148. Rothstein, Susan 1995 Pleonastics and the interpretation of pronouns. Linguistic Inquiry 26: 499-529. Starke, Michal 1995 On the format for small clauses. In: Anna Cardinaletti and Maria Teresa Guasti (eds.), Syntax and Semantics: Small Clauses, 237269. London: Academic Press. Stowell, Tim 1981 Origins of phrase structure. Ph.D. dissertation, MIT. 1983 Subjects across categories. The Linguistic Review 2: 258-312. 1991 Small Clause restructuring. In: Robert Freidin (ed.), Principles and Parametres of Comparative Grammar, 182-218. Cambridge, Mass.: MIT Press. Suner, Avellina 1990 La predicaciön secundaria en espanol [Secondary predication in Spanish], Ph.D. dissertation, Universität Autönoma de Barcelona. Vallduvi, Enric 1990 The informational component. Ph.D. dissertation, University of Pennsylvania, Philadelphia.

Lexical particles, semi-lexical postpositions Jochen Zeller

1.

Introduction

In this paper 1 ,1 suggest that the notion of semi-lexicality can be defined in morphological terms. I assume that a semi-lexical head is "half lexical, half functional", by which I mean that it is a morphologically complex element that consists of a lexical node and a functional suffix. I elaborate this hypothesis through a detailed discussion of the properties of postpositions and particles in German and Dutch. It can be shown that postpositions are functional heads that project functional phrases, whereas the projection of a particle is a lexical PP without functional structure. Although there are a number of differences between particle phrases and postpositional phrases that follow from this fact, particles and postpositions also show certain parallels that cast doubt on the claim that the functional status of postpositions mirrors that of other functional elements like tense affixes or determiners. Rather, I show that postpositions are not "genuine" functional heads, but semi-lexical elements. They are derived from lexical prepositions via suffixation of a zero-operator that alters the thematic properties of the P-element. Since this operator is a functional element, the derived postposition is a complex functional head. However, it inherits the semantics of the lexical preposition from which it is derived. Therefore, postpositions are semi-lexical elements; they have the semantic content of a lexical element, but the categorial properties of a functional suffix. The paper is organized as follows. In section 2,1 discuss prepositional, postpositional and particle phrases, and I show in what respects particles and postpositions differ and how they behave similarly. Section 3 provides an analysis of the morphological structure and the thematic properties of postpositions. From this analysis I conclude that postpositions are semilexical elements. In section 4,1 use the results of the previous sections to

506 Jochen Zeller compare postpositional phrases in Dutch and post- and circumpositional phrases in German. I first introduce some terminology and some notations. I treat local transitive prepositions as expressing relations between two (individualtype) entities. In (1), for example, the preposition in provides the information that the bread stands in the IN-relation to the closet: (1)

The bread is in the closet

I refer to the external argument of a preposition (the DP the bread in (1)) as the theme; its internal argument (the closet), I call the reference object. I assume that in examples like (1), the theme is generated in SpecVP and moves to SpecIP, whereas the reference object is the complement of the preposition and receives case from this lexical head. I follow Wunderlich and Herweg (1991), Stiebeis (1996), Olsen (1999), and others in expressing the semantics of prepositions through the general form in (2): (2)

Xy λχ [LOC(x) c R(y)]

According to (2), transitive prepositions are represented as two-place functions. The first argument corresponds to the reference object, the second argument corresponds to the theme. The argument structure of a preposition is expressed through the lambda operators that bind the respective variables inside the formula. The predicates in the formula provide the information that the location of the theme intersects with a particular region R of the reference object y. The specific properties of R are defined by the respective preposition. The preposition in, for example, is represented as in (3) (cf. Olsen 1999: 117): (3)

λy λχ [LOC(x) c IN(y)]

(3) represents the situation in which the referent of the x-argument is located in the internal region of the referent of y. Notice that (3) is the locative variant of in. The directional version is represented as in (4): (4)

Xy λχ [FIN(LOC(x)) c IN(y)]

Lexical particles,

semi-lexical postpositions

507

The predicate FIN maps the location of χ onto a final part of a path traversed by x. The directional preposition in provides the information that the final part of this path ends in the interior region of y. 2. 2.1.

Prepositional phrases and functional structure The extended projection ofprepositions

Van Riemsdijk (1990, 1998) and Koopman (1993) suggest that PPs are selected by functional prepositional heads that license certain lexical and grammatical properties of the lexical preposition. In terms of Grimshaw's (1991) theory, this means that prepositions, like the lexical categories Ν and V, have extended projections. I adopt this idea and assume with Van Riemsdijk (1990) that the structure of prepositional phrases in German looks as in (5) (cf. Zeller 1999):2 (5)

FPprep



DP

As (5) shows, the maximal projection of P° is the complement of a functional prepositional head that I simply label F°prep (it corresponds to p° in Van Riemsdijk (1990)). In constructions with transitive or intransitive prepositions, F V P is phonologically unrealized: (6)

transitive P°:

Youri steigt auf den Berg Y. climbs up/on the mountain [FP [F [ PP auf [DP den Berg]] 0 ]]

(7)

intransitive P°: Youri wohntoben Y. lives upstairs [FP [F [ PP oben] 0 ]]

In (6), the PP is headed by the transitive preposition auf that takes its internal argument as a complement to its right. In (7), the PP consists solely of its head, the intransitive preposition oben.

508 Jochen Zeller F 0 ^ may also host lexical material. Following Van Riemsdijk (1990, 1998), I assume that in circumpositional phrases like those in (8), F 0 ^ is realized by a postposition, printed in boldface: (8)

a.

b. c.

(9)

Mirco geht unter der Brücke durch Μ. goes under the bridge through 'Mirco passes under the bridge.' Ein Tourist steigt auf den Berg hinauf a tourist climbs up/on the mountain A-up/on Der Hund kommt aus dem Haus heraus the dog comes out-of the house A-out

[FP [ F [ppauf [Dp den Berg]] hinauf]]

As (8) shows, German has both complex and simple postpositions. Whereas the postposition durch in (8a) is homophonous with the preposition durch, the postpositions hinauf and heraus in (8b) and (8c) consist of the prepositional elements auflaus and a prefix her/hin (henceforth h-prefix) which has a deictic function (towards/away from the speaker). In the following, I will refer to these complex postpositions as h-postpositions\ phrases headed by Α-postpositions always receive a directional interpretation.3 Notice that in (8b) and (8c), the prepositional parts of the A-postpositions {auf in hinauf, aus in hinaus) are identical to the lexical heads of the respective PP-complements. Constructions with this property are called pleonastic circumpositional phrases; their PP-constituents are referred to as "cognate" PPs (cf. Olsen 1996; Mclntyre 1998). Interestingly, Α-postpositions may also appear without a complement: (10)

a. b.

(11)

a. b.

(12)

Ein a Der the

Tourist tourist Hund dog

Johan J. Johan J.

steigt climbs kommt comes

schickt sends bringt takes

hinauf A-up/on heraus A-out

einen Touristen hinauf the tourist A-up/on den Hund heraus the dog A-out

[YP [DP ein- Tourist] [V [FP hinauf ] schick- ]]

Lexical particles, semi-lexical postpositions 509 As I will show in detail in section 3.2, the PP-complement of a postposition helps identify the internal argument of the prepositional relation expressed by the FPprep. The ^-postpositions in (10) and (11), however, are used intransitively. They occupy the F0Prep-position, but do not take PP-complements. Therefore, the internal argument of the relation expressed by the hpostpositions remains implicit; the constructions in (10) and (11) only provide information about the external arguments of hinauf and heraus, which are expressed by the DPs ein Tourist and der Hund (represented as derived subjects in (10) and as direct objects in (11)). For example, in (10b) and (lib), it is clear that the dog moves out of something, but this something, the reference object, is not mentioned explicitly. In the nominal domain, a functional head may occur without a lexical complement as well. Since Abney (1987), it is widely assumed that pronouns are realizations of a functional D°-head with no NP-complement (cf. e.g. Stowell 1991; Longobardi 1994). Analogously, we can analyze intransitive postpositions like those in (10) and (11) as prepositional proforms; they are F°ptep-elements without PP-complements. Some Α-postpositions also project postpositional phrases·. (13)

Ein Tourist steigt den a tourist climbs the

Berg hinauf mountain A-up/on

Note first that postpositional structures like (13) must not be confused with examples like (11a). Although both sentences include a DP bearing accusative case, the DP einen Touristen in (11a) is the theme of the prepositional relation and is located outside the FPpreP (cf. (12)), whereas the DP den Berg in (13) identifies the internal argument of the relation expressed by hinauf and is hence generated inside the FPprep. Evidence for this latter assumption is provided by the following contrast: (14)

a. b.

Den Berg hinauf steigt ein Tourist Einen Touristen hinauf schickt Johan

In (14a), FPprep-topicalization has moved the Α-postposition and the accusative DP to SpecCP. In (14b), however, topicalization of the accusative DP and the Α-postposition is ungrammatical, because the Α-postposition is a proform, and the DP is the theme which is not generated inside the FPp„.p.4 It has been suggested that the DP in postpositional structures like (14) is the complement of an empty preposition that projects the PP-complement of the A-postposition:

510 Jochen Zeller

(15)

[FP [F [ PP 0 [DP den Berg]] hinauf ]]

Opinions differ with respect to the status of the empty preposition in (15). Van Riemsdijk (1990, 1998) argues that the head of the PP in (15) is the trace of the preposition auf that has moved to F^ep where it combines with the /i-prefix. In contrast, Mclntyre (1998) suggests that the head of the PP is an independent element, a preposition without a phonological realization, but with a specific meaning. I will return to this issue in section 4 where I compare the two proposals in light of a discussion of postpositional structures in Dutch and circumpositional phrases in German. For now, it suffices to assume that the structure of postpositional phrases includes a PP whose head is not realized phonologically. So far, I have only provided examples from German. However, the structure in (5) also accounts for the properties of pre-, post-, and circumpositional phrases in Dutch: (16)

a. b. c. d.

omdat Jan op de berg reed because J. up the mountain drove Hij is binnen/buiten/boven/beneden he is inside/outside/upstairs/downstairs Het vliegtuig is onder de brug door gevlogen the airplane is under the bridge through flown omdat Jan de berg op reed because J. the mountain up drove

The transitive and the intransitive prepositions in (16a) and (16b) are heads of PPs selected by an empty F0preP; in the circumpositional structure in (16c), the PP onder de brug is the complement of the postposition door, and in the postpositional phrase in (16d), the postposition op selects a PPcomplement with an empty P°. (Notice that Dutch does not have complex postpositions that correspond to Α-postpositions in German.) I will come back to postpositional structures like (16d) in section 4.

