Selected Papers from SinFonIJA 3 [1 ed.] 9781443846738, 9781443840804

Citation preview

Selected Papers from SinFonIJA 3

Selected Papers from SinFonIJA 3

Edited by

Sabina Halupka-Rešetar, Maja Markoviü, Tanja Miliüev and Nataša Miliüeviü

Selected Papers from SinFonIJA 3, Edited by Sabina Halupka- Rešetar, Maja Markoviü, Tanja Miliüev and Nataša Miliüeviü This book first published 2012 Cambridge Scholars Publishing 12 Back Chapman Street, Newcastle upon Tyne, NE6 2XX, UK British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library Copyright © 2012 by Sabina Halupka- Rešetar, Maja Markoviü, Tanja Miliüev and Nataša Miliüeviü and contributors All rights for this book reserved. No part of this book may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior permission of the copyright owner. ISBN (10): 1-4438-4080-7, ISBN (13): 978-1-4438-4080-4

TABLE OF CONTENTS

Introduction Sabina Halupka-Rešetar, Maja Markoviü, Tanja Miliüev and Nataša Miliüeviü ................................................................................... 1 Part One: Syntax, Phonology and Language Analysis When are Negative Imperatives Banned? Željko Boškoviü............................................................................................ 6 Some Observations on Multiple X Henk van Riemsdijk ................................................................................... 18 On the Tense-less Future in Polish Joanna Báaszczak, Patrycja JabáoĔska, Dorota Klimek-Jankowska and Krzysztof Migdalski ............................................................................ 35 Development of TAM Categories in Turkic: Feeling Free in the Deterministic System Pavel Grashchenkov.................................................................................. 58 On the Slovenian Demonstrative Reinforcers and the Internal Structure of Demonstratives Franc Marušiþ and Rok Žaucer................................................................. 76 Word Order(s) and Clitic Doubling in Bulgarian Teodora Radeva-Bork ............................................................................. 105 The Property Analysis of Control Constituents: An Argument from Modal Existential Wh-constructions Radek Šimík ............................................................................................. 126 Nonfinal Information Focus in Serbian: Some Experimental Data Sabina Halupka-Rešetar.......................................................................... 159

vi

Table of Contents

Clitic Placement and the Properties of the Intonational Phrase in Serbian Maja Markoviü and Tanja Miliüev .......................................................... 175 Part Two: Natural Language Processing Names Entity Recognition in the System for Information Extraction Staša Vujiþiü Stankoviü............................................................................ 206 Prosody Prediction in Speech Synthesis based on Regression Trees Milan Seþujski, Darko Pekar, Dragan Kneževiü and Vladimir Svrkota............................................................................... 224 Attention and Linguistic Encoding of Motion Events in Human-Machine Interaction Milan Gnjatoviü and Vlado Deliü............................................................ 237 Contributors............................................................................................. 258 Index........................................................................................................ 266

INTRODUCTION SABINA HALUPKA-REŠETAR, MAJA MARKOVIû, TANJA MILIûEV AND NATAŠA MILIûEVIû

The current volume is a selection of papers presented at the SinFonIJA 3 conference, held at the University of Novi Sad in October 2010. A wide range of linguistics-related topics was covered on this occasion, and the selection before you represents the variety and quality of research we wish to promote. It aims at upholding sound linguistic theorizing as the basis of natural language analysis and cutting-edge applied-linguistics concerns. In addition to the novel linguistic topics in the fields of syntax, phonology, semantics and natural language processing, the research presented here shows why the assumptions regarding the interfaces of these domains, which we often take for granted, require solid empirical grounding. The formal rigor and clarity of modern syntactic theory, once again, prove to make an adequate framework for pushing forward our understanding of the various language phenomena. Thus, When are negative imperatives banned? by Željko Boškoviü illustrates how a wellknown affix hopping analysis (Chomsky 1957), based on the Head Movement Constraint, which is commonly taken to account for the verbal syntax and morphology in English, can be extended to the formation of imperatives in other languages. More precisely, the author offers an account of the observation made by Zeijlstra (2004) regarding the impossibility of negative imperative formation in languages with a negative head marker (Xo). Building on the insights in Miyoshi (2002), Boškoviü argues that the ban on true negative imperatives is the consequence of the negation marker blocking affixation of the imperative suffix to the verb, which in turn means that either the (non-)head status or the (non-)affix status of the imperative marker will determine the availability of true negative imperatives cross-linguistically. That the current, wide empirical coverage of the syntactic research in fact opens new challenges and uncharted territories is brought up by Henk van Riemsdijk in Multiple X. This paper outlines all the striking similarities

2

Introduction

between adjunct free relative clauses and wh-questions across languages. It also shows that the interpretation of all wh-adjuncts is parallel to that of conditional clauses, and that these constructions allow the occurrence of multiple wh-elements. Following the implications of these observations Van Riemsdijk further suggests that there is an identifiable class of constructions with multiple occurrences of the semantically interdependent indefinite expressions, or multiple X, which includes also comparative and consecutive clauses. On the tense-less future in Polish by Joanna Báaszczak, Patrycja JabáoĔska, Dorota Klimek-Jankowska and Krzysztof Migdalski takes up the challenge of defining the syntax and semantics of the simple and periphrastic future forms in Polish, extending its main conclusions to facts from Slovenian. The central claim of the paper is that the future constructions in Polish represent a combination of the present tense and perfective aspect features, which are responsible for the derived future time reference. Pavel Grashchenkov’s Development of TAM categories in Turkic: feeling free in the deterministic system searches for a consistent account of the grammatical function development of Turkic verbs that behave both as lexical and functional elements. This paper examines the syntactic reanalysis that follows the occurrence of the functional (auxiliary) role of such verbs, as well as their novel semantics. It also addresses the deeper question regarding the nature of the observed change. On Slovenian demonstrative reinforcers and the internal structure of demonstratives by Franc Marušiþ and Rok Žaucer discusses the doubling of the spatio-temporal deictic marker (or, a demonstrative reinforcer), and the case markers. The authors draw comparisons between these facts and similar phenomena in other languages (specifically, Czech and Slovak), suggesting a way of analyzing, at first glance, such a peculiar case of doubling demonstratives. Teodora Radeva-Bork’s Word order(s) and clitic doubling in Bulgarian investigates the interaction of the range of possible word orders and direct object clitic doubling, discussing the contexts of both the optional and obligatory reduplication in Bulgarian. The analysis presented here treats all (in)direct objects that co-occur with pronominal clitics as instances of clitic doubling, regardless of the position of the objects. It is also argued that the obligatory vs. optional presence of a doubling clitic in Bulgarian is dependent on the syntactic structure rather than just the specific predicate types. In the paper entitled For the property analysis of control constituent An argument from modal existential wh-constructions Radek Šimík offers

Sabina Halupka-Rešetar, Maja Markoviü, Tanja Miliüev and Nataša Miliüeviü 3

a novel approach to the issue of the mechanism behind the control relation. The author claims that the control constituents should be semantically viewed as properties and, therefore, the control relation requires a semantic resolution. His main argument for this position comes from the modal existential wh-constructions. Pointing out a strong parallelism between obligatorily controlled PRO and wh-expressions, he argues for treating them both as logical lambda-operators. Sabina Halupka-Rešetar's Nonfinal information focus in Serbian – some experimental data challenges the standard minimalist claim that PF has to pronounce the highest member of a nontrivial chain. The analysis put forward in Halupka-Rešetar (2009), which rests, on the one hand, on viewing [+f] as a lexical feature, and on the other hand, on a version of Boškoviü’s (2008) Pronounce Lower Copy, allowing the pronunciation of a lower copy if doing so does not violate any phonological constraints is put to a test by examining the prosodic properties of senteces uttered by native speakers of Serbian, with the goal of corroborating the assumption that non-final information focus needs to be prosodically marked. The well-know generalization that the second position clitics in Serbian are sensitive to the Intonational Phrase (IP) boundaries is reexamined in Tanja Miliüev and Maja Markoviü’s Clitic placement and the properties of the Intonational Phrase in Serbian. The goals of this paper are to establish the exact set of properties required for a constituent to be realized as an IP in Serbian, and consequently, delay clitic positioning in a clause. Basing their conclusions on the relevant production tests, the authors find that in addition to the specific phonological traits, the cliticdelaying constituents have distinct semantic/pragmatic properties. It is, therefore, their distinct interpretation that sometimes allows the relevant phonological features to be somewhat obliterated in the production/ perception of I- boundaries. In a session on Natural language processing, which was organized for the first time at SinFonIJA 3, there were several issues addressed that illustrate the kind of linguistic challenges the developers of modern technologies in the region have to deal with. Staša Vujiþiü Stankoviü’ s Named entity recognition in the system for information extraction discusses the problems related to the Serbian language in the domain of named entity recognition in information extraction. The author proposes a solution to these problems based on the language resources developed for Serbian, using the GATE1 and the Unitex 2 systems. Prosody prediction in speech synthesis based on regression trees by Milan Seþujski, Darko Pekar, Dragan Kneževiü and Vladimir Svrkota examines the possibility of automatic prediction of prosodic parameters,

4

Introduction

namely f0 contour and phonetic segment durations, in the context of textto-speech synthesis in the Serbian language. The prosodic parameters are predicted using regression trees trained on a speech corpus containing approximately 4 hours of a recorded and annotated speech. The speech corpus annotation contains both phonological and phonetic markers, as well as markers related to prosody (lexical pitch accents, prosodic phrase boundaries and prosodic prominence). The results, in terms of values of objective measures, are shown to be comparable with those reported for other languages. Milan Gnjatoviü and Vlado Deliü’s paper Attention and linguistic encoding of motion events in human-machine interaction introduces a computational model of a schematic system of attention in spatial language. It expands upon the previous work – the focus tree computational model of attentional state in task-oriented human-machine interaction (HMI). The study extends the concept of the focus tree with an insight from the cognitive semantics into a dichotomy between two fundamental cognitive functions in language – of the Figure and of the Ground. The paper also discusses the possibility of the application of the model in the field of therapeutic HMI on children with developmental disorders.

References Boškoviü , Ž. 2008. On the Clausal and NP-structure of Serbo-Croatian. In Formal Approaches to Slavic Linguistics: The Toronto Meeting, 2006, edited by R. Compton et al., 42-75. Ann Arbor, MI: Michigan Slavic Publications. Chomsky, N. 1957. Syntactic Structures. The Hague: Mouton. Halupka-Rešetar, S. 2009. Strukturna artikulacija fokusa u engleskom i srpskom jeziku: uporedna analiza strukturalnog i generativnog pristupa. [The structural articulation of focus in English and Serbian: a comparative analysis of the structural and the generative approach.]. Doctoral Dissertation. University of Novi Sad. Miyoshi, N. 2002. Negative imperatives and PF merger. Ms. University of Connecticut, Storrs. Zeijlstra, H. 2004. Sentential Negation and Negative Concord. Doctoral Dissertation. University of Amsterdam.

PART ONE. SYNTAX, PHONOLOGY AND LANGUAGE ANALYSIS

WHEN ARE NEGATIVE IMPERATIVES BANNED? ŽELJKO BOŠKOVIû

Outline Zeijlstra (2004) observes that only a subset of the set of languages with a negative marker Xo bans true negative imperatives. This paper shows that the generalization can be deduced under the affix hopping analysis of the ban on negative imperatives, where in languages where the ban holds negation blocks affixation of a null imperative affix to the verb.

1. Ban on negative imperatives and different types of negation It is well-known that languages differ regarding their treatment of sentential negation. Some languages, for example, Italian and Russian, have Xo negation, where the negation heads a NegP,1 while other languages, e.g. Icelandic, have adverbial adjunct negation, where negation is adjoined to an independent phrase (e.g. VP). I will refer to this distinction as head vs adjunct negation (see Zeijlstra 2004 for a more extensive language survey and relevant references). Languages also differ regarding the ban on negative imperatives. Some languages disallow negative imperatives, switching to a different verbal form in the context in question, while other languages allow negative imperatives. Among the languages with independent imperative forms, the ban on negative imperatives, illustrated by Greek (1), holds for Greek, Romanian, Spanish, Italian, Portuguese, Catalan, Sardinian, Hungarian, Hebrew, and Latin. On the other hand, Serbo-Croatian, Slovenian, Bulgarian, Macedonian, Russian, Polish, Czech, Albanian, German, Swedish, Norwegian, Yiddish, Berber, and Basque allow negative imperatives.

1 In some languages of this type negation may undergo head movement outside of NegP, which I ignore here.

Željko Boškoviü

(1)

a. b.

7

Diavase! read.Imp *Den/mi diavase! Neg read.Imp ‘Don’t read!’

Instead of an imperative verb form, Greek uses a subjunctive in a negative imperative context, i.e. as a surrogate imperative.2 (2)

Na mi Subj.Mark Neg ‘Don’t read!’

diavazis! read.Subj

Zeijlstra (2004:165) makes a very interesting observation that correlates the two cross-linguistic differences noted above (see also Zanuttini 1997). In particular, generalizing Zanuttini’s discussion of Romance, Zeijlstra establishes the following generalization. (3)

Only a subset of the set of languages with a negative marker Xo bans true negative imperatives.

In other words, what Zeijlstra observes is that the ban on negative imperatives is found only in Xo negation languages. Indeed, all the ban-onnegative-imperatives languages noted above are head negation languages. Note, however, that we are dealing here with a one-way correlation: not all head negation languages have negative imperatives. In this paper I will show that one particular analysis of the ban on negative imperatives provides a principled explanation why the ban in question is found only in head negation languages. In the next section I discuss the analysis in question, summarizing a couple of arguments for it (since the analysis does not appear to be widely-known). In section 3 I show how the analysis in question captures Zeijlstra’s generalization (for relevant discussion, see also Zanuttini 1997; Zeijlstra 2004).

2

Languages differ with respect to which verbal forms they use as surrogate imperatives. The options are subjunctive, infinitive, indicative, and gerund. For relevant discussion, see especially Zanuttini (1997).

When are Negative Imperatives Banned?

8

2. The affix hopping analysis of the ban on negative imperatives There are a number of accounts of the ban on negative imperatives in the literature, (see, e.g., Boškoviü 2004; Han 1999; Isac & Jakab 2001; Laka 1994; Miyoshi 2002; Rivero 1994; Rivero & Terzi 1995; Tomiü 2001; Zanuttini 1994, 1997). Interestingly, a similar phenomenon is found in English. Just like the languages in question, English has a verbal form that cannot co-occur with negation. Whereas the languages in question disallow negative imperative verbs (I will use Greek as a representative of these languages), English disallows negative finite verbs, more precisely, finite main verbs. (I will refer to them as indicatives.) Like Greek, English switches to another verbal form in the context in question, namely, infinitive. (4)

a. b.

*John not laughed. John did not laugh.

Abstractly, we have the same pattern in both Greek and English. Both languages disallow a particular verbal form to co-occur with negation. In the relevant negative context, they switch to another verbal form. The parallelism between Greek and English is generally not noted in the existing accounts of the ban on negative imperatives, which appear to have nothing to say about it. (That is, they are not readily extendable to the ban on negative indicatives in English.) One exception is Miyoshi (2002), further developed in Boškoviü (2004), where a uniform account is provided for the ban on negative imperatives in Greek and the ban on negative indicatives in English. This is implemented by extending a particular account of the ban on negative indicatives in English to the ban on negative imperatives in Greek, more precisely, Chomsky’s (1957) affix hopping analysis, which was revived recently in Hale and Marantz (1993), Bobaljik (1994, 1995), and Lasnik (1995), and extended to several other phenomena in Boškoviü (2001a,b). In the recent instantiations, affix hopping, often referred to as PF/morphological merger (I will use the terms interchangeably), is treated as a morphophonological rule that involves merger between an affix and its host in PF under adjacency. Merger/affix hopping is blocked by intervening phonologically realized elements, but not by phonologically null elements such as traces and pro. To illustrate the mechanism, consider (5a-c), whose pre-PF merger and Do-Support structures are given in (6).

Željko Boškoviü

(5)

a. b. c.

John laughed. *John not laughed. John did not laugh.

(6)

a. b.

[IP Johni I (ed) [vP ti laugh]] [IP Johni I (ed) [NegP not [VP ti laugh]]]

9

English I is a verbal affix, which must merge with a verbal element in PF under adjacency. The adjacency requirement is not met in (6b) due to the intervening negative head, which blocks PF merger. Do-Support, a last resort operation, then takes place to save the stranded affix, deriving (5c). In (6a), the merger is not blocked since no phonologically realized element intervenes between I and the verb. I then merges with the verb, deriving (5a). The crux of the analysis is that indicatives cannot co-occur with negation in English because the co-occurrence results in a violation of the Stranded Affix Filter, which rules out constructions with stranded affixes. Miyoshi (2002) extends this analysis to the ban on negative imperatives. He proposes that imperatives in languages like Greek contain a functional head, the precise identity of which is not important for our purposes (for Miyoshi, it is an imperative C), which is a PF affix that must merge with a verb under adjacency. Affix hopping can proceed without any problems in (1a), where the verb and the functional head in question, referred to as F, are adjacent. However, in (1b) the negation disrupts the adjacency relation between F and the verb. Affix hopping is then blocked and the example is ruled out due to the presence of a stranded affix, just like (6b). (7)

F [+affix]

den/mi diavase.

Greek does not have the language specific rule of Do-Support, which English employs in (5c) to save the stranded affix. Instead, Greek uses a different verbal form, namely subjunctive. We can assume either that the affix head F is not present in subjunctive imperatives or that it is supported by the subjunctive marker na.3 Regarding languages that allow negative imperatives, they either do not have F or, perhaps more plausibly, F is not an affix in such languages. The difference between languages with and those without a ban on negative imperatives is then treated in terms of a rather straightforward 3

Notice, however, that the subjunctive marker is optional in constructions like (2).

When are Negative Imperatives Banned?

10

lexical difference regarding the PF status of the imperative head. Miyoshi provides evidence for the affix hopping analysis of (1) by showing that it also accounts for the often observed difference in clitic placement in imperative and non-imperative contexts. It is well-known that Greek clitics generally precede the verb in indicatives, but follow it in imperatives. (8)

a.

(9)

b. a. b.

To diavasa. it read.Ind ‘I read it.’ *Diavasa to. Diavase to! read.Imp it ‘Read it!’ *To diavase!

Miyoshi proposes a uniform account of (1) and (9) based on Franks’s (1998, 2000) (see also Abels 2001, Bobaljik 1994, 1995, 2002, Boškoviü 2001a, 2002, Boškoviü & Franks 2002, Boškoviü & Nunes 2007, Hiramatsu 2000, Lambova 2001, Landau 2003, Nunes 2004, Pesetsky 1998a,b, Reglero 2004, Stjepanoviü 2003) proposal that a lower copy of a non-trivial chain can be pronounced instead of the head of the chain iff this is necessary to avoid a PF violation. I will illustrate how the proposal works with respect to the analysis of the basic cliticization pattern in Bulgarian (B) and Macedonian (Mac) from Boškoviü (2001a). Consider (10).4 (10)

a. b. c.

Petko mi go Petko me.Dat it.Acc ‘Petko gave it to me.’ Mi go dade. Dade mi go.

dade. gave

B:OK

Mac: OK

B:* B: OK

Mac: OK Mac: *

The contrast between Bulgarian and Macedonian (10b) indicates Bulgarian clitics must encliticize, while Macedonian clitics procliticize. Macedonian clitics always precede the verb in the context in question. Bulgarian clitics 4

I ignore here certain non-finite clauses in Macedonian, which raise additional questions. For relevant discussion, see Boškoviü (2001a), Caink (1998), Franks (1998, 2000), Franks & King (2000), Legendre (1999), and Tomiü (1996, 1997), among others. The works in question also contain a more general discussion of clausal cliticization in Bulgarian and Macedonian.

Željko Boškoviü

11

precede the verb unless preceding it would result in a violation of their enclitic requirement. In that case they follow the verb. This state of affairs can be accounted for under Franks’s (1998) proposal that a lower copy of a non-trivial chain can be pronounced instead of the head of the chain iff this is necessary to avoid a PF violation, given that a copy of pronominal clitics is present both above and below the verb (see Boškoviü 2001a for discussion of the precise position of these copies). This approach straightforwardly captures the generalization that the verb can precede a clitic in Bulgarian only when no other lexical material is located in front of the clitic. Only in this situation will we be able to pronounce the lower copy of the clitic, which is located below the verb. (Pronunciation of the head of the clitic chain in (11b) would lead to a PF violation since the clitic, which must encliticize, cannot be properly prosodically supported.) If there is lexical material preceding the clitic in its raised position, the head of the chain of clitic movement can be, hence must be pronounced. (11)

a. b.

X clitic V clitic clitic V clitic

Since in Macedonian nothing goes wrong in PF if the head of the clitic chain is pronounced, the head of the clitic chain, located above the verb, has to be pronounced. As a result, the V-clitic order is underivable in Macedonian. (12)

(X) clitic V clitic

The contrast in the acceptability of (10b-c) in Bulgarian and Macedonian, as well as the role of phonology in the possibility of the V-cl order in Bulgarian, is thus straightforwardly captured. Returning to (8)-(9), Miyoshi (2002) observes that given that a lower member of a non-trivial chain can be pronounced if this is necessary to avoid a PF violation, the affix hopping analysis provides a straightforward account of the V-clitic switch in (9). Assume that imperatives and indicatives in Greek do not differ regarding clitic placement in the syntax. They both have the clitic-V order, with a lower copy of the pronominal clitic following the verb. In indicatives, the higher copy of the clitic can be, hence must be pronounced. On the other hand, in imperatives pronunciation of the higher copy of the clitic would lead to a Stranded Affix Filter violation, i.e. a PF violation, since the clitic disrupts the adjacency between F and V, necessary for F to hop onto the verb. The violation can be avoided if we pronounce a lower copy of the clitic, which follows the verb.

12

(13)

When are Negative Imperatives Banned?

F to diavase to

Since the verb and F are adjacent in (13), affix hopping can take place. Lower pronunciation of the clitic is licensed in (13), just as in Bulgarian (10c), because it is necessary to avoid a PF violation.5 The affix hopping analysis thus provides us with a principled account of the clitic-V switch in languages that have a ban on negative imperatives. In fact, the clitic-V switch and the ban on negative imperatives are accounted for in the same way. 6 In Boškoviü (2004) I provide an additional argument in favor of the affix hopping analysis. In particular, I show that the affix hopping analysis provides us with a principled account of a peculiar clitic switch in Greek imperatives, noted in Joseph and Philippaki-Warburton (1987), Terzi (1999), and Warburton (1977); more precisely, the fact that the accusativedative clitic order is available in the postverbal clitic position in imperatives, in contrast to the preverbal clitic position in indicatives, where only the dative-accusative clitic order is possible. I argue that the clitic switch is a PF phenomenon that arises through lower copy pronunciation. More precisely, while the highest position of the clitics, which is pronounced in indicatives, yields the dat-acc order, lower copy pronunciation in imperatives yields the acc-dat order. Here is the gist of the analysis: I argue that, with respect to clitic placement, imperatives are derived just like indicatives in that in the highest syntactic position, the order of clitics in a double object clitic construction is dative-accusative. 5

Bobaljik (1995) and Boškoviü (2001a) also propose analyses in which a lower copy of X is pronounced in order to prevent X from blocking affix hopping. 6 The analysis does not posit a two-way correlation between the clitic-V switch and the ban on negative imperatives. Thus, Miyoshi argues that in some Italian surrogate imperatives, which do not contain the F affix, the V-clitic order arises as a result of V-movement. (The same seems to hold for Cypriot Greek indicatives that have V-clitic order, as discussed in Terzi 1999. Note that the accusative-dative clitic switch, discussed directly below, is not expected to occur in constructions in which the V-clitic order arises as a result of V-movement since in such constructions the clitics are pronounced in the highest position (see the discussion below). As shown in Terzi 1999, it indeed does not occur in Cypriot Greek indicatives.) It is also worth noting that in Boškoviü (2001a) I argue that the affix hopping analysis also leaves room for an F-affix language that has the clitic-V switch but no ban on negative imperatives. (The language in question is Macedonian. See Boškoviü 2001a for explanation why the presence of the F affix does not make negative imperatives impossible in Macedonian; see also Miyoshi 2002 for an alternative analysis of Macedonian that does not involve affix hopping.)

Željko Boškoviü

13

Under the affix hopping analysis, all we need in order to accomplish clitic switch is that in a lower position the order can be accusative-dative. Since in indicatives the highest copy of the pronominal clitics must be pronounced, we still get only the dative-accusative order in indicatives. On the other hand, since in imperatives lower copies of the pronominal clitics are pronounced for reasons discussed above, we can get the accusativedative order.7 (14)

a. b.

dat acc V acc dat… dat acc V acc dat…

(indicatives) (imperatives)

3. Deducing Zeijlstra’s generalization I now return to Zeijlstra’s generalization, demonstrating that it can be straightforwardly captured under the affix hopping analysis of the ban on negative imperatives. We have already seen how the ban on negative imperatives in head negation languages can be captured under this analysis. Furthermore, we have seen that the analysis does not force the ban on all languages of this type; i.e. it leaves room for languages with head negation to differ regarding the ban on negative imperatives. What we need to do now is prevent the ban on negative imperatives from ever applying in adverbial negation languages. This is actually straightforward. Bobaljik (1994, 1995) gives a number of arguments that adjuncts (i.e. adjoined elements) do not interfere with affix hopping even when they intervene between the elements involved in the merger. To give just one illustration of this, the adjunct never in (15)-(16) differs from the negative head in 0)-0) in that it does not block affix hopping. As a result, DoSupport does not take place in (15), in contrast to (4). 8 7

See Boškoviü (2004) for details of the analysis; as discussed in this work, the lower copy pronunciation analysis in fact allows for both the acc-dat and the datacc order in imperatives, depending on which lower copies of the clitics are activated in PF. Note that the clitic switch is not possible in all languages where clitics are pronounced in a lower position in imperatives. E.g., it is not possible in Spanish. The reader is referred to Boškoviü (2004) for an account of this difference between Spanish and Greek, where the difference is tied to another independently motivated difference between the two languages, namely the height of verb movement. 8 While Bobaljik (1994, 1995) simply stipulates the “invisibility” of adjuncts with respect to affix hopping, Bobaljik (2002) and Ochi (1999) offer a deduction of the different behavior of adjoined and non-adjoined elements in this respect. Thus, assuming Lebeaux’s (1988) acyclic adjunct insertion and multiple spell-out, Ochi

14

(15) (16)

When are Negative Imperatives Banned?

John never laughed [IP Johni I (ed) [VP never [VP ti laugh]]]

Since in adjunct negation languages the negative marker is an adjunct, just like never in English, it then follows that languages of this type will not show the ban on negative imperatives. I therefore conclude that the affix hopping analysis deduces Zeijlstra’s generalization in its entirety: it allows, but does not force the ban on negative imperatives in head negation languages, and prevents the ban from applying in adjunct negation languages. To the extent that the analysis is successful, it also provides additional evidence for the mechanism of affix hopping.

References Abels, K. 2001. The predicate cleft construction in Russian. In Formal Approaches to Slavic Linguistics: The Indiana Meeting 2000, edited by S. Franks, T. King and M. Yadroff, 1-19. Ann Arbor, MI: Michigan Slavic Publications. Bobaljik, J. 1994. What does adjacency do?. MIT Working Papers in Linguistics 22. Cambridge, Mass.: MITWPL, Department of Linguistics and Philosophy, MIT, 1-32. —. 1995. Morphosyntax: The Syntax of Verbal Inflection. Doctoral dissertation. MIT, Cambridge, Mass. —. 2002. A-chains at the PF interface: Copies and “covert” movement. Natural Language and Linguistic Theory 20, 197-267. Boškoviü, Ž. 2001a. On the Nature of the Syntax-Phonology Interface: Cliticization and Related Phenomena. Amsterdam: Elsevier Science. —. 2001b. PF Merger in Scandinavian: Stylistic Fronting and Object Shift. Working Papers in Scandinavian Syntax 68, 75-115. —. 2002. On multiple wh-fronting. Linguistic Inquiry 33, 351-383. —. 2004. On the clitic switch in Greek imperatives. In Balkan syntax and semantics, edited by O. Mišeska Tomiü, 269-291. Amsterdam: John Benjamins.

argues that the reason why adjuncts do not interfere with affix hopping is that they can be acyclically inserted into the structure after the structure, with the elements involved in affix hopping adjacent to each other, has already been sent to PF. Affix hopping/morphological merger is then licensed derivationally, at the point when the relevant elements are adjacent (and before the adjunct is acyclically inserted into the structure). I refer the reader to Ochi (1999) for independent evidence for this analysis.