13.2. Particles In this section, I will briefly introduce the analysis of particle verbs that I develop in Zeller (1999). Based on proposals made by Koopman (1993) and Haiden (1997), I suggest that particles are represented as phrasal com-

Lexical particles, semi-lexical postpositions 511 plements of their verbs that lack functional structure. Thus, the prepositional particle verbs in (17) have a structure like (18): (17)

a.

German:

b.

Dutch:

(18)

weil Youri aufsteigt because Y. Prt-climbs 'because Youri moves up, rises.' omdat Jan mijn broer op beide because J. my brother up called

V' PP

V° steig-

po

auf Furthermore, I argue that particles do not undergo incorporation in order to combine with the verb.5 Instead, I assume that the adjacency of a particle and the verb in verb-final clauses like (17) is the result of the structure of the VP in Dutch and German. Since both languages are SOV, the PP-complement which includes the particle precedes the verb (see Zeller 1999, chapter 2, for detailed discussion): (19)

[ c weil [Π» Youri [VP [V [PP auf] steigt ]] ]]

Since particles do not move, the two terminal nodes that a particle verb consists of (i.e. P° and V°) do not form a syntactic word (a complex V°). However, the relation between the two nodes is nevertheless strictly local, due to the absence of functional structure. I define the relevant local relation as "structural adjacency": (20)

Structural adjacency A head X and the head Y of its complement YP are structurally adjacent.

The central claim that I make in Zeller (1999) is that the particle status of a prepositional element is caused by its syntactic environment. If a lexical head P° is structurally adjacent to a functional head F0Prep, it is a regular

512 Jochen Zeller

locational preposition; however, if it is structurally adjacent to a verb, it is a particle. Following the extended Distributed Morphology-framework outlined in Marantz (1997), I assume that special meanings are associated with terminal nodes according to particular syntactic environments; meaning is determined on the basis of a syntactically determined locality domain. I suggest that the locality domain relevant for the thematic and semantic properties of prepositional elements is defined by structural adjacency. If a terminal node of category P° that has a particular phonological form (say, e.g., /an/) is structurally adjacent to a functional head, it is a regular preposition and is associated with the respective semantic and thematic properties. However, if an is structurally adjacent to a verb, it is a particle whose meaning may differ in crucial respects from the semantics of the preposition. For example, the preposition an in (21) is obligatorily transitive and expresses a location: (21)

Die Leute stehen an der Ecke the people stand at the corner

In contrast, the^article an in (22) is intransitive; its reference object may be left unrealized: (22)

Die Leute stehen an the people stand Prt 'The people queue up.'

Moreover, the particle an can have aspectual meanings which are entirely unrelated to the meaning of the preposition in (21): (23)

a.

b.

Peter brät das Fleisch an Peter fries the meat Prt 'Peter fries the meat lightly.' Der Zug rollt an the train rolls Prt 'The train starts to roll.'

In (23a), an expresses that the frying-event is only carried out lightly or partially; an in (23b) is an inchoative operator that focuses on the start of the rolling-event. Both uses of the particle an are at least semi-productive, but different from the prepositional meaning of an in (21). According to my

Lexical particles, semi-lexical postpositions 513

analysis, the intransitivity or the aspectual meanings of a prepositional element like an are only licensed if P° is structurally adjacent to a verb. The different lexical properties of particles and regular prepositions are a result of their different structural contexts. In a nutshell, the notion "particle" refers to a prepositional element that is structurally adjacent to a verb, i.e. to the head of a PP without functional structure. In contrast, a "regular" prepositional phrase is an FPprep. In the next section, I will show that there are certain systematic differences between particle verbs and verb + FPpreP-complement-constructions that follow from this difference.

2.3.

2.3.1.

Differences between particle phrases and fimctional preprositional phrases P° as a case-assigner

Compare (24) and (25): (24)

a.

b. (25)

a. b.

Peter hat sein Bier aus der Flasche getrunken Peter has his beer out-of the bottleDAT drunk 'Peter drank his beer from the bottle.' [Ai« der Flasche\ hat Peter sein Bier tj getrunken Peter hat (*sein Bier) die Flasche ausgetrunken Peter has his beer the bottleAcc out-drunk *Die Flasche aus hat Peter getrunken

In (24), we have a regular FPprep with an empty F0Prep and a PP-complement aus der Flasche. The reference object, the DP die Flasche, receives case from aus, which assigns dative case. The preposition and its DP-complement can be topicalized together, as shown in (24b). However, if the prepositional element aus is used as a particle, the situation changes. First, if the reference object of aus becomes the object of the particle verb, the theme of the particle aus cannot be realized, (25a).7 Second, the reference object of aus now occurs to the left of P°, and it bears accusative case. Moreover, as (25b) shows, topicalization of P° together with the reference object is now impossible. These differences follow from the different structural environments of prepositions and particles. On standard assumptions, case-assignment prop-

514 Jochen Zeller erties of lexical nodes are licensed through functional structure (cf. Koopman 1993; Borer 1998). This means that in (24), due to the presence of a functional head F0Prep, P° can assign case to its internal argument-DP. Therefore, this DP stays inside the PP to the right of P°, it receives the case which is assigned by P° according to P°s lexical specification, and both elements can be moved together. In contrast, particle phrases lack functional structure; therefore, P° is not able to assign case if it is a particle. Consequently, the reference object cannot be located inside the PP. It must either be generated in or moved to a position where it cang receive the structural case assigned by the verb, which is accusative. As a result, the particle and its reference object in (25) do not form a constituent that can be topicalized. 2.3.2.

Complements of Ν

In German, regular FPpreP-projections may also occur as complements of N. This is illustrated for prepositional and circumpositional phrases in (26b) and (27b): (26)

a.

b.

(27)

a.

b.

auf die Mauer springen on the wall jump 'jump on the wall.' der Sprung auf die Mauer the jump on the wall (in die Stadt) hinein fahren (in(to) the city) h-in drive 'drive into the city.' die Fahrt (in die Stadt) hinein the drive (in(to) the city) A-in

In contrast, as first observed by Van Riemsdijk (1978), particles are not licensed as complements of N: (28)

a.

aufspringen up-jump 'jump up.' b. *der Sprung auf

Lexical particles, semi-lexical postpositions

(29)

515

a.

{in den Hafen) einfahren (into the harbor) in-drive 'sail into the harbor.' b. *die Fahrt (in den Hafen) ein

(30) illustrates the different structures of (26b) and (28b): (30)

a.

auf die Mauer b.

*NP N° Sprung

PP po

auf The contrast between (26)-(27) and (28)-(29) follows from the specific conditions that must be fulfilled in order for an element to become a particle. A prepositional element becomes a particle only by virtue of a verbal context; i.e. if it is structurally adjacent to a verb. Therefore, (28b) and (29b) are excluded by definition, since auf and ein are structurally adjacent to N. In (26) and (27), however, these elements are used as prepositions and are structurally adjacent to a functional head. Since an FPpreP is allowed as a complement of both V and N, (26b) and (27b) are grammatical. Again, the proposed analysis accounts straightforwardly for this difference between particle phrases and functional prepositional phrases.

2.3.3.

Referentiality

A third difference between particle verbs and FPpreP-constructions is best illustrated through a comparison between (31) and (32):

516 Jochen Zeller

(31)

a.

b.

(32)

a.

b.

Hier strömt Gas heraus here streams gas A-out 'Gas escapes (out of some contextually given entity).' Peter will einen Kreis heraus schneiden Peter wants a circle A-out cut 'Peter wants to cut out a circle (out of some contextually given entity).' Hier strömt Gas aus here streams gas Prt 'Gas escapes here.' Peter will einen Kreis ausschneiden Peter wants a circle Prt-cut 'Peter wants to cut out a circle.'

The examples in (31) show verbs with FP^-complements; the heads of these complements are intransitive Α-postpositions. In (32), we have particle verbs. The particles correspond to the prepositional elements that form the Α-postpositions in (31). The Α-postpositions in (31) are prepositional proforms; the reference object of the prepositional relation remains implicit (cf. section 2.1). In (32), it is the verbal context that licenses the intransitive use of the prepositional elements (as was illustrated in (21) vs. (22) in section 2.2); the internal argument of the prepositional relation that is expressed by the particles also remains unrealized. Therefore, we have an implicit reference object in both (31) and (32). Nevertheless, there is a systematic semantic difference between examples like (31) and (32). Mclntyre (1998) observes that in spite of an implicit reference object, an FPfrep headed by an intransitive A-postposition still expresses a referential path. Like nominal pronouns that refer to a specific individual, the prepositional proforms in (31) refer to particular instances of a path-concept. The referentiality of a path can best be tested by considering the interpretation of the implicit reference objects in the examples in (31). As Mclntyre (1998) notes, a referential path requires a referential reference object. This means that the internal arguments of the Α-postpositions in (31), although not mentioned explicitly, must be contextually given. Whenever a speaker uses a prepositional proform, she also talks about a specific instance of a reference object. For example, although it is not specified in (31a) from where exactly gas is escaping, the source of the OUT-OF-relation expressed by the Α-postposition must be inferable

Lexical particles, semi-lexical postpositions 517 from the context. (31a) can only be uttered felicitously if the discourse allows the hearer to discover the respective source of the escaping gas (a whole in the wall; a pipe) from the context. On the basis of this contextually given reference object, the hearer can construct a referential, tokenlike interpretation of the path denoted by the respective FPprep. In contrast to the referential interpretation of the FPprep-constructions in (31), the particle phrases in the examples in (32) are interpreted non-referentially. They express non-specific types of paths characterized by the respective prepositional relation. Again, this can be tested through the interpretation of the implicit reference object. The implicit reference objects of the particle verbs in (32) are not contextually given. For example, if a speaker utters (32a), it may not be clear at all from where gas is escaping. The relevant information here is only that gas is escaping. In Zeller (1999) I discuss the well-known correspondence between the referential interpretation of a phrase and functional structure. It is widely assumed that e.g. noun phrases can only be interpreted referentially if functional structure is present (cf. Stowell 1991; Longobardi 1994). According to the analysis presented above, it is not surprising that the prepositional phrases in (31) are referential, since the Α-postposition is a proform and realizes a functional head. By the same token, the non-referential interpretation of particle phrases follows directly from the analysis of particle verbs presented in section 2.2. Particle phrases lack functional structure by definition. However, if a referential interpretation is only available if functional structure is present, we expect that particle phrases can only express non-referential path-types. Again, the different interpretations of the FPprePconstructions in (31) and the particle verbs in (32) follow from the differences discussed in sections 2.1 and 2.2 above.

2.4.

Particle-like properties of postpositions

The data discussed in the preceding section show that a number of differences between particle phrases and pre-, post-, or circumpositional phrases follow from the assumption that the latter, but not the former, are functional projections. Particle phrases do not include arguments of the prepositional particle, they cannot occur as complements of N, and they always express non-referential paths. In contrast, functional structure licenses case-assignment to a DP inside the PP; a functional prepositional phrase can occur as a complement of N, and it is always interpreted referentially.

518 Jochen Zeller

However, the structure of particle verbs and the structure of verbs that take FPprep-complements show an interesting parallel: (33)

a.

VP PP

b.

VP

^PP^



F°prep

DP

The lexical head P° in (33a) is structurally adjacent to the verb, whereas P° in (33b) is not, because functional structure intervenes. However, in (33b), there is also a head that is structurally adjacent to the verb: the functional head of the FPpreP itself. This is an important fact in light of the observation that F°prep can be filled with lexical material like postpositions. If the proposal illustrated in section 2.2 is on the right track, and the particleproperties of a preposition follow from its local verbal environment, then we might expect to find similarities between postpositions and particles, because both heads are structurally adjacent to the verb. In fact, there are certain parallels between particles and Ä-postpositions which I will discuss in the following sections. However, I will also show that, although Ä-postpositions are structurally adjacent to verbs, an additional assumption is needed to account for the particle-like properties of postpositions. This assumption is explored in section 3.

2.4.1.

Meaning variations

In section 2.2 I noted that the meaning of many particles differs from the meaning of the corresponding preposition, and I suggested that these semantic peculiarities of particles follow from the fact that they are structurally adjacent to a verb. Interestingly, similar semantic differences between postpositions and the basic prepositions from which they are

Lexical particles, semi-lexical postpositions 519 derived are also attested. Mclntyre (1998) observes that not all A-postpositions are systematically related to their corresponding prepositions: (34)

a.

b.

c.

d.

herum (um = around) action without immediate accomplishment of its purpose; inappropriate or inept activity: herumdoktern, 'try one's hand at'; herumschreien, 'shout for no reason', herumalbern, 'fool around' heran (an = at, to) expression of temporal approach: heranrücken, heranbrechen, heranziehen, 'draw near, approach, advance' herab (ab = off, from) expression of downward movement (not part of the meaning of ab): herabkommen, 'come down'; herablassen, 'let down' hervor (vor = in front of) expression of emergence from concealment: hervorkommen, hervorgehen, 'emerge from', hervorblicken, 'peep from behind', hervorheben, 'emphasize'

The examples in (34) hint that at least some Α-postpositions have special meanings that are not derived entirely on the basis of a combination of the meanings of their parts. Like particles, postpositions are structurally adjacent to the verb. Let us assume for the moment that this structural environment is responsible for the semantic peculiarities of Α-postpositions that are illustrated in (34) and turn to another striking parallel between particles and postpositions.

2.4.2.

Word formation

An important property of particle verbs is that they can form the input to further operations of derivational morphology: (35)

a. b. c. d.

einfuhren, 'introduce' —> Einführung, 'introduction' ausgraben, 'excavate' —> Ausgrabung, 'excavation' aufblasen, 'inflate' —> aufblasbar, 'inflatable' annehmen, 'accept' —> unannehmbar, 'unacceptable'

(35) raises the question of whether the structure of these derived nominale and adjectives is morphological (= the respective words are derived ex-

520 Jochen Zeller

clusively from minimal projections like stems and affixes) or syntactic (= the structure of the derived nominale and adjectives in (35) includes a maximal projection, for example a VP). The analysis of particle verbs that I proposed above seems to favor the latter conclusion. If particles have to be structurally adjacent to verbs, this requirement automatically implies that particle verbs are V's or VPs. This seems to suggest that the derivational suffixes -ung and -bar in (35) combine with a VP-structure like (33a) that corresponds to the particle verb. The affixal properties of -ung and -bar may then be satisfied via incorporation of the verb into the structurally adjacent N- or A- node. However, a syntactic derivation of nouns and adjectives makes certain predictions that are not realized with all words derived from particle verbs. For example, Borer (1993) argues that the well-known differences between process and result nominale are caused by a structural difference. According to her proposal, process nominals are derived syntactically and include a VP-projection of the base verb; this VP gives rise to the "verbal" properties of process nominals. In contrast, no such VP is present in result nominals. If all word formation with particle verbs were syntactic, we would predict that nouns derived from particle verbs are always process nominals. However, (36) shows that this prediction is not borne out: (36)

a. b. c.

die Anweisung wurde ihm zugeschickt 'the instruction was sent to him.' die Ausarbeitung lag auf dem Schreibtisch 'the elaboration lay on the desk.' die Auszahlung bestand nur aus 100-Dollar Scheinen 'the disbursement consisted only of 100-dollar bills.'

Since the nouns in (36) are result nominals, it is extremely unlikely that they include a phrasal projection of the verb. In Zeller (1999, chapter 6) I provide more evidence that particle verbs are non-phrasal constructions when they appear inside derived adjectives or nouns. Of course, this creates a problem for the analysis suggested in section 2.2. If particles must be structurally adjacent to their verbs, the structural representation of a particle verb is phrasal; it is a V- or VP-node that dominates the verb and the particle phrase. I therefore suggest in Zeller (1999) that structural adjacency of the particle and verb is not required in morphological contexts. I argue that the local domain that is established between a particle and a verb through structural adjacency can be reanalyzed when morphological operations are applied to the particle verb.

Lexical particles, semi-lexical postpositions

521

The basic idea behind this reanalysis-proposal is simple. I assume that two syntactic heads that stand in a local (but non-morphological) relation to each other can be combined to form a morphologically complex word if this representation is required by independent conditions. The attachment of a derivational affix is such a condition; it may require its input to be a word. Since the relation between a particle and a verb is strictly local, the particle and the verb can be reanalyzed as a complex morphological object if a derivational affix is attached.10 Importantly, the particle and verb can only be reanalyzed because of the structural-adjacency relation that is the result of the absence of functional structure. If a lexical head has an extended projection which is the complement of a verb, reanalysis of this head and the verb is impossible, because the relation between the two nodes is non-local.11 As shown in (33), a verb and an Λ-postposition are also structurally adjacent. Therefore, we expect reanalysis to be possible with these nodes as well. This expectation is borne out: (37)

a. b. c.

herausfordern, 'challenge' —» Herausforderung 'challenge' hinzufügen, 'add' —> Hinzufügung, 'addition' herausnehmen, 'take out' —> herausnehmbar lit.: 'out-take-able'

(37) shows that the Α-postposition and its base verb can also be the input to further morphological operations. Again, we can tentatively assume that the structural adjacency of these two nodes allows reanalysis. Thus, the derivational affixes in (37) can attach to the words built from the ^-postposition and the verb.

2.4.3.

Verb Raising in Dutch

Another phenomenon which illustrates that particles and postpositions have similar properties is Verb Raising in Dutch. In Verb Raising, an embedded infinitive raises and attaches to the right of the matrix verb. If the embedded infinitive is a particle verb, two options exist. Either only the verbal part moves, and the particle is stranded, (38a), or the particle verb moves as a whole, (38b) (cf. Evers 1975; Van Riemsdijk 1978):

522 Jochen Zeller (38)

a. b.

dat that dat that

Jan J. Jan J.

zijn his zijn his

moeder op tj wil [bellen]i mother up wants phone moeder tj wil [opbellen]i mother wants up-phone (Neeleman 1994: 24)

Notice that transitive and intransitive prepositions, and (most) resultative predicates cannot be moved with the verb: (39)

a.

*dat zij de jas de stoel keen hebben [over gelegd] that they the coat the chair (heen) have over put (Koopman 1993: 33) b. *dat Jan tj wil [boven wonen]\ that J. wants upstairs live (Den Dikken 1995: 30) c. *dat Jan de deur tj wil [violet verven]j that J. the door wants violet paint (Neeleman 1994: 23)

Curiously, postpositions, like particles, can move together with the verb in Verb Raising constructions, as shown in (40b): (40)

a. b.

omdat because omdat because

hij de boom in tj is [geklommen], she the tree in is climbed hij de boom tj is [in geklommen]\ she the tree is in climbed (Van Riemsdijk 1978: 98)

Since Dutch does not allow for Verb Projection Raising, the raised constituents in (38b) and (40b) must be complex V°s that consist of the verb and the particle or postposition, respectively.12 In Zeller (1999), I argue that these complex verbs have been derived via reanalysis, just like the complex verbs that form the input to operations of word formation in (35) and (37) above. The following data provide evidence for this view:

Lexical particles, semi-lexical postpositions 523 (41)

a.

datJan that J. b. dat Jan that J. c. *dat Jan that J.

de bal [vlak over]fp tj heeft geschoten\ the ball right over has shot de bal tj heeft [over geschoten], the ball has over shot de bal [vlak tj]n>ti heeft [[over\ geschoten], the ball right has over shot (cf. Den Dikken 1995: 108)

The postposition over in (41a) is modified by the adverb vlak. In (41b), the complex verb consisting of postposition and verb has been raised into the matrix clause. If this V° had been derived via incorporation of the postposition into V°, we would expect that the adverbial modifier could be stranded inside the maximal projection of the postposition. However, (41c) shows that the modifier is not licensed if the postposition has undergone raising with the verb. This provides evidence that examples like (40b) do not include a maximal projection of the postposition. Complex verbs that undergo Verb Raising are derived via reanalysis of the verb and the postposition or particle. Structural adjacency is a necessary condition for reanalysis; postpositions, like particles, can be reanalyzed with the verb.

2.4.4.