Željko Boškoviü

15

Boškoviü, Ž. and S. Franks. 2002. Phonology-syntax interactions in South Slavic. Balkanistica 15, 49-74. Boškoviü, Ž. and J. Nunes. 2007. The copy theory of movement: A view from PF. In The copy theory of movement on the PF side, edited by N. Corver and J. Nunes, 13-74. Amsterdam: John Benjamins. Caink, A. 1998. The Lexical Interface: Closed Class Items in South Slavic and English. Doctoral dissertation. University of Durham. Chomsky, N. 1957. Syntactic Structures. The Hague: Mouton. Franks, S. 1998. Clitics in Slavic. Paper presented at the Comparative Slavic Morphosyntax Workshop, Spencer, Indiana. —. 2000. Clitics at the interface: An introduction to Clitic Phenomena in European Languages. In Clitic Phenomena in European Languages, edited by F. Beukema and M. den Dikken, 1-46. Amsterdam: John Benjamins. Franks, S. and T. King. 2000. A handbook of Slavic Clitics. Oxford: Oxford University Press. Halle, M. and A. Marantz. 1993. Distributed morphology and the pieces of inflection. In The View from Building 20: Essays in Linguistics in Honor of Sylvain Bromberger, edited by K. Hale and S. J. Keyser, 111176. Cambridge, Mass.: MIT Press. Han, C.-H. 1999. Cross-linguistic variation in the compatibility of negation and imperatives. Proceedings of the West Coast Conference on Formal Linguistics 17. Stanford, Calif.: CSLI Publications, 1-17. Hiramatsu, K. 2000. Accessing Linguistic Competence: Evidence from Children’s and Adults’ Acceptability Judgments. Doctoral dissertation. University of Connecticut, Storrs. Isac, D. and E. Jakab. 2001. A unified syntactic analysis of Balkan imperatives with distinctive morphology. Paper presented at the Acme Balkanica Conference, Concordia University, Montreal. Joseph, B. and I. Philippaki-Warburton. 1987. Modern Greek. London: Croom Helm. Laka, I. 1994. On the Syntax of Negation. New York: Garland. Lambova, M. 2001. On A’-movements in Bulgarian and their interaction. The Linguistic Review 18, 327-374. Landau, I. 2003. Modular recoverability: Chain resolution in Hebrew V(P)-fronting. Ms. Beer-Sheva, Israel. Lasnik, H. 1995. Verbal morphology: Syntactic Structures meets the Minimalist Program. In Evolution and Revolution in Linguistic Theory: Essays in Honor of Carlos Otero, edited by H. Campos and P. Kempchinsky, 251-275. Washington, D. C.: Georgetown University Press.

16

When are Negative Imperatives Banned?

Lebeaux, D. 1988. Language acquisition and the form of the grammar. Doctoral dissertation. University of Massachusetts, Amherst. Legendre, G. 1999. Morphological and prosodic alignment at work: The case of South-Slavic clitics. Proceedings of the West Coast Conference on Formal Linguistics 17. Stanford, Calif: CSLI Publications, 436-450. Nunes, J. 2004. Linearization of chains and sideward movement. Cambridge, Mass.: MIT Press. Miyoshi, N. 2002. Negative imperatives and PF merger. Ms. University of Connecticut, Storrs. Ochi, M. 1999. Multiple spell-out and PF adjacency. Proceedings of NELS 29. University of Massachusetts, Amherst: GLSA, 293-306. Pesetsky, D. 1998a. Optimality theory and syntax: Movement and pronunciation. In Optimality Theory: An Overview, edited by D. Archangeli and T. Langendoen, 134-170. Malden: Blackwell. —. 1998b. Some optimality principles of sentence pronunciation. In Is the Best Good Enough?, edited by P. Barbosa et al., 337-383. Cambridge, Mass.: MIT Press and MITWPL, Department of Linguistics and Philosophy, MIT. Reglero, L. 2004. On A’ dependencies in Spanish and Basque. Doctoral dissertation. University of Connecticut, Storrs. Rivero, M. L. 1994. Negation, imperatives, and Wackernagel effects. Rivista di Linguistica 6, 39-66. Rivero, M. L. and A. Terzi. 1995. Imperatives, verb movement, and logical mood. Journal of Linguistics 31, 301-332. Stjepanoviü, S. 2003. A word-order paradox resolved by copy deletion at PF. Linguistic Variation Yearbook 3, 139-177. Terzi, A. 1999. Clitic combinations, their hosts, and their ordering. Natural Language and Linguistics Theory 17, 85-121. Tomiü, Mišeska O. 1996. The Balkan Slavic clausal clitics. Natural Language and Linguistic Theory 14, 811-872. —. 1997. Non-first as a default clitic position. Journal of Slavic Linguistics 5, 1-23. —. 2001. Negation and imperatives. In Current issues in formal Slavic linguistics, edited by G. Zybatow et al., 159-168. Frankfurt am Main: Peter Lang. Warburton, I. 1977. Modern Greek clitic pronouns and the “Surface Structure Constraint” hypothesis. Journal of Linguistics 13, 259-181. Zanuttini, R. 1994. Speculations on negative imperatives. Rivista di Linguistica 6, 119-142. —. 1997. Negation and Clausal Structure: A Comparative Study of Romance Languages. Oxford: Oxford University Press.

Željko Boškoviü

17

Zeijlstra, H. 2004. Sentential negation and negative concord. Doctoral dissertation. University of Amsterdam.

SOME OBSERVATIONS ON MULTIPLE X HENK VAN RIEMSDIJK

Outline The point of departure for the present article is the question of whether adjunct free relatives and adjunct questions can or should be distinguished. It is suggested that there are no strong grounds for distinguishing them. Part of the discussion involves the issue of multiple wh structures. A link between wh-adjuncts and conditional clauses is suggested, raising the question whether multiple wh generalizes to multiple X, where X could include multiple targets in conditionals as well as multiple degree phrases in comparative and consecutive clauses. There is some evidence that the answer should be yes, but the jury is still out.

1. Multiple puzzles The goal of this paper is a modest one, viz. to formulate a question that has bothered me for a considerable time and to place it in a somewhat wider context in the hope of thereby clarifying a few issues and, hopefully, improve the prospects for a solution that remains to be found. More specifically, I propose to explore, in section 2, the boundaries between indirect questions and adjunct free relatives (Van Riemsdijk, 2006:360f). In section 3, I discuss the absence of multiple wh in some languages and suggest that conjoined wh-constructions may constitute an alternative strategy to express multiple X. Section 4 explores the range of constructions in which multiple X appears to manifest itself, and section 5 adds yet another construction, comparative correlatives, to this inventory.

2. Free relative adjuncts or indirect questions? Most discussions of free relative clauses (henceforth FRs) center around FRs in argument positions. Indeed, all the important questions surrounding FRs such as categorial and case matching arise in FRs of this

Henk van Riemsdijk

19

kind. Observe, however, that in many cases what looks like a FR shows up in a non-argument, that is in an adjunct position: (1) (2)

Whatever you do, don’t cross the road Whenever you want to leave, just tell me

In both examples we appear to have the free choice interpretation typical of (the universal subtype of) FRs. The fact that the wh-word in these examples carries the suffix –ever reinforces this conclusion as –ever does not occur in questions, modulo some marginal exceptions such as whatever are you talking about? In contrast with –ever there are other whsuffixes that are typically found with questions and not in FRs such as the hell and the fuck: (3) (4) (5)

a. b. a. b. a. b.

I wonder what the hell they will tell me Have you got any idea who the fuck this is? *I tend to eat what the hell they serve me *I will talk to who the fuck I want *What the hell you do, don’t cross the road *When the fuck you want to leave, just tell me

The presumed adjunct free relatives in (5) pattern with the standard FRs in (4) and not with the embedded questions in (3). Unfortunately, this is not all there is to say about this problem. There is another property that is generally attributed to questions and not to FRs: multiple wh. Indeed, it appears that presumed adjunct FRs such as those in (1/2) can be constructed with multiple wh: (6) (7)

Whatever book you buy in whatever store, you always end up paying too much Whenever you want to buy whoever a birthday present, you never know what to get

The interpretations of such examples is very similar to the typical pair-list readings familiar from multiple wh-questions, except that here we have a pairing of free choices: it does not matter what book you buy, and it does not matter in what bookstore you buy it, you end up paying too much. That is, once you have made a choice out of the set of possible books, and once you have chosen a store from the set of possible stores, the resulting pairing will lead to your spending more than you should or want to spend.

Some Observations on Multiple X

20

There is, however, a quite plausible reason why multiple wh does not occur in standard FRs, as shown in the following examples: (8)

a. b.

I wonder who danced with whom *I met who danced with whom

It is generally agreed that in a simple standard FR the wh-word has a double function. On the one hand it plays a role in the semantics of the relative clause, as regular relative pronouns do, and on the other hand it plays a role in the matrix clause, as the head of a headed relative clause tends to do: (9)

a. b.

I met the girl who danced with Joe I met who Joe danced with

In (9a), the girl is the object of met, and who is the subject of dance. In (9b), however, who is both the object of met and the object of the preposition with. That is, it does double duty. This much is uncontroversial, though the proper treatment of such sharing phenomena remain somewhat controversial (cf. Van Riemsdijk, 2006 and references cited there). Returning now to (8b), the problem is what the status of the second wh-word could possibly be. If it is only the first who that is shared with the matrix clause, then there is no conceivable way in which the second wh-word whom could be either interpreted as a question word or as a relative pronoun. The question word interpretation does not make any sense anyway, since we are looking at a FR. And we independently know that headed relative clauses with multiple wh-words do not exist. If we tried to force the issue, we would have to assume that the wh-word in situ could somehow undergo the absorption process typical of multiple whquestions (cf. Dayal, 2006 for extensive discussion). But then this would presuppose that absorption (normally thought of as covert movement or some other process in LF) results in a syntactic object that can be shared by the matrix structure. In the case of (8b) this would result in a meaning roughly corresponding to ‘I met the pair of persons x and y such that x danced with y’. But both x and y would be represented by the single whword who. That would imply a kind of across-the-board movement of both the subject who and the prepositional who into the Spec, CP position of the relative clause, clearly an impossibility. I conclude from these considerations that it is the absence of sharing and the absence of ATB-movement in examples like (6/7) that permits the

Henk van Riemsdijk

21

multiple wh property. Hence we can still maintain the possibility that such examples are adjunct FRs and not indirect questions. There are other problems for the adjunct FR analysis, however. Consider the fact that whether can also be used to introduce semantically similar adjuncts: (10) (11)

Whether you believe it or not, it’s going to rain Whether or not he has enough money isn’t the issue

The difference between these two examples is that in (11) the wh-clause introduced by whether is the subject of the matrix clause while in (10) the whether-clause is an adjunct. Indeed the matrix predicate in (11) is one that typically takes an indirect question as a subject. A predicate that typically does not, yields an ungrammatical result with a whether-clause, as shown in (12). (12)

*Whether or not he has enough money is a dubious fact

In (10) on the other hand, there is no such sensitivity to the choice of the matrix predicate. Put differently, the matrix predicate must be saturated independently of the adjunct. Suppose we take this to mean that (10) can still be an (adjunct) FR, we must then ask again why whether cannot show up in regular FRs. Or can it? Ostensibly it cannot. That is, every time a whether-clause is selected by some matrix predicate, we call it an indirect question. That makes sense to the extent that the interpretation of whether-questions differs from that of FRs in at least one crucial respect. FRs are either definite (I ate what you prepared – what is ‘the (food) thing(s) that you prepared) or free choice universals (I will eat whatever you prepare – whatever means anything that you happen to prepare). Whether (or not) does not fit either of these two meanings in any obvious way. Adjunct FRs belong to the free choice subcase of FRs. Might we not say, then, that whether is special because, being the element that introduces Yes/No questions, the choice is limited to two options but is otherwise as free as can be. So we again find ourselves in a position where it is not clear at all whether at the level of adjuncts there is any systematic distinction between questions and FRs. It is as if the distinction which, in selected positions, is clear enough, is neutralized in adjunct positions. This suspicion is reinforced when we consider what type of adjuncts we are actually talking about. In many cases these adjuncts appear to be very close to conditionals. This is indeed suggested by the fact that the

Some Observations on Multiple X

22

word if occurs both in conditionals and in questions. The generalization is stated explicitly in Bhatt and Pancheva (2006): (13)

Interrogative adjunct clauses are interpreted as conditionals (= Bhatt and Pancheva, 2006:(42))

Many authors have contributed to this generalization, see their article for references. To illustrate, take example (2) above. This example is easily paraphrasable as (13). (14)

If (at any time) you want to leave, just tell me

If this line of reasoning is on the right track, then it should not come as a surprise that conditional adjuncts of this type can also be ‘multiple’ in the sense that the computation to determine whether the condition is met involves more than one target. Take an example like the following. (15)

a. b.

If I come across any article on Lezgian in any linguistics journal, I get very excited If you want to drink any alcoholic beverage in any city of Pakistan, you had better do so in a private home with the curtains closed

Clearly what gets me excited in (15a) is not any old article on Lezgian, nor an arbitrary article in a linguistics journal, but the combination, that is, an article in a linguistics journal that discusses Lezgian. Similarly, the caution expressed in (15b) holds for the combination of alcoholic beverages and cities in Pakistan, not for non-alcoholic beverages, and not for cities in many other countries. Without venturing any thoughts on how to formalize this, it seems quite clear that combining two noun phrases in a conditional clause in this way is essentially the same thing as combining two (or more) wh-phrases by absorption in multiple questions. If the above considerations are on the right track, they raise a number of questions about what does and what does not belong in the realm of what we may call ‘multiple X’, in other words constructions in which more than one phrase constitutes the target for the calculation of the values that must be obtained in order for some condition or question to hold in a significant way. These are predominantly semantic questions, which I should and will leave to semanticists to answer. My sole purpose here is to contribute a few syntactic observations that may be useful in drawing the

Henk van Riemsdijk

23

lines of what does and what does not belong to the species ‘multiple X’. This is what I will take up in the next section.

3. Absence of multiple wh and alternative strategies Some languages, such as Italian, lack the possibility of using the multiple wh strategy, see (2008) for discussion: (16)

(17)

*Non so qual libro ho comprato non I-know which book I-have bought dove where ‘I don’t know which book I bought where’ *Quando lo hai trovato dove, when it you-have found where, questo anello? this ring ‘When did you find it where, this ring?’

There is, however, an alternative strategy that such languages can use to produce the same effect: conjoined wh (example from Haida and Repp (to appear)). (18) (19)

Quando e dove esci normalmente? when and where you-get-out normally ‘When and where do you normally get out?’ Quando e dove lo hai trovato, when and where it you-have found, questo anello? this ring ‘When and where did you find it, this ring?’

Note that the use of this strategy is not limited to languages that lack multiple wh constructions such as Italian. English also likes to use conjoined wh-phrases. Note further that Italian FRs differ in other ways from Germanic ones (cf. Caponigro, 2002, 2003): (20)

C’ è chi dice sempre there is who says always ‘There is someone/are people who always say(s) yes’

si yes

24

(21)

Some Observations on Multiple X

Non aveva dove nascondersi in caso not he-had where hide in case di pericolo of danger ‘He did not have anyplace to hide in case of danger’

The difference appears to be that these FRs have an essentially indefinite interpretation, as opposed to FRs of the Germanic kind, as discussed briefly in section 2 above. Other languages that are like Italian (according to Caponigro) include Spanish, Portuguese, French, Romanian, Russian, Serbo-Croatian, Bulgarian, Hungarian, Modern Greek, Modern Hebrew, and Yiddish. It is tempting to suppose that this property is somehow related to the (non-)occurrence of multiple questions. This issue must unfortunately be relegated to future research. Turning back to the more central topic of this paper, is it possible to use the conjoined wh strategy also in adjunct FRs? (22) (23)

To whichever doctor you go in this country and with whatever type of ailment, you are likely to be wrongly diagnosed and ripped off You will learn to become a food detective and control what and when you eat (fromWhitman, 2005:73 ex (87))

(22) is a typical FR adjunct, (23), however, is interesting because it appears to be at least in part a selected FR in that what is the object of control. And the when part may well be a selected FR as well to the extent that the meaning approximates control at what time you eat / control the time at which you eat. Or could this be a case of a conjoined adjunct question? The question may simply be the wrong question. As suggested at the end of section 2, there may not be any real difference between adjunct FRs and adjunct questions.

4. Towards multiple X There are, of course, many constructions that the general term ‘multiple X’ could apply to. Take, for example, the fact that (headed) relative clauses can take multiple heads, as in: (24)

We tend to team up boysi with girlsj in our school [that have a high IQ]i+j

Henk van Riemsdijk

(25)

25

Sports carsi tend to go together well with handbagsi, generally, [that have been created by Italian designers]i+j

Similarly, multiple superlatives can occasionally take a single PP adjunct. (26)

[Of the icons on your display]i+j, the most useful onesi should be separated from the cutest onesj

In both cases, it would appear that there is a set of items (pupils, objects designed by Italians, icons). The relative clause or adjunct roughly describes that set. The multiple heads to which those relative clauses or PP-adjuncts apply represent subsets that entertain a certain relation described in the main proposition (team up x with y, x goes together well with y, x should be separated from y). These cases may very well be genuine instances of what I call multiple X. But in order to not be led too far astray, I will limit myself here to two cases involving degree words, comparative and consecutive clauses. Consider the following multiple comparatives with a single than-clause, and multiple so-phrases with a single consecutive that-clause. (27) (28) (29) (30)

More peoplei do crazier thingsj at higher speedsk on the McGrath Highway [than they do other places]i+j+k (example from Andrews, 1985, attributed to Mark Liberman) John has more girl friendsi in more citiesj [than anyone else I know]i+j John has so many girl friendsi in so many citiesj [that he tends to forget which one is where]i+j Bill drank so many glasses of whiskyi in such a short timej [that he died of alcohol poisoning]i+j

Do these have anything in common with one another and with the previously discussed wh-adjuncts? It would appear that they do in the sense that a kind of absorption is involved in all these cases. That is the multiple comparatives must be computed in combination in order to evaluate whether the combined degree indeed exceeds the standard of comparison described in the than-clause. Similarly, the (elevated) number of girl friends and cities must together attain a certain value in order for the consequence described in the that-clause to ensue, and similarly it is the number of glasses of whisky combined with the relative shortness of the interval in which they got drunk that leads to the consequence of alcohol poisoning. For some relevant literature, see Corver (1993), Hendriks (1994) Meier (2001), Oda (2008), Von Stechow (1984).

Some Observations on Multiple X

26

Given this similarity, it should not come as a surprise that in many of these cases, an alternative with coordinated phrases is available: (31) (32)

More people do crazier things and at higher speeds on the McGrath Highway than they do other places (cf. (27)) Bill drank so many glasses of whisky and in such a short time that he died of alcohol poisoning (cf. (30))

Note also that, as in the case of conjoined wh-questions the coordinated Xs can sometimes be ‘adjacent’ or ‘distant’, which, depending on one’s analysis may correlate with phrasal vs. (reduced) clausal. In the case of resultative constructions, however, it appears that the phrasal (non-distant) conjoined variant is ungrammatical for reasons that remain obscure.1 (33) (34)

a. b. c. a. b.

When and where did you find it, this ring? When did you find it and where, this ring? When did you find it, this ring, and where? *So many glasses of whisky and in such a short time did Bill drink that he died of alcohol poisoning So many glasses of whisky did Bill drink and in such a short time that he died of alcohol poisoning

At this point we should also ask whether we are not deluding ourselves that any kind of absorption is involved in such cases. What about multiple topicalization, for example. This is generally held to be impossible in English, as shown in (35). (35)

*To Johnj, that booki, (Bill said that) Mary handed ti tj (Boškoviü, 2004:(16)p621)

And conjoining the two topicalized phrases does not help: (36)

*That book and to John, (Bill said that) Mary handed

Essentially the same facts, though perhaps marginally better, are found with argument+adjunct and with two independent adjuncts:2

1

Thanks to Jonathan Bobaljik, Joe Emonds and Edwin Williams for providing the judgment on example (34). 2 Note that the intended reading here is the one in which both adjuncts modify the most deeply embedded clause, i.e. Mary’s handing Bill that book.

Henk van Riemsdijk

(37)

27

?*On a sunny Saturday morning for no apparent reason (Bill said that) Mary handed him that book

Observe that conjoining the two topicalized phrases does not improve these examples either: (38)

a. b.

*That book and on a sunny Saturday morning (Bill said that) Mary handed him ?*On a sunny Saturday morning and for no apparent reason (Bill said that) Mary handed him that book

It appears, then, that we are justified in singling out cases in which some kind of absorption needs to be assumed to apply for semantic reasons. It should be kept in mind, however, that many languages do allow multiple topicalization. German and Dutch, for example, frequently allow more than one phrase to “share” the first position, i.e. the position preceding the finite verb in main clauses. Sie Müller (2003) for ample illustration in German:3 (39)

(40)

[Alle Träume] [gleichzeitig] lassen sich all dreams simultaneously let themselves nur selten verwirklichen only rarely realize ‘You rarely get to realize all dreams simultaneously’ (Müller’s (3b)) [Die Kinder] [nach Stuttgart] sollst du The children to Stuttgart should you bringen bring ‘You should bring the children to Stuttgart’ (Müller’s (14f))

But even if these cases appear to have something in common with the constructions discussed under the heading absorption, it should be noted that they are absolutely ungrammatical when the two topicalized phrases are conjoined.

3

Müller shows that many of the examples he cites can be analyzed in different ways and points out that some, such as (40), while attested, are not accepted widely.

Some Observations on Multiple X

28

(41)

a. b.

*Alle Träume und gleichzeitig lassen sich nur selten verwirklichen *Die Kinder und nach Stuttgart sollst du bringen

Therefore, I tentatively exclude multiple topicalization cases from the discussion. To conclude, let us now turn to a last case of conditional adjuncts, comparative correlatives.

5. Comparative correlatives A last construction that I propose to look at in the present context is that of comparative correlatives, often also referred to as comparative conditionals. For some selective literature on this construction, see Beck (1997), Den Dikken (2005), McCawley (1988). The issue of the construction’s name is discussed in some detail in cf. Brasoveanu (2008), and for a different view as well as data on Spanish and French, see Abeillé and Borsley (2008). Comparative correlatives are exemplified in (42). (42)

The faster you drive, the riskier it gets

As it happens, comparative correlatives in English are somewhat special in some ways, therefore it is good to start with a look at Dutch (43/44) and German (45/46). (43)

(44)

Hoe vaker je vitamines neemet how oftener you vitamins take hoe gezonder je wordt how healthier you get ‘The more frequently you take vitamins, the healthier you get’ Hoe harder je rijdt, deste meer how faster you drive the more benzine gebruik je gas use you ‘The faster you drive, the more gas you use’

Henk van Riemsdijk

(45)

(46)

29

Je öfter man in die Disko geht, desto the oftener one in the disco goes, the schneller wird man taub fast becomes one deaf ‘The more frequently you go to the disco, the faster you become deaf’ Je weniger Bücher du liest, desto mehr Zeit the fewer books you read the more time du hast für die Musik you have for the music ‘The fewer books you read, the more time you have for music’

In view of the fact that comparative correlatives contain degree operators (cf. the discussion on multiple comparatives and consecutives in section 4 above), it is tempting to investigate whether they share the other properties that we have found: the semantics of conditional adjuncts, multiple X and a conjoined X alternative. On the first property, we can be brief. A sentence like (42) is straightforwardly paraphrasable as (47). (47)

If you drive faster, it gets riskier

Similarly, a German example like (45) can be paraphrased as (48). (48)

If you go to the disco more frequently, you will become deaf more quickly

And indeed, it does not come as a surprise that there are multiple X comparative correlatives as shown in (49) for German. (49)

Je mehr Bücher weniger Seiten the more books fewer pages besser ist das für die better is that for the ‘The more books have fewer pages the environment’

haben desto have the Umwelt environment better it is for the

The meaning of this example, very roughly, is that given the total number of books, if the subset of books with a smaller number of pages (than average) is larger (than average), then it is better for the environment. And the more the combined numbers/extents deviate from the averages, the better.

Some Observations on Multiple X

30

(50/51) exemplify the same property in Dutch. (50)

(51)

Hoe meer taalkundigen lezen over uiteenlopendere how more linguists read about more-diverse talen deste beter zullen ze worden languages the better will they get ‘The more linguists read about more diverse languages, the better they will become’ Hoe meer taalkundigen over uiteenlopendere how more linguists about more-diverse talen lezen dan gebruikelijk, deste beter languages read than usual, the better zullen ze worden will they get ‘The more linguists read about more diverse languages than usual, the better they will become’

Note that (51) is identical to (50), but with a (reduced) than-clause explicitly stating that the evaluation is in terms of the average. The meaning here amounts to saying that if the number of linguists reading about more languages than the average linguist reads about exceeds a certain average, then those that do will become better linguists. Again, we find that computing the meaning of such a sentence involves a quite complex process of absorption of the individual evaluations of the degree phrases. Native speakers of English may well have noticed that the translations of (49-51) are not grammatical.4 Indeed, it appears to be the case that English lacks multiple comparative correlatives. In assessing this fact, we need to be cautious, for an example such as (52) does not appear to be bad. (52)

The more people read more books, the lower the crime rate

Note however that (52) is potentially ambiguous. The first more can either be a phrasal modifier of the DP people. In that case it means something like ‘if it is the case that an above average number of people read an above average number of books, then the crime rate will be lower’. Alternatively it can be an independent DP signifying an above average number (or frequency) of events, i.e. ‘if it is more often the case that people read more books than average, then the crime rate will be lower. 4

Thanks to Edwin Williams for pointing this fact out to me, as well as the relevant observation regarding (52).

Henk van Riemsdijk

31

Indeed, (52) is ungrammatical on the frequency of events reading but ungrammatical on the ‘number of people’ reading. That this is so is confirmed by the ungrammatical (53) because fewer cannot be interpreted as Degree Phrase denoting a (low) frequency of events. (53)

*The fewer people read fewer books, the higher the crime rate

In Dutch and German, however, the independent degree phrase reading evaluating the frequency of events is not available, hence meer/mehr in (49-51) must be interpreted as a phrasal modifier of the head noun (i.e. ‘more books’, ‘more linguists’). Still, these examples are grammatical. Clearly, computing the averages that are involved here is a far from trivial semantic issue which requires a higher degree of semantic sophistication than I am able to provide here. The intuitive idea seems fairly straightforward, though. The meaning of an example like (50) could be visualized as follows: (54)

Turning back to the Dutch example in (51), it should be observed that relative clauses, consecutive clauses and comparative clauses can be multiply headed. Indeed, the (reduced) dan-clause in (51) may be taken to be multiply headed by the two comparative phrases. This type of multiple construal of the comparative clause can be seen as a kind of manifestation of absorption. Going one step further, we might say that it is the absorptive construal that licenses multiple heads of this kind. This way of thinking would be, if I understand correctly, compatible with the proposal of late merger of degree clauses advanced in Bhatt and Pancheva (2004).

32

Some Observations on Multiple X

The impossibility of multiple comparative correlatives in English remains as mysterious as does the absence of multiple questions in Italian. Whether or not it is legitimate to suspect there to be a connection between these two cases, it does raise the question as to whether the restriction in English can be circumvented by making use of conjoined comparative correlatives. (55)

The fewer people read books, and the fewer books *(they read), the higher the crime rate (cf. ex (52))

There is a constraint on multiple wh in English: the verb must be (pseudo-) intransitive, and at least one of the two conjuncts must be an adjunct (cf. Haida and Repp, to appear). While (55) does not appear to be impossible, (56) may well be somewhat better in that it meets the same constraint. (56)

The more articles you read, and the more quickly you do so, the better your chances of getting admitted at MIT

Be that as it may, it does appear that conjunction must take place at the clausal level and cannot be phrasal. Hence it is not clear that conjoined comparative correlatives can be assimilated to conjoined wh-questions or wh/FR-adjuncts.

6. Multiple puzzles This short note raises a number of questions that deserve to be investigated in more detail. Rather than suggest possible solutions tying the various, apparently interrelated, properties together in some way, I have given priority here to outlining what the (potentially) relevant properties are, and how they are distributed in a number of languages. We first need to understand whether ‘multiple X’ is a genuine generalization or not. And if it is, we need to know more about the constructions it occurs in – e.g. does conditionality truly play a role. Third, we need to gather data on what the actual variability across languages is. Then, perhaps, we can start worrying about the best way to relate the variation to independent structural factors in the languages in question. This, in other words, was just a small beginning.

Henk van Riemsdijk

33

References Abeillé, A. and R. D. Borsley. 2008. Comparative correlatives and parameters. Lingua 118, 1139-1157. Andrews, A. D. 1985. Studies in the syntax of relative and comparative clauses. New York: Garland Pub. Beck, S. 1997. On the semantics of comparative conditionals. Linguistics and philosophy 20, 229-271. Bhatt, R. and R. Pancheva. 2004. Late merger of degree clauses. Linguistic Inquiry 35/1, 1-45. —. 2006. Conditionals. In The Blackwell Companion to Syntax, Vol. I, edited by M. Everaert and H. C. v. Riemsdijk, 638-687. Oxford: Blackwell Publishing. Boškoviü, Ž. 2004. Topicalization, focalization, lexical insertion, and scrambling. Linguistic Inquiry 35/6, 613-638. Brasoveanu, A. 2008. Comparative and equative correlatives as anaphora to differentials. Handout, University of Massachusetts, Amherst. Caponigro, I. 2002. On the semantics of indefinite free relatives. Proceedings of Console X (2001), edited by M. v. Koppen, J. Sio and M. de Vos. Leiden University. . —. 2003. Free not to ask: on the semantics of free relatives and wh-words cross-linguistically. Doctoral dissertation. University of California. Corver, N. 1993. A note on subcomparatives. Linguistic Inquiry 24/4, 773781. Dayal, V. 2006. Multiple wh-questions. In The Blackwell Companion to Syntax, Vol. III, edited by M. Everaert and H. C. v. Riemsdijk, 275326. Oxford: Blackwell Publishing. Dikken, M. d. 2005. Comparative correlatives comparatively. Linguistic Inquiry 36/4, 497-532. Haida, A. and S. Repp. (to appear). Mono-clausal question word coordinations across languages. Proceedings of NELS 39. Amherst, MA: University of Massachusetts Graduate Student Association. Hendriks, P. 1994. Multiple head comparison and infinite regress. Proceedings of the Western Conference on Linguistics (WECOL) 5, edited by J. Ashmore Nevis and V. Samiian, 117-131. Fresno, CA: Department of Linguistics, California State University. McCawley, J. D. 1988. The comparative conditional construction in English, German, and Chinese. Proceedings of BLS, 176-187.