Functional elements as particles: The problem

The preceding section has shown that postpositions share certain properties with particles. As with particles, the meaning of an Α-postposition may differ from the meaning expressed by the corresponding preposition, and like particle verbs, the combination of a postposition and the verb may be reanalyzed to form the input to morphological derivations or syntactic movement rules. As the two structures in (33) above show, both particles and postpositions are structurally adjacent to the verb. In line with the idea that properties of terminal nodes are determined by properties of their (syntactically defined) locality domains, the fact that both postpositions and particles occur in non-functional (namely, verbal) environments may be taken as one reason for the observed parallels. However, whereas particles are lexical heads, postpositions are functional elements. If postpositions were genuine functional heads like, for example, complementizers or determiners, we would face a serious problem. The particle-like properties of postpositions are properties that are usually not attested in the functional domain. For example, if postpositions are functional heads, we expect a certain "semantic stability" - functional

524 Jochen Zeller elements in the verbal or nominal domain, like, for example, a past tense morpheme or an indefinite article, have fixed meanings and are not ambiguous. As the discussion in section 2.4.1 has shown, however, functional postpositions behave differently. A similar problem arises with respect to reanalysis. If any two structurally adjacent heads can be reanalyzed in order to undergo operations of derivational morphology or Verb Raising, one wonders why reanalysis does not also form words that consist of other functional heads and the heads that select their maximal projections. For example, we would expect that a determiner and the verb that selects the respective DP can be reanalyzed and move as a complex verb in Dutch Verb Raising constructions. Of course, this option is absolutely excluded: (42)

a.

dat hij een huis tj wilde [kopen], that he a house wanted buy b. *dat hij huis wilde [een-kopen], that he house wanted a-buy

Furthermore, it is commonly assumed that derivational affixes do not attach to words that include functional material. Therefore, it should be impossible to reanalyze a functional and a lexical head and apply operations of derivational morphology to the resulting word. If an Α-postposition like heraus is an F0prep, then a noun like e.g. Herausforderung should not exist. The nominalizing suffix -ung should not be allowed to attach to the complex verb herausforder-, because this verb includes functional material. The solution to this problem that I will offer is based on the intuition that in a particular sense, postpositions are not "as functional" as e.g. determiners and complementizers. This particular sense will be specified in the next section. What I suggest is that postpositions are semi-lexical heads, and I will define semi-lexicality as a property of functional heads that are morphologically derived by attaching a functional suffix to a lexical node. The semi-lexicality of postpositions makes them elements of category Fprep, which gives rise to the "functional" properties of their projections. At the same time, however, they incorporate the semantic content of a lexical preposition. I suggest that their semi-lexical status allows postpositions to behave like particles, i.e. like fully lexical heads that are structurally adjacent to a verb.

Lexical particles, semi-lexical postpositions 525 3. 3.1.

The semi-lexicality of postpositions The morphological structure of postpositions

A first step in investigating the internal structure of postpositions is to look at the structure of complex Α-postpositions like German heraus or hindurch. I will argue below that the deictic prefixes hin and her are functional elements. Therefore, one might assume that these prefixes are the source of the functional status of Α-postpositions. However, there are two obvious problems with this idea. First, recall that some postpositions in German and most postpositions in Dutch do not require a prefix and are therefore homophonous with an existing preposition: (43)

a. b.

(44)

a. b.

auf den Wald on the forest unter der Brücke under the bridge de berg op the mountain up het water in the water in

zu to durch through

(cf. zu dem Wald) (cf. durch den Wald) (cf. op de berg) (cf. in het water)

If the functional status of an Α-postposition were determined by its functional prefix, the functional status of the postpositions in (43) and (44) would remain unaccounted for. There is a second problem with the idea that the functional status of an Α-postposition is determined by its Α-prefix. The categorial status of a complex word is determined by its morphological head. According to Williams' (1981) Right-hand Head Rule, the head of a word is its rightmost element. However, this implies that prefixes can never be category-changing. In German, the validity of the Right-hand Head Rule can be illustrated by looking at combinations of a verb and a prepositional prefix. Besides separable particles, German has a class of non-separable prefixes that correspond to existing prepositions like über-, um-, unter- etc. and therefore are of category P. If they are combined with a verb, the resulting word is still a verb, because it inherits its categorial status from the morphological head:

526 Jochen Zeller (45)

a. b. c.

unter (Ρ) + schreiben (V) under write um (Ρ) + fahren (V) around drive durch (Ρ) + schneiden (V) through cut

unterschreiben (V) sign umfahren (V) 'drive around' durchschneiden (V) 'cut through'

Since the Right-hand Head Rule holds in German, the functional status of an ft-postposition cannot be determined by the functional prefix. The alternative proposal that I make now assumes that apparently simple postpositions like e.g. German durch in (43b) or Dutch op in (44a) are in fact morphologically complex. They are derived from simple prepositional elements through the suffixation of a functional head which is phonologically zero: (46) po

op durch

F°Prep

0

According to my proposal, a postposition consists of a bare preposition of category P° that is combined with a functional suffix F 0 ^ . Accord, this suffix is the morphological head of the complex word; therefore, the derived postposition is a functional head as well. This assumption has the welcome consequence that postpositions like those in (43) and (44) can be analyzed as functional heads even though no /i-prefix is present. If an hprefix is added, however, the postposition becomes the morphological head of the resulting /i-postposition:13

The assumption that postpositions are derived by adding a functional suffix to a lexical preposition raises questions about the status of this suffix. In the

Lexical particles, semi-lexical postpositions 527 following section, I will therefore discuss its semantic role in the derivation of postpositions.

3.2. 3.2.1.

The thematic properties of postpositions Simple postpositions

Olsen (1999, 1999a) discusses the relation between the preposition durch, 'through', in (48a) and its use as a postposition in (48b): (48)

a. b.

Mirco fährt durch den Tunnel Μ. drives through the tunnel Mirco geht unter der Brücke durch Μ. goes under the bridge through 'Mirco passes under the bridge.'

The preposition durch is obligatorily transitive. Olsen (1999, 1999a) assigns it the lexical representation in (49): (49)

preposition durch:

λγ λχ [MID(LOC(x)) c IN(y)]

The preposition durch expresses that the middle part of a path traversed by χ intersects with the interior region of the reference object y. In (48a), this reference object is the DP den Tunnel, 'the tunnel', the complement of P°. The DP Mirco is the theme of the motion; it saturates the external argument of durch in (49). Olsen notes that the crucial semantic difference between (49) and the postposition durch is that the internal argument of the latter is not linked to syntax, but remains implicit. Instead, the postposition takes a predicative argument: (50)

postposition durch: λΡ λχ 3y [MID(LOC(x)) c IN(y) λ P(x)]

In (50), the y-argument, which corresponds to the reference object of durch, is not linked to syntax. In Olsen's original representation, this situation is captured by leaving the y-variable inside the formula unbound. In (50), I have instead represented the implicit internal argument as bound by an existential quantifier (I return to this point below). The fact that the internal argument of durch is not linked to syntax is directly tied to the

528 Jochen

Teller

presence of a predicative argument which is saturated in (48b) by the PPcomplement of the postposition (unter der Brücke, 'under the bridge'). The interpretation of the FPpreP in (48b) is as follows: (51)

unter der Brücke durch: λχ 3y [MK)(LOC(x)) c IN(y) λ LOC(x) c UNDER(the-bridge)]

(51) says that the middle part of x's path intersects with the interior region of an implicit y while χ is under the bridge. This means that the region characterized by IN(y) can now be identified through the location expressed by the PP-complement of the postposition. On the basis of conceptual knowledge, the path traversed by χ is interpreted as intersecting with the region "under the bridge". As Olsen (1999a) notes, the PP-complement of a postposition helps to recover the information that gets lost because the reference object is implicit. It is important to note that I understand the range of the existential quantifier as being implicitly restricted by the conceptual condition that the PP-complement of the postposition identifies the region IN(y). In other words, y in (50) and (51) cannot just be any entity, but rather must be something whose internal region can be defined by the newly added predicate. This restricted use of the quantifier in (50) and (51) is just another means of capturing the conceptual recoverability of the implicit argument, but it will be relevant in section 3.2.2 below.14 Olsen (1999) argues that (49) and (50) are two variants of one single lexical representation of durch. This is where I depart from her analysis. What I suggest instead is that the representation in (50) is derived from (49) through the addition of the functional zero-suffix discussed in section 3.1. This suffix's semantics is given in (52): (52)

0-FopreP: XQ λΡ λχ By [Q(y)(x) λ P(x)]

(52) is an operator whose main function is to change the argument structure of the prepositional element with which it combines. It existentially binds one argument of the preposition and adds a predicative argument that helps identify this implicit argument. If the semantics of a preposition like durch is combined with (52), it saturates the Q-argument, and we derive the semantics of the respective postposition. The operator in (52) has an interesting parallel in the verbal domain: the passive operator PASS, which is realized in English through the participle suffix -ed plus the auxiliary be. Semantically, PASS is a function that

Lexical particles, semi-lexical postpositions 529 changes the argument structure of the verb to which it attaches (cf. Chierchia 1989). Like the operator in (52), PASS existentially binds an argument of the verb (which may be then be identified through a prepositional fry-phrase). Notice that the passive morpheme is an inflectional element; it is commonly assumed that this morpheme is associated with Infl° (cf. Baker 1988; Baker, Johnson, and Roberts 1989). Given the semantic parallel between PASS and the operator in (52), it is reasonable to assume that the latter is also a functional suffix. Therefore, the similarity between the passive morpheme and the operator in (52) can be taken as evidence that the morphological structure of a postposition looks in fact as in (53), as was argued in section 3.1. (53)

Ρο durch

3.2.2.

rjo F Prep

0

Complex postpositions

The analysis can now be extended in order to capture the derivation of hpostpositions. A first question that must be addressed concerns the morphological and semantic status of the deictic Α-prefix. The examples in (54) show that Α-elements also occur in isolation: (54)

a.

b.

Olaf lief {zur Stadt) hin Ο. ran (to-the city) hin Olaf ran towards the city.' Olaf kommt (vom Bahnhof) her Ο. comes from-the station her Olaf walks towards the speaker, approaches the speaker, from the station.'

The Α-elements in (54) take optional PP-complements. The question is whether hin and her are of category P° or of category F 0 ^ . I will henceforth assume that the latter holds. If her and hin are used intransitively, their maximal projections express referential paths, due to the deictic information which is part of their semantics. As argued in section 2.3.3, referentiality indicates the presence of a functional projection; therefore, I assume that her and hin are functional heads that project an FPprep. How-

530 Jochen Zeller

ever, notice that their categorial status is not crucial for the categorial status of the /i-postposition, because the latter is not determined by the A-prefix, but by its morphological head. I formulate the semantic representation of hin as follows (the semantics of her can be formulated analogously): (55)

hin:

λχ 3y [FIN(LOC(x)) c PROX(y) λ -ι [y = speaker] a P(x)]

(55) is based on the representation given in Olsen (1999, 1999a).15 According to (55), hin expresses a path traversed by χ whose final part intersects with the proximity region of an implicit argument y which is not the speaker. The (proximity region of the) implicit argument of hin is identified through an optional predicative argument. Let me now turn to the question of how an Α-postposition like hindurch is derived. I assume that the semantics of an Α-element is not changed when it is used as a prefix. Therefore, I suggest that hin in (55) may saturate the predicative argument of the postposition durch in (50): (56)

a. b.

c.

durch (function); λΡ λχ, 3y, [MID(LOC(xi)) c DSTT(yi) a P(xj)] hin (argument): λχ 2 3y2 [HN(LOC(x2)) c PROX(y2) a - i [y2 = speaker] a P(x2)] hindurch ((56a), (56b), and function composition) Xxj 3yi 3y2 [MID(LOC(xi)) c INT(yi) a FIN(LOC(xi)) c PROX(y2) a - . [y2 = speaker] a P(xi)]

Semantically, the element hin in (56b) takes over the function otherwise performed by a PP-complement of a postposition. It saturates the respective argument of durch-, at the same time, the resulting Α-postposition hindurch inherits the optional predicative argument of the Α-prefix. (As with simple postpositions, this PP-complement of an Α-postposition is represented as the sister of the whole functional Α-postposition; cf. the structure in (77) below). It follows that the PP-complement of an Α-postposition is always optional, whereas the PP-complements of non-prefixed postpositions like durch and zu are always obligatory.16 Let me emphasize again that although the Α-prefix and the PP-complement of a simple postposition fulfill the same semantic function, they are morphosyntactically quite different. Whereas a PP is a phrasal sister of

Lexical particles, semi-lexical postpositions 531 F°PreP, the prefix is an F°prep- node that combines with the postposition morphologically (cf. (47) above).17 The claim that prefixes can saturate arguments of their hosts that are usually saturated by full phrases, thereby transferring their own arguments onto the complex predicate, is nothing new. Similar cases are attested, for example, with the prefix ver- in German: (57)

a. b.

Sie stellen Kisten vor die Einfahrt they put boxes in-front-of the driveway Sie verstellen die Einfahrt (mit Kisten) they Pref-put the driveway (with boxes)

In (57a), the predicative argument of the base verb stellen is saturated by its PP-complement vor die Einfahrt. In (57b), the prefix ver- fulfills the same function morphologically. The direct object in (57b) is an argument introduced by ver- and inherited by the complex verb (cf. Stiebels 1996: 107). Coming back to ^-postpositions, notice that, according to (56c), hindurch has two implicit arguments. One is contributed by durch, the other one by hin. The Α-postposition expresses that the middle part of the path traversed by χ intersects with the interior region of the first implicit argument (y0 and that the end part of this path intersects with the proximity region of the second implicit argument (y2). Interestingly, if hindurch combines with a PP-complement, the location expressed by this PP identifies the (interior region of the) internal argument of the postposition durch, as shown in (58a). Although the P-predicate is introduced by the prefix hin, PP-complements that identify the internal argument of the Α-element are not perfectly acceptable when combined with the complex A-postposition, (58b): (58)

a. b.

Mirco geht unter der Brücke hindurch Μ. goes under the bridge A-through Yves geht zum Bahnhof hindurch Y. goes to-the station A-through

(58a) is interpreted as Mirco going away from the speaker towards some unspecified region (= the internal argument of hin remains implicit), thereby passing under the bridge. (58b) means that Yves walks towards the station, which is not where the speaker is, thereby crossing some unspecified region (= the internal argument of durch remains implicit).

532 Jochen Zeller Let me finally turn to Α-postpositions in pleonastic circumpositional phrases like (59): (59)

a. b.

aus out auf on

dem Haus the house den Berg the mountain

hinaus A-out hinauf A-on

The derivation of the Α-postpositions in (59) follows the pattern illustrated above. The functional operator in (52) is attached to the preposition aus in (60a) and derives the postposition aus in (60b). Prefixation of hin derives the Α-postposition hinaus·. (60)

a. b. c.

aus (preposition): λy λχ [FEM(LOC(x)) c NOT-IN(y)] aus (postposition): λΡ λχ 3y, [FIN(LOC(x)) c NOT-IN(yi) λ P(x)] hinaus (A-postposition): λχ 3y, 3y2 [FIN(LOC(x)) c NOT-IN(yi) λ FIN(LOC(x)) c PROX(y2) λ -> [y2 = speaker] a P(x)]

The Α-postposition hinaus in (60c) expresses that the final part of a path traversed by χ ends in the outer region of an implicit reference object yi and in the proximity region of another implicit argument y2, which is not the speaker. Now consider how the PP-complement of hinaus in (59a) directly identifies the implicit argument yj. Crucially, the PP in (59a) is a cognate PP. Therefore, the local relation expressed by the preposition aus is introduced once again: (61)

aus dem Haus hinaus: λ χ By, 3y 2 [FIN(LOC(x)) c N O T - I N ( Y I ) a FIN(LOC(x)) c P R O X ( y 2 ) α —ι [y 2 = speaker] a FIN(LOC(x)) c NOT-IN(the-

house)] Recall that I have represented the implicit argument(s) of a postposition as being bound by an existential quantifier whose range is restricted by the interpretation of the PP-complement. It is clear that in interpreting pleonastic circumpositional phrases, the DP-complement of the preposition (dem Haus in (61)) directly identifies the implicit reference object of the original postposition, because the preposition's complement and the postposition's

Lexical particles, semi-lexical postpositions 533 reference object are arguments of identical predicates. Therefore, we can simplify the representation of the FPpreP aus dem Haus hinaus: (62)

aus dem Haus hinaus: λχ 3y2 [FIN(LOC(x)) c NOT-IN(the-house) Λ HN(LOC(x)) c PROX(y2) Λ -, [y2 = speaker]]

(62) says that the path of the individual χ ends outside the house and in the proximity region of some individual y who is not the speaker. (62) only differs from the meaning of the directional FPprep aus dem Haus in that it includes the deictic information contributed by the Α-prefix. Otherwise, pleonastic circumpositional phrases have the same meaning as the corresponding prepositional phrases. This fact will be highly relevant for the discussion in section 4.18

3.3.

The semi-lexical status of postpositions

The discussion in sections 3.1 and 3.2 has shown that the morphological and thematic properties of postpositions justify an analysis according to which they are functional elements derived from lexical prepositions. Postpositions in German and Dutch inherit the categorial status of their morphological heads (= the operator in (52)) and are therefore functional. If they project a phrase, this phrase is an FPpreP with all the properties of fully referential phrases. However, postpositions are not simply functional elements. Rather, they are "inflected" prepositions; they consist of a lexical element with a functional suffix. I suggest using the term "semi-lexical" to refer to morphologically complex functional heads that are derived from a lexical node. Postpositions are semi-lexical elements, and I claim that it is this property that explains why postpositions, although they are F°Prep-heads, may behave like lexical particles when they are structurally adjacent to a verb. The interaction between a verb and a structurally adjacent head Η of the sort discussed in section 2.4 is only licensed if Η is lexical or semi-lexical. Only lexical heads (i.e. particles) or functional heads derived from lexical heads (i.e. postpositions) may be reanalyzed with a structurally adjacent verb, and only lexical particles and semi-lexical postpositions are expected to show meaning variations caused by their local verbal context.

534 Jochen Zeller 4.

German vs. Dutch

In this final section I want to compare post- and circumpositional constructions in Dutch and German. Let me start by looking at German. (63a) shows a regular prepositional phrase; (63b) the corresponding postpositional phrase: (63)

a. b.

Niels steigt auf den Berg N. climbs on the mountain Niels steigt den Berg hinauf Ν. climbs the mountain A-up/on

Following Van Riemsdijk (1990) and Mclntyre (1998), I suggested in section 2.1 that the DP in (63b) is the complement of an empty preposition. The postpositional structure is in fact a circumpositional structure where the head of the postposition's PP-complement is phonologically unrealized. Notice that this assumption is necessary on both syntactic and semantic grounds. Syntactically, postpositions realize the extended projection of prepositions and hence are expected to have PP-complements. Furthermore, I argued that postpositions take predicative arguments; they therefore need to combine with PP-complements for interpretative reasons as well. It can be shown that the semantics of the empty preposition is different from, and unrelated to, the semantics of the Α-postposition. This can be seen if we compare the thematic interpretation of the DP den Berg, 'the mountain', in (63a) and in (63b). In (63a), this DP is the goal of Niels' trip; the climbing ends on the mountain. However, den Berg in (63b) does not denote the goal of the event, but rather the actual path that is traversed by the theme. (63a) means that Niels climbs on top of the mountain, (63b) means that Niels climbs up the mountain. The apparent similarities between (63a) and (63b) result from the fact that the mountain can define both the goal of the motion and the path that leads to it. However, if we choose different DPs, the difference between prepositional and postpositional phrases becomes immediately obvious: (64)

a. b.

Youri stieg auf die Leiter Y. climbed on the ladder Youri stieg die Leiter hinauf Y. climbed the ladder h-up

Lexical particles, semi-lexical postpositions 535 (65)

a.

Youri Y. b. * Youri Y.

stieg auf den Stuhl climbed on the chair stieg den Stuhl hinauf climbed the chair ft-up

Whereas (64a) is true as soon as Youri has his feet on the ladder, (64b) is only true if Youri uses the ladder as a path, which means that he has to take more than one step. The thing denoted by the DP den Stuhl, 'a chair', in (65) can be used as a possible goal, but not as a path. Therefore, the use of this DP in a postpositional phrase is impossible. Mclntyre (1998) captures the difference illustrated through (64) and (65) by postulating the existence of an empty preposition VIA that denotes a path. He argues that in postpositional constructions, this preposition is the head of the PP-complement of the A-postposition: (66)

P° 0 VIA

DP den Berg P° auf

F°Prep

The tree in (66) is interpreted as follows: (67)

a. b.

c.

(VIA) den Berg λχ [MID(LOC(x)) c ON(the-mountain)] hinauf: λχ 3y, 3y2 [FIN(LOC(x)) c ON(yO λ FIN(LOC(x)) c PROX(y2) λ -ι [y2 = speaker] a P(x)] (VIA) den Berg hinauf λ χ By, 3y 2 [FIN(LOC(x)) c ON( Y I ) λ FIN(LOC(x)) c

PROX(y2) a - . [y2 = speaker] a MID(LOC(x)) c ON(themountain)] I have represented the meaning of the VIA-PP in (67a) as expressing that the middle path traversed by χ intersects with the ON-region of the mountain. The derivation of the Α-postposition hinauf and its combination with the PP in (67a) proceeds along the lines discussed in section 3.2. The denotation of the FPpreP den Berg hinauf is the set of individuals χ that

536 Jochen Zeller

move along the mountain to become located in the ON-region of some individual yi and in the proximity region of a second individual y-ι who is not the speaker. Let me now turn to Dutch. The question is whether the interpretation of postpositional phrases in Dutch patterns with that of their German counterparts. Consider (68): (68)

a.

b.