34

Some Observations on Multiple X

Meier, C. 2001. Multihead comparatives and result clauses with “split antecedents”. In Audiatur vox sapientiae: A festschrift for Arnim von Stechow, edited by C. Féry and W. Sternefeld, 348-371. Berlin: Akademie Verlag. Müller, S. 2003. Mehrfache Vorfeldbesetzung. Deutsche Sprache 31/1, 29-62. Oda, T. 2008. Degree constructions in Japanese. University of Connecticut. Riemsdijk, H. C. v. 2006. Free relatives. In The Blackwell Companion to Syntax, Volume II, edited by M. Everaert and H. C. v. Riemsdijk, 338382. Oxford: Blackwell. Stechow, A. v. 1984. Comparing semantic theories of comparison. Journal of Semantics 3, 1-77. Stoyanova, M. 2008. Unique Focus: languages without multiple whquestions. Amsterdam: John Benjamins. Whitman, N. 2005. Category neutrality: a type logical investigation. New York: Routledge.

ON THE TENSE-LESS FUTURE IN POLISH JOANNA BàASZCZAK, PATRYCJA JABàOēSKA, DOROTA KLIMEK-JANKOWSKA AND KRZYSZTOF MIGDALSKI

Outline This study focuses on the syntax and semantics of simple and periphrastic future forms in Polish. There are three central aspects of the proposed analysis. First, it is argued that both in simple and periphrastic future constructions, it is a combination of present tense and perfective aspect which is responsible for the expression of future time reference. Second, the study presents an account of some new semantic contrasts between simple and periphrastic future forms. Finally, this paper contains a preliminary comparative analysis of periphrastic be-future forms in Polish and Slovenian.

1. Introduction It is traditionally assumed that in Polish there are two future forms: (i) a simple future form expressed by means of a perfective form of a present tense verb (cf. (1a)) and (ii) a periphrastic future form which is expressed by using the be-auxiliary bĊdzie and the so-called l-participle (or infinitive) of an imperfective verb (cf. (1b)). (1)

a.

b.

zje Ö“simple future” eat.prs.perf.3sg (| ‘He/she will eat / will have eaten.’) bêdzie jad³ Ö“periphrastic future” be.prs.perf.3sg eat.prt.impf.sg.m / jeœæ / eat.inf.impf (| ‘He/she will eat / will be eating.’)

36

On the Tense-less Future in Polish

It has been a standard assumption so far that the two forms cited above are “future tenses”. However, the question is: how much “tense” or more specifically how much “future tense” is there really in such “future forms”? The main claim of this study is that there is no “future tense form” as such in Polish (and North Slavic in general). Instead, future time reference is expressed through aspectual distinctions. The following questions will be addressed: (i) (ii) (iii) (iv) (v)

What is the status of the be-auxiliary bêdzie in the periphrastic future? What is the status of the l-participle in the periphrastic future? How can the meaning of future reference be derived in “simple” and “periphrastic” future constructions? What is the role of aspect in expressing future time reference? Are there any differences in the meaning of “simple future” and “periphrastic future” and if so, how can they be explained?

1.1. Outline of the article This article is divided into two parts. First, in sections 2.1 and 2.2 several assumptions concerning the formal syntactic and semantic differences between “simple” and “periphrastic” future constructions are presented. In the second part, these assumptions are put together into a coherent analysis of the syntax and semantics of “simple future” and “periphrastic future” forms. More precisely, section 3.1 focuses on the role of perfective aspect in deriving future time reference. Section 3.2 discusses the function of bêdzie in deriving future time reference and its interaction with the l-participle (infinitive) in the periphrastic future. Section 3.3 explains how perfective aspect and present tense form of the main lexical verb interact in expressing future time reference. Section 3.4 accounts for the different behaviour of “simple” and “periphrastic” future constructions in the contexts discussed in section 2.2. Finally, section 3.5 presents a preliminary analysis of the difference between future constructions in Slovenian and Polish.

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

37

2. Assumptions on the syntax and semantics of simple and periphrastic future forms in Polish 2.1. Syntactic difference between the “future forms” in Polish The first assumption made in this study is that both “simple future” and “periphrastic future” forms are monoclausal in the sense that they do not contain 2 TPs or 2 CPs (they introduce one eventuality). The evidence for this is based on negation facts. In Polish the position of sentential negation (understood as event negation) is below TP but above the verbal projection (see B³aszczak 2001a, 2001b and the references cited there). Using two diagnostics for sentence negation in Polish, the licensing of negative pronouns and the licensing of Genitive of Negation (cf. (2)), it can be demonstrated that the lower position of negation in (3b) does not represent sentence negation. In this respect, “future construction” in (1b) clearly differs from biclausal constructions, such as control (and also (deontic) modal) constructions; cf. the contrast between (3) and (4). (2)

a.

b.

(3)

a.

b.

Przeczyta gazetê. / *gazety read.prs.perf.3sg newspaper.ACC / *newspaper.GEN / *nic. / *nothing (| ‘He/she will read a newspaper.’) przeczyta Nie gazety read.prs.perf.3sg newspaper.GEN NEG / * gazetê / niczego / *newspaper.ACC / nothing (|‘He/she will not read a newspaper.’ / ‘He/She will not read anything.’) bêdzie czyta³ Nie NEG be.prs.perf.3sg read.prt.impf.sg.m / niczego gazety newspaper.GEN / nothing (| ‘He/she will not be reading a newspaper /anything.’) *Bêdzie czyta³ nie be.prs.perf.3sg NEG read.prt.impf.sg.m gazety / niczego. newspaper.GEN / nothing

On the Tense-less Future in Polish

38

(4)

a.

b.

Nauczyciel pozwoli³ Janowi nie teacher allowed John.DAT NEG czytaæ ¿adnej ksi¹¿ki. read.inf [no book].GEN ‘The teacher allowed John not to read any book.’ Jan mo¿ jutro czytaæ nie e John can tomorrow NEG read.inf ¿adnej ksi¹¿ki. [no book].GEN ‘John is allowed not to read any book tomorrow.’

Summarizing the above facts, we have a difference between a modal/control verb and bêdzie in terms of negation. Unlike “a modal or control verb” the complement of bêdzie cannot be preceded by negation (so if negation occurs above vP), the complement of bêdzie is something smaller than vP. This in turn means that the complement of bêdzie is smaller than the complement of a modal or a control verb. Whether this means that “bêdzie + VP” is monoclausal while a modal/control verb + complement are biclausal is an open issue (see Wurmbrandt 2001 for a general discussion).1 For the time being it does not matter whether bêdzie is part of VP or vP or some other kind of functional aspectual projection dominating the VP proper. The second issue concerns the morphological make-up of “simple future” and “periphrastic future” forms. Simple future forms are morphologically speaking perfective forms of present tense. However, it is not a trivial task to establish what is what in morphologically complex “periphrastic future” forms consisting of two parts: bêdzie and l-participle (or infinitive). What is bêdzie? There are two facts which indicate that bêdzie is a perfective form of the verb ‘be’. First, it originates from the perfective present tense paradigm of the Old Church Slavonic verb byti ‘to be’ (van Schooneveld 1951). Second, it can be argued that perfective/imperfective morphological opposition demonstrated by the majority verbs in Polish is also realized in the case of the verb ‘be’; cf. (5). (5)

1

a.

pisze write.prs.impf.3sg ‘s/he writes/is writing’

vs.

napisze write.prs.perf.3sg ‘s/he will write’

We use this term ‘biclausal’ just as a descriptive term.

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

b.

jest be.prs.impf.3sg ‘s/he is’

vs.

39

bêdzie be.prs.perf.3sg ‘s/he will be’

What is the status of the l-participle? The l-participle can be used in a variety of environments in Polish: “future constructions”; cf. (6a), “past constructions”; cf. (6b), “counterfactual conditional constructions”; cf. (6c), “subjunctive constructions”; cf. (6d). (6)

a. b. c. d.

bĊdzie jadá ‘be.prs.perf.3sg eat.prt.impf.sg.m’ (» ‘He will be eating.’ / ‘He will eat.’) jadá ‘eat.prt.impf.3sg.m’ (» ‘He ate. / ‘He was eating.’) jadáby ‘eat.prt.impf.3sg.m + COND’ (» ‘He would eat.’) Maria chce, Īeby on wiĊcej Mary wants that+SUBJ he more jadá. eat.prt.impf.3sg.m (| ‘‘Mary wants that he eat more.’ / ‘Mary wants him to eat more.’)

Hence, it seems reasonable to assume that the l-participle is a completely tenseless form.

2.2. Semantic differences between the “future forms” in Polish One of the assumptions made in the previous section was that both “simple future” (1a) and “periphrastic future” (1b) contain perfective aspect and present tense morphology. Given that language usually does not “like” redundancy, the question that arises concerns the difference in meaning/use between “simple future” (1a) and “periphrastic future” (1b). The answer to this question is not trivial. Given what we know about languages possessing more than one future marker (e.g. East Javanese), one might expect that there is (a kind of) a clear-cut semantic distinction between individual future constructions. For example, one future marker in East Javanese conveys intention while another future marker in this language seems to express prediction (see Van der Klok 2010). However, the situation in Polish seems to be more complicated because both simple and periphrastic futures are compatible with a variety of contexts. For example, both forms can express intention and prediction; cf. (7) and (8).

On the Tense-less Future in Polish

40

(7)

a.

b.

(8)

a.

b.

ObiecujĊ, Īe ci jutro promise.prs.1sg that you.dat tomorrow pomogĊ. help.prs.perf.1sg ‘I promise to help you tomorrow.’ (‘I promise that I will help you tomorrow.’) ObiecujĊ, Īe ci jutro promise.prs.1sg that you.dat tomorrow bĊdĊ pomagaá przez caáy be.prs.perf.1sg help.prt.impf.sg.m for whole dzieĔ w sprzątaniu. day in cleaning ‘I promise to help you the whole day tomorrow with cleaning.’ (‘I promise that I will help you the whole day tomorrow to do cleaning.’) Patrz na jego twarz. Zaraz siĊ look.imp.2sg at his face immediately refl rozpáacze. burst-into-tears.prs.perf.3sg ‘Look at his face. He is going to/will burst into tears right now.’ Patrz na jego twarz. Zaraz look.imp.2sg at his face immediately bĊdzie páakaá. be.prs.perf.3sg cry.prt.impf.sg.m ‘Look at his face. He is going to/will cry right now.’

In spite of the fact that the distribution of “simple future” and “periphrastic future” is not complementary, there seem to be contexts in which there is a clear contrast/difference in meaning between the two future constructions. The first difference in meaning of “simple future” and “periphrastic future” forms can be observed in ‘warning contexts’ presented in (9a) and (9b): (9)

a.

Uwaga, spadniesz! caution fall-down.prs.perf.2sg ‘‘Be careful/Watch out: You are going to fall down (otherwise)! (warning – the hearer can still do something to prevent falling)

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

b.

41

Uwaga, bêdziesz spada³! caution be.prs.perf.2sg fall-down.prt.impf.sg.m ‘Caution: you will be falling down (now).’ (‘Caution: you are now beginning to fall down.’) (announcement – the falling is prearranged)

In (9a) “simple future” expresses a strong warning and it implies that the hearer can still do something to prevent the action of falling. In other words, the action of falling in (9a) is not prearranged at the moment of speaking. By contrast, the “periphrastic future” used in a warning context in (9b) announces an action which is already settled at the moment of speaking. (9b) would most likely be used by a bungee jumping instructor talking to someone who is just about to jump. Native speakers of Polish share a strong intuition that the difference beween (9a) and (9b) is not aspectual but rather there is a difference in whether the future eventuality is settled or not at the moment of speaking. The second semantic contrast between “simple future” and “periphrastic future” forms can be observed in ‘question contexts’ presented in (10a) and (10b): (10)

a.

b.

Kto naprawi mi samochód? who repair.prs.perf.3sg me.dat car.acc ‘Who will repair my car?’ (the decision is to be made) Kto bĊdzie mi naprawiaá who be.prs.perf.3sg me.dat repair.prt.impf.sg.m samochód? car.acc ‘Who will be repairing my car?’ (the speaker has already decided that somebody will be repairing his/her car; the question is who is going to do it)

In (10a), in which “simple future” is used, we actually ask two questions: who will perform a future action and whether the future action will take place at all. It is not certain at the moment of asking whether the action of repairing a car will take place in the future or not. (10a) does not presuppose the existence of a plan of a future action. On the other hand, the question with the “periphrastic future” in (10b) contains one question: who will perform a future action. In (10b) it is certain that someone will be repairing my car but it is still unknown who it will be. This happens so because (10b) presupposes the existence of a plan of a future action.

On the Tense-less Future in Polish

42

Interestingly, this contrast is made sharper when we change these questions into negative-bias questions such as those in (10’): (10')

a.

b.

Kto mi kiedykolwiek naprawi takie Who me.dat ever repair.prs.perf.3sg such stare zardzewiale auto? OczywiĞcie, Īe nikt. old rusty car.acc Of course that nobody ‘Who will ever repair such an old rusty car? Nobody, of course.’ #Kto mi kiedykolwiek bĊdzie Who me.dat ever be.prs.perf.3sg naprawiaá takie stare zardzewiaáe auto? repair.prt.impf.sg.m such old rusty car.acc OczywiĞcie, Īe nikt. of course that nobody ‘Who will ever repair such an old rusty car? Nobody, of course.’

Given that only a "simple future" form implies that it is not certain at the moment of asking whether the action of repairing a car will take place in the future or not, it should be acceptable in a negative-bias question in which the implication is that the future event will never take place. This prediction is borne out, as shown in (10’a). On the other hand, a "periphrastic future" form implies that it is certain that someone will be repairing my car due to the existence of a plan of this future action. This means that a “periphrastic future” form should be incompatible with a negative-bias question which implies that nobody will ever repair such an old rusty car. In fact, “periphrastic future” sounds awkward when used in a negative-bias context, as shown in (10’b). The third semantic contrast between “simple future” and “periphrastic future” forms can be observed in ‘offering contexts’ (see Copley 2002, 2009) presented in (11a) and (11b): (11)

a.

JeĞli chcesz, nasza firma naprawi if want.prs.2sg our company repair.prs.perf.3sg ci samochód. you.dat car.acc ‘If you want, our company will repair your car.’

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

b.

43

#JeĞli chcesz, nasza firma bĊdzie if want.prs.2sg our company be.prs.perf.3sg ci naprawiaü samochód. you.dat repair.inf.impf car.acc ‘#If you want, our company will repair your car.’ (implausible under an episodic reading but okay under a habitual reading)

Only “simple future” forms are suitable in such contexts since offering entails that the decision as to a future action has not been made yet; cf. (11a). By contrast, “periphrastic future” is not suitable in offering contexts under an episodic interpretation because bĊdzie-constructions presuppose that the decision has been made earlier and the hearer has no say on the offered issue, which explains the implausibility of the context in (11b). The action being offered cannot be itself already prearranged at the moment of speaking because the hearer, to whom the offer is made, should still have a chance to decide whether he or she wants the offer to be realized in the future, as is the case in (11a) but not in (11b). The fourth contrast in the meaning of “simple future” and “periphrastic future” forms arises in idiomatic ‘I cannot believe that’ contexts (see Copley 2002, 2009) presented in (12a) and (12b):2 (12)

2

a.

Nie chce mi siĊ wierzyü, Īe Janek not wants me.dat refl believe.inf that John wykona tak odpowiedzialne zadanie. fulfil.prs.perf.3sg so responsible task ‘I cannot believe that John will fulfill/perform such a responsible task.’ (#‘I am amazed that John will fulfill/perform such a responsible task.’) (only literal meaning, no idiomatic meaning)

Both “simple” and “periphrastic” futures are compatible with a literal use of I cannot believe that, in which they both mean that the speaker does not believe that the action will take place at all.

On the Tense-less Future in Polish

44

b.

Nie chce mi siĊ wierzyü, Īe Janek not wants me.dat refl believe.inf that John bĊdzie wykonywaá tak be.prs.perf.3sg fulfil.prt.impf.sg.m so odpowiedzialne zadanie. responsible task ‘I cannot believe that John will be fulfilling/performing such a responsible task.’ (‘I am amazed that John will be fulfilling such a responsible task.’) (literal meaning + idiomatic meaning)

In (12a), “simple future” is incompatible with an idiomatic use of I cannot believe that meaning ‘I am amazed that’ since one can only be amazed by something which is already presupposed to be true, i.e., by something which is preplanned or prearranged and “simple future” is not compatible with contexts which express prearranged actions. On the other hand, “periphrastic future” in (12b) is compatible with an idiomatic use of I cannot believe that meaning ‘I am amazed that’ since “periphrastic future” can express prearranged actions. All these observations lead to a conclusion that (i) periphrastic future is preferably used to express future actions which are settled at the moment of speaking; (ii) simple is preferably used to express future events which are not settled at the moment of speaking.

3. Putting the pieces of the puzzle together 3.1. Perfective aspect and its role in expressing future time reference In section 2.1, it was demonstrated that both “simple” and “periphrastic” future constructions contain present tense and perfective aspect. Obviously, it is the combination of present tense morphology and perfective aspect which gives us future reference since present tense morphology alone can only refer to present time (cf. the contrast between (13a) and (13b)). In contrast, perfective present tense can only refer to future, as this form cannot be used in a context which implies that the eventuality is taking place at the moment of speaking or was taking place in the past. (cf. (14a) and (14b) respectively). Finally, perfective aspect without present tense morphology does not have any future time reference; cf. (15).

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

(13)

a.

b. (14)

a.

b.

(15)

a. b.

45

Ewa robi porządki. Eve do.prs.impf.3.sg cleaning ‘Eve is cleaning.’ ‘*Eve will be cleaning.’ Ewa robi jutro porządki. Eve do.prs.impf.3.sg tomorrow cleaning ‘Eve is cleaning tomorrow.’ *ObserwujĊ, jak Ewa zrobi Observe.prs.impf.1sg how Eve do.prs.perf.3.sg porz¹dki. cleaning ‘*I am observing how Eve will be cleaning.’ *Obserwowa³am, jak Ewa zrobi Observe.prt.impf.1sg how Eve do.prs.perf.3.sg porz¹dki. cleaning ‘*I was observing how Eve will be cleaning.’ Ewa zrobi³a porządki. Eve do.prt.perf.sg.f cleaning ‘Eve cleaned / has cleaned.’ *Ewa zrobi³a jutro porz¹dki. Eve do.prt.perf.sg.f tomorrow cleaning ‘*Eve has cleaned tomorrow.’

There exists independent evidence that the combination of present tense and perfective aspect (in periphrastic constructions) is sufficient to trigger future time reference; cf. (16). (16)

a.

Muszê myæ tê klatkê schodow¹ za must.1sg clean.impf this staircase.acc as karê. punishment ‘I have to clean this staircase as a punishment.’ (‘I am cleaning a staircase because I have to.’) present reading

On the Tense-less Future in Polish

46

b.

Muszê umyæ tê klatkê schodow¹ za must.1sg clean.perf this staircase.acc as karê. punishment ‘I have to clean this staircase as a punishment.’ (‘I will clean a staircase because I have to.’) future reading

The only formal difference between (16a) and (16b) is the aspectual form of an infinitival complement of a present tense modal verb. This difference results in a semantic contrast. (16a) means by default (when no adverbial modifiers are used) that a speaker is cleaning a staircase at the moment of speaking because it is his or her obligation. On the other hand, (16b) means that a speaker will clean a staircase after the moment of speaking because he or she has such an obligation at the moment of speaking. This points to a conclusion that perfective aspect has the function of forwardshifting of the reference time (see Condoravdi 2001).3 A relevant question that arises at this point is how future reference is derived in “simple” and “periphrastic future” forms? Is it derived in the same manner in both cases?

3.2. The function of ‘bêdzie’ in deriving future time reference and its interaction with the l-participle (infinitive) in the periphrastic future In this study, it is argued that bêdzie is a semi-lexical (semi-functional) verb. Being semi-functional bêdzie (present tense perfective BE) only forward-shifts the reference time and locates it right after the speech time. Being semi-lexical bêdzie (BE) introduces a state. Notice that in Polish bêdzie means ‘will be’. We would like to suggest that bêdzie does not introduce its own eventuality argument. The only eventuality argument in the whole periphrastic future form is introduced by a lexical complement. One may ask at this point how it is possible that bêdzie introduces a state but it does not introduce an eventuality argument. In a neo-Davidsonian paradigm the assumption is that any verbal predicate (including states) has 3

Perfective verb in the past tense also forward shifts the reference time. Perfective aspect usually introduces events into discourse. Kamp and Reyle (1993) claim that event clauses update the current reference time, and therefore give rise to sequential relation, as presented in (i): (i) My friend picked up my glove (e1), crossed the ice rink (e2) and gave it to me (e3) smiling.

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

47

an underlying Davidsonian event argument. However, Maienborn (2001, 2004) argues that two kinds of states should be distinguished. While verbs such as sit, stand, sleep refer to eventualities in the sense of Davidson (= Davidsonian states), the states denoted by such stative verbs like know, weigh, and own, as well as any combination of copula plus predicate are of a different ontological type (= Kimian states). Maienborn argues that Kimian states do not introduce a typical eventuality argument but rather they introduce a referential argument for a temporally bound property exemplification. Summarizing the above discussion, bêdzie introduces a state but essentially it does not introduce an event argument, hence the only role of perfective aspect on bêdzie is to forward-shift a reference time and to locate it after the speech time. Perfective aspect does not have any lexical access to an eventuality argument, hence it does not have any other semantic effect. Since bêdzie denotes a state BE which is durative it can only be complemented with [+durative] eventualities, i.e., states and processes. Since states and processes are compatible only with imperfective aspect, we expect bêdzie to select a [+imperfective] complement. This expectation is corroborated by the examples in (17). (17)

a.

b.

bêdzie

jad³

/jeϾ imperfective /eat.inf.impf

be.prs.perf.3g eat.prt.impf.sg.m (| ‘He/she will eat / will be eating.’) perfective *bêdzie zjad³ /zjeœæ be.prs.perf.3g eat.prt.perf.sg.m /eat.inf.perf

In Polish, bêdzie can be followed only by an imperfective lexical verb, which takes the form of either an l-participle or an infinitive4. The lparticiple (infinitive) introduces an eventuality argument. It provides the lexical content specifying the meaning of the [+durative] complement of the stative BE. The selectional properties of bêdzie are in fact expected in our account since we assume that bêdzie heads a semi-functional projection right above the lexical verbal projection. The whole periphrastic future form ends up being durative as a result of which it can be modified only by ‘for an hour’ but not by ‘in an hour’ adverbials e.g. Jan bêdzie jad³ godzinê/*w godzinê ‘John will be eating for an hour/*in an hour’. The 4

The choice of an infinitive or an l-participle as a complement in a periphrastic future form does not cause any semantic difference. The only formal difference between a periphrastic future form with an infinitive and the one with an lparticiple is that only the latter marks gender agreement on its complements.

48

On the Tense-less Future in Polish

interaction between bêdzie and the l-participle/infinitive can be understood as schematically represented in (18): (18)

bêdzie semi-functional BE (state) forward-shifting of the reference time × perfective

l-participle/infinitive lexical content of the eventuality /introducing e-argument [+durative] Â unbounded event states / processes × imperfective

3.3. Interaction of perfective aspect and present tense in expressing future time reference in “simple future” constructions in Polish “Simple future” forms obviously differ from “periphrastic future” forms in that they do not contain any be-auxiliary. Consequently, perfective aspect is marked directly on the main verb and because of this it has a direct access to the event argument introduced by the lexical verb and to the lexical content of the event denoted by the verb. In addition, perfective aspect forward-shifts the reference time of the event and it situates it after the speech time just as it happens in the case of bĊdzie in “periphrastic future” constructions. In contrast to “periphrastic future” constructions, in the case of “simple future” there is no stative BE. Operating directly on a lexical verb, the perfective aspect can develop its usual meaning. More specifically, it locates the temporal trace of an eventuality argument within the forward-shifted reference time. The role of perfective aspect in deriving future time reference in “simple future” constructions is schematically presented in (19): (19)

perfective forward-shifting of the reference time

lexical verb lexical content of the eventuality /introducing e-argument [-durative] Â bounded event Û Ü Perfective

Given that in “periphrastic future” the state denoted by BE is coextensive with the reference time and it is complemented by a [+durative] ([+unbounded]) eventuality and (i.e., the point of speech and the beginning of the reference within which the state BE + a durative eventuality is

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

49

situated may abut), we predict that the “periphrastic future” but not “simple future” should be compatible with a ‘still’ context. This prediction is fulfilled, as shown in (20a). By contrast, “simple future” is not compatible with a ‘still’ context, as shown in (20b), because either the focus is on the final boundary (and hence there has to be a gap between the utterance time, which is an instant, and the final point of an event), or the change of state is implied, hence it cannot serve as a natural (“seemingly temporally uninterrupted”) continuation of a situation which is going on at the point of utterance. By contrast in the case of the periphrastic future, the eventuality it describes can but does not have to be understood as an uninterrupted continuation of the present eventuality. (20)

a.

b.

Janek czyta gazetĊ i nadal John read.prs.impf.3sg newspaper and still  bĊdzie ją czytaá. be.prs.perf.3sg it read.prt.impf.sg.m ‘John is reading a newspaper and he will still be reading it.’ *Janek czyta gazetĊ i nadal John read.prs.impf.3sg newspaper and still ją przeczyta. it read.prs.perf.3sg ‘*John is reading a newspaper and he will still have read it.’

3.4. Towards explaining the semantic difference between “simple” and “periphrastic future” forms in Polish The intermediary conclusion reached in section 2.2 was that only “periphrastic future” but not “simple future” is compatible with contexts in which the future action is settled, predetermined at the moment of speaking. This conclusion is actually expected in the proposed account. Why? In the contexts in which “simple future” is good (i.e., warning contexts or offering contexts) there is still a possibility of intervention (the future action can be prevented or changed). In other words, there is still time between the speech time and the final (or the initial) boundary of the future event. This corresponds to the fact that “simple future” focuses either on the final or the initial boundary of an event, signalling a change of state, so there is “enough” time (i.e., there exists an interval) between the moment of speaking and the point of change of state during which possible changes can occur. Simple future forms are less compatible with contexts in which a future eventuality forms an immediate causal chain between a plan made at the moment of speaking and a future outcome. In

On the Tense-less Future in Polish

50

“periphrastic future” the beginning of the reference time within which a durative state BE + predicative content [+unbounded] is situated can be located immediately after the moment of speaking, therefore there is “not enough” time (i.e., the time interval between the moment of speaking and beginning of the state BE + predicative content [+unbounded]) to prevent or change the future event. As a result, “periphrastic future” is more compatible with contexts, which announce a settled, prearranged or predetermined event, idiomatic ‘I can’t believe’ contexts) because their semantics allows a speaker to express the fact that they form an immediate causal link between a plan made at the moment of speaking and the future outcome. In other words, only periphrastic future forms can be used in contexts in which there is seemingly no temporal gap between the moment of speaking and the beginning of a future eventuality and this makes them more suitable for the expression of the immediate causal relation between a plan made at the moment of speaking and a future eventuality which is a realization of this plan. Essentially, we do not want to say that periphrastic future forms in Polish presuppose the existence of a plan. All we want to say is that these forms are more compatible with contexts in which there exists a plan or in which a future eventuality is settled/predetermined.

3.5. The problem of Slovenian Recall the contrast concerning the incompatibility of bĊdzie and perfective complement pointed out in section 3.3 and repeated here for convenience as (21). (21)

a. b.

bĊdzie jadá be.prs.perf.3sg eat.prt.impf.sg.m (| ‘He/she will eat / will be eating.’) *bĊdzie zjadá be.prs.perf.3sg eat.prt.perf.sg.m

/ jeĞü  imperfective / eat.inf.impf / zjeĞü perfective / eat.inf.perf

In the proposed analysis, we explained this by assuming that bĊdzie is a semi-lexical verb whose lexical content is reduced just to BE. By virtue of denoting a STATE it is only compatible with [+unbounded] predicates. This selectional property of bĊdzie also follows from its low position in the syntactic tree. On the basis of the evidence discussed in section 1.1 (recall the negation facts) we assumed that bĊdzie heads a semi-functional projection immediately above the VP proper and as such it can select [+unbounded] complement. Given this we would expect that in other

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

51

(Slavic) languages which have an element corresponding to the Polish bĊdzie the following should be true: (22)

a. b.

bĊdzie + imperfective bĊdzie + perfective



However, there is one language which does not fulfill this expectation, namely Slovenian. In Slovenian it is possible to use a perfective verb in combination with the “future auxiliary”, which directly resembles the Polish bĊdzie; cf. (23). (23)

a. b.

bom pohavalil be.prs.1sg praise.prt.perf.sg.m ‘I shall praise.’ bom + perfective



(contra (22b))

There are reasons to believe that the Slovenian equivalent of the Polish bĊdzie is not exactly the same element. According to Whaley (2000), the Slovene bo + l-participle stems from the Old Slavic Future Perfect; in contrast, the infinitival be-future (like that one in Polish or Russian) cannot be linked to the Common Slavic future perfect. It thus followed a different path of development. The participial future is rarely attested in the earliest Polish texts, which might suggest that we are dealing here with an innovative construction. Next to these diachronic facts, there is syntactic evidence related to the position of negation in Polish and Slovenian which indicates that the the Slovenian equivalent of the Polish bĊdzie is different. At first glance it seems that there is no difference between Polish and Slovenian as negation in both cases precedes the BEauxiliary; cf. (24). However, it is a standard assumption in Slavic linguistics that there is a difference in the position of negation (NegP) in the syntactic tree between Polish and Slovenian (see Rivero 1991, Borsley and Rivero 1994); cf. (25). Unlike bĊdzie in Polish, which is a VP-related semi-functional element (“low auxiliary”), bo in Slovenian is a TP-related functional element (“higher auxiliary”). The future auxiliary in Slovenian is a second position clitic (see Franks and Holloway King 2000, Migdalski 2010). (24)

a.