Hij springt op de trap he jumps on the stairs 'In one jump, he landed on the stairs.' Hij springt de trap op he jumps the stairs up 'He jumps all the way up, via the stairs.'

The interpretation of (68) suggests that, as in German, a postpositional phrase in Dutch includes a PP with an empty VIA-preposition that determines the DP's thematic properties. Whereas (68a) is true if there is only one jump that ends on the first step of the stairs, (68b) is only true if the agent jumps all the way up the stairs. This difference follows from Mclntyre's proposal on which (66) and (67) are based. This account, however, is different from the analysis of postpositional phrases presented in Van Riemsdijk (1990, 1998), according to which (68b) would have been derived from (68a) via head movement of the preposition to F°prep: (69)

Move

Adopting the copy-theory of movement (cf. Chomsky 1995; Brody 1995), I represent the base position of head movement through a copy which is not spelled-out at PF. According to (69), a postposition is derived by moving a preposition and adjoining it to the functional head F 0 ^ . 1 9 Clearly, a head movement account is inadequate for the example in (68), since it does not capture the semantic difference between (68a) and (68b). Nevertheless, I want to argue now that some postpositional phrases in

Lexical particles, semi-lexical postpositions 537 Dutch are derived as in (69). There are two observations that provide support for this idea. Consider the examples in (70) and (71): (70)

a. b.

(71)

a. b.

uit het huis out the house het huis uit the house out in de gevangenis in the jail de gevangenis in the jail in

First, note that it seems implausible to assume that the DPs het huis and de gevangenis in (70b) and (71b) are arguments of a preposition VIA, because this would yield the implausible interpretation "the theme moves along the house/the jail in order to get out of something/into something". The DPs in (70b) and (71b) do not qualify as complements of a path-denoting preposition; rather, they seem to receive the same θ-role that is assigned by the prepositions in the (a)-examples. The DP het huis denotes a source; the DP de gevangenis denotes a goal. This means that, in contrast to the DP in (68b), the DPs in the (b)-examples in (70) and (71) seem to be arguments of the prepositional elements uit and in, respectively. Second, notice that the Α-postpositions hinaus/heraus and hinein/herein in German, which correspond to the postpositions uit and in, do not occur in postpositional constructions, as is shown by the (a)-examples in (72) and (73). Instead, in order to express the meaning of the phrases in (70b) and (71b), a German speaker would use a pleonastic circumpositional phrase, viz. (72b) and (73b). This possibility does not exist in Dutch, as the (c)examples illustrate: (72)

a.

*das Haus hinaus the house A-out b. aus dem Haus hinaus out the house A-out c. *uit het huis uit

538 Jochen Zeller (73)

a.

*das Gefängnis hinein the jail h-in b. in das Gefängnis hinein in the jail A-in c. *in de gevangenis in

It seems that what is expressed through pleonastic circumpositional phrases in German is expressed through postpositional structures in Dutch. Based on this observation, I now suggest the following. The postpositional structures in (70b) and (71b) are in fact derived via head movement: (74) PP po

D

p

het huis

po

Prep

uit

As noted above, the chain created by P°-movement is only spelled-out in its head position. However, what I propose here is that both links of the P°chain in a postpositional construction like het huis uit are interpreted semantically. The higher link is attached to F 0 ^ and derives the postposition uit whose meaning is based on the semantics of the preposition uit. Crucially, the same semantic value is assigned to the copy of P°. In its base position, the preposition now takes the DP het huis as its internal argument; the semantics of the PP in (74) corresponds to the semantics of the PP uit het huis. This PP saturates the predicative argument position of the postposition uit: (75)

a. b. c.

uit (preposition): λy λχ [FIN(LOC(x)) c NOT-IN(y)] uit (postposition): λΡ λχ 3y [HN(LOC(x)) c NOT-IN(y) A P(X)] 0 (= copy of uit) het huis: λχ [FIN(LOC(x)) C NOT-IN(the-house)]

Lexical particles, semi-lexical postpositions 539 d.

d'.

het huis uit: λχ By [FIN(LOC(x)) c NOT-IN(y) Λ FIN(LOC(x)) c NOT-IN(the-house)] het huis uit: λχ [FIN(LOC(x)) c NOT-IN(the-house)]

The PP in (74) is interpreted as in (75c) and saturates the predicative argument of the postposition. Since the predicates introduced by the preposition and the postposition are identical, (75d) can be simplified as (75d'). The implicit argument of the postposition uit is identified through the argument introduced by the PP, which is the complement of the (copy of the) preposition uit. Therefore, the meaning of (75d') is identical to the meaning of the phrase uit het huis. However, in certain cases, it does not seem that postpositional phrases in Dutch that are derived by movement always receive exactly the same semantic representation as the corresponding prepositional phrases. Compare (76a) and (76b): (76)

a. b.

Hij loopt op de berg he runs on the mountain Hij loopt de berg op he runs the mountain on/up

According to my Dutch informants, the postpositional phrase de berg op does not have the same interpretation as the postpositional phrase den Berg hinauf in German. This excludes an analysis according to which de berg in (76b) is the complement of a VIA preposition. This leaves the second option, i.e. that (76b) is derived by head movement of op. However, this predicts that (76a) and (76b) essentially have the same semantics; a prediction which is not borne out either. The FPprep op de berg in (76a) has a locative interpretation, whereas the postpositional structure in (76b) is interpreted directionally. However, I think this problem can be solved by assuming that the locative or directional character of a preposition is determined by properties of the functional head to which it is structurally adjacent. (76a) and (76b) involve different functional heads. The F°PreP-head in (76a) does not trigger movement and yields a locative interpretation of op. In contrast, the functional head in (76b) triggers head movement of the preposition and thereby assigns a directional interpretation to both links of the P°-chain.20

540 Jochen Zeller

Now compare (74) to the structure of the German pleonastic circumpositional phrase aus dem Haus hinaus: (77)

In both (74) and (77), the lexical preposition occurs twice, once as the head of the PP, and once as part of the postposition. The only difference is that the second occurrence of uit in (74) is created by movement, whereas in 21

(77), the P-element aus is just "used" twice. However, if we disregard the contribution of the deictic prefix for the moment, we find that (77) receives exactly the same interpretation as the postpositional phrase het huis uit (cf. (75) and (60)-(62) in section 3.2). Let me summarize my proposal. In both Dutch and German, the argument structure of postpositions requires a PP-complement. First, the head of this complement and the prepositional base of the postposition may be entirely different elements. In that case, no movement relation exists between both prepositional elements, and we derive circumpositional phrases like (78a) and (78b) that exist in both languages: (78)

a. b.

Mirco fährt unter der M. drives under the Hetvliegtuig is onder de the airplane is under the

Brücke durch (German) bridge through brug door gevlogen (Dutch) bridge through flown

Second, the head of the PP might be the same prepositional element that forms the lexical base of the complex postposition. Dutch and German have two different ways of creating a structure which corresponds to this situation. Whereas in German, the preposition is merged into the tree twice (and is therefore spelled-out in both positions), Dutch simply moves the prepositional head of the PP and uses the head of the chain to derive the postposition:

Lexical particles, semi-lexical postpositions 541 (79)

a. b.

aus dem Haus hinaus out the house A-out ti het huis uit[ the house out

(German) (Dutch)

In both (79a) and (79b), the head of the PP is semantically interpreted in exactly the same way as the P-head that is part of the postposition, regardless of whether it is a fully spelled-out element as in German or the phonologically unrealized copy of a moved preposition (represented by the trace tj in (79b)) as in Dutch. Finally, not all simple postpositional phrases in Dutch are generated by movement. If the phonologically unrealized head of the PP-complement and the P-head that is part of the postposition have different meanings, the head of the PP-complement must be analyzed as a zero-preposition whose semantics is different from the prepositional part of FVep· These postpositional structures also exist in German: (80)

a. b.

Youri stieg die Leiter hinauf Y. climbed the ladder h-up Hij springt de trap op he jumps the stairs up 'He jumps all the way up, via the stairs.'

(German) (Dutch)

In both (80a) and (80b), the head of the PP is a phonologically unrealized VIA-preposition. 5.

Conclusion

The properties of particles, prepositions, and postpositions discussed in this paper suggest that the differences and similarities between these elements are best captured in a theory that takes into account the local environment of a lexical P-node. If P° is structurally adjacent to a verb, it is a particle; if it is structurally adjacent to a functional head, it is a preposition. If P° is suffixed with a functional head, it is a postposition. This explains why postpositions are mixed creatures. On the one hand, the presence of functional structure causes them to adopt properties of functional heads. Therefore, post- and circumpositional phrases are functional projections. On the other hand, postpositions are also structurally adjacent to the verb. Therefore, their syntactic environment is the same as that of particles.

542 Jochen Zeller

However, the fact that postpositions behave like particles with respect to phenomena like word formation or Verb Raising does not follow from their local environment alone. It requires the additional assumption that postpositions, although they are functional categories, also have properties of lexical elements. I argued that these properties follow from the idea that postpositions are derived from lexical P-elements through the suffixation of a functional zero-morpheme, and I elaborated this idea by investigating the systematic difference between the thematic properties of postpositions and those of prepositions. The argument structure of the former was analyzed as being derived from the argument structure of the latter through an operator that corresponds to the functional suffix that derives postpositions morphologically. This explains why postpositions are semi-lexical elements; they have the lexical status of prepositions, but inherit the functional status of their morphological heads. I had to leave open many questions about the precise relation between the particle-like properties of postpositions and their semi-lexicality. Nevertheless, I hope that my analysis has opened an interesting empirical window which provides a new perspective on issues concerning lexical and functional nodes and their interaction in morphosyntactically defined locality domains.

Notes 1. I would like to thank the participants of the workshop on semi-lexical heads and an anonymous referee for their valuable comments. I would also like to thank Dori Posel for proof-reading this article. 2. Van Riemsdijk (1998) argues that the whole prepositional phrase only includes one maximal projection, the ΡΡρκρ- This means that there is no maximal lexical projection in his theory. 3. Besides Λ-postpositions, German has other complex postpositions that consist of a prepositional element and a prefix dr- or r- (cf. Mclntyre 1998). Some of the dr-postpositions may also be used non-directionally. 4. (14b) might be analyzed as an instance of remnant topicalization of the phrase which includes the object-DP and the ΡΡρπφ, but not the verb. This might explain why for some speakers, (14b) is not entirely excluded. However, notice that the topicalized phrase still includes the unbound trace of the verb and therefore still contrasts clearly with (14a). 5. An anonymous reviewer reminds me that a syntactic analysis of particle verbs which does not assume incorporation has first been proposed by Taraldsen (1983) for Norwegian.

Lexical particles, semi-lexical postpositions 543

6. Examples like (22) have led researchers like Van Riemsdijk (1978) and Den Dikken (1995) to consider particles as particular instances of intransitive prepositions. Notice, however, that the particle verb in (22) is not simply an intransitive version of the construction in (21), but has adopted an idiomatic meaning. From a lexicalist point of view, according to which special meanings are only associated with words, the idiosyncratic properties of particle verbs seem to raise a problem for the phrasal analysis of particle verbs. However, this lexicalist perspective has been criticized in recent work (cf. Jackendoff 1997; Marantz 1997). In Zeller (1999), I propose an analysis of particle verbs which is based on an alternative model of grammar. Since this model allows for the association of special meanings with non-minimal syntactic structures, the idiosyncratic semantics of many particle verbs does not raise a problem for a phrasal analysis. 7. The particle aus licenses so-called Objektvertauschung or Objekt-umsprung (cf. e.g. Kühnhold 1973; Hundsnurscher 1968), i.e. it may also realize its Theme as the direct object of the particle verb. In that case, the reference object must remain implicit, cf.: (i) Peter hat sein Bier (*die Flasche) ausgetrunken Peter has his beerAcc the bottle out-drunk Since it will be argued below that the direct object of the particle verb austrinken receives case from the verb, and since the verb trinken has only one objective case to assign, it follows that the theme and reference object in this example are mutually exclusive. 8. Some prepositions (like zu, 'to', or nach, 'after') that assign dative case to their reference objects may transfer this property onto a structurally adjacent verb, such that the particle verb also assigns dative case to its object. See Zeller (1999) for further discussion. 9. Similar assumptions about terminal nodes in complex syntactic structures, and about their relation to words formed by directly merging these nodes, are made in Marantz (1989) and Borer (1991, 1993). See Zeller (1999), chapter 6, for discussion. 10. So-called root formations are sometimes analyzed as being derived by combining a particle with a noun derived from the base verb (cf. Stiebels and Wunderlich 1994): (i) der Absprung, 'the take-off [Ab [sprung]] (ii) der Ausguss, 'the gutter' [Aus [guss]] If root formations are indeed combinations of a particle and a noun, they pose a problem for my claim that particles always require local verbal contexts. Therefore, as an anonymous referee points out, one might be forced to assume that root formations are syntactically derived. According to this view, they would include a full VP which consists of the verb and the particle. The verb incorporates into an abstract nominalizing head; the complex noun is then affected by the phonological process characteristic of root nominalizations. However, this analysis wrongly predicts that all root formations are process

544 Jochen Zeller

nominate. I therefore tentatively assume that root formations like those in (i) and (ii) are genuine nominalizations of (reanalyzed) particle verbs and that the Ablaut-process that triggers the phonological change may also effect complex verbs (see Zieller (in preparation) for some discussion): (iii) [[ab spring] + N] —> [Absprung] 11. As shown in (i), adjectival resultative constructions differ from particle verbs in that they do not allow for further word formation; cf. Neeleman and Weerman (1993); Neeleman (1994); Kratzer (1994): (i) particle verb: abwaschen, 'wash off derived adjective: abwaschbar resultative construction: sauber waschen, 'wash clean' derived adjective: * sauberwaschbar Assuming that adjectival resultatives involve an extended projection of the resultative predicate, the analysis presented in the text explains the impossibility of having word formation with resultative constructions. 12. Booij (1990) argues that particle verbs are verbal projections of a specific kind that can be raised in Dutch. However, this claim is based on Booij's questionable notion of a V*-node which is formed in the lexicon and includes the particle and the verb. Since the postulation of this special node seems rather ad hoc and does not capture the fact that Verb Raising with postpositions and certain kinds of resultatives is also possible, I reject Booij's assumption. 13. As an anonymous reviewer notes, Α-postpositions in Austrian dialects may show the order [postposition-A-prefix], cf. auf-i (= auf-hin) instead of the German hin-auf. The problem for the present proposal is posed by the fact that the Austrian examples apparently violate the Right-hand Head Rule. However, it might be possible that the functional deictic element in Austrian dialects has been reanalyzed as an inflectional suffix and has adopted properties quite similar to those of the functional zero-suffix discussed in example (46). This assumption gains support from the observation that the deictic element occurs in a phonologically reduced form, a characteristic typical of grammaticalization (cf. Hopper and Traugott 1993). 14. If the range of the existential quantifier in (50) and (51) were not restricted, its interpretation would be too weak. For example, the sentence in (i) would be true in the situation in which Mirco crawls through a pipe that lies under the bridge: (i) Mirco kriecht unter der Brücke durch Μ. crawls under the bridge through If Mirco crawls through the pipe, there is a y such that the path traversed by Mirco intersects with y's interior region while Mirco is under the bridge. However, (i) is not a correct statement to describe this situation. In order to prevent (i) from being true in the aforementioned context, I assume that the reference of y is restricted to entities whose interior regions can plausibly be defined by the PP-complement of the postposition.

Lexical particles, semi-lexical postpositions

545

15. In contrast to Olsen, I have marked the predicative argument of hin as optional (the triangle brackets); as (54) shows, the PP-complement of an Α-element can be omitted. Again, I have represented the implicit argument as being bound by an existential quantifier whose range is restricted in the same way as was noted for the quantifier in (50) above. 16. In other words, a postposition in German either needs a PP-complement or a prefix. As far as I can see, this generalization seems correct. However, notice that it is not always easy to test, because durch and zu may also appear as particles. According to my theory, whenever these elements appear without a PP-complement, they must be particles, i.e. P°-heads of PPs without functional structure (cf. Olsen 1997 who draws a similar conclusion): (i) Ebbe geht auf mich zu: zu = functional postposition 'Ebbe walks towards me' (ii) Die Tür geht zu: zu = particle 'The door closes' 17. If used without a host, the Α-element may also project a full FP^p, as shown by the examples in (54). 18. One important question that I have to leave unanswered is why only some postpositions in German, like durch, can appear without a prefix, whereas others, like auf and an, require a prefix. Mclntyre (1998) argues that the prefix is needed in order to absorb the case assigned by the prepositional element. However, although Mclntyre's proposal explains why German postpositions like auf, ein etc. cannot occur without a prefix, it still leaves open the question of how case is absorbed with simple postpositions. It seems that the zerooperator turns certain prepositional elements (like e.g. a«/and an) that are free morphemes into bound morphemes that require a prefix. Interestingly, there are other constructions in German with similar properties (cf. Höhle 1982: 97): (i) zweitklassig, 'second-rate';ganztägig, 'all-day'; südländisch, 'southern' The adjectives in (i) are derived from the nominal compounds *Zweitklass-, *Ganztag-, and *Südland-. The heads of these compounds are free morphemes (Klasse, Tag, Land)·, however, the compounds themselves do not occur independently. They are only licensed in combination with the adjectival suffixes in (i). Höhle (1982) labels words that only occur as parts of other words Zusammenbildungen. According to my theory, postpositions like auf and an are Zusammenbildungen as well. 19. This assumption implies that head movement can derive words, and that these words obey morphological rules. For a discussion and a defense of this assumption, see Zeller (1999), chapter 1. 20. The idea that functional heads determine the directional or locative interpretation of prepositions is defended in Zeller (1999) to which I refer the reader for details. 21. One may wonder whether there is a substantial syntactic difference between (74) and (77) with respect to the relation between the two P°-heads. Notice that "head movement" is a metaphor for "copy and merge". The

546 Jochen Zeller

copying of a moved element might be nothing else than accessing a lexical element twice and merging it into the structure in two different positions. In that case, the "movement" operation in Dutch postpositional phrases is identical to the formation of pleonastic circumpositional constructions in German. It might therefore be possible also to analyze the relation between the two prepositions in (75) as a chain. Formally, the difference between Dutch postpositional phrases and German pleonastic circumpositional phrases would then reduce to the fact that only in German, the copy of the moved P° is phonologically interpreted.

References Abney, Steve 1987 The English Noun Phrase in Its Sentential Aspect. Ph.D. dissertation, Massachusetts Institute of Technology. Baker, Mark 1988 Incorporation. A Theory of Grammatical Function Changing. Chicago/London: The University of Chicago Press. Baker, Mark, Kyle Johnson and Ian Roberts 1989 Passive Arguments Raised. Linguistic Inquiry 20: 219-251. Booij, Geert 1990 The Boundary between Morphology and Syntax: separable complex verbs in Dutch. Yearbook of morphology 3: 45-63. Borer, Hagit 1991 The Causative-Inchoative Alternation: A Case Study in Parallel Morphology. The Linguistic Review 8: 119-158. 1993 Derived nominals. Manuscript, University of Massachusetts, Amherst. 1998 Passive without Theta Grids. In: Steve Lapointe (ed.), Morphology and Its Relation to Phonology and Syntax. Stanford: CSLI Publications. Brody, Michael 1995 Lexico-Logical Form. Cambridge: MIT Press. Chierchia, Gennaro 1989 A Semantics for Unaccusatives and its Syntactic Consequences. Manuscript, Cornell University. Chomsky, Noam 1995 The Minimalist Program. Cambridge: MIT Press. Den Dikken, Marcel 1995 Particles. On the Syntax of Verb-Particle, Triadic, and Causative Constructions. New York/Oxford: Oxford University Press.

Lexical particles, semi-lexical postpositions

547

Evers, Arnold 1975 The Transformational Cycle in Dutch and German. Ph.D. dissertation, University of Utrecht/Indiana University Linguistics Club. Grimshaw, Jane 1991 Extended Projections, Manuscript, Brandeis University. Haiden, Martin 1997 Verbal Inflection and the Structure of IP in German. Groninger Arbeiten zur germanistischen Linguistik 41: 77-106. Höhle, Tilman 1982 Über Komposition und Derivation: zur Konstituentenstruktur von Wortbildungsprodukten im Deutschen. Zeitschrift für Sprachwissenschaft 1: 76-112. Hopper, Paul J. and Elizabeth Closs Traugott 1993 Grammaticalization. Cambridge: Cambridge University Press. Hundsnurscher, Franz 1968 Das System der Partikelverben mit A US in der Gegenwartssprache. Göppingen: Kümmerle. Jackendoff, Ray 1997 The Architecture of the Language Faculty. Cambridge: MIT Press. Koopman, Hilda 1993 The structure of Dutch PPs, Manuscript, University of California at Los Angeles. Kratzer, Angelika 1994 The event argument and the semantics of Voice, Manuscript, University of Massachusetts, Amherst. Kühnhold, Ingeburg 1973 Präfixverben. In: Ingeburg Kühnhold and Hans Wellmann (eds.), Deutsche Wortbildung: Das Verb, 141-362. Düsseldorf: Schwann. Longobardi, Giuseppe 1994 Reference and Proper Names: A Theory of N-Movement in Syntax and Logical Form. Linguistic Inquiry 25: 609-665. Marantz, Alec 1989 Clitics and Phrase Structure. In: Mark Baltin and Anthony Kroch (eds.), Alternative conceptions of phrase structure, 99-116. Chicago: The University of Chicago Press. 1997 No escape from Syntax: Don't try Morphological Analysis in the Privacy of Your Own Lexicon. Penn Working Papers in Linguistics, Volume 4/2: 201-225. Mclntyre, Andrew 1998 Double Particles of the Type Heran, Hinzu, Raus and the German Preverb System. A Study in Morphology and Conceptual Semantics. Ph.D. dissertation, University of Sydney.