Janez ne bo pisal. Slovenian Janez NEG BE-aux write.prt.impf.sg.m ‘Janez will not write.’ (‘Janez will not be writing.’)

On the Tense-less Future in Polish

52

b. (25)

a. b.

Jan nie bĊdzie pisaá. Jan NEG BE-aux write.prt.impf.3sg.m ‘Jan will not write.’ (‘Jan will not be writing.’) NegP > TP > VP TP > NegP > VP

Polish Slovenian Polish

The question which arises now is exactly what kind of element the Slovenian bo is. We assume that unlike the Polish bĊdzie, the Slovenian bo is a functional element. Just like other tense auxiliaries, bo in Slovenian has just a temporal function (in this case that of locating the reference time after speech time). The whole aspectual distinctions/interpretational differences in “periphrastic future” constructions in Slovenian are due to the aspectual marking on l-participle; cf. (26) (due to Franc Marušiü (p.c.)). (26)

a.

Pismo bom pisal letter be.aux.prs.3sg write.prt.imp.sg.m / *v 3 ure. /*in 3 hours ‘I’ll write the letter for three hours.’

3 ure 3 hours

b.

Pismo bom napisal letter be.aux.prs.3sg write.prt.perf.sg.m / *3 ure. / *3 hours ‘I’ll write the letter in three hours.’

v 3 ure in 3 hours

Being purely functional, bo does not have the lexical content of the stative BE. Therefore, it does not need to take a specific complement: both [+unbounded] and [bounded] complements are allowed. Moreover, the Slovenian BE-aux, which lacks the lexical content, might co-occur with another BE (the lexical BE spelling out the lower part of the tree). This is completely unacceptable in Polish; cf. (27) (due to Frank Marušiü (p.c.)) and (28). (27)

a.

bom be.aux.1.sg ‘I will be.’

bil be.prt.sg.m

Slovenian

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

b.

(28)

a. b.

*bĊdĊ byá be.aux.1.sg be.prt.sg.m byü be.inf (intended: ‘I will be’) bo BE bĊdzie BE

/

* bĊdĊ be.aux.1.sg

53

Polish

Slovenian  Polish

Another relevant observation is that in Polish it is possible to complement the BE-aux in “periphrastic future” by an imperfective infinitive (instead of the l-participle). This seems to be even the older form (according to Whaley 2000). This makes bĊdzie similar to the so-called ‘phase’ verbs like, for instance, ‘to begin’, ‘to finish’, etc. which also take imperfective infinitival complement in Polish; cf. (29). Such ‘phase’ verbs are also sort of semi-functional (semi-lexical) elements whose function is to indicate the beginning or the end of an event (see also Veselovska 1995). (29)

a. b.

bĊdzie pisaü be.aux.3.sg write.inf.impf ‘(s)he will write’ ((s)he will be writing’) zacznie pisaü begin.3.sg write.inf.impf ‘(s)he will begin to write’

The difference between future constructions in Slovenian and Polish is schematically summarised in (30). (30)

a.

Polish “periphrastic future”

On the Tense-less Future in Polish

54

“simple future”

b.

Slovenian “periphrastic future”

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

55

4. Final conclusion Coming back to the initial question: how much “tense” or more specifically how much “future tense” is there really in “future forms” in Polish?, the answer seems to be as follows: even though the Polish “future forms” do not contain any TP-related (temporal / future) auxiliary, they are not completely tense-less since both the “periphrastic” and the “simple” future consist of a present tense element (the semi-lexical/semi-functional bĊdzie in the former case and a lexical verb in the latter case) which in combination with perfective aspect gives rise to future time reference.

56

On the Tense-less Future in Polish

References Báaszczak, J. 2001a. Investigation into the Interaction between the Indefinites and Negation. [=Studia grammatica 51]. Berlin: AkademieVerlag. —. 2001b. Covert Movement and the Genitive of Negation in Polish. [=Linguistics in Potsdam 15]. Potsdam: Universität Potsdam. Borsley, R. D. and M.-L. Rivero 1994. Clitic Auxiliaries and Incorporation in Polish. Natural Language and Linguistic Theory 12, 373-422. Condoravdi, C. 2001. Temporal interpretation for modals. Modals for the present and modals for the past. In Stanford Papers on Semantics, edited by D. Beaver et al., 59-87. Stanford: CSLI. Copley, B. 2002. The Semantics of the Future. Doctoral dissertation. MIT. To appear in Routledge’s Outstanding Dissertations in Linguistics series. —. 2009. The Semantics of the Future. New York: Routledge. Franks, S. and T. Holloway King. 2000. A Handbook of Slavic Clitics. New York: Oxford University Press. Kamp, H. and U. Reyle. 1993. From Discourse to Logic. Dordrecht: Kluwer. Klein, W. 1994. Time in Language. London, New York: Routledge. Klok, J. Van der. 2010. On the Semantics of Future Markers in East Javanese. Paper presented at AFLA-17 (Austronesian Formal Linguistics Association XVII), Stony Brook University, May 7-9, 2010. Maienborn, C. 2001. On the Position and Interpretation of Locative Modifiers. Natural Language Semantics 9, 191-240. —. 2004. On the Limits of the Davidsonian Approach: The Case of Copula Sentences. To appear DVWDUJHWDUWLFOHin Theoretical Linguistics. Mezhevich, I. 2006. Featuring Russian Tense: A Feature-Theoretic Account of the Russian Tense System. Doctoral dissertation. University of Calgary, Alberta. —. 2008. A Feature-Theoretic Account of Tense and Aspect in Russian. Natural Language and Linguistic Theory 26, 359-401. Migdalski, K. 2010. On the Loss of Tense and Verb-Adjacent Clitics in Slavic. Paper presented at the DiGS XII, Cambridge University, July 14-16, 2010. Rivero, M.-L. 1991. Long Head Movement and Negation: Serbo-Croatian vs. Slovak and Czech. The Linguistic Review 8, 319-351.

Báaszczak, JabáoĔska, Klimek-Jankowska and Migdalski

57

Schooneveld, van C. H. 1951. The Aspect System of the Old Church Slavonic and Old Russian verbum finitum byti. Word 7/2, 96-103. Veselovska, L. 1995. Phrasal Movement and X0-Morphology. Word Order Parallels in Czech and English Nominal and Verbal Projections. Doctoral dissertation. Palacky University Olomouc, Czech Republic. Whaley, M. L. 2000. The Evolution of the Slavic ‘BE(COME)’-Type Compound Future. Doctoral dissertation. The Ohio State University. Wurmbrand, S. 2001. Infinitives: Restructuring and clause structure. Berlin, New York: Mouton de Gruyter.

DEVELOPMENT OF TAM CATEGORIES IN TURKIC: FEELING FREE IN THE DETERMINISTIC SYSTEM PAVEL GRASHCHENKOV

Outline Turkic languages have about two dozens of verbs that can be used either as lexical verbs or as auxiliaries expressing TAM categories. The paper discusses the process of grammatical function development in such verbs. I will consider details of syntactic reanalysis that underlie novel TAM items appearance and take a closer look on the grammatical semantics of novel auxiliaries. I will also concern a question of whether these changes were structural or semantic by nature.

1. Introduction I will use the term “grammaticalization” without any ideological background: “grammaticalization” will mean just the process of emergence of new grammatical elements. In this paper grammaticalization is treated as a syntactic process but the only reason for its syntactic nature is that I consider evolution of a syntactic construction. I will not touch upon the problem of whether a grammaticalization is a complex phenomenon including phonological, morphological, syntactic and semantic changes or whether there are different types of changes and each of them can be called “grammaticalization”, see Newmeyer (2001) for the discussion. There are from twenty to thirty verbs in Modern Turkic that can be used either as independent lexical items or as TAM auxiliaries. As regular verbs they do not differ from another verbal lexemes. Being auxiliaries, they have fixed position just after the meaning verb, share the stress with it and express significantly different meaning, see Johanson (1995), Erdal (2004) a.o. Sometimes such items give rise to verbal morphemes with TAM functions. I will propose the detailed scenario for the shift from

Pavel Grashchenkov

59

lexical verb into TAM markers and discuss some aspects of new grammatical meanings with respect to the source verbs. Then, there are two main trends in historical syntactic researches. According to the first one, see Heine et al. (1991), Haspelmath (1998), Traugott (2003), Bybee (2003), Hopper & Traugott (2003) a.o., grammaticalization is driven mostly by semantic bleaching that is the reason for its vague, gradual nature. As stated in Haspelmath (1999):1062 “Thus, in my model it is not so much that semantic bleaching and phonological reduction go hand in hand, but semantic generalization is in a sense the cause of the other processes of grammaticalization.” Under the structural treatment of grammaticalization, started in Lightfoot (1979), the driving force of grammatical change is restructuring of syntactically ambiguous constructions that leads to changes in the grammatical system, see Harris & Campbell (1995), Roberts & Roussou (2003) a.o. This perspective presupposes that changes are discrete since they are dependent on concrete syntactic constructions and vocabulary items on the one hand and give rise to the pre-defined grammatical elements on the other. I will mostly adhere the structural scenario here but will however compare the two approaches discussing development of Turkic auxiliaries.

2. Auxiliary Verb Construction in Turkic 2.1. Function of converbs The combination of a converb and a finite form in modern Turkic is regularly used to introduce coordinated or multiple events. All Turkic verbs freely participate in such constructions. In the examples below we see two events both of which can be either simultaneous or follow one after another: (1)

Tubalar a. Wasja uxta-p tur-dy Vasya sleep-conv stay-pst i. ‘Vasya stood when sleeping.’ ii. ‘Vasya slept and (then) stood.’ b. Wasja uxta-p kajla-dy Vasya sleep-conv sing-PST i. ‘Vasya sang when sleeping.’ ii. ‘Vasya slept and (then) sang.’

Shluinskij (2006b)

60

Development of TAM Categories in Turkic

2.2. Auxiliary Verb Constructions Turkic auxiliary verb (AV hereafter) constructions have the following properties: (i) they are created as a sequence of two or more verbs, the lexical verb coming first and the AV follows it; (ii) such verb chain can not be split; (iii) it has the common phrasal stress; (iv) the meaning of such construction is defined by the first, lexical item; (vi) only verbs from a very limited group can serve as a AV. In the example below verbs ‘stay’ and ‘see’ create AV constructions that introduce the action duration and possibility (correspondingly): (2)

Tubalar a. Wasja uxta-p tur-dy Vasya sleep-conv stay- PST i. ‘Vasya stood when sleeping.’ ii. ‘Vasya slept and (then) stood.’ iii. ‘Vasya was sleeping.’ b. ?þaška tül-üp kör-di cup fall-conv see-PST ‘The cup could fall.’

Shluinskij (2006b)

In (2.a), that repeats (1.a), we provide one more reading available for it apart of (i) and (ii) in (1.a). Under the reading (iii), the verb tur-, ‘stand’, loses its lexical meaning and has a function of tense / aspect marking. In (2.b) the verb kör-, ‘see’, denotes an epistemic probability for an action with this subject. (2.b) has an inanimate subject used with the verb ‘see’ that also supports the idea of loss of initial lexical meaning when this verb used as an auxiliary. Another evidence for the fact that AVs lose their meanings and serve just as TAM markers can be seen in the example below where the AV expresses negation for the event introduced by the lexical verb (olu-, ‘die’)1, see also past tense reference on (2) that is attributed to the meaning of the lexical verbs:

1

As was noted by an anonymous reviewer, the AV “carries it (=negation, PG), not expresses it (which may be weaker evidence)”. But Turkic indeed have a negative counterpart of the -p converb, but in the construction under issue only -p converbs can be used, so negation should be expressed by the serial, not lexical verbs.

Pavel Grashchenkov

(3)

61

Azerbayjani Azerbayjani (1971) kitab jandyrylyrsa, šairin ΩsΩrlΩri mΩhv olu-p book if.burned poet.GEN works die-CONV ket-mi-r go-NEG-3 ‘Even if a book is burned, the poet’s work doesn’t die.’

2.3. Sources and grammatical outcomes of AVs The earliest data observed in Old Turkic show that even a thousand years ago2 AVs already existed in the language and had more or less the same shapes and meanings as they have nowadays. In spite of this fact, we can argue that the process of AV grammaticalization can be clearly tracked down, see for instance, Johanson (1995) a.o. In brief, the pathways of grammaticalization can be stated as following. Embedded clauses3 always share at least their subjects with the main clause: (4)

Kazakh V1 [Nurlan koldin maȘynda [suga Ιara-p ] Nurlan lake near water look-Conv ‘Looking at the water, Nurlan stayed near the lake.’

V2 tΡr-dy] stay-Pst

Such configurations are not easy to parse, see Hawkins (1994) a.o. To recognize that ‘near the lake’ is attributed to ‘stay’, the hearer should wait while the embedded clause finishes. The clause headed by V1 should be parsed first and only then the hearer can attach ‘near the lake’ to its host. The more material we have in such sentences, the more significantly increases parsing difficulty. The early closure becomes actual, that calls to syntactic reinterpretation: (5)

2 3

Kazakh V1 [Nurlan [koldin maȘynda suga Ιara-p ] Nurlan lake near water look-Conv ‘Looking at the water near the lake, Nurlan stayed.’

V2 tΡr-dy] stay-Pst

Old Turkic Documents can be attributed to the period of VII – XIII cc. V1 clause, since Turkic is strictly head-final.

Development of TAM Categories in Turkic

62

In (5) the dependent stuff is already attached to V1. The next step is whole sentence reanalysis. The subject and all other material become attributed not to the V2 item but to the whole V1+V2 complex: (6)

Kazakh V1 [Nurlan koldin maȘynda suga Ιara-p Nurlan lake near water look-Conv ‘‘Nurlan was looking at the water near the lake.’

V2 tΡr-dy] stay-Pst

It is crucial for the latter switch that V1+V2 complex was ambiguous with another monoclausal structure. In Old Turkic there was an instance of a grammaticalized verb,4 är-, ‘to be’ that could be preceded by different verb forms: preterit, -mIš-nominalizations, -madOk, -gAy etc, and had different grammatical functions depending on this, Erdal (2004). This verb can also follow -p converbs and “appears to convey post-terminal meaning”, Erdal (2004:252): (7)

(Old) Uygur Erdal (2004:401) olar burxan kutïn bulu-p ärtmištä basa anta… they Budhahood become-Conv be after ‘after they will have reached Budhahood, ...’

Thus, the combination “converbial V1 + main V2” had an equivalent among analytical verbal forms and could in principle be (re-)interpreted as a “verb + auxiliary”. According to Roberts & Roussou (2003), this is a stimulus that helps to reinterpret current structure into the novel one and thus supports V2 grammaticalization to auxiliary. After the Old Turkic period, Modern Turkic languages proceeded further with grammaticalization and turned some AVs into affixes, see novel Perfective, Present and Praesens Historicum markers below (note that Kyrgyz and Turkmen had not much influence on each other and developed the Present marker from the verb ‘lie’ independently): (8)

a. b.

4

pary-b-ys go-CONV-PFV bara-žata-myn come-pres-1.sg

-is (Perf) Å us- (‘send’)

Khakass Khakass (1953) -žata- (Pres) Å žata- (‘lie’) Kyrgyz (http://Kyrgyz.lugovsa.net)

Such forms with auxiliary-only verbs (like edi) are even more productive in nowadays Turkic, cf.: keli-p edi, come-Conv AUX, ‘to have come’.

Pavel Grashchenkov

c. d.

63

gel-jar-yn -jat-(Pres)Å jor- (‘move’) Turkmen come-pres-1.sg SIGTY-M (1988) bere-p.tir-men -tir- (PresHist) Å tur- (‘stay) Tyva come-pres.hist-1.sg Tyva (1961)

This is the way how embedded clause construction gave rise to novel grammatical categories in Turkic: (9)

Lexical items Æ 1st step Æ Verbs with grammatical functions Æ 2nd step Æ TAM markers

Thus, the emergence of new TAM categories in Turkic started in the Proto-Turkic period as lexical to auxiliary switch and proceeds nowadays as auxiliary to morphology switch. The rest of the paper considers in detail the first step in this process and discuss particular conditions under which lexical verbs in Turkic acquired auxiliary functions. I will concentrate on which novel grammatical meanings have been developed and how they correspond to initial lexical meanings of verbs.

2.4. AVs as functional head fillers Taking into account the grammaticalization process described above, I propose the following scenario. A new role of auxiliary was possible due to the fact that the clause structure had functional head positions that had not been filled by overt material and served a potential landing site for V2: (10)

Grammaticalized AVs end up in functional head positions

Development of TAM Categories in Turkic

64

So, in this section I proposed how the novel TAM markers arose in Turkic. In brief, syntactic reanalysis is a mechanism that eroded clause boundaries and let regular verbs occupy functional head positions. This was accompanied by the reanalysis of the main clause internal material as attached to the embedded verb and supported by the existing models of lexical plus auxiliary combinations. In what follows I will discuss how and which new TAM meanings appeared. I argue that this scenario is much more plausible as opposed to the “semantic bleaching” approach. If the semantic bleaching was prerequisite to reanalysis, we would expect to find it not only in ‘V1 + V2’ configurations but in other instances where AV are used. However, this is not the case: in the chaining example below the verb ‘stay’ clearly lacks grammatical meaning: (11)

Kazakh V1 V2 Nurlan suga Ιara-p koldin maȘynda tΡr-dy Nurlan water look-Conv lake near stay-Pst i. ‘Nurlan stood near the lake and looked at the water.’ *ii. ‘Nurlan was looking at the water near the lake.’

3. Meaning and functions of novel grammatical items What we observe in the result of AV appearance is: (i) the array of lexical meanings participating is AV development; (ii) the array of grammatical functions served by AVs in Modern Turkic. I am going to implement a kind of set theoretic analysis of the (elements of the) two arrays. What are expectations concerning the volume of the arrays and the relations between elements in the arrays? With respect to the volume, an array can represent an open or closed set of elements. As for relations among elements, every element from the first array can correspond to a single or multiple elements in the second array. This is exemplified in the figure below: (12)

Array 1 (Lexical Meaning) A, B Æ CÆ DÆ …

Array 2 (Grammatical Function) Å A’ Å C’ Å D’, E’ …

Pavel Grashchenkov

65

In our case the Array 1 will be represented by the set of lexical meanings and the Array 2 – by the set of grammatical functions. As for the volume, both grammatical and lexical arrays can potentially constitute either open or closed sets of meanings. Then, many-to-one relations between the two arrays will create synonymous AVs, one-to-one relations will create non-synonymous AVs. Below I will discuss what shapes took the interplay between the lexical and the grammatical arrays.

3.1. Which verbs become auxiliaries? The structural approach to grammaticalization proposed in 2. allows us to make the following generalization. As far as AVs arose from the main clause heads, we can argue that the best candidates for them were matrix predicates and predicates, denoting background actions. Indeed, both types of items are met very often among Turkic AV. Thus, the first couple of auxiliaries below exemplifies matrix AVs and another couple – regular background action predicates. (13)

(14)

(15)

(16)

Tubalar Shluinskij (2006b) þaška tülü-p kör-di cup fall-Conv see-Pst ‘The cup could (~tried to) fall.’ Kyrgyz balam oku-p bil-di child.my read-Conv know-Pst ‘My child can read.’ Kyrgyz Kurmanbek süzü-p žat-at Kurmanbek swim-Conv lie-Prs ‘Kurmanbek is swimming.’ Chuvash xer aþa-n piþe šyš’-sa tă-þ-Ɵ girl child-Gen face swell-Conv stay-Pfv-3 ‘The girl’s face was swelling.’

However, the structural approach says nothing about how a new grammatical meaning depends on the lexical semantics of the source item. Having in mind just structural pattern one cannot predict, why, for instance the first couple of examples above have modal meaning and the second one have durative meaning.

66

Development of TAM Categories in Turkic

Considering semantics of source items, one can find regular correspondences among it and the novel grammatical categories: verbs of position and non-directional movement grammaticalize into durative / progressive markers, verbs of directional movement – into resultative markers, perception verbs into modals and so on. Under the semantic or functional treatment, we can speculate that verbs like ‘know’ often have additional meaning ‘know how to do something’ and the verbs like ‘stay’ are often used not to describe the exact position but to underlie the long duration of another process (cf. Piles of rubble where houses stood and smoke rising, not from bombs but from cracks in the ground feeding oxygen to the subterranean coal fire5) etc.6

3.2. (Un)restricted meaning and functions There is a general agreement on the fact that the major lexical classes constitute open-type sets of meanings. If so, from a purely semantic point of view we would expect that no semantic restrictions are imposed on grammaticalization. On the contrary, the structural approach prescribes a limited list of meanings in accordance to the fact that the amount of (grammatical features of) functional heads is restricted. Thus, if we address the list of meanings described by the AV from the point of view of the set theory, it would create an open set from the semantic perspective and a closed set according to the structural one. However, meanings like ‘to do something staying / coming somewhere / seeing something’ are not attested. In spite of the fact that the verbs ‘stay’, ‘come’, ‘see’ are AVs, their meanings changed into those similar to English progressive (‘stay’), perfective (‘come’) and the expression of attempts (‘see’). But what kind of semantics do we observe in the AV domain? We can detect the following most regular grammatical meanings among Turkic AVs: duration and perfectivity (aspectual), attemptive and possibillitive (modal), and applicativity. Let me briefly exemplify them all.

5

http://www.offroaders.com/album/centralia/Fire-in-the-hole_Kristie-Betts.htm As the anonymous reviewer pointed it: “… the verb that grammaticalizes into an ability modal head is 'know' and the verb that grammaticalizes into a durative aspect head is a posture verb (e.g. 'stand'). On the ForS [=structural – P.G.] approach, is there any reason that it is exactly 'know' that becomes an ability modal head and a posture verb that becomes a durative aspect head?”

6

Pavel Grashchenkov

67

One type of aspectual meaning regularly expressed by AVs deals with action duration. Lexemes used in this function are: position verbs ‘stay’, ‘lie’, ‘sit’ and the verb of undirected movement, bar- ‘move’: (17)

(18)

Kazakh Medvedev kazakša sΫyle-p otyr Medvedev in.Kazakh speak-CONV sit.3.SG ‘Medvedev is speaking Kazakh.’ Chuvash xer aþa-n piþe šyš’-sa girl child-GEN face swell-CONV ‘The girl’s face was swelling.’

tă-þ-Ɵ stay-PFV-3

Different instances of perfective meaning are provided by verbs of arrival / departure like ‘go’, ‘leave’, ‘come’, ‘remain’ and others or verbs of change of possession / location like ‘put’, ‘send’, etc: (19)

(20)

(21)

Kyrgyz Žer basyp al-uu kurþu-gan muktaždyk-tan soil push take-NMN surround-PFCT need-ABL keli-p þyk-ty. come-CONV go-past ‘The seizure of lands happened (~arrived) due to constant need.’ Kazakh avtomobil' žoldary basƷarmasy-nyƾ bastyƫy Abay, auto ways department-gen head Abay, AyagΫz audan-dar-yna kelɿ-p ket-tɿ. Ayagԧz region-PL-DAT come-CONV leave-PAST ‘The chief of the road department (has) visited Abay and Ayagoz regions.’ Tyva Tyva (1961) Aþa-m zavod-ka ürgülþü ažylda-p kel-gen Father-1.SG factory-DAT all.time work-CONV come-PFCT ‘My father has always been working at the factory.’

Verbs ‘take’ and ‘give’ are used across Turkic in the applicative function. ‘Take’ introduces an action benefactive for the subject and ‘give’ denotes benefactivity for somebody else’s sake:

68

(22)

(23)

Development of TAM Categories in Turkic

Kyrgyz Ežem žigit-ke bijle-p sister guy-DAT dance-CONV ‘The sister danced for (some) guy.’

ber-di gave-PST

Kyrgyz Kurmanbek koj soju-p al-dy Kurmanbek ram cut-CONV take-PST ‘Kurmanbek cut (=slaughtered) a ram for himself.’

Two more cases are modal meanings of ability and attempt. Ability is expressed by the verb ‘know’: (24)

Kyrgyz bala-m oku-p child-1.sg read-CONV ‘My child can read.’

bil-di know-PAST

Attempts to do something are regularly introduced in Turkic by the AV ‘see’ and sometimes – by ‘show’. Below we find the situation in which the causer tried to realize some action (but whether she succeeded is unclear): (25)

Tubalar Maša þoþko-ny þeri-p Masha piglet-ACC chase-CONV ‘Masha tried to chase the piglet away.’

Shluinskij (2006b) kör-di see-PST

Thus, Turkic AVs grammaticalize only those meanings that are regularly found in the languages of the world, namely aspectual (perfective and durative), modal (attemptive and possibillitive), and applicative categories. No meanings like ‘to do something staying / coming somewhere / seeing something’ arose in Turkic AV in spite of the fact that these very lexical items were used in AV grammaticalization. The novel grammatical meanings attested among Turkic AVs are perfectly consistent with the functional heads semantics, see, for instance Cinque (1999, 2007); Cinque & Rizzi (2009); Pylkkanen (2002) a.o. So, it seems that structural approach is more accurate here since any lexical item converted to AV acquires specific grammatical meaning that has been previously reported to be a part of UG.

Pavel Grashchenkov

69

3.3. (In)finite Process of Grammaticalization In this section I consider the question of whether the grammaticalization process stopped in Turkic AV or it still proceeds resulting in new AVs with new meanings. As mentioned in Marcel Erdal’s description of Old Turkic, see Erdal’s (2004:245): “When converbs are used as first elements in analytical constructions, the products always express actionality, intention, ability or version.” All these functions are just those we observed in the previous section for the AVs semantics in Modern Turkic. Moreover, the list of auxiliary remains close to the original during the last thousand years and we find the very same set of grammaticalized verbs (with slight deviations) in different Modern Turkic, cf. (26.a) for Tyva and (26.b) for Kyrgyz below: (26)

a.

al-, bar-, ber-, düš-, kal-, kag-, kör-, kel-, olur-, tur-, þit-, …

Tyva (1961)

b.

al-, bar-, ber-, tuš-, kal-, koj-, kör-, kel-, oltur-, sal-, tur-, žiber-, …

Kyrgyz (1985)7

What we see thus is that neither new grammatical meanings arose during the last thousand years, nor new lexical items developed grammatical function. It is a bit strange taking into account that the source construction for the AV did not become less productive – the use of converbs is still the main tool for expressing event coordination in Turkic. If grammaticalization is driven mostly by semantics, it is strange that no new AV arise when the source construction itself did not loose productivity. This situation can be better described under the structural approach. We can suppose that some grammatical “lacunas”, namely, the absence of an overt marker for some feature supports the changes. So, one can argue that grammaticalization should stop when the features matrix is “satisfied” to some extent.8

7

Some items are not the same in these two languages, as was pointed me by the anonymous reviewer. This is true, but if we consider the broader list of Turkic, we will see that some items misses sporadically rather than appear sporadically (this is to say, a verb is present in 90% of Turkic languages and is absent in 10% of them). This argues for the fact that some languages have lost this of that AV rather than develop any from the Old Turkic era.

70

Development of TAM Categories in Turkic

To summarize: no new items or grammatical meanings arose in Turkic AV that is more consistent with the structural treatment of grammaticalization and is a bit hard to describe in purely semantic terms.

3.4. Synonymous AVs In terms of set theory, synonymy is a situation when more than one element from one set (lexicon in our case) corresponds to a single element in another set (the domain of grammatical meaning). Indeed, as we saw in section 5, many lexical items seem to provide the same grammatical meaning. Perfective and durative aspectual domains are the best examples here. But how can we make sure that the AV with these items are synonymous? Grammatical descriptions are usually quite vague and address such AV meanings as “long lasted actions”, “actions in progress” etc. To show that we really observe synonymy among AVs, let me consider verbs with the durative functions in Kyrgyz and Kazakh. In both languages (as in the most other Turkic) the durative meaning is provided by verbs ‘sit’, ‘stay’, ‘lie’ and ‘move’. Below I present results of a quick corpus study of such verb on the text arrays of 700 thousand words for Kyrgyz and 1.7 million words for Kazakh. After creating these text collections I made a search of the sequences of two adjacent converbs followed by a auxiliary verb. The question was whether we observe sequences of two verbs from the durative domain (for instance, ‘sit’ and ‘stay’, ‘sit’ and ‘lie’, ‘sit’ and ‘move’ etc.) If we do, this would doubt the synonymy of these items in AV function. The results are present in the table below. As we can see, cooccurrence of different durative AVs are extremely rare that can argue in favor of identity of their grammatical meaning and function.

8

I’m not aware of any research on how many grammatical categories can a language possess. As was noted be an anonymous reviewer, “people working in the “cartographic" program argue for tens if not hundreds of distinct grammatical categories.” Turkic has about two or three dozens of AVs, so in sum with regular, morphological categories, “tens” level is satisfied indeed.

Pavel Grashchenkov

(27) Durative AVs žat ‘lie’, žür ‘move’, tur ‘stay’, ΫltΟrΟp ‘sit’ otyr ‘sit’, žür ‘move’, tur ‘stay’

71

first AV 183

second AV 148

in total 331

cooccur 1

Kyrgyz

167

189

356

6

Kazakh

Thus, the grammatical synonymy is what we really find among Turkic AVs.9 If then semantic changes are a driving force for grammaticalization, it is not quite clear why semantic simplification of verbs ‘sit’, ‘stay’, ‘lie’ and ‘move’ results in one and the same grammatical meaning. It would be much more likely from the functional views that each verb creates its own semantic outcome, no matter how “abstract” or “concrete” meaning such an outcome must have. From the structural point of view we can explain the fact that different lexical items have turned into the same grammatical element: the real driving force for all these verbs is a gap in a feature matrix. The feature {durative} of the aspectual functional head “attracts” verbs like ‘sit’, etc. to be an overt marker of it. If we look again at the table above, it is also hard to explain from semantic or functional perspectives, why the two verbs (even with the same grammatical function) cannot co-occur in spite of their relative frequency in AV constructions. At the same time, the lack of co-occurrence of the items with the same grammatical functions follows quite naturally from the structural approach:

9

As was noted by an anonymous reviewer, “even if the verbs are not synonymous we probably don't have to expect them to cooccur. Morphemes that realize heads from the same domain (say, mood) but not the same values sometimes can't cooccur (e.g. *John must can go), and even when they can, it may not happen very often in actual use, and relying on grammaticality judgements may be safer.” I surely agree with it, moreover, some time ago I have been started to research the co-occurrence possibilities of AV both on text corpora and native speaker judgments. It appears that constraints on co-occurrence are even stronger: not only for instance two durative or two modal AVs are not allowed with the same lexical verb, but even durative and some other aspect serials are not ok, etc. AV pairs are not very often, but not totally excluded. One of the most interesting phenomena here is that when more than one serial allowed in a clause, such items seem to form a hierarchy, see Bridges (2008) for similar research.

Development of TAM Categories in Turkic

72

all the verbs ‘sit’, ‘stay’, ‘lie’ and ‘move’ occupies the same position and thus are mutually excluded. To summarize: AV synonymy is a problem for the semantic or functional approach, since it is not quite clear how and why the semantic drift move different verbs in the same direction. Moreover, synonymous AVs are in complementary distribution that can be interpreted as a result of the concurrence for the same position in the structure.

4. Conclusion The process of grammaticalization lexical verbs in Turkic into novel TAM markers acted as following: the embedded clauses were reanalysed as auxiliary constructions and then some auxiliaries turned to morphemes. I tried to show that it was syntactic reanalysis and not the semantic bleaching that gave rise to multiple auxiliary items in Modern Turkic. As for the structural vs semantic perspectives on grammaticalization, we found that the semantic approach can easily accommodate the fact that there are good correlations between the semantic type of a verb and a grammatical meaning expressed by this verb as auxiliary. The structural view on the semantic changes can explain the following facts: (i)

many AVs are either matrix predicates or verbs denoting background actions; (ii) AVs have been derived as a closed set of grammatical meanings that corresponds to grammatical meanings attested in the language of the world; (iii) grammaticalization stopped at some moment in spite of the productivity of converb plus finite verb construction, since the system of features got “satisfied”; (iv) synonymy is widespread among AVs, since more than one verb may target the same functional head. The claim is that in case of AV reanalysis “spoils the program” of normal functioning of lexical items. They thus have little free choice and their future is quite determined.

Pavel Grashchenkov

73

Acknowledgements I’m very grateful for my good friend in Kyrgyzstan, Fatima Sulejmanovna, for her long and careful work on the Kyrgyz data. Many theoretical topics of the paper was discussed with Sergei Tatevosov, I would thank him and Ekaterina Lyutikova and Vita Markman for their help. I’m also very grateful to Andrey Shluinskij, who supplied me with many Turkic data carefully collected and analyzed by him during his field studies. Comments and corrections of the anonymous reviewers were also extremely useful and helped me to significantly reconsider the paper. All faults are mine.

References Aikhenvald, A. Y. 2006. Serial verb constructions in typological perspective. In Serial Verb Constructions: A Cross-Linguistic Typology, edited by A. Y. Aikhenvald and R. M. W. Dixon, 1–68. Oxford: Oxford University Press. Anderson, G. 2005. Auxiliary Verb Constructions. Oxford: Oxford University Press. Baker, M. 1985. The Mirror Principle and morphosyntactic explanation. Linguistic Inquiry 16, 373–415. Baskakova, N. A. 1953. Khakassko-russskij slovar' s grammaticheskim ocherkom. Moskva. Bridges, M. 2008. Auxiliary Verbs in Uyghur. Paper, submitted to the graduate degree program in Linguistics and the Faculty of the Graduate School of the University of Kansas in partial fulfillment of the requirements for the degree of Master’s of Arts. Bybee, J. 2003. Mechanisms of Change in Grammaticization: The Role of Frequency. In The Handbook of Historical Linguistics, edited by B. Joseph and R. D. Janda, 602-623. Oxford: Blackwell. Cinque, G. 1999. Adverbs and Functional Heads. A cross-linguistic perspective. New York: OUP. —. 2009a. Again on Tense, Aspect, Mood morpheme order and the “Mirror Principle”. In Functional Structure form Top to Toe, edited by P. Svenonius. New York: OUP. . —. 2009b. The fundamental left-right asymmetry of natural languages. Studies in Natural Language and Linguistic Theory, Vol. 76, 165-184. Cinque, G and L. Rizzi. 2009. The Cartography of Syntactic Structures. CISCL Working Papers on Language and Cognition 2, 43-59.

74

Development of TAM Categories in Turkic

Durie, M. 1997. Grammatical structures in verb serialization. In Complex Predicates, edited by Alex Alsina et al., 289–354. Stanford: CSLI Publications. Erdal, M. 2004. A Grammar of the Old Turkic. Leiden: Brill Academic Publishers. Givón, T. 1975. Serial verbs and syntactic change: Niger-Congo. In Word order and word order change, edited by C. Li, 47-112. Austin and London: University of Texas Press. —. 1991. The Evolution of Dependent Clause Morphosyntax in Biblical Hebrew. In Approaches to Grammaticalization, Vol. 2, edited by E. C. Traugott and B. Heine, 257-310. Amsterdam and Philadelphia: John Benjamins. Harris, A. C. and L. Campbell. 1995. Historical Syntax in Cross-Linguistic Perspective. Cambridge: Cambridge University Press. Frazier, L. and J. Fodor. 1979. The Sausage Machine: A New Two-Stage Parsing Model. Cognition, Vol. 6, 291-325. Haspelmath, M. 1998. Does grammaticalization need reanalysis?. Studies in Language 22, 315-351. —. 1999. Why is grammaticalization irreversible?. Linguistics 37/6, 10431068. Hawkins, John A. 1994. A performance theory of order and constituency. Cambridge: Cambridge University Press. Heine, B., U. Claudi and F. Hünnemeyer. 1991. Grammaticalization: A conceptual framework. Chicago: University of Chicago Press. Hopper, P. and E. C. Traugott. 2003. Grammaticalization. Cambridge: Cambridge University Press. Isxakov, I. B. and A. A. Palmbax. 1961. Grammatika tuvinskogo yazyka. Moskva. Johanson, L. 1995. On Turkic converb classes. In Converbs in crosslinguistic perspective, edited by M. Haspelmath and E. Kaunig, 313– 347. Berlin and New York: Mouton. Judakhin, K. K. 1985. Kirgizsko-russkij slovar. Frunze. Laenzlinger, C. 2000. More on Adverb Syntax. In Linguistics in Potsdam 6, edited by A. Alexiadou and P. Svenonius, 103-132. Potsdam: University of Potsdam. Lehmann, C. 1995. Thoughts on Grammaticalization. Munich: Lincom Europa. Lightfoot, D. W. 1979. Principles of Diachronic Syntax. Cambridge: Cambridge University Press. —. 2006. How New Languages Emerge. Cambridge: Cambridge University Press.

Pavel Grashchenkov

75

McDonough, Q. 2004. Head parameterization at the multiclausal level: clause-chaining as episodic macroevents in Altaic-area languages. Talk at WAFL-2, October, Istanbul. McGinnis, M. 2001. Phases and the syntax of applicatives. Proceedings of NELS 31, 333-349. —. 2005. UTAH at Merge: Evidence from multiple applicatives. In MITWPL 49: Perspectives on Phases, edited by M. McGinnis and N. Richards, 183-200. Newmeyer, F. J. 2001. Deconstructing grammaticalization. Language Sciences 23, 187-229. . Pylkkänen, L. 2002. Introducing arguments. Doctoral dissertation. MIT, MITWPL. Roberts, I. and A. Roussou. 2003. Syntactic Change. A minimalist Approach to Grammaticalization. Cambridge: Cambridge University Press. Shluinskij, A. B. 2000. The report of a field-research trip to the Mishar Dialect of Tatar. (Ms) Moscow: MSU. —. 2006a. Aktsional'nost' i aspektual'nyj pokazatel': konstruktsiya so vspomogatel'nym glagolom il- v chuvashskom yazyke. Vestnik MGU. —. 2006b. Biverbal'nye konstruktsii v tubalarskom dialekte i ih leksicheskie ogranicheniya. In Tubalarskie etyudy, edited by S. G. Tatevosov, 6-53. Moskva: IMLI RAN. Širaliev, M. Š. and E. V. Sevotjan. 1971. Grammatika azerbaidjanskogo yazyka (fonetika, morfologiya i sintaksis). Baku. Traugott, E. C. 2003. Constructions in Grammaticalization. In The Handbook of Historical Linguistics, edited by B. Joseph and R. D. Janda, 624-647. Oxford: Blackwell.

ON SLOVENIAN DEMONSTRATIVE REINFORCERS AND THE INTERNAL STRUCTURE OF DEMONSTRATIVES FRANC MARUŠIý AND ROK ŽAUCER

Outline This paper presents some peculiar Slovenian data that exhibit doubling of certain functional morphemes in a single word, specifically, of a spatiotemporal deictic marker/demonstrative reinforcer and of case markers. The paper offers an overview of the basic paradigms and a first attempt at an analysis. We suggest that when analyzed correctly, our data actually turn out not to be that unexpected. We also show that in exhibiting such data, Slovenian is actually not that peculiar, since similar phenomena have been found in several languages.

1. Introduction Even though Slovenian demonstratives may appear quite ordinary and uninteresting elements at first sight, they actually participate in some rather peculiar morphological paradigms. Specifically, some of the morphemes internal to demonstratives can be doubled, and sometimes even tripled within a single demonstrative. As shown in (1a) below, for example, the demonstrative reinforcer le can be doubled; and even more surprisingly, as shown in (1b), case morphology can also be doubled, appearing both before and after the reinforcer morpheme le. (1)

a.

i.

tega this-GEN ‘of this one’

ii.

tegale this-GEN-LE ‘of this very one’

iii.

tegalele this-GEN-LE-LE ‘of this very one’

Franc Marušiþ and Rok Žaucer

b.

i.

tega this-GEN ‘of this one’

ii.

tegale this-GEN-LE ‘of this very one’

iii.

77

tegalega this-GEN-LE-GEN ‘of this very one’

While our native-speaker awareness of (1a), whose use seems generalized across Slovenian dialects, is nothing recent, and the option of a doubled le from (1aiii) is also noted in Toporišiþ (2000: 341), we were quite astonished when recently coming across the pattern in (1biii). Nevertheless, running a series of searches on the internet to determine whether this is a real pattern or just a one-time coinage or even slip of the tongue, we found that the form in (1biii) is actually not even particularly rare. Subsequently we also found some such forms reported in Logar (1967). This paper reports some of our first findings related to the curious case-doubling phenomenon, while leaving the questions raised by the doubling of le itself, displayed in (1aiii), for future research. We should add that even though we talk about Slovenian, of which we are native speakers, the speech of neither of us includes all of the patterns presented below. We collected the data that we discuss as Slovenian from the internet, from texts that have clearly been written by native speakers; subsequently, we were also able to confirm many of the reported patterns, including several of those that our varieties do not include, with native speakers. Where necessary we also provide some quantitative data. In section 2, we give some background on Slovenian demonstratives. Section 3 presents the demonstrative reinforcer le. Section 4 provides the details regarding case doubling and section 5 offers an analysis. Section 6 concludes this paper.

2. Slovenian demonstratives Like in most other Slavic languages, Slovenian nominal demonstrative pronouns inflect for case, gender and number, as shown in (2). (2)

a.

i. ii. iii.

ta this-F.SG.NOM te this-F.SG.GEN tej this-F.SG.DAT

slika picture-F.SG.NOM slike picture-F.SG.GEN sliki picture-F.SG.DAT

On Slovenian Demonstrative Reinforcers

78

b.

i. ii. iii.

to this-N.SG.NOM tega this-N.SG.GEN temu this-N.SG.DAT

mesto town-N.SG.NOM mesta town-N.SG.GEN mestu town-N.SG.DAT

In this respect, Slovenian nominal demonstratives behave like adjectives, which is indeed how comparable demonstratives in the closely related Bosnian/Croatian/Serbian are analyzed in Zlatiü (1997) and Boškoviü (2007). And if they are adjectival, Slovenian nominal demonstratives will also get their agreement - just as is typically assumed for adjectives - via ‘concord’ (see e.g. Giusti 2008). Furthermore, if demonstratives are adjectival, it seems natural to treat them as a phrase in the specifier position of the relevant functional head in the nominal domain (in the spirit of Cinque 2005, 2010, Roehrs 2010), as in (3); following Roehrs (2010), we label this functional head IndexP. (3)

Apart from the adjectival demonstratives - the Slovenian equivalents of ‘this’ and ‘that’ - there are also many other demonstrative elements in Slovenian. What they all share is the demonstrative morpheme t-/s-/on-, which is followed by a morpheme expressing location/time/quantity/etc. and by the final morpheme determining the word’s grammatical category. (4)

a. b. c. d. e.

nominal demonstratives:

ta, tisti, oni this that that locatives: tukaj, tam here-LOC there-LOC directionals: sem, tja here-DIR there-DIR adjectives: tak, takšen that-type/kind that-type/kind numerals: toliko that many/much

Franc Marušiþ and Rok Žaucer

f.

manner adverbs:

g.

temporal expr.:

tako that-way sedaj tedaj, now, then (formal),

79

takrat then (spoken)

3. The demonstrative-reinforcer-like le Before getting into the discussion of the double-case-marking data, we need to determine the nature of le, the element appearing between the two instances of case morphology. Slovenian demonstratives can, at first sight optionally, occur with a particle-like element le affixed to them, as in (5). Without going into any detailed description, the literature describes le (in this use) as a reinforcing particle that raises the definiteness of the demonstrative (Logar 1967, Toporišiþ 2000: 340; but see 3.1 below for our own description of le). Its historical source is presumably the imperative form of the verb gledati “look” (Snoj 2003, according to Logar 1967 also Štrekelj 1906), not at all an unlikely candidate for the origin of a demonstrative reinforcer-like morpheme. (5)

a. b. c.

Ta(le) žoga je pa res this-LE ball AUX PTCL really 'This ball (here) is really too fast.' Tisti(le) je pa res dober. that-LE AUX PTCL truly good 'As for that one (there), it's really good.' Tam(le) je padel Robert Koren. there-LE AUX fell Robert Koren '(Just) over there is where Robert Koren fell.'

prehitra. too-fast

This element can appear on essentially any demonstrative, whether it is an adnominal, pronominal or adverbial, (6); excluded from this - predictably, as we will explain below - are only the two temporal demonstratives in (6g). (6)

a. b. c.

nominal demonstratives:

tale, tistile this-LE that-LE locatives: tukajle tamle here-LOC-LE there-LOC-LE directionals: semle tjale here-DIR-LE there-DIR-LE

On Slovenian Demonstrative Reinforcers

80

d. e. f. g.

adjectives:

takle, takšenle that-type/kind-LE that-type/kind-LE numerals: tolikole that many/much-LE manner adverbs: takole that-way-LE temporal expr.: sedajle *tedajle, *takratle now-LE, then-LE, then-LE

Before continuing, we should add that there exist other elements in Slovenian which may at least superficially seem quite close to the demonstrative reinforcer le, but which we will not deal with in this study; we will briefly mention them now only to properly delimit the scope of our study. Firstly, we assume that the morpheme -lej, which appears in temporal expressions potlej (“afterwards”), dotlej (“until then”), poslej (“from now on”), doslej (“until now”), vselej (“always”), is synchronically a separate element even though it may be etymologically related to le, i.e. likewise a derivative of the imperative of “look”, (g)lej; the distribution of these -lej forms differs from the distribution of the counterpart -le-forms, and the distribution of -lej itself differs from that of -le (the latter, for example, cannot attach to vse “all”). Secondly, the -le that combines with vendar and edino to form vendarle and edinole (“nonetheless” and “only”) can most likely be treated as separate from our -le in (5) for the following reasons. The compound edinole is synonymous and interchangeable with both edino by itself and lè by itself, with all three meaning “only”, (7a). Also, both the compound edinole and the stand-alone lè (in its use of “only”) are restricted to formal Slovenian, whereas the -le from (5) is rampant in spontaneous colloquial Slovenian; similarly, keeping the -le from (5) separate from the -le in edinole seems corroborated by looking at one of Slovenian’s close relatives, Slovak, where “only” is len, but the demonstrative reinforcer - the counterpart of Slovenian le from (5) - is the clearly separate hle. In a similar vein, vendarle can also be replaced, at least to some degree, by the stand-alone stressed lè, (7b)-(7c), which suggests that it should be treated separately from the -le from (5). (7)

a. b

Edino Peter bi lahko … = Lè Peter bi lahko … = Edinole Peter bi lahko … “Only Peter could …” i. Pa ti je vendarle uspelo. PTCL you AUX nonetheless made it

Franc Marušiþ and Rok Žaucer

ii.

81

Pa

ti je lè uspelo. you AUX nonetheless made it “You made it after all!” Jaz pa vendarle mislim, da to I PTCL nonetheless think that this ni isto. not same Jaz pa lè mislim, da to I PTCL nonetheless think that this ni isto. not same “I nevertheless think this is not the same thing.” PTCL

c.

i.

ii.

And thirdly, we believe that the demonstrative reinforcer le is synchronically separate from the ‘ever’-like wh-word reinforcer in cases like Kje lè? “Wherever?/Where on earth?”. While all of these elements might be historically related to our le from (5) (cf. Snoj 2003), their synchronically sufficiently different behaviours fully warrant leaving them out of the present study; but this does not mean that at least in principle, an account could not be found later which could explain their commonalities and differences synchronically as well.

3.1. Semantic contribution of le As mentioned above, le combines with demonstratives and seems to act as a kind of reinforcer. In this section, we will show that this reinforcing interpretation comes from le’s narrowing a regular demonstrative to a strictly spatio-temporal deictic demonstrative, essentially restricting its deictic potential to reference to the physical visual field. Le cannot be used with a demonstrative to refer to a person that is not present in the immediate physical context. That is, if a person is present only in the linguistic context, e.g. through the previous utterance, (8), we cannot use le on the demonstrative that refers back to this person from the previous utterance; in other words, le prevents what is sometimes called the discourse-deictic use of demonstratives.

On Slovenian Demonstrative Reinforcers

82

(8)

a.

b.

Peter se je vþeraj navduševal nad Peter REFL AUX yesterday rave over Valterjem Birso. Valter Birsa “Peter was all enthusiastic about Valter Birsa yesterday.” A tega(#le) tipa ti poznaš? Q this-LE guy you know “Do you know the guy?” (intended reference – Valter Birsa)

Note that in formal Slovenian, the combination of a demonstrative and a preposed particle with the same phonological form can be used in contexts comparable to (8), as shown in (9a). However, not only is this only acceptable with the preposed version and not with the postposed one, as most clearly shown with the minimal-pair contrast between tale and le-ta (both “this”) in (9), this preposed instance of le is also more like an independent word, carrying its own stress. (9)

a.

b.

Novakoviü poda do Birse, le-ta poda Novakoviü passes to Birsa, LE-this passes naprej do Cesarja ... on to Cesar “Novakoviü passes to Birsa, the latter passes on to Cesar ...” Novakoviü poda do Birse, ta(#le) poda Novakoviü passes to Birsa, this-LE passes naprej do Cesarja ... on to Cesar “Novakoviü passes to Birsa, the latter passes on to Cesar …”

Furthermore, in spontaneous Slovenian only tale exists and does not have the use of the formal-Slovenian le-ta. So even if they may be related, there is enough reason to keep tale and le-ta apart. In the remainder of this paper we will leave le-ta aside. Revealing restrictions also appear when we combine le with temporal demonstratives. As mentioned above, le can appear on temporal-deictic adverbs like 'now', (10a), but not on the discourse-dependent 'at that point', (10b), and even less so on non-demonstrative adverbials like 'later', (10c), or even demonstrative-containing relative adverbials like 'afterwards', (10d). (10)

a.

Zdajle ga še ne now-LE he-ACC still not “Right now I don't see him yet.”

vidim. see

Franc Marušiþ and Rok Žaucer

b.

c.

d.

83

Tedaj(*le) je prišel sam pred gol that-time-LE AUX come alone in-front-of goal in zadel. and scored “At that point he found himself in front of the goal all alone and scored.” Kasnej(*le) je prišel sam pred gol that-time-LE AUX come alone in-front-of goal in zadel. and scored “At that point he found himself in front of the goal all alone and scored.” Potem(*le) je prišel sam pred gol after-that-LE AUX come alone in-front-of goal in zadel. and scored “Afterwards, he found himself in front of the goal all alone and scored.”

The only other temporal demonstrative that can combine with le is prej “before”, in prejle “before-le”, but this form can only be used relative to now, that is, when it means “before now”/“just now”. It is impossible when it is used relative to some other point, e.g. “beforehand, prior to x”. In sum, of the temporal demonstratives, only those can combine with le which refer to a point in time to which direct referral is possible. Only two such temporal demonstratives exist in Slovenian, “now” and “just (before) now”, and as we have shown, these are the only ones that allow the addition of le. (11)-(13) offer some more (minimal) pairs showing the semantic import of le. (11)

a.

b.

Tinþek je preveþ razdražljiv. Tinþek AUX over sensitive A Tone je tud tak(#le)? Q Tone AUX also like-this-LE “Tinþek is over sensitive. Is Tone also like this?” Lej ga, takle mi gre najbolj look he-ACC like-this-LE I-DAT goes most na jetra. on liver “Look at him, I really hate him when he is like this.”

On Slovenian Demonstrative Reinforcers

84

(12)

(13)

a.

Peter Peter si

je

šel vþeraj na Triglav. A ti gone yesterday on Triglav Q you že bil tam(#le)? AUX already were there-LE “Yesterday, Peter climbed Mt. Triglav. Have you been there yet?” b. Lej, na tisti hrib tamle gremo. look onto that mountain there-LE go “Look, that hill over there is where we're going.” a. Vid je kupil 5 kil mesa. A misliš, Vid is bought 5 kilos meat. Q think da bo tok(#le) zadost? that will that-much-LE enough “Vid bought 5 kilos of meat. Do you think this will be enough?” b. [at a butcher's] Tokle tistgale mesa prosim, that-much that-LE meat please pa semle mi ga dejte. PTCL here-LE me it give That much of that kind of meat please. Give it here.” AUX

In (11a), le cannot be used to refer to “being sensitive” unless there is someone in visible sight which seems sensitive and to whom we can point. Similarly, (11b) suggests a scenario where the person uttering the sentence is simultaneously pointing at someone visibly in sight. In (12a), the demonstrative tamle cannot be used to refer to Mt. Triglav from the previous utterance, unless Mt. Triglav is in visible sight from where the interlocutors are standing (or else both interlocutors know exactly in which direction Triglav is) and at the same time the speaker points towards it. Just like in (11b), (12b) suggests a scenario where a simultaneous pointing gesture is needed. (13) shows the same for the demonstrative of quantity. While in (13a) we cannot use tokle to mean the previously mentioned 5 kilos, we can use it in (13b) if we accompany the utterance with a hand-gesture that shows how much meat we want to buy. At the same time, as shown in (11a), (12a) and (13a), using the plain (i.e. le-less) demonstrative to refer back to an entity from the discourse works fine. To summarize, the apparently reinforced interpretation of the lesuffixed demonstrative variants comes from le’s narrowing a regular demonstrative to a strictly spatio-temporal deictic, essentially restricting its deictic potential to reference to the physical visual field and stripping it

Franc Marušiþ and Rok Žaucer

85

of its discourse-deictic use. The le-variants thus cannot be used to refer to an entity that is present only in the linguistic context, such as through the previous utterance. If le evolved from an imperative form of the verb “look”, as has been suggested, this semantic narrowing is not at all unexpected. .

3.2. Possible analyses for le? As we have seen above, le appears on all types of demonstratives. Further, le is impossible with any other determiner, as well as with numbers, adjectives, nouns or any other noun phrase-internal element, (14). (14)

a.

b.

c.

mnogo(*le) / nekaj(*le) / malo(*le) / vse(*le) / a-lot-LE some-LE few-LE all-LE pet(*le) / vsak(*le) five-LE every-LE Prinesel mi je mnogo(*le) / nekaj(*le) / brought I-DAT AUX a-lot-LE some-LE malo(*le) / vse(*le) / pet(*le) / žog. few-LE all-LE five-LE balls “He brought many/some/few/all/five balls to me.” Prinesel mi je vsako(*le) žogo. brought I-DAT is every-LE ball “He brought me every ball.”

So with respect to an analysis, the first option that comes to mind is that le should be analyzed as the spell-out of some head or phrase that is specific to demonstratives. Some such elements have been independently proposed in the literature, so it is worth checking if they can be adopted for a successful analysis of le. Leu (2008) proposes that demonstratives combine a determiner of some extended AP projection (xAP) and a silent HERE or THERE inside this xAP, so that every demonstrative looks like (15). (15)

[xAP D [AP HERE/THERE ]]

However, le cannot be the overt realization of these null elements, for the simple reason that we have only one element, namely le, which is found on both proximal and distal demonstratives, (16), whereas the silent

On Slovenian Demonstrative Reinforcers

86

element is argued to come in two flavors for the two types of demonstratives – HERE vs. THERE. (16)

a. b.

ta = proximal, “this” ĺ tale tisti = distal/medial, “that” ĺ tistile oni = distal, “that” ĺ onile Tale/tistile/onile nogometaš mi this-LE/that-LE/that-LE footballer me še kar všeþ. still rather like “I rather like this/that football player.”

je

pa

AUX

PTCL

Kayne’s (2004) analysis of demonstratives also involves a demonstrative-specific null element, but unlike Leu (2008), Kayne argues that demonstrative adverbials contain a null noun PLACE. However, trying to invoke this account for analyzing le also results in problems. Le cannot be the realization of this null noun inside the demonstrative because it occurs with all kinds of deitic elements, not just with demonstrative place adverbials; as shown above, it can combine with manner, temporal, quantitative and adjectival demonstratives. All these demonstratives are composed ot t-/s- + place/time/manner + AGR + le, so we would need more than just one single empty noun. For the demonstrative adverbial in (17), we would thus presumably need something like an empty noun WAY. So just as was the case above with multiple adjectives, it does not seem reasonable to assume that the one and only le is the overt realization of several null nouns. (17)

Tkole se igra fuzbal na prvenstvu. this-way-LE REFL play football at world-cup “This is how football is played at the world cup.”

In order for such an analysis to work, we would need to assume that the null element inside demonstratives is not HERE/THERE or PLACE/WAY but something less specific. But if the null noun/adjective in demonstratives is more general, we lose the original motivation for the null noun/adjective. Moreover, such a null noun/adjective analysis cannot explain the possibility of having multiple occurrences of le inside a single demonstrative, an example of which was given in (1a) above and two more are offered in (18) below ((18b) is from the internet).

Franc Marušiþ and Rok Žaucer

(18)

a. b.

87

Tkolele se igra fuzbal na prvenstvu. this-way-LE-LE REFL play football at world-cup “This is how football is played at the world cup.” Hod tistlele na leu, k se ga go that-LE-LE on left that REFL him sam pou vid, na lesnoo?? only half see on wooden “Does the partly hidden guy on the left go to the woodindustry high school?”