548 Jochen Zeller

Neeleman, Ad 1994 Complex predicates. Ph.D. dissertation, Utrecht University. Neeleman, Ad and Fred Weerman 1993 The Balance between Syntax and Morphology: Dutch Particles and Resultatives. Natural Language and Linguistic Theory 11: 433475. Olsen, Susan 1996 Pleonastische Direktionale. In: Harras, G. and Manfred Bierwisch (eds.) Wenn die Semantik arbeitet, 303-329. Tübingen: Niemeyer. 1997 Zur Kategorie Verbpartikel. Beiträge zur Geschichte der deutschen Sprache und Literatur 119: 1-32. 1999 Durch den Park durch, zum Bahnhof hin: komplexe Präpositionalphrasen mit einfachem direktionalem Kopf. In: Heide Wegener (ed.), Deutsch kontrastiv, 111-134. Tübingen: Stauffenberg. 1999a Komplexe Präpositionalphrasen mit postponiertem direktionalem Kopf. Linguistische Berichte 180: 389-408. Riemsdijk, Henk van 1978 A Case Study in Syntactic Markedness. Dordrecht: Foris. 1990 Functional Prepositions. In: Η. Pinkster and I. Gende (eds.): Unity in Diversity, 229-241. Dordrecht: Foris. 1998 Categorial Feature Magnetism: The Endocentricity and Distribution of Projections. Journal of Comparative Germanic Linguistics 2: 1-48. Stiebels, Barbara 1996 Lexikalische Argumente und Adjunkte. (Studia Grammatica 39). Berlin: Akademie Verlag. Stiebels, Barbara and Dieter Wunderlich 1994 Morphology feeds Syntax: the Case of Particle Verbs. Linguistics 32: 913-968. Stowell, Tim 1991 Determiners in NP and DP. In: Katherine Leffel and Denis Bouchard (eds.), Views on Phrase Structure, 37-56. Amsterdam: Kluwer Academic Press. Taraldsen, Knut Tarald 1983 Parametric Variation in Phrase Structure. Ph.D. Dissertation, University of Tromso. Williams, Edwin 1981 On the Notions 'Lexically Related' and 'Head of a Word'. Linguistic Inquiry 12: 245-274.

Lexical particles, semi-lexical postpositions 549

Wunderlich, Dieter and Michael Herweg 1991 Lokale und Direktionale. In: Arnim von Stechow and Dieter Wunderlich (ed.), Handbuch der Semantik, 758-785. Berlin: de Gruyter. Zeller, Jochen 1999 Particle verbs, local domains, and a theory of lexical licensing. Ph.D. dissertation, University of Frankfurt, in prep. Complex verbs in Bare Phrase Structure Theory. Manuscript, University of Frankfurt.

Index absolute, 164 absolutive construction, 350 adverb hierarchy, 398 Äfarli, 15 agentivity, 342 agreement heterogeneous double agreement pattern, 294 homogeneous double agreement pattern, 291 Akmajian, 163 Allopenna, 118 Alsina, 333 American English, 371, 382,402 andative verb, 400 animacy restriction, 395 argument transfer, 424 as, 475 aspectogenesis, 334 asymmetric, 419 auxiliary, 81, 374 Babby, 296 ^-construction, 428 Baltin, 29 Bangla, 191 bare phrase structure, 396 Bayer, 79 be, 128,152 bekommen, 84 Bengali, 79 Bhattacharya, 12 body-part term, 253 Bouchard, 464 boundedness, 260 bounded event, 342 Bovalino Marina, 402 Bowerman, 115 Bowers, 455 Braine, 106 branching constraint, 75

Burzio, 44,51 Butt, 14, 324 Cardinaletti, 14 cardinals, 13,280 Case, 154 Case adjacency effect, 154 Case Filter, 165 Case Hierarchy, 297 Case NP-Agreement, 290 configurational Case, 296 lexical case, 296 Categorial Identity Thesis, 11, 23 causative, 44,50,401 child language, 11 Chinese, 226,496 Chomsky, 32,208 Cinque, 398 circumpositional phrase, 506 pleonastic circumpositional phrase, 532 classifier, 12, 191 class language, 191 mensural classifier, 224 numeral classifier, 223 clitic climbing, 372, 388 cognate PP, 508 Collins, 207 comitative, 164 complex complex predicate, 324,418 complex small clause, 15,482 compositionality, 459 control, 487 Coordinate Structure Constraint, 418 copula, 456 copy-theory of movement, 536 Corver, 38, 142,234 Creole language, 467 c-selection, 101 Czech, 273

552 Index

deep insertion, 24, 32 demonstrative, 279 derivational morphology, 519 directionality, 11 Distributed Morphology, 128, 512 do so, 161 Dowty, 355 dummy subject, 89 Dutch, 70,99,119, 505 dynamic enrichment, 460

gapping, 143 gender language, 191 German, 70, 80,99, 103,119, 196, 243,505 Geuder, 14 Giusti, 14 give, 338 grammaticalization, 444 Grimshaw, 424,489 group noun, 273

Eastern Sicilian, 373 Eckardt, 360 Eide, 15 ellipsis, 282 elliptic construction, 258 phrasal ellipsis, 47 VP-ellipsis, 151, 156 Elsewhere Principle, 171 Emonds, 6, 10,68, 127, 164,166, 193,298 English, 99,119,196,338, 348 entity-denoting noun, 256,260 event event augmentation, 360 event binding, 156 event modification, 361 event summation, 359 extended projection extended nominal projection, 9 extended projection, 489,507

Hagemeijer, 14 Haider, 11,75,101 have, 168 Hebrew, 197 Higginbotham, 355 Hindi, 191, 328 Hook, 335 Huang, 496

flat structure, 11 for, 475 Franks, 297 French, 33, 36,50, 55, 103, 119, 392 Full Identification, 464 functional functional extension, 78 functional operator, 15 functional preposition, 105 functional suffix, 505 functor, 113

Italian, 27,44, 50, 212, 371 I-to-C movement, 70 Jackendoff, 29, 38, 299 Japanese, 33,43, 80,226,229 Jespersen, 323 Kayne, 53 Kitahara, 226 kriegen, 84 Kubo, 43,229 Kwawu, 427 Larson, 396 last resort expression, 127 late insertion, 32 Lebeaux, 114 Lefebvre, 427 Left Branch Constraint, 38 lexical noun, 273 Li, 226 light verb, 14, 323,424 Linear Correspondence Axiom (LCA), 197

Index

Lobeck, 27 Löbel, 12 Mad Magazine sentence, 161 Mam, 200 Mandarin, 427 Marsalese, 372 mass-denoting noun, 256,260 Mclntyre, 516 measure phrase, 243 merge, 110, 206 meronomy, 254 Mester, 424 metaphorical transfer, 353 Milazzese, 374 monoclausal structure, 385 more, 101 Morgan, 118 motion verb, 14, 371, 374,430 multiple specifiers, 214 Ν gapping, 132 Newman, 346 Ninio, 105 Norwegian, 455 nouniness, 5 numeral, 193 object-drop, 420 object-shift, 197 of, 166 Old Indo-Aryan, 350 Old Norse, 462 one, 128, 129 operator predication operator, 457 zero-operator, 505 Pali, 350 Parsons, 361 participle agreement, 55 particle, 505 partitive, 7, 39,205 direct partitive construction, 435

553

partitive Case, 154 pseudo-partitive, 39 pseudopartitive construction, 223 passive morpheme, 529 perfectivity, 334 pivot grammar, 106 plurality, 228 possessive, 279 postposition, 15, 505 h-postposition, 508 Powers, 11 predicate cleft, 421 predicate inversion, 154 process, 520 Procrastinate principle, 32 property, 240 Pr-projection, 458 pseudo-relative, 495 quantifier, 13, 193 existential quantifier, 274 floating quantifiers, 390 quantifier-float, 196 universal quantifier, 274 quantity-designating noun, 247 Rafel, 15 reanalysis, 521 recoverability of deletion, 140 reduplication, 331 relational nouns, 246 relative clause, 275 restrictive appositive, 243 restructuring, 44 result nominal, 520 Riemsdijk, 7,23,68,233,435,489, 507 Right-hand Head Rule, 525 Rizzi, 27,46 Roeper, 112 Ross, 4,26, 38,418 Rothstein, 157, 162

554 Index

Salentino, 374 Sanskrit, 330, 350 Säo-Tomense, 415 Saramaccan, 425,438 Schütze, 12 secondary theta-role, 394 Selkirk, 38 semi-lexicality, 524 Serbo-Croatian, 200 serial verb construction, 14, 374,415 directional serial, 440 locative serial, 439 take-serial, 427 set-merge, 208 Shi, 118 Shopen, 386 small clause, 457,477 so, 128,142 som, 455 sortal classifier, 224 Southern Apulian, 373 Southern Calabrian, 373 Spanish, 50,480,495 Spec-Head agreement, 475 specificity, 214 squishy, 4 s-selection, 100 stranded, 493 structural adjacency, 511 subcategorisation, 204

subject raising, 55 Swedish, 201, 371,404,458 syntacticon, 298 taxonomy, 254 transnumerality, 252 Tremblay, 164 underspecification, 171 Urdu, 323, 327,426 V2,70 V-Agreement, 290 Verb Projection Raising, 522 Verb Raising, 521 verbiness, 5 Veselovskä, 13 Vietnamese, 223,250 VP-shell, 396,406 V-to-I movement, 70 Wagiman, 358 weigh, 239 Williams, 248 Wilson, 358 with, 128, 164 Zagona, 52 Zeller, 15,510 Zwicky, 199