3.3. Le as the head of a functional projection In this section, we will show that le forms a constituent together with the demonstrative and propose to analyze le as the head of the functional projection that tops off the phrase containing the demonstrative. Firstly, the demonstrative ta and le cannot be separated. When they do appear separated, as in (19b), we are not dealing with the same le (cf. section 3 above); the interpretation that le gets in such cases is “only”, so that (19b) means “only these three books”. And as was explained in section 3.1 above, the formal-Slovenian construction le-ta is also separate from our use of le, so does not constitute a counterexample to our le’s forming a constituent. Secondly, the demonstrative+le complex behaves like an independent unit since it can appear in different positions inside the noun phrase, as shown in (20), where it appears either before or after the numeral; since (20a) and (20b) receive different interpretations, they are likely not related by movement. (19)

a. b. c.

(20)

a.

* te these # le

tri le knjige three LE books te tri knjige LE these three books * te tri knjige le these three books LE (intended reading for (a)-(c): “these three books/these here three books”) tele tri hruške b. tri tele hruške these-LE three pears three these-LE pears “these three pears” “three pears of this type”

Next, since le can appear on various kinds of demonstratives, and in particular, on at least three kinds of noun phrase-internal elements (regular

On Slovenian Demonstrative Reinforcers

88

demonstratives, quantity demonstratives, and kind demonstrative), positing a dedicated functional projection (e.g. LeP/DeicSTP) in the main DP-NP projection line does not appear to be an option, since these three types of elements should be hosted in three separate functional projections. Furthermore, the adjectival demonstrative+le can stand for any type of adjective, as shown in (21), where takale ‘this kind’ stands for size in the (a) example and for color in its (b) counterpart. (21)

a.

ena takale rdeþa žoga one this-kind-LE red ball “a red ball of this type” (e.g. pointing to an object of a particular color) b. ena velika takale žoga one big this-kind-LE ball “a big ball of this type” (e.g. pointing to an object of a particular size) Given that different types of adjectives, such as size and color, are hosted in different functional projections (e.g. Scott 2002), this suggests that takale in (21a) and (21b) is hosted in different projections, and hence an analysis with le occupying a single functional projection is untenable. Therefore, we could say that le occupies the same dedicated functional projection of the NP-DP projection line only if all demonstratives were part of a noun phrase with a null noun PLACE/KIND/WAY/etc., which is itself situated somewhere inside the main noun phrase, as in (22). Now, a similar line of analysis was rejected in 3.2 above; note, however, that what we argued against there was the possibility that le represents an overt realization of the otherwise null nominal element. But in (22), le is not an overt counterpart of this otherwise null nominal element; and at the same time, in accordance with what we have just demonstrated to be the case, (22) correctly has the demonstrative and le forming a constituent to the exclusion of any other noun phrase-internal element. (22)

[DP … [DP [DemP ta [LeP/DeicP le [NP PLACE/KIND/…]]]] … [NP žoga ]] this LE ball

Finally, note that the sequence DEM+AGR+le, which was the only pattern taken into consideration till now, is actually not the only option for combining le with the demonstrative. As was the case in the data discussed so far and as is shown in (23b) below, several Slovenian dialects and standard Slovenian have le on the outside of case morphology; but in other

Franc Marušiþ and Rok Žaucer

89

Slovenian dialects, le appears on the inside of case morphology, as shown in (23c) (cf. Logar 1967).1 When the latter is the case, the constituency of the demonstrative and le is even more unquestionable (there is no independently established case of productive infixation in Slovenian). (23) NOM ACC GEN DAT LOC INST

a. le-less form (no variation) ta tega tega temu v tem(u) s tem

> > > > > >

b. Standard and Colloquial Slov. tale tegale tegale temule v tem(u)le s temle

| | | | | |

c. Colloquial Slovenian tale tel(e)ga tel(e)ga tel(e)mu v telem s telim/telem

So if we accept the constituency we argued for above, the structure of a noun phrase with any of the demonstrative+le complexes will be along the lines of (24). Following Brugè (2002) and Cinque (2005), (24a) places the demonstrative in DemP, which, in turn, is located in the specifier position of some dedicated functional head in the main NP-DP projection line. Following Roehrs (2010), we label this dedicated functional projection IndexP. (24)

Other demonstrative elements - locatives, directionals, numerals, etc. - are located in the relevant position that hosts adverbs, numerals (‘that-much’), adjectives (‘that-type’) etc. With respect to the internal structure of the demonstrative, however, the situation may be less clear. Since whatever follows t- seems to determine the grammatical category of the demonstrative (that is, whether it is an adverb, a numeral or an adjective), this should presumably be the 1

The brackets in (23) generally indicate the existence of alternative spellings of the same form; tel(e)ga thus means that the form can be found spelled either as telega or as telga.

90

On Slovenian Demonstrative Reinforcers

highest head of the demonstrative phrase. We will call this AgrP (as a cover term for various projections). The internal structure of a regular demonstrative could then be as in (25a), and that of other, e.g. adverbial demonstratives as in (25b). (25)

Next, we have to determine where le merges. As we have argued above, le only attaches to demonstratives, which could suggest that -le and the demonstrative head are merged together, with one selecting the other. But since in at least some varieties of Slovenian, -le is found on the outside of agreement morphology, we can conclude that at least in these varieties, -le merges with AgrP, and the relevant demonstrative feature then percolates up from DemP to AgrP. Furthermore, it is clear that in more complex demonstratives, le never merges directly with their demonstrative subpart, because we never get combinations like *tleako/*taleko (dem-le-ak-agr) and *tleakšen/*talekšen (dem-le-ak-š-adj-agr) but always only takole and takšenle (“this way” and “this kind”). So the internal structure of the demonstrative may be - depending on its components - something like (26); since we claimed above that le turns a deictically underspecified demonstrative into a strictly spatio-temporal deictic, we will label the projection that introduces le DeicSTP. (26)

In-between DemP and AgrP, there may also be additional structure, such as projections that determine a determiner’s proximal vs. distal character, etc.

Franc Marušiþ and Rok Žaucer

91

3.4. Further data and their structures So far we have only discussed the simple cases of demonstrative+le combinations and have not said much about the structural position of case morphology. As already mentioned in the previous section and summarized for one the masculine singular paradigm in (27b) and (27c) below, case can appear either above or below the DeicSTP of our structure in (26a) above. In this section we present the third option of the three possible patterns with respect to the placement of le and case morphology, i.e., the option - attested in some varieties of colloquial Slovenian - in which case morphology can be doubled, appearing on both sides of le. The column in (27d) gives the entire masculine singular paradigm for this option. The number to the right of a colloquial form is the number of hits we found for the form by searching the internet with Google (search restricted to Slovenian pages, numbers reported from January/February 2011); the purpose of the numbers is mostly to convince the reader that we are dealing with real patterns. The numbers reported are combined totals for all spelling alternatives (as indicated by brackets), so that, for example, temulemu, temulmu and temlemu have all been included in the figure reported for the dative form of (27d). Occurrences of the same form, when it covers more than one value (e.g. tegal(e)ga, which covers both accusative and genitive), have not been kept apart; we just checked to make sure the occurrences attest the use of the form for both values. a. le-less form NOM ta > tega > ACC GEN tega > temu > DAT LOC v tem(u) > s tem > INST c. Colloquial NOM tale/tela* noise tel(e)ga 1800 ACC GEN tel(e)ga tel(e)mu 400 DAT LOC v tel(e)m(u) 830 s telim 4400 INST * tela avto “this car” gives 2 hits † tala avto “this car” gives 3 hits

(27)

| | | | | |

b. Stand. & Coll tale tegale tegale temule v tem(u)le s temle d. Colloquial tala† tegal(e)ga tegal(e)ga tem(u)l(e)mu v tem(u)lem(u) s temlem

| | | | | | noise 4000+ 70 1200+ 8000+

On Slovenian Demonstrative Reinforcers

92

Note that although (27) gives the three paradigms of the demonstrative pronoun only for masculine singular, the same three options are restricted neither to this value of gender nor to this value of number. The neuter paradigm differs from the masculine only in nominative case, while in plural all genders share the same paradigm; (28) below offers the feminine singular paradigm, and (29) offers the plural paradigms. (28)

FEMININE NOM ACC GEN DAT LOC INST NOM ACC GEN DAT LOC INST

a. ta to te tej v tej s to c. tela* telo† tele†* telej/teli‡ v telej/teli s telo‡*

b. tale tole tele†* tejle v tejle s tole d. tala‡† tolo‡†* tele†* tejli v tejlej s tolo

> > > > > > noise noise noise noise 2 noise

| | | | | |

| | | | | | noise noise noise 1000+ 2 800+

* tistala “that one” (Fem.Nom) gives 800+ hits † tistalo “that one” (Fem.Acc) gives 140+ hits †* tele strani “this page”/“these pages” gives 8000+ hits ‡ teli punci “this girl” gives 250+ hits ‡* we found s tistalo “with that one” (Fem.Inst) on the internet ‡† tala stran “this page” gives 5 hits ‡†* tolo stran “this page” gives 520+ hits

(29)

PLURAL NOM ACC GEN DAT LOC INST NOM ACC GEN DAT

a. te/ti te teh tem v teh s temi c. tele/teli* tele† teleh†* telem‡

> > > > > > noise noise noise noise

| | | |

b. tele/tile tele† tehle temle v tehle s temile d. tele/tili‡* tele† tehleh temlem

| | | | | | noise noise 1200+ 8000+

Franc Marušiþ and Rok Žaucer LOC INST

v telih 1000+ | v tehleh s telemi 43 | s temilemi * teli tipi “these guys” gives 2500+ hits † tele punce “these girls” gives 5000+ hits †* teleh slik “these pages” 2 hits ‡ we found telem stvarem “these things” ‡* tili tipi “these guys” gives 9 hits

93 84 (na) 200+

As shown in (28) and (29) above, essentially all forms are attested on the internet, i.e. in a corpus of written Slovenian. We have to stress again that the counts we report above are really just to show that these forms do exist (an actual example for each of the forms is given in the Appendix); doing reliable counts that would make it possible to compare the relative frequencies of individual forms unfortunately seems more or less impossible.2 So what could be the structure for the third, case-doubling pattern? We suggest that the doubling paradigm - which may well be in free variation with the standard/(b)-column alternatives - arises as a combination of the other two other patterns. In (30), we sketch the proposed structures for all three patterns.3

2

As has already been mentioned, two of the three patterns are colloquial only. Accordingly, the internet is currently the only corpus that can be used to search for these forms. For example, whereas tegal(e)ga and tel(e)ga prove to be among the more frequent forms of their respective paradigms on the internet, both the FidaPlus (http://www.fidaplus.net/Pisni/Pisni_index.html) and the Nova beseda (http://bos.zrc-sazu.si/s_beseda.html) corpora of written Slovenian return no hits for tegal(e)ga and one single hit for tel(e)ga, and the emerging GOS corpus of spoken Slovenian (http://www.korpus-gos.net/) returns no hits for either form (all checks from March 1, 2011); admittedly, the GOS corpus currently contains only about 280,000 words, of which only 13% are from registers where one may expect these forms, and furthermore, only an (unspecified) part of this share includes speech from the dialects where one may expect these forms. 3 As pointed out by a reviewer, it is unusual to have a variable ordering of AgrP and DeicSTP even if these are three different dialects/systems. One option to address this problem is to try and analyze the two le’s as heads of two different projections (e.g. DeicST1P vs. DeicST2P). Another option to explore is to put le in either the specifier or else the head of DeicSTP. We discuss these options (and their possible problems), offer further argumentation for positing an AgrPs on each side of DeicP and a discussion of the implications for the theory of agreement in Marušiþ and Žaucer (in preparation).

94

On Slovenian Demonstrative Reinforcers

(30)

4. Other languages, same phenomena Patterns similar to the ones we presented above for Slovenian can be found in Czech and Slovak. Both of these languages use the particle -hle as a reinforcer that appears to be functionally parallel to the Slovenian -le. Just as has been suggested for Slovenian -le (Snoj 2003), the Czech and Slovak -hle is also claimed to derive historically from the imperative form of the verb "look" (Janda & Townsend 2002). A doubling pattern that seems to be best known for Czech and Slovak hle-forms is the one in (31), in which the adnominal demonstrative allows the demonstrative to reduplicate, yielding a form that does not seem to exist in Slovenian.4 (31)

4

tenhleten / thisNOM-hle-thisNOM techhletech thisGEN.PL-hle-thisGEN.PL

togohletogo, thisGEN-hle-thisGEN

/ etc.

According to Logar (1967), some Upper Soþa Valley dialects show DEM-le-DEM forms, but from what we can tell from the brief mention, these forms might differ from the Czech and Slovak forms in (31) in not doubling an inflected demonstrative form but rather combining an uninflected demonstrative with le and an inflected demonstrative (the only form given is tltie, which we reconstruct as DEM-le-DEM-AGR).

Franc Marušiþ and Rok Žaucer

95

However, just like in the case of Slovenian, colloquial Czech and Slovak apparently also exhibit forms that double only the case morphology, as in (32) (data again from the internet, culled from Czech sentences confirmed as native by Pavel Caha). Just like in Slovenian, the entire paradigm is attested on the internet. (32)

tohohleho tomuhlemu techhlech ...

= to-ho-hle-ho = to-mu-hle-mu = t-ech-hle-ech

»DEM-AGR-hle-AGR« »DEM-AGR-hle-AGR« »DEM-AGR-hle-AGR«

- GEN.SG - DAT.SG - GEN.PL

We have not looked at these cases in any detail, but at first sight, it seems safe to assume that these forms are structurally parallel to the Slovenian double-case forms discussed above. Furthermore, Haspelmath (1994) reports on similar cases of agreement doubling from Georgian and other languages, and Harris and Halle (2006) and Nevins (2009) discuss doubling of agreement in Spanish. In all of these cases a clitic is claimed to first get reinterpreted as a derivational suffix, which thus finds itself at the external edge of the word, outside the inflectional morpheme. According to Haspelmath, variation between the three patterns (reinterpreted derivational suffix outside inflection, inflection doubled inside and outside the reinterpreted derivational suffix, inflection only outside the reinterpreted derivational suffix) reflects a change in progress which is driven by the universal preference for inflectional morphology to be located outside derivational morphology. The Slovenian demonstratives exhibit a surprisingly similar pattern to the patterns described for Georgian, Spanish, etc. So the coexisting Slovenian patterns we presented might also involve a historical change in progress. In fact, Logar (1967) does consider the form with case morphology trapped between the stem and -le to be the older form, although without much discussion. We will have to leave the historical aspect, and the question of how the three synchronically coexisting patterns are historically related, for future research.5 5

The literature documents many other kinds of data that are sometimes mentioned as case doubling but whose potential (un)relatedness to our cases is, unfortunately, not very clear to us. One such case is the phenomenon that is often also termed suffixaufnahme, in which a nominal ends up with two different-value case suffixes as a result of the case of the entire noun phrase repeating on every noun in it, despite the fact that the nominal complement to the noun already carries its own genitive case, (i); another comes from substandard Hungarian, where the case suffix on the demonstrative can be doubled or marginally even tripled, with no

96

On Slovenian Demonstrative Reinforcers

5. Conclusion This paper constitutes an initial investigation into a Slovenian demonstrative-reinforcer-like construction in which the attachment of the particle le was shown to restrict a demonstrative's use to one of strict spatio-temporal deixis, preventing a discourse-deictic use. Confronting our construction with the theoretical literature on demonstratives, we first tested and rejected two potential analyses and then - having demonstrated that le forms a constituent with the demonstrative - proposed that it originates in a dedicated functional projection (DeicSTP) on the projection line of the demonstrative, which in turn sits in the dedicated demonstrative-hosting functional projection in the projection line of the relevant category (e.g. for an adnominal demonstrative, the IndexP of the the NP-DP line). Furthermore, we presented the three variants that our construction comes in (DEM-AGR-le, DEM-le-AGR, and the case-doubling DEM-AGR-le-AGR), and given the lack of awareness of the two substandard patterns, we also equipped the paradigm tables with some corpus counts and added some corpus data in the appendix. We suggested three variations of the proposal to capture the three distinct patterns, and concluded by briefly showing that the peculiar case-doubling pattern also exists in the possibly perfectly parallel demonstrative-reinforcer-like construction in Czech and Slovak. We had to leave diachronic aspects of the three construction variants for future work.

intervening material, (ii); another comes from Chichewa, where the universal quantifier -ri -onse 'each/every' carries two instances of a class prefix (-ch(i)-), (iii); and another comes from English (McIntyre 2009), (iv). (i) ba:ba-gu junuy-gundi-yu Gumbaynggir father-Erg child-Gen-Erg 'the child's father (ergative)' (Moravcsik 1995: 458) ? (ii) a. ez-t-et b. ez-t-et-et Hungarian this-acc-acc this-acc-acc-acc 'this' 'this' (cf. standard ez-t this-acc) (Moravcsik 1995: 456-7) (iii) Pa-li nchenche pa-chi-seko chi-ri-ch-onse. Chichewa 16-be 10-fly 16-7door 7every 'There are flies on every door.' (Carstens 1997: 384) (iv) a. sisters-in-laws in addition to sister-in-laws and sisters-in-law b. governors-generals in addition to governor-generals and governorsgeneral

Franc Marušiþ and Rok Žaucer

97

Acknowledgments We thank the audience and three anonymous reviewers of SinFonIJA 3 for valuable comments and suggestions, and Pavel Caha for helpful discussions of the Czech data.

References Boškoviü, Ž. 2007. On the Clausal and NP Structure of Serbo/Croatian. In Formal Approaches to Slavic Linguistics 15, The Toronto Meeting, edited by R. Compton, M. Goledzinowska and U. Savchenko, 42-75. Ann Arbor: Michigan Slavic Publications. Brugè, L. 2002. The position of Demonstratives in the Extended Nominal Projection. In Functional Structure in DP and IP, edited by G. Cinque, 15-53. Oxford: Oxford University Press. Carstens, V. 1997. Empty nouns in Bantu locatives. The linguistic review 14, 361-410. Cinque, G. 2005. Deriving Greenberg’s Universal 20 and Its Exceptions. Linguistic Inquiry 36/3, 315-332. —. 2010. The syntax of adjectives. Cambridge, Mass.: MIT Press. Giusti, G. 2008. Agreement and Concord in Nominal Exressions. In The Bantu-Romance Connection, edited by C. de Cat, 201-238. Amsterdam, Philadelphia: Benjamins. Janda, L. and C. E. Townsend. 2002. Czech. SEELRC. Kayne, R. S. 2004. Here and there. In Lexique, Syntaxe et LexiqueGrammaire/Syntax, Lexis and Lexicon-Grammar, edited by E. Laporte et al., 253-275. Amsterdam: John Benjamins. Leu, T. 2008. The Internal Syntax of Determiners. Doctoral dissertation. New York University. Logar, T. 1967. Kazalni zaimek v slovenskih nareèjih [The demonstrative pronoun in Slovenian dialects]. In Zadravec, F. et al. (eds.), Seminar slovenskega jezika, literature in kulture III. Ljubljana: Filozofska fakulteta. [Reprinted in Slavia orientalis 17 (1968), 347-350 and in Logar, T. 1996. Dialektološke in jezikovnozgodovinske razprave. Ljubljana: SAZU, 324-327.] Marušiè, F. and R. Žaucer. (in preparation). On case doubling in Slovenian. Ms. University of Nova Gorica. McIntyre, A. 2009. Synthetic compounds and argument structure. Messages from a bandwagon-jumper-onner and a two-cents'-worththrower-inner.Manuscript/handout. . Moravcsik, E. A. 1995. Summing up Suffixaufnahme. In Double Case: Agreement by Suffixaufnahme, edited by F. Plank, 451-484. New York: Oxford University Press. Roehrs, D. 2010. Demonstrative-reinforcer constructions. J. of Comparative Germanic Linguistics 13/3, 225-268. Scott, G..-J. 2002. Stacked adjectival modification and the structure of nominal phrases. In Functional Structure in DP and IP. The Cartography of Syntactic Structures, Vol.1, edited by G. Cinque, 91120. Oxford: Oxford University Press. Snoj, M. 2003. Slovenski etimološki slovar. Ljubljana: Modrijan. Štrekelj, K. 1906. Vermischte Beiträge zum slavischen etymologischen Wörterbuch. Berlin: Weidmann. Toporišiè, J. 2000. Slovenska slovnica. Maribor: Založba Obzorja. Zlatiü, L. 1997. The Structure of the Serbian Noun Phrase. Doctoral dissertation. University of Texas, Austin.

Franc Marušiþ and Rok Žaucer

99

Appendix 1 In this appendix we give examples with the nonstandard demonstratives from tables (33)-(35) below, which repeat the tables (27)-(29) from the main text. In the examples that follow the tables, the examples in (a) give masculine forms, the examples in (b) give feminine forms and examples in (c) give plural forms. Examples i. contain the demonstrative from column (c) in the cited tables and examples ii. contain the demonstrative from column (d). Due to excessive noise we include examples with tisti 'that' instead of ta 'this' in a couple of cases. Since the forms for the plural accusative and the feminine genitive are the same across the three paradigms in columns (b), (c), and (d), we do not give any examples in (37c) and (38b). The examples are quoted in their original form and thus typically contain nonstandard spelling. (33) Nom Acc Gen Dat Loc Inst (34) Nom Acc Gen Dat Loc Inst (35) Nom Acc Gen Dat Loc Inst

Masculine a. ta tega tega temu v tem(u) s tem

> > > > > >

b. tale tegale tegale temule v tem(u)le s temle

| | | | | |

c. tale/tela/tel tel(e)ga tel(e)ga tel(e)mu v tel(e)m(u) s telim

| | | | | |

d. tala tegal(e)ga tegal(e)ga tem(u)l(e)mu v tem(u)lem(u) s temlem

Feminine a. ta to te tej v tej s to

> > > > > >

b. tale tole tele tejle v tejle s tole

| | | | | |

c. tela telo tele telej/ teli v telej/ teli s telo

| | | | | |

d. tala tolo tele tejli v tejlej s tolo

Plural a. te/ti te teh tem v teh s temi

> > > > > >

b. tele/tile tele tehle temle v tehle s temile

| | | | | |

c. tele/ teli tele teleh telem v telih s telemi

| | | | | |

d. tele/ tili tele tehleh temlem v tehleh s temilemi

100

On Slovenian Demonstrative Reinforcers

(36)

NOMINATIVE a. i. Zdena, tal tip ti je pa padel Zdena this guy you aux ptcl fell v oko. in eye 'Zdena, it seems like you really like this guy.' ii. Kakor vem, je bil tala avto kar as know aux was this car quite resno "vandaliziran“ [...] seriously vandalized 'As far as I know, this car was seriously vandalized.' b. i. Kaj pomen tela beseda? what means this word 'What does this word mean?' ii. Tala stran je kar bogato založena this page aux quite rich stacked z njihovimi izdelki [...] with their products 'This page seems to be fairly abundant with their products.' c. i. Joj, teli tipi so tko smešni... boy these guys were so funny 'Boy are these guys funny!' ii. tili tipi pa rulajo!!!! these guys ptcl rock 'These guys rock!'

(37)

ACCUSATIVE a. i. drugaþe pa podpiram telega wikija otherwise ptcl support this-acc wiki.acc 'Otherwise I support this wiki.' ii. Tegalega bi se pa ustrašil, þe this would refl ptcl scared if bi ga sreþal nekje na samem! would him meet somewhere in private 'This guy would definitely scare me, if I met him in some deserted place.'

Franc Marušiþ and Rok Žaucer

b.

i.

ii.

c.

101

Seveda bi tistalo zgornjo zadevo sure would that upper thing lahko skrþil v par vrstic, […] can compres into pair lines 'Sure, I could compres that thing above into a couple of lines.' Zdravo, mene znima þe je že hello me interests if aux already kdo probal tolo zadevo? anyone test this 'Hello, I'd like to know if anyone has already tried this thing.' The three options are indistinguishable - tele

(38)

GENITIVE a. i. Poskusiš preko telega linka try over this link 'Try it over this link.' ii. Kaj ni gradnja tegalega modela Q not construction this model nekje na forumu..... somewhere on forum 'Isn't the construction of this model somewhere on the forum.' b. The three options are indistinguishable - tele c. i. Ej od teleh tvojih koktajlov bi hey from these your coctails would bili pa malo pijani ane? were ptcl little drunk right 'We'd be a little drunk with these your coctails, wouldn't we.' ii. Tolk o tehleh servisih. this-much about these car-shop 'This much about these car shops.'

(39)

DATIVE a. i. Pa kaj telemu Nikotu ne bodo ptcl what this Niko neg will priredili nobene zabave? organize no party 'Aren't they going to throw a party for Niko?'

On Slovenian Demonstrative Reinforcers

102

ii.

b.

i.

ii.

c.

i.

ii.

(40)

Z Anico sva se vþeraj iz with Anica aux refl yesterday from srca nasmejale temulemu stricu. hearth laugh this dude 'Yesterday, Anica and me laughed a lot at this guy.' Teli punci pa domišljija laufa sto this girl ptcl imagination runs 100 na uro. on hour. 'This girl has wild imagination.' Ideja je narediti nekaj podobnega idea aux make something similar tejli stvari this thing 'The idea is to make something similar to this thing.' jst telem stvarem zlo mal vrjamm, I these things very little believe ker usak pravi drugaþe because every says different 'I don't trust these things, since everyone says it differently.' Tanaja tud þe se boš kdaj poroþila, Tanaja even if refl will ever married tega temlem fantom ne povej. this these guys not tell 'Tanaja, even if you ever get married, don't tell this to these guys.'

LOCATIVE a. i. Komad na telmu posnetku (reklami) song on this recording (commercial) me zanima: […] me interests 'I'm interested in the song on this recording.' ii. Predvidevam da se je to zgodilo suspect that refl aux this happened na temlem prehodu [...] on this crossing 'I suspect it happened on this crossing.'

Franc Marušiþ and Rok Žaucer

b.

i.

ii.

c.

i.

ii.

(41)

103

Na telej fotki se vidi, kako se on this photo refl see how refl izza table vije dim from-behind board goes smoke 'On this photo, one can see the smoke coming out from behind the blackboard.' Jaz pa še nisem bil na I ptcl still neg-aux were on tejlej lokaciji nikoli this location never. 'I've never been to this location.' a teb normalno laufa v telih Q you normally runs in these mrzlih pogojih? cold conditions 'Do you have problems running it in these cold conditions.' V tehleh besedah ginekologinje ne vidim in these words gynecologist not see nobenega obsojanja no accusations 'I don't see any accusations in these words of the gynecologists.'

INSTRUMENTAL a. i. S telim je bilo pa kar with this aux was ptcl quite precej dela a-lot work 'There was quite a lot of work with this.' ii. sem pa full zadovoljna s temlem aux ptcl very satisfied with this forumom, ker sem dobila [...] forum, since aux got 'I'm very pleased with this forum, since I got ...' b. i. a ti resn misliš s tistalo kopþijo. Q you seriously think with this deal 'Are you serious about this deal?'

On Slovenian Demonstrative Reinforcers

104

ii.

c.

i.

ii.

s tolo izjavo se pa popolnoma with this statement refl ptcl completely strinjam. agree 'I completely agree with this statement.' sploh pa morfin se ne da anyway ptcl morphine refl not can primerjat s telemi švoh zdravili compare with these weak pills 'Anyway, you can't compare morphium with these weak medications.' najprej so bli vsi blogi iz ZDA first aux were all blogs from USA poplavljeni s temlemi ajfouni, [...] swamped with these iphones 'First all phones in the USA were swamped with Iphones, ...'

WORD ORDER(S) AND CLITIC DOUBLING IN BULGARIAN TEODORA RADEVA-BORK

Outline This paper explores the nature of clitic doubling in Bulgarian and investigates the interaction of the range of possible word orders and direct object clitic doubling, discussing contexts of optional vs. obligatory reduplication. In my analysis, all (in)direct objects that co-occur with pronominal clitics are instances of clitic doubling, regardless of the position of the objects in relation to the verb (i.e. in the Left Periphery or to the right of the verb). Based on the presented data, I show that syntactic structure rather than merely predicate types triggers the obligatory vs. optional presence of a doubling clitic in Bulgarian.

1. Introduction The present paper considers the interaction of the range of possible word orders and direct object clitic doubling in Bulgarian, discussing contexts of optional vs. obligatory clitic doubling (henceforth CD). In my analysis, all instances of (in)direct objects co-occurring with pronominal clitics are cases of genuine CD, and for Bulgarian I consider the distinction between “true CD” and dislocation structures incompatible with data from this language. Further on, I establish dependence between syntactic structure and use of CD contexts where the doubling clitic mainly functions as means of object and topic identification. If predicate choice, irrespective of the used construction, induces obligatory CD (as suggested in Krapova & Cinque 2008), it is expected that CD appears without exception when used with such predicates. The data in the paper does not confirm this prediction (cf. also Radeva-Bork 2010b). The paper is organized as follows. Chapter 2 starts with a general presentation of the CD phenomenon, discussing its spread across the Balkan languages. The problems that arise as a consequence of first, the

106

Word Order(s) and Clitic Doubling in Bulgarian

uneven distribution of the phenomenon in the Balkan Sprachbund and second, the influence of analyses of dislocation structures in Romance are discussed in Chapter 3. Here some conclusions about the analysis of genuine CD in Bulgarian are drawn. Chapter 4 is devoted to the analysis of Bulgarian clitic doubling. I survey the factors that guide the obligatory use of CD, and outline three main types of doubling constructions in Bulgarian in accordance to the function of the doubling clitic. The main conclusions of the paper are presented in Chapter 5.

2. The phenomenon of clitic doubling Clitic doubling refers to the overt doubling of a verbal argument (i.e. the associate) by a weak pronoun – a clitic, inside the same clausal domain. The clitic bears the same phi-features and case as the associate, when this is a pronoun or a DP. When it is a CP, the clitic is a singular (neuter) accusative.1 In this paper I restrict my analysis of CD constructions to cases with a full DP associate. Both direct and indirect objects can be doubled in this way as shown in (1) and (2).2 (1) (2)

Meþkata ja xvana Borko. bearDEF herCL.ACC caught Borko ‘Borko caught the bear.’ Poštadžijata mu dostavi pismoto na postmanDEF himCL.DAT delivered letterDEF to Ivan s goljamo zakăsnenie. Ivan with great delay ‘The postman delivered the letter to Ivan with a great delay.’

Although CD is one of the most pervasive phenomena in the languages of the Balkan Sprachbund, it is displayed in varying degrees and is governed by different conditions across the individual languages.3 Obligatory CD is triggered by the following environments in the single languages of the expansion area:

1

The associate can also be a wh-word. Examples throughout the paper are from Bulgarian unless indicated otherwise. 3 Cf. Dimitrova-Vulchanova & Hellan (1999), Tomiü (1996, 2008), Anagnostopoulou (1999), Kallulli (2000), Tasmowski (1987), Franks & King (2000), Rudin (1997), Guentchéva (1994), Friedman (2008) a.o. 2

Teodora Radeva-Bork

(3)

Macedonian: Albanian: Romanian:

Greek: Bulgarian:

107

all definite direct objects (DO) and indirect objects (IO) all IOs, 1st and 2nd pronoun DO, non-focal/nonrhematic DO DPs all full personal and definite pronouns, preverbal IO and specific DPs, postverbal specific DO DPs introduced by pe, postverbal specific and/or human Goal IO DPs facultative CD except with olos ‘all, everyone/thing’ CD is prescriptively not required except with Experiencer objects, objects of ima, njama ‘there is (not)’4

We find strict grammatical constraints on CD in the West and the South of its area of expansion. Moving to the North and the East, discourse -pragmatic factors condition the use of the phenomenon. This inhomogeneous situation can be shown on a scale, cf. (4), presenting a grammaticalization continuum5 for Balkan CD: at one end CD is grammatically constrained becoming freer and pragmatically significant as one proceeds to the other end. (4)

Macedonian > Albanian > Romanian > Greek > Bulgarian

3. The “genuineness” problem We already saw that although CD is one of the most salient features of the Balkan Sprachbund, it is unevenly distributed and dependent on a number of specific conditions. Doubling constructions, however, are not only restricted to the Balkan Sprachbund. They can also be found in Romance and Semitic; a fact which has yielded a number of influential generative analyses with the aim to arrive at a uniform analysis of the main properties of the phenomenon in both Romance and Slavic. It might be due to such factors that to date the constitution of the true nature of CD has remained controversial. A still ongoing debate in the literature on

4

Note that the description of the distribution of CD across the Balkan is based on normative accounts for the standard languages. In Chapter 4 I will show that these are not the only environments in Bulgarian that induce obligatory CD. 5 The idea of such a grammaticalization continuum dates back to Lopašov (1978).

Word Order(s) and Clitic Doubling in Bulgarian

108

clitics6 investigates the question whether CD and other related constructions such as Clitic Left Dislocation (henceforth CLLD) and Clitic Right Dislocation (henceforth CLRD) have to be considered as separate structures characterized by distinct properties, or whether they are not manifestations of a unitary phenomenon; and even more, whether a strict distinction between CD, CLLD and CLRD is not superfluous for Slavic. So the question arises: what constitutes genuine CD and what factors guide the distribution of the phenomenon?

3.1. The Romance influence on the analysis of Balkan clitic doubling Inspired by the Romance tradition (cf. Cinque 1984, 1990), constructions that typically are considered to show CD, i.e. cases when direct and indirect objects moved to the left periphery invariably trigger the surfacing of a doubling clitics (cf. Dimitrova-Vulchanova & Hellan 1999 for Bulgarian and Alexopoulou & Kolliakou 2002 for Greek), have been analyzed as Clitic Left Dislocation (CLLD).7 CLLD is seen to represent a distinct construction type, different from “true” CD. Other constructions such as Clitic Right Dislocation (CLRD), Hanging Topic and Focus Movement have been introduced to describe constructions types that share some properties with CLLD and CD but yet represent different constructions (cf. e.g., Krapova & Cinque 2008 a.o.). Many of the Romance languages, such as French and Italian, lack object clitic doubling, as evidenced by the ungrammaticality in (5),8 but typically allow for co-occurrences of clitics with objects to their left, i.e. CLLD constructions as in (6). (5)

a. b.

6

*Je Le vois I himCL.ACC see ‘I see John.’ *l’ho vista herCL.ACC have1SG. seen ‘I have seen the girl.’

Jean. John la the

ragazza. girl

(French) (Italian)

For references see Chapter 3.1. See, for example, Iatridou (1990), Anagnostopoulou (1994), Arnaudova (2003). 8 Cinque (1990: 178 fn4 and fn5) mentions limited cases of clitic doubling in Italian. 7

Teodora Radeva-Bork

(6)

a. b.

Pierre, je le verrai Peter I himCL.ACC will see ‘I will see Peter tomorrow.’ Gianni, lo vedrò John himCL.ACC will see1SG. ‘I will see John tomorrow.’

109

demain. tomorrow

(French)

domani. tomorrow

(Italian)

Similarly Peninsular Spanish and Catalan exhibit CLRD, a construction in which a clitic co-occurs with a phrase to its right, but lack clitic doubling of direct objects. The typology of the main Romance languages with respect to their (object) clitic doubling properties and the availability of clitic left- and right dislocation is shown in the table below: Table 1. Typology of Romance languages: distribution of CD, CLLD, CLRD. Language Italian French Peninsular Spanish Standard Spanish Catalan Romanian

CD no no no yes no yes

CLLD yes yes yes yes yes yes

CLRD yes yes yes yes yes yes

As the table shows, only the minority of Romance languages exhibit CD. It is thus dubious whether a strict distinction between CD, CLLD and CLRD should be introduced in Slavic on the basis of analysis of languages that have constituent dislocation but no constituent doubling. In Chapter 3.3. I argue that such an analysis, mainly inspired by research on languages with no clitic doubling such as most of the main Romance languages, is not applicable to data from Bulgarian.9 My claim is based on two grounds: first, the doubling of the object by a clitic is not dependent on the position of the object in the clause, and second, intonational breaks are not always identification for CLLD and CLRD constructions in Bulgarian.

9

For a similar view cf. Guentchéva (2008) and Tomiþ (2008).

110

Word Order(s) and Clitic Doubling in Bulgarian

3.2. Krapova & Cinque’s analysis In a recent analysis of clitic doubling constructions in Bulgarian, Krapova & Cinque (2008) suggest that CLLD, CLRD, Hanging Topic and Focus Movement are structures distinct from genuine clitic doubling, and reject a unification of the phenomena on the basis of distinct pragmatic, prosodic and structural properties. Accordingly, only the structure in (7) exhibits proper doubling. In (8) and (9) the doubling clitic is optional (in their view) and therefore the structures do not show true CD but simply CLRD and CLLD, respectively. (7) (8)

(9)

Ivan *(go) boli kraka. Ivan himCL hurts legDEF ‘Ivan’s leg hurts. Genata go polivaše cveteto vseki womanDEF itCL.ACC watered flowerDEF every vtori den. second day ‘The woman used to water the flower every second day.’ Lekarkata ja nabljudavaxa vsiþki ot doctorDEF herCL.ACC watched everyone from otdelenieto, za da mogat obektivno wardDEF for to can objectively da oceniat rabotata í to evaluate workDEF herCL.POSS ‘The doctor was being watched by everyone in the ward so that they could objectively evaluate her work.’

The authors’ main claim is that true doubling is obligatory and crucially depends on the predicate used irrespective of the structure, in which it appears. As a consequence, CD proper is reserved to special cases of clauses with certain types of predicates (e.g. psych and physical perception predicates, modal predicates, predicates with possessor datives, etc.). The syntactic construction here is irrelevant and it is rather the relation between types of predicates and the obligatoriness of the doubling clitic that distinguish CD constructions. Under this analysis again only (7) but not (8), (9) exemplifies genuine CD. Krapova & Cinque suggest additional properties of what they define as genuine CD: the reduplicated element can carry new information, it can be wh-moved and contrastively focused; moreover, a (non-specific) indefinite quantifier can be used as an associate. Due to space limitations, at present

Teodora Radeva-Bork

111

I mainly focus on the claim that the obligatoriness of the doubling clitic with certain types of predicates irrespective of word order is as a distinguishing feature of CD. In Chapter 3.3. I briefly touch upon on the doubling environments in which a reduplicated DP can be (contrastively) focused and wh-moved.

3.3. The present analysis In the literature it is generally assumed that doubled objects and dislocated constituents majorly differ in terms of distribution and intonation. I will try to show that these two conditions are not applicable for Bulgarian. First, labels such as CLLD and CLRD may well be adequate for languages in which the occurrence of the clitic is dependent on the occurrence of the object in the Left- or Right Periphery, but it is not for a language, such as Bulgarian, where clitic doubling is not contingent on the position of the object in the sentence. Consider the sentences in (10): (10)

a. b. c.

Maþa go speþeli po-dobriat matchDEF itCL.ACC won betterDEF Po-dobriat otbor go speþeli betterDEF team itCL.ACC won Speþeli go po-dobriat otbor won itCL.ACC betterDEF team ‘The better team won the match.’

otbor. team maþa. matchDEF maþa. macthDEF

In my analysis, the co-occurrences of all direct and indirect objects with pronominal clitics, regardless of the position of the objects (i.e. in the Left Periphery or to the right of the verb), are instances of genuine CD.10 The second expectation, that a dislocated element is distinguished by an intonational break, is not borne out either, cf. (11), since there is no intonational break necessary11 after and before the constituent noga in (11a) and (11b), respectively. (11)

10

a.

Noga knifeDEF

gotvaþa chefDEF

go itCL.ACC

slogi put

v in

þekmedgeto drawerDEF

This analysis is in line with Assenova (2002), Guentchéva (1994), Leafgren (1997), Franks & Rudin (2004, 2005) for Bulgarian and Tomiü (2008) for Macedonian. 11 Or at least an intonational break is not necessary in the way it may be in Romance.

Word Order(s) and Clitic Doubling in Bulgarian

112

b.

Gotvaþa go slogi v þekmedgeto ChefDEF itCL.ACC put in drawerDEF ‘The chef put the knife in the drawer.’

noga. knifeDEF

Guentchéva (2008) presents similar arguments in favour of the view that clitic doubling in Bulgarian and clitic dislocation in Romance languages (especially French) should be kept apart. I agree with her that intonational breaks, marked by a comma before a dislocated DP, characteristic of French right dislocation for example, do not function in the same way in constructions in Bulgarian. Moreover, I do not share the intuition that there is an intonation ‘break’ between the clitic go and its associate tova þuvstvo in an example given by Krapova & Cinque (2008: 271): (12)

Poznavam go tova know1SG. itCL.ACC this ‘I know this sentiment.’

þuvstvo. sentiment

As noted by Guentchéva (2008), doubled objects (direct or indirect) encoded by personal pronouns are not prosodically detached, as shown below: (13)

a. b.

Poznavam go know1SG. himCL.ACC ‘I know him.’ Pokazax mu showed himCL.DAT ‘I showed it to him.’

nego. himACC go itCL.ACC

na to

nego. himDAT

Furthermore, it is hard to distinguish between the structures in (14a) and (14b) and see them as representing two distinct structures, i.e. CLRD vs. CD proper, as there is no obvious difference between the sentences, either in terms of intonation or in distributional terms. (14)

a. b.

Uþitelkata go narisuva deteto. teacherDEF itCL.ACC drew childDEF ‘The teacher drew the child.’ Uþitelkata ja narisuva deteto. teacherDEF herCL.ACC drew childDEF ‘The teacher was drawn by the child.’/ ‘The child drew the teacher.’

Teodora Radeva-Bork

113

Similar constructions cannot be analyzed by analogy with CLLD or CLRD in Romance. Two reasons support this view: first, CD in languages such as Bulgarian and Macedonian is not dependent on the position of the object in the clause, and second, intonational breaks are not always necessary with doubled arguments. Recall at this point that Krapova & Cinque suggest it is only in genuine CD constructions, a term they reserve to obligatory CD with certain types of predicates, that the duplicated element can be (contrastively) focused as in (15a) and wh-moved as in (15b). (15)

a. b.

Jad *(go) e samo Ivan. anger himCL.ACC is only Ivan ‘Only Ivan is angry.’ Kogo *(go) e jad? whom himCL.ACC is anger ‘Who is angry?’ (from Krapova & Cinque 2008: 269)

However, there are cases where it is in fact possible to put contrastive focus on the doubled DP: (16)

Samo bezalkoxolnata bira ja only non-alcoholicDEF beer itCL.ACC Ivan, (a drugata ja Ivan (and otherDEF itCL.ACC ‘Ivan drank only the non-alcoholic beer, (and one).’

izpi drank ostavi). left) left the other

In (16), which according to Krapova & Cinque’s analysis would be CLLD and not CD, it is possible to contrastively focus the DP bezalkoxolnata bira just as the doubled DP Ivan in the proper CD construction in (15a). Similarly, the prediction that only doubled elements in genuine CD contexts as in (15b) can be wh-moved is not borne out in (17a-c): (17)

a. b.

Kogo go poznavaš? whom himCL.ACC know2SG ‘Whom do you know?’ Kogo go obiþaš poveþe, whom himCL.ACC love2SG. more Jonny Depp ili Al Pacino? Jonny Depp or Al Pacino ‘Whom do you like more, Jonny Depp or Al Pacino?’

Word Order(s) and Clitic Doubling in Bulgarian

114

c.

Kakvo go sănuva djado what himCL.ACC dream2SG.. grandpa ‘What did you dream about your grandpa?’

ti?12 yourCL.POSS

To summarize, I have shown that in clitic doubling environments in Bulgarian: first, the distribution of the object does not play a role and any configurations are possible and second, reduplicated elements are not necessarily prosodically distinct. Additionally, contractive (focus) and whmovement are not restricted solely to reduplicated elements in true CD constructions in the sense of Krapova & Cinque, but can as well occur in doubling constructions with other predicates. On the basis of the data presented above, the distinction between genuine CD, CLLD and CLRD seems to be superfluous for Bulgarian. For my analysis I consider all the instances in (7) to (9) above to represent a unitary phenomenon, i.e. genuine CD.13 Additionally, if there does exist some relation between obligatoriness of the doubling clitic and the appearance of CD, the source of this obligatoriness is the syntactic structure rather than the predicate itself. The following chapter explores the arguments in support of this view.

4. Clitic doubling in Bulgarian This chapter explores the interaction of the range of possible word orders and direct object clitic doubling in Bulgarian, discussing contexts of optional vs. obligatory clitic doubling and establishing a relationship between choice of syntactic structure and CD rather than predicate choice. Instead of singling out one particular environment, e.g. the type of CD with psych and physical perception predicates, as suggested by Krapova & Cinque (2008), I propose a classification of various CD environments that can account for the realization of different types of a uniform phenomenon.

4.1. Between optionality and obligatoriness CD is rare in formal and written Bulgarian. As a consequence of a strong prescriptive tradition, some speakers avoid the use of doubling

12

From the spontaneous speech of a monolingual Bulgarian native speaker. The view that Bulgarian CD can be analyzed as a syntactically uniform phenomenon is compatible under an analysis of pronominal clitics as K° suggested in Franks and Rudin (2004). 13

Teodora Radeva-Bork

115

constructions even in colloquial speech14 and often it seems that doubling is a purely optional phenomenon since the absence of a doubling clitic does not always lead to ungrammaticality and structures without clitics are productive.15 Generally, there are three situations in which CD is obligatory in Bulgarian16 (cf. Franks & Rudin 2004, Jaeger & Gerassimova 2002, Jaeger 2003 a.o.): (I) when the associate is an oblique subject, as in (18); (II) when it is a topic, as in (19); and (III) when whmovement appears to violate Superiority, as in (20). (18) (19) (20)

Ivan *(go) sărbi Ivan himCL.ACC itches ‘Ivan’s arm is itching.’ Marija nikoj ne Maria nobody not ‘Nobody loves Maria.’ Kogo koj *(go) whom who himCL.ACC ‘Who beat whom?’

răkata. armDEF *(ja) herCL.ACC

obiþa. loves

natupa? beat

Bulgarian is a language, characterized by great syntactic flexibility and structure information-driven word order despite a lack of a case marking system. Clitic doubling and the range of possible word orders in Bulgarian are often dependent on each other, and indeed there are cases when CD licenses certain word orders: (21)

14

Knigite *(gi) izgori booksDEF themCL.ACC burnt ‘Maria burnt the books.’

Marija. Maria

The elicitation of CD contexts in Bulgarian is difficult as indicated in Jaeger & Gerassimova (2002: footnote 6). Leafgren (2002) provides a valuable source of analyzing a corpus of Bulgarian colloquial data that allows making judgements based on actual usage rather than on speakers’ judgements of what they think they say. 15 For instance, structures in which the clitic precedes the associate are possible without the clitic: Obiþat (go) mnogo tozi ministar. love3PL. himCL.ACC much this minister ‘They love this minister very much.’ Note, however, that Obiþat mnogo tozi minister has a different meaning and information structure. 16 I will not deal with the third situation here.

Word Order(s) and Clitic Doubling in Bulgarian

116

(22)

Izgori *(gi) burnt themCL.ACC ‘Maria burnt the books.’

Marija Maria

knigite. booksDEF

In fact, as it has been previously discussed (cf. Rudin 1986, Werkmann 2003), if the preferred S-V-DO-IO surface order is not followed, CD is necessary to identify the syntactic roles of object vs. subject. If this is not done, the correct interpretations for (23) and (24) are grammatically excluded since if not doubled, the fronted objects dvete nevinni žertvi and Boris will be wrongly interpreted as subjects: (23)

(24)

Dvete nevinni žertvi *(gi) izjali twoDEF innocent victims themCL.ACC ate vălzi tazi sutrin. wolves this morning ‘The two innocent victims were eaten by wolves this morning.’ Boris izvednăž *(go) svali bolestta Boris suddenly himCL.ACC knock down sicknessDEF na legloto. onto bedDEF ‘Boris was knocked down by a sudden sickness.’

This is not to say that we can reduce CD to a constituent order repair (or a case-marking device) but in my view there are certain contexts, which necessitate a doubling clitic as an object marker (cf. Chapter 4.2. for a typology of possible realizations of CD in Bulgarian). Now recall that Krapova & Cinque’s (2008) analysis of CD is based on the idea of a relation of certain types of predicates and obligatoriness of a doubling construction (cf. Chapter 3.2.). A closer inspection shows that not all of the listed predicates induce obligatory CD irrespective of the construction used (cf. Radeva-Bork 2010b). If oblique subjects co-occur with nominative arguments, CD is not obligatory (cf. (25a) vs. (25b), (26a) vs. (26b), and (27a) vs. (27b)). (25)

a.

Omrăzna *(í) da gleda got tired herCL.ACC to watch (na Marija). (to Maria) ‘Maria got tired of watching TV.’

televizia TV

Teodora Radeva-Bork

b.

(26)

a.

b.

(27)

a. b.

117

Televiziata/ Gledaneto na televizia (í) TVDEF/ watchingDEF of TV herCL.ACC omrăzna bărzo na Marija. got tired quickly to Maria ‘TV/Watching TV quickly got Maria tired.’ V poslednia moment *(mu) xrumna, in lastDEF moment himCL.ACC occurred þe e zabravil da izkluþi utijata. that is forgot to switch-off ironDEF ‘It occurred to him in the last minute that he had forgotten to switch off the iron.’ (Tova) ce e zabravil da izkluþi (this) that is forgot to switch-off utijata (mu) xrumna na Ivan v ironDEF himCL.ACC occurred to Ivan in poslednia moment. lastDEF moment ‘That he had forgotten to switch off the iron occurred to Ivan in the last moment.’ Na Ivan *(mu) dosažda pesenta. to Ivan himCL.ACC bothers songDEF ‘Ivan is bothered by the song.’ Pesenta (mu) dosažda na Ivan. songDEF himCL.ACC bothers to Ivan ‘The song bothers Ivan.’

It is possible to assume that in (25b-27b) CD is not necessary as there is no mismatch between the syntactic positions of the arguments in the neutral SVO order and their syntactic roles. Additionally, the data in (28) show that even predicates such as boli ‘hurt’, which are expected to induce CD all over, show a different behaviour with a generic associate, cf. (28a,b) or with a quantified DP, cf. (28c) as doubling is not necessary.

Word Order(s) and Clitic Doubling in Bulgarian

118

(28)17

a.

b.

c.

Ušite þesto boliat decata prez earsDEF often hurt childrenDEF during zimata. winterDEF ‘The ears often hurt children in winter.’ Ušite mnogo boliat decata earsDEF much hurt childenDEF (pri văzpalenie). (at inflammation) ‘The ears hurt children badly (when inflamed).’ Glavata boli mnogo vseki þovek headDEF hurts much every person pri padane ot visoko. after falling from high ‘The head hurts badly after falling down from a high place.’

The sentences in (28) were accepted as grammatical by ten monolingual Bulgarian native informants without a linguistic background. When the constituents decata and vseki þovek were fronted, they opted for the cliticdoubled object option. These observations can be integrated in the present analysis if facts about word order are taken into account. Structures such as (29) do not differ much from sentences such as (23), repeated for convenience in (30), where the fronted objects have to be clitic-doubled. (29)

Decata þesto *(gi) boliat childrenDEF often themCL.ACC hurt prez zimata. during winterDEF ‘The ears often hurt children in winter.’

ušite earsDEF

17 It needs mentioning that the sentences in (28) make available a different semantic interpretation than the one in the equivalents with a clitic-doubled associate, apparently due to semantic conditions on the used DPs. This paper cannot include a discussion of the semantic factors at play here, but what is important for the analysis is that the predicate itself cannot be the only defining factor for use vs. non-use of CD. A mixture of syntactic and semantic conditions plays a role in determining the optionality vs. obligatoriness of clitic doubling. Another interesting issue is that apparently semantic factors also play a role in determining the kind of nominative arguments that can be used in such structures without doubling.

Teodora Radeva-Bork

(30)

119

Dvete nevinni žertvi *(gi) izjali vălzi twoDEF innocent victims themCL.ACC ate wolves tazi sutrin. this morning ‘The two innocent victims were eaten by wolves this morning.’

When decata appears in initial position, it has to be doubled, cf. (29), which is not the case in (28a). The same is true for (30) where dvete nevinni žertvi is the fronted, doubled object of the sentence. So it is possible that there are cases when fronted oblique subjects behave like fronted direct object and therefore need to be doubled. This parallel indicates again that predicate types in clitic doubling constructions in Bulgarian alone are not a sufficient trigger of obligatoriness.

4.2. Types of clitic doubling Although the phenomenon of CD in Bulgarian is syntactically a uniform phenomenon and no strict distinction between proper CD, CLLD and CLRD is necessary for Bulgarian (for argumentation cf. Chapter 3.3. as well as Guentchéva 2008 and Tomic 2008), the data presented above show that Bulgarian clitic doubling is in many ways multi-faceted and a subject to various conditions. In certain cases CD is necessitated by the syntactic structure of the sentence, in others the information structure or the used predicate play a role. In order to account for the various CD environments that are conditioned by distinct factors in Bulgarian, below I present a brief classification of three18 main types of clitic doubling environments. o

Type I. Object identification

Object identification is typically found in transitive constructions, in which both nominal constituents are animate and of the same number, cf. (31). The unmarked word order and the interpreted order is SVO. If the initial constituent functions as a direct object, CD is obligatory. (31)

18

Uþitelkata *(ja) narisuva teacherDEF herCL.ACC drew ‘The child drew the teacher.’

I am excluding doubling of interrogatives here.

deteto. childDEF

Word Order(s) and Clitic Doubling in Bulgarian

120

(32)

Marija nikoj ne Maria nobody not ‘Nobody kissed Maria.’

*(ja) herCL.ACC

celuna. kissed

Here it is the particular word order, i.e. SUBJ not first, that triggers the obligatory use of CD as means of a disambiguation between uþitelkata and Marija being subjects as non-clitic-doubled arguments or objects as cliticdoubled arguments. Empirical evidence, supporting the analysis of this type of CD, comes from first language acquisition. In an elicited comprehension task with 16 monolingual Bulgarian children aged 2;5-4;2, the children showed sensitivity to the object identification function of CD constructions as they successfully identified fronted clitic-doubled arguments to be the syntactic objects in sentences similar to (31) above (cf. Radeva-Bork 2010a). These findings are in line with the results of Jaeger & Gerassimova’s (2002) online study showing that fronted, topical objects are always doubled. o

Type II. Topic marking

As already mentioned, Bulgarian has a flexible word order theoretically allowing for all possible word order combinations, i.e. SVO, OVS, VSO, etc. The constituent order is to great extent dependent on information packaging, namely topicalization and focalization. Topicalized objects can appear in different positions depending on the context, as shown in (33a) and (33b) where the doubled object occurs in initial and postverbal position, respectively. (33)

a. b.

Knigite *(gi) booksDEF themCL.ACC Deteto *(gi) childDEF themCL.ACC ‘The child tore the books.’

nakasa tore nakasa tore

deteto. childDEF knigite. booksDEF

We see that object CD constructions are compatible with different constituent orders. In both cases, the clitic gi signals the topical function of the argument knigite,19 in other words CD has a topic marking function here. This is in line with Leafgren (2002) who demonstrates that object reduplication in Bulgarian is almost always used as an overt marker of 19 Concerning examples such as (33b), Guentchéva (2008) suggests that the doubled argument in final position is interpreted as a secondary topic.

Teodora Radeva-Bork

121

topicality.20 Dimitrova-Vulchanova & Vulchanov’s (2008) also show that even in Old Bulgarian contrastive topic was a trigger for the surfacing of doubling clitics. o

Type III. Oblique subjects conditioned on word order

When the associate is an oblique subject it has to be doubled, however, often conditioned on the constituent order. To illustrate this I repeat examples (27) and (28, 29a) for convenience in (34) and (35), respectively: (34)

a. b.

(35)

a.

b.

Na Ivan *(mu) dosažda pesenta. to Ivan himCL.DAT bothers songDEF ‘Ivan is bothered by the song.’ Pesenta dosažda na Ivan. songDEF bothers to Ivan ‘The song bothers Ivan.’ Decata þesto *(gi) boliat ušite childrenDEF often themCL.ACC hurt earsDEF prez zimata. during winterDEF Ušite þesto boliat decata earsDEF often hurt childrenDEF zimata. winterDEF ‘The ears often hurt children in winter.’

prez during

Similarly to the other two types, doubling of oblique subjects is in certain cases dependent on the kind of structure in which they appear. Thus predicate choice and word order play a role here as shown by the difference between (34, 35a)- obligatory clitic doubling, and (34, 35b), where CD is not necessary in the presence of a nominative argument in the subject position. This is not to say that doubling is optional across the board when a nominative argument is present in oblique subject constructions. What I aim to show here is that even in the case of oblique subject constructions, which are typically considered to induce CD all 20

The idea of obligatory CD for topicalized objects has also been discussed in Alexandrova (1997), Dimitrova-Vulchanova & Hellan (1995/99), Leafgren (1997), Rudin (1997), Kallulli (1999, 2000), Jaeger & Gerassimova (2002), Jaeger (2003), Guentchéva (2008).

122

Word Order(s) and Clitic Doubling in Bulgarian

over, syntactic structure plays a role.

5. Conclusion Two are the main contributions of this paper. The first is more general and concerns the nature and constitution of clitic doubling constructions in Bulgarian. Drawing an analogy to languages which do not have clitic doubling but double constituents in the right or the left periphery of the clause, makes us suspicious as to whether we should disregard Balkan doubling constructions (or at least doubling in Bulgarian) with an object to the left or the right of the clitic as true CD and see such structures as cases of CLLD and CLRD. The distinction between genuinely doubled DPs and left- or right dislocated objects seems superfluous in the light of Bulgarian data. In this language it is only the distribution of the doubling clitic that is restricted (i.e. to the right of left of the verb, no intervening elements), but the distribution of the object does not play a role and any configurations are possible. Additionally, replicated elements in allegedly dislocation constructions do not always prosodically differ from “truly” doubled elements. Therefore, in my view, Bulgarian CD encompasses all cases of (in)direct objects co-occurring with a clitic. The second contribution made here is associated with the kinds of CD triggers in Bulgarian. We saw that conditions, such as predicate choice and obligatoriness of the clitic, are relevant but not sufficient for the occurrence of CD, and the interaction of clitic doubling with constituent order and information structure needs to be highlighted. The empirical observations made on the basis of data from colloquial Bulgarian strongly support the existence of three main types of CD constructions: I) A mismatch between the syntactic positions of the arguments in the neutral SVO order and their syntactic roles induces CD as means of object identification; II) Objects as topics need a doubling clitic for topic marking; III) CD with oblique subjects often conditioned on the structure used. In my view, these types are more widely spread in colloquial Bulgarian than formerly discussed in the literature,21 and are on the way of becoming obligatory in a greater number of Bulgarian dialects. It is then legitimate to predict a further and deeper distribution of the doubling phenomenon in the spoken language and possibly a grammaticalization-like process. An open issue, an object of further analysis, is the role of semantic 21 But see Leafgren’s (2002) study of a corpus of colloquial Bulgarian, which is very illuminating in this respect.

Teodora Radeva-Bork

123

features on reduplicated elements and on nominal constituents in oblique subject constructions as well as their interplay with the conditions imposed on Bulgarian clitic doubling by syntactic and information structure.

Acknowledgements The author is recipient of a DOC-fellowship of the Austrian Academy of Sciences at the Institute of Linguistics.

References Alexandrova, G. 1997. Pronominal clitics as g(eneralized) f(amiliarity) licensing agro. In Formal Approaches to Slavic Linguistics: The Cornell Meeting, edited by Browne et al., 1-31. Ann Arbor, MI: Michigan Slavic Publications. Alexopoulou, T. and D. Kolliakou. 2002. On linkhood, topicalization and clitic left dislocation. Journal of Linguistics 38, 193-245. Anagnostopoulou, E. 1994. Clitic Dependences in Modern Greek. Doctoral dissertation. Universität Salzburg. —. 1999. On the representation of clitic doubling in Modern Greek. In Clitics in the Languages of Europe, edited by H. v. Riemsdijk, 761798. Berlin: Mouton de Gruyter. —. 2002. Case 17: Clitic Doubling. Syntax Companion Project (SYNCOM). . Arnaudova, O. 2003. Focus and Bulgarian Clause Structure. Doctoral dissertation. University of Ottawa. Assenova, P. 2002. Balkansko ezikoznanie. Osnovni problemi na Balkanskija ezikov săjuz. Sofia: Fabr. Cinque, G. 1984. Clitic Left Dislocation in Italian and the ‘Move-Į’ parameter. MS. Universitá di Venezia. —. 1990. Types of A’-dependences. MIT, Cambridge, Mass. Dimitrova-Vulchanova, M. and L. Hellan. 1995/99. Clitics and Bulgarian clause structure. In Clitics in the Languages of Europe, edited by H. v. Riemsdijk, 469-514. Berlin: Mouton de Gruyter. —. (eds.). 1999. Topics in South Slavic Syntax and Semantics. Amsterdam: John Benjamins. Dimitrova-Vulchanova, M. and V. Vulchanov. 2008. Clitic doubling and Old Bulgarian. In Clitic doubling in the Balkan languages, edited by D. Kallulli and L. Tasmowski, 105-132. Amsterdam: John Benjamins.

124

Word Order(s) and Clitic Doubling in Bulgarian

Franks, S. and C. Rudin. 2004. Bulgarian clitics as K0 heads. Presented at FASL 13, South Carolina. —. 2005. What makes clitic doubling obligatory. MS. Franks, S. and T. H. King. 2000. A Handbook of Slavic Clitics. Oxford: OUP. Friedman, V. A. 2008. Balkan object reduplication in areal and dialectological perspective. In Clitic doubling in the Balkan languages, edited by D. Kallulli and L. Tasmowski, 35-6. Amsterdam: John Benjamins. Guentchéva, Z. 1994. Thématisation de l’objet en bulgare. Bern: Peter Lang. —. 2008. Object clitic doubling constructions and topicality in Bulgarian. In Clitic doubling in the Balkan languages, edited by D. Kallulli and L. Tasmowski, 203-223. Amsterdam: John Benjamins. Iatridou, S. 1990. Clitics and island effects. UPenn Working Papers in Linguistics 2, 11-38. Jaeger, F. 2003. Topicality and superiority in Bulgarian wh-questions. Presented at FASL 12, Ottawa. Jaeger, F. and V. A. Gerassimova. 2002. Bulgarian word order and the role of the direct object clitic in LFG. In Proceedings of the LFG Conference 2002, edited by M. Butt and T. H. King, 97-219. Stanford: CSLI Publications. Kallulli, D. 1999. The comparative syntax of Albanian: on the contribution of syntactic types to propositional interpretation. Doctoral dissertation. University of Durham. —. 2000. Direct object clitic doubling in Albanian and Greek. In Clitic Phenomena in European Languages, edited by F. Beukema and M. den Dikken, 209-248. Amsterdam: John Benjamins. Krapova, I. and G. Cinque. 2008. Clitic reduplication constructions in Bulgarian. In Clitic doubling in the Balkan languages, edited by D. Kallulli and L. Tasmowski, 257-287. Amsterdam: John Benjamins. Leafgren, J. 1997. Bulgarian clitic doubling: Overt topicality. Journal of Slavic Linguistics 5, 117-143. —. 2002. Degrees of Explicitness. Amsterdam: John Benjamins. Lopašov, J. A. 1978. Mestoimennye povtory dopolneija v balkanskix jazykax. Leningrad: Nauka. Radeva-Bork, T. 2010a. An elicitation study on the comprehension of clitic doubling constructions by young Bulgarian children. University of Vienna. —. 2010b. Clitic doubling in Bulgarian: Between optionality and obligatoriness. In Formal Approaches to South Slavic and Balkan

Teodora Radeva-Bork

125

Languages 7, edited by M. Tadiü, M. Dimitrova-Vulchanova and S. Koeva, 89-95. Zagreb: Croatian Language Technologies Society. Rudin, C. 1986. Aspects of Bulgarian syntax: Complementizers and WH constructions. Columbus: Slavica. —. 1997. AgrO and Bulgarian pronominal clitics. Formal Approaches to Slavic Linguistics: The Indiana meeting 1996. Ann Arbor, MI: Michigan Slavic Publications, 224-252. Tasmowski, L. 1987. La reduplication clitique en roumain. In Rätoromanisch und Rumänisch. Akten der Theodor Gartner-Tagung, edited by G. Plangg and M. Iliescu, 377-399. Innsbruck: Amae. Tomiü, Mišeska O. 1996. The Balkan Slavic clausal clitics. Natural Language and Linguistic Theory 14, 811-872. —. 2008. Towards grammaticalization of clitic doubling. In Clitic doubling in the Balkan languages, edited by D. Kallulli and L. Tasmowski, 65-87. Amsterdam: John Benjamins. Werkmann, V. 2003. Objektklitika im Bulgarischen [= Studia grammatica 57]. Berlin: Akademie Verlag.

THE PROPERTY ANALYSIS OF CONTROL CONSTITUENTS: AN ARGUMENT FROM MODAL EXISTENTIAL WH-CONSTRUCTIONS RADEK ŠIMÍK

Outline Recent discussion of obligatory control in the literature concentrates on the issue of which syntactic operation (movement, agreement, etc.) is responsible for the establishment of the control relation. This paper looks at the issue of control from a higher order perspective. Abandoning the presupposition that control constituents denote propositions and that, therefore, control must be syntactic, I deliver an argument in favor of the property analysis of control constituents and, by transitivity, for a semantic resolution of the control relation. The argument comes from modal existential wh-constructions and in particular from a strong parallelism between obligatorily controlled PRO and wh-expressions. It is revealed that PRO and wh-words form a natural class, to the exclusion of all other types of nominal expressions. This is then turned into an argument of treating PRO and wh-words essentially as logical lambda-operators, naturally leading to the property theory of control.

1. Introduction This paper is intended as a contribution to the discussion of the syntax and semantics of obligatory control. Drawing from the empirical domain of modal existential wh-constructions, I will put forth some novel evidence supporting the view that control constituents map to properties rather than propositions. By transitivity, the evidence supports approaches in which obligatory control is resolved on the basis of the lexical semantics of control predicates, rather than by interplay of syntactic conditions. The argument is based on a strong parallelism between PRO

Radek Šimík

127

and wh-words and as such, it also supports a particular analysis of whfronting in which fronted wh-words map to logical lambda-operators (as opposed to indefinites or quantifiers). The parallelism will be formulated in terms of the PRO-wh generalization, which states that if a language has modal existential wh-constructions whose empty subject is an obligatorily controlled PRO (as in Russian or Spanish), then a structurally analogous modal existential wh-construction can contain an overt subject if and only if the subject is a wh-expression. Section 2 sets the theoretical stage by laying out the landscape of theories of obligatory control. The PRO-wh generalization will be introduced in section 3, along with the basic typology of modal existential whconstructions. In section 4, I will argue that the PRO-wh generalization can be captured in an elegant way if one adopts the conjunction of (i) the property theory of control constituents (Chierchia 1984, 1989), (ii) the logical lambda-operator theory of wh-words (Groenendijk and Stokhof 1984, Heim and Kratzer 1998), and (iii) a strict type-theoretic construal of control predicates. In that case, the PRO-wh generalization falls out as a natural consequence of the success and failure of functional application at the syntax-semantics interface. In section 5 I provide a detailed analysis of modal existential wh-constructions that exhibit obligatory control and propose an explicit treatment of the PRO-wh complementarity. Section 6 concludes the paper.

2. The landscape of theories of obligatory control In this section, I will first lay out the set of assumptions about control that I will take for granted and then introduce those that will be subject to testing. As far as I can tell, none of the adopted assumptions is intrinsically tied to any of the tested ones, so the argument to be made is not biased from this perspective. Control is broadly defined as a particular way of determining the reference of phonologically empty subjects, designated as PRO, which typically (but not necessarily; cf. Landau 2004 and section 3) appear in non-finite clauses. I will assume the usual (though simplified) dichotomy (going back to Williams 1980) of obligatory control (OC), illustrated by (1a), in which the reference of PRO is grammatically fixed, and nonobligatory control (NOC), (1b), in which it is fixed contextually. The bracketed part is called the control constituent. (1)

a. b.

They persuaded himi [PROi/*j to remain in the cabin]. Maryi wonders [PROi/j how to feed kids in Africa].

The Property Analysis of Control Constituents

128

In OC, the primary focus of this paper, the reference of PRO is determined by one of the arguments of the control predicate (persuade in (1a)), so called controller (him in (1a)). The characteristic properties of PRO in prototypical obligatory control are (i) exhaustive (as opposed to partial) determination of the reference by the controller, (2a), (ii) sloppy (as opposed to strict) readings under ellipsis, (2b), and (iii) de se (as opposed to de re) readings, (2c). (2)

a. # b. c. #

They persuaded himi [PROi to gather at the square]. Ii tried [PROi to win] and so did Johnj (try [PRO*i/j to win]) Mary is watching a video, not recognizing herself in it and she is saying that brown eyes would fit that person [i.e. herself] better. Mary actually loves her own blue eyes. Maryi wants [PROi to have brown eyes].

Concerning the choice of the controller among the arguments (i.e. why him and not they determines the reference of PRO in (1a)), I will assume that it is, by default, the argument that is merged with the control predicate immediately after the control constituent is merged, i.e. the one that is the “closest” to PRO (Rosenbaum 1967, Bach 1979, Bach and Partee 1980, Larson 1991, Hornstein 1999).1 This is by no means the only possible approach to controller choice and, as suggested by Landau (2000, 2003), perhaps not even the most generally accepted one. Yet, it will be adequate for the limited set of data discussed here. Having set the basic working assumptions let me now introduce some parameters along which theories of control vary and which will be subject to the test imposed by the newly observed PRO-wh generalization. The dispute most relevant to the present purposes concerns the semantic type of the control constituent. Closely related is then the issue of OC PRO, in particular its semantic type and nature. There are two basic approaches to this issue. The propositional approach assumes that control constituents are semantically propositions, i.e. expressions of type t (or ¢s,t² in an intensional system), and PRO is a variable of type e (usually regarded as an anaphor). The property approach assumes that control constituents are semantically properties, i.e. expressions of type ¢e,t² (or ¢s,et²/¢e,st²; cf. Stephenson 2010), and PRO is either non-existent or reduces to a logical 1

The minimality-based account of controller choice will converge here with a theta-role-based account (Jackendoff 1972, Chierchia 1984), since I will also assume a strict locality condition (or in fact one-to-one head-spec mapping) on socalled theta role assignment.

Radek Šimík

129

lambda-operator (see section 4). Let me give a simple example for clarity. On the proposition approach, the infinitive to remain in the cabin in (1a) denotes a truth-value, i.e. is either true or false (in some situation), depending on the value eventually assigned to the PRO variable. On the property approach, the infinitive characterizes a set of individuals which remain in the cabin (in some situation). Importantly, this dispute touches upon a more general issue in the theory of control, namely whether the control relation is primarily syntactic or semantic. The property theory quite naturally couples with semantic approaches and the proposition theory with syntactic approaches. Let us see why. Remember that the goal of a theory of control is to explain why the empty subject in a control constituent is obligatorily coreferent with some argument of the control predicate.2 For a theory in which control constituents denote properties it is very natural to assume that the coreference relation is a consequence of the lexical semantics of the control predicate. This is because both the denotation of the controller and the denotation of the subject of the control constituent are perfectly accessible to the predicate: the former directly - by being one of the arguments, and the latter indirectly - by being lambda-bound in the representation of its other argument (the control constituent). In effect, the control predicate introduces a predicative relation between the two. A simplified and schematic lexical entry of a control predicate pred under a semantic theory of control is in (3), where P corresponds to the control constituent, x to the controller, and pred’ to the denotation of the control predicate:3 (3)

||pred/property|| = ȜPȜx[pred’(x) & P(x)]

Now, in the proposition approach, there is no chance for the control predicate to resolve the controller-PRO coreference simply by predication.4 The reason is that the variable introduced by PRO, being “buried” inside of the proposition, is inaccessible for compositional manipulation from the 2

Note that I use the term coreference in a non-technical sense, encompassing (accidental) coreference and binding. 3 For the first full-fledged analysis of control predicates along these lines, see Chierchia (1984). See also section 5 of the present paper. 4 Wurmbrand (2002) is of a different opinion. She assumes a proposition analysis of (a subclass of) OC and at the same time a semantic resolution of the controllerPRO coreference. Unfortunately, she does not clarify what the “inherent semantic properties of the selecting (OC) verbs”, allegedly responsible for this coreference, should be.

130

The Property Analysis of Control Constituents

control predicate. A lexical entry for a control predicate under the proposition approach therefore looks like in (4), where p corresponds to the control constituent, x to the controller, and y to PRO. The notation p[y] should be read as a proposition containing a free variable y (i.e. y is not an argument of p, as opposed to (3), where x is an argument of P). (4)

||pred/proposition|| = ȜpȜx[pred’(x) & p[y]]

It follows that some other module than (lexical) semantics must be responsible for the obligatory coreference between the controller and PRO in the proposition approaches. There is wide agreement that this module is syntax (though see Farkas 1988). What submodule of syntax this should be is subject to continuing controversy. The existing accounts are based on (i) a designated control module of the grammar, where PRO has special properties ([+pronominal, +anaphor]) and its reference is determined by special rules (e.g. Chomsky 1980, 1981), (ii) binding, where PRO is considered a subtype of a reflexive anaphor (e.g. Koster 1984, 1987), (iii) movement, where PRO is treated as a trace after A-movement (e.g. Hornstein 1999, 2001) and (iv) agreement, where PRO has to agree with a functional head associated with the control predicate, such that the functional head in turn agrees with the controller (e.g. Landau 2000, 2004). These submodules, in particular binding, movement, and agreement, naturally lend themselves to locality conditions and hence contain a seed of accounting for the relatively restricted occurrence of the phenomenon of obligatory control. Notice that no reference to syntactic locality is needed in the semantic approach, where the relevant restriction is captured by general principles of semantic compositionality: PRO, mapping to a lambda-operator, must be at the edge of the control constituent in order for the control predicate to be able to establish the coreference between PRO and its argument (the controller). The last parameter to be considered is syntactic and carves out two subclasses within the class of property theories. These correspond to two ways of arriving at the property semantics of the control constituent. In the first case, PRO is an operator that undergoes operator movement, much like relative clause operators, for instance (e.g. Aoun and Clark 1985, Clark 1990). This movement then maps onto lambda-binding of the trace and the control constituent - presumably a CP - maps onto a property. In the second case, which is entertained more often, PRO does not exist at all (e.g. Bresnan 1978, Chierchia 1984, Culicover and Wilkins 1986, Jones 1991, Boškoviü 1996, Babby 1998). The control constituent is a subject-

Radek Šimík

131

less VP, mapping to a property in the semantics.5 These two types of property approaches to control are schematically illustrated below: (5)

a. b.

PRED [CP OPi … [vP ti [VP …]]] PRED [VP …]

In summary, I considered three parameters that have shaped the landscape of theories of obligatory control. They are (i) the semantic type of the control constituent (proposition vs. property), (ii) the module in which control is established (syntax vs. semantics), and (iii) the very existence of PRO (yes or no). We will see that the PRO-wh generalization most clearly relates to the first parameter and provides an argument in favor of the property theory of control constituents. To the extent that the property analysis entails something about the second parameter, the PRO-wh generalization also supports the semantic resolution of control. With respect to the third parameter, the generalization provides an inconclusive support for the existence of PRO and hence - in combination with the preceding two - its operator-hood.

3. Types of MECs and the PRO-wh generalization Modal existential wh-constructions are primarily infinitival and secondarily subjunctive.6 Their subject is mostly empty. Consider an example from Spanish: (6)

Tengo con quién e hablar. have.1sg with who speak.inf ‘There is somebody I can speak with.’

What is the nature of the MEC subject (marked as e in (6))? A crosslinguistic investigation reveals that there is no universal constraint on MEC subjects; the whole range of subject types are compatible with MECs - a trace after raising, obligatorily or non-obligatorily controlled PRO, pro, as well as overt nominative-marked subjects. Most languages 5 A similar division can in principle be applied to the proposition approaches, though it is not really attested. In virtually all proposition approaches, the control constituent is a CP (or a TP). The movement of PRO is then not motivated by its operator-hood (though see Clark 1990 for a notable exception), but rather by some formal requirements, relating to government, case-checking, or agreement. 6 For a detailed description of the cross-linguistic distribution of the infinitive and subjunctive mood in MECs, see Šimík (2011:Ch2).

The Property Analysis of Control Constituents

132

can make use of more than one of these strategies, depending on various factors, such as the mood of the embedded verb or the type of the embedding predicate. Let us illustrate these types one by one.

3.1. Types of MECs with respect to their subject o

Raising MECs

That an MEC empty subject can be a trace after subject raising was first observed by Ceplová (2007) for Czech. The clearest argument for this comes from the possibility to use impersonal predicates in MECs, as in (7a). Both the examples below are therefore unified under an analysis where the subject - an empty expletive in (7a) and Honzík in (7b) - is base generated in the embedded clause, raises into the matrix clause, and agrees with the matrix verb. (7)

a. b.

expl1

NemČlo tady kdy t1 pršet. neg.had.3sg.neut here when rain.inf ‘There was no time when it could rain here.’ Honzík1 si nemČl kde t1 hrát. H.3sg.masc refl neg.had.3sg.masc where play.inf ‘Honzík had nowhere to play.’

Obviously, the overt raising of the empty expletive (and even its very existence) is difficult to prove. However, nothing hinges on it - arguably, overt raising is not obligatory in Czech anyway, as it can be replaced by long-distance agreement. The important point is that the matrix verb nemČlo ‘neg.had’ in (7a) - is not thematically related with what appears to be its subject. Besides Czech, raising MECs also exist in Polish, Slovak, and Slovenian. o

Control MECs

Most languages that make use of infinitival MECs cannot use impersonal predicates in the MEC. This suggests that the selecting predicate cannot be a raising predicate, like in Czech. Consider the Spanish example in (8). Neither the personal predicate tuvo ‘had’ (meant to express agreement with the empty expletive, as in (7a) above), nor the impersonal hubo ‘had’ (displaying default agreement) are capable of embedding a weather predicate.

Radek Šimík

(8)

133

*No tuvo/hubo cuando llover. neg had.3sg/imprs when rain.inf ‘There was no time when it could rain.’

Yet, the reference of the embedded subject in Spanish MECs selected by the verb tener ‘have’ is strictly determined by the matrix subject, suggesting that it is an obligatorily controlled PRO. (9)

Tienesi con qué PROi/*j escribir? have.2sg with what write.inf ‘Do you have anything that you/*I/*one can write with?’

The following examples further confirm that the empty subject qualifies as an OC PRO: it is exhaustively controlled (no partial control is possible), (10a), and ellipsis yields sloppy readings only, (10b).7 (10)

a. b.

*Pabloi aún no tiene donde PROi+ reunirse. Pablo still neg has where gather.inf ‘Pabloi still doesn’t have a place where theyi+ could gather.’ Juani no tiene a quién PROi pedir J. neg has a who ask.inf consejo, y sus amigosj tampoco advice and his friends neither (have who PRO*i/j ask for advice). ‘Juan has nobody to ask for advice and neither do his friends (have anybody that Juan/they could ask for advice).

Control MECs constitute the major MEC-strategy in nearly all the languages that make use of the infinitive mood (e.g. Portuguese, Italian, and Russian). An interesting case are subjunctive MECs in languages that also have infinitival MECs, in particular Czech and Hungarian. In these languages, the embedded subject of a subjunctive MEC must be coreferent with the matrix subject, despite the fact that the embedded context is finite and there should therefore be no formal obstacle for hosting referentially independent subjects such as pro.

7 Testing the availability of de se vs. de re readings is not easy due to the fact that the matrix subject is not an attitude holder.

The Property Analysis of Control Constituents

134

(11)

a.

b.

Péteri van kit {ei/*j / * Anna} küldjön P. be.imprs who.acc A. send.sbj.3sg a postára. Hungarian the post.office.to ‘Peter has somebody who he/Anna can send to the post office.’ Kareli nemČl koho by {ei/*j / * Petr} K. neg.had.3sg who.acc sbj.3 P. pozval na veþeĜi. Czech invite for dinner ‘Karel had nobody who he/Petr could invite for dinner.’

I will assume that (11) represent examples of genuine obligatory control into finite complements and hence use the standard notation PRO for the empty category above. This stance is further supported by the Czech examples in (12), which are analogous to (10) above. (12)

a.

b.

*Kareli nemá kde by PROi+ se K. neg.has where sbj.3 rflx shromáždil(*i). gather.pst.ptcp.sg(pl) ‘Karel has no place where he (and others) could gather.’ Kareli nemá koho by se zeptal K. neg.has who.acc sbj.3 rflx ask.pst.ptcp na radu a Petrj bohužel taky for advice and P. unfortunately also ne (have who PRO*i/j ask for advice). not ‘Karel has nobody who he could ask for advice and unfortunately, Petr doesn’t either (have anybody who Karel/Petr could ask for advice).’

OC into finite structures is not unattested, though typically limited to languages which lack the infinitive mood (e.g. Serbian, Bulgarian, and Greek; see Landau 2004 and the numerous references cited therein). Admittedly, the presently observed cross-linguistic pattern is somewhat unexpected, since languages that normally exhibit OC into finite structures do not do so in MECs (see the next section), while languages without this trait (Czech and Hungarian) do so in MECs. This paradox will have to stay unresolved here. However, notice that it provides a tentative argument in favor of treating control (in MECs) as a semantic phenomenon, in

Radek Šimík

135

accordance with the presently made argument, rather than a phenomenon restricted by syntax. In some sense, these cases of control into finite constituents are akin to Chierchia’s (1989) cases of “control” of overt pronouns by attitude-holders. This means that they could be reanalyzed as containing a pro, which is obligatorily bound by a lambda-operator (effectively an OC PRO, cf. section 4) in the left periphery of the MEC. (See also footnote 12.) 8 o

Non-control MECs

The last type of MECs is one where the MEC-subject is referentially independent of any matrix argument. There are three basic cases to distinguish. Firstly, non-control MECs are the only strategy in languages which only have subjunctive MECs (due to the complete lack of the infinitive in that language), such as Greek: (13)

Den exo ti na foresi i Vassiliki neg have.1sg what sbj wear.3sg the Vassiliki sti jiorti tis. at.the name.day her.gen ‘I don’t have anything that Vassiliki could wear on her name-day.’

Other members of the non-control class are those MECs which are selected by predicates other than ‘be’ or ‘have’, irrespective of the grammatical mood of the embedded predicate (infinitive in (14a,b) and subjunctive in (14c)). Notice that the embedded PRO can be coreferent 8

An anonymous reviewer is wondering whether this type of MECs retains its control behavior in a situation where the MEC is not introduced by a morphologically interrogative wh-operator, but rather by a (relative) complementizer combined with a resumptive pronoun. In my typological survey (see Šimík 2011:Ch2) I did not encounter such a language. Perhaps the closest correlate to what the reviewer is looking for would be Hungarian, in which MECs can be introduced either by interrogative-like wh-operators or by relative operators. The example in (i) (from Anikó Lipták, p.c.) shows that OC behavior is observed even if relative operators (boldfaced) are used. (i)

Nekem van {* küldjél / küldjek} akit I.dat be.imprs who.rel.acc send.sbj.2sg / 1sg a postára the post.office.to I have somebody who you / I can send to the post office.’

The Property Analysis of Control Constituents

136

with one of the matrix arguments, as in (14a), but need not be, as in (14b,c). (14)

a. b. c.

Dei-lhei o que PROi fazer. Portuguese gave.1sg-to.him.cl the what do.inf ‘I give him something (work) that he can do.’ Jai našel þem {PROi/j / tebek} pisat’ Russian I found what.instr you.dat write.inf ‘I found something that I/you/one can write with.’ Karel ještČ hledá kde by se K. still look.for where sbj.3 rflx (jeho kolegové) sešli. Czech (his colleagues) meet.pst.ptcp.pl ‘Karel is still looking for a place where they/his colleagues could meet.’

Finally, there are MECs that are selected by impersonal predicates, typically the existential ‘be’, with no overt controller, as in the Hungarian example (15). Yet, these examples still allow for an analysis where the obligatory control is implicit (to the extent that our definition of OC is flexible enough to include these cases, as e.g. in Wurmbrand 2002). Indeed, the analysis put forth in section 5 will rely on the assumption that predicates like the Hungarian van ‘be:imprs’ involve implicit arguments.9 (15)

Van kivel beszélni. be.imprs who.instr talk.inf ‘There is somebody who one could speak with.’

This finishes a brief excursus into the variation in (empty) subject realization and interpretation in both cross- and intralinguistic types of MECs. We saw that every type of (empty) subject can appear in at least some type of MEC. Starting with the next subsection, which introduces the PRO-wh generalization, I will zoom in onto the behavior of obligatory control MECs.

3.2. The PRO-wh generalization The core empirical contribution of this paper is embodied in the following generalization: 9

In some languages, e.g. Hungarian and Russian, the relevant argument can be expressed overtly; it typically surfaces with dative marking.

Radek Šimík

(16)

137

The PRO-wh generalization Obligatorily control MECs are in complementary distribution with MECs containing wh-subjects.

In other words, it is possible to disrupt the obligatory control relation between the matrix and the embedded subject but only if the embedded subject takes a wh-form. As I show below, the PRO-wh generalization cuts across a number of typologically quite different languages (Slavic, Romance, Hungarian, Hebrew) as well as the two verbal moods that MECs can make use of (infinitive, subjunctive) - just as long as it targets OC MECs. Perhaps the most straightforward illustrations of (16) come from Czech and Hungarian subjunctive MECs, where neat minimal pairs of OC MECs vs. MECs with wh-subjects can be formed. This is because both OC PRO and nominative wh-subjects can be licensed by the subjunctive mood in MECs. The examples from Czech in (17) show that despite the acceptability of a subject disjoint from the matrix subject, (17b), this disjointness obtains always and only if the subject is a wh-word, cf. (17c).10 (17)

a.

b.

10

Kareli nemČl koho by PROi pozval K. neg.had who.acc sbj.3 invite na veþeĜi. for dinner ‘Karel had nobody who he could invite for dinner.’ Kareli nemČl kdoj by ho K. neg.had who.nom sbj.3 him pozval na veþeĜi. invite.pst.ptcp for dinner ‘Karel had nobody who could invite him for dinner.’

The pattern in (17) is strongly supported by corpus findings (Google, November 25, 2010). Three conditions were tested, corresponding to the types in (17), the factors being the value of the matrix and embedded subject: (i) a congruent condition (1sg+1sg), corresponding to (17a), (ii) an incongruent condition (1sg+wh-subject), corresponding to (17b), and (iii) an incongruent condition (1sg+2sg/3[not wh-subject]), corresponding to (17c). Condition (i) yielded 8240 hits, while out of the first 40, all were relevant examples of MECs. Condition (ii) yielded 28.400 hits, while out of the first 40, 29 were relevant examples of MECs. Condition (iii) yielded 18 hits, none of which were examples of MECs.

The Property Analysis of Control Constituents

138

c.

*Kareli nemČl koho by Petrj K. neg.had who.acc sbj.3 Petr pozval na veþeĜi. invite.pst.ptcp for dinner ‘Karel had nobody who Petr could invite for dinner.’

An interesting situation obtains in languages with infinitival OC MECs, such as Spanish or Portuguese. The illustrations in (18) are from Spanish. These languages allow for an exceptional use of the subjunctive, but just in case this finite mood is introduced to license the nominative on whsubjects, (18b). Again, non-wh-subjects are ungrammatical, (18c). (18)

a. b. c.

No tienesi a quién {PROi ayudar / *ayudesi}. neg have.2sg a who help.inf / help.sbj.2sg You don’t have anybody who you could help.’ No tienesi quiénj te {ayude / *ayudar}. neg have.2sg who you.cl help.sbj.3sg / help.inf ‘You don’t have anybody who could help you.’ *No tienesi a quién (Juani) ayude. neg have.2sg a who (J.) help.sbj.3sg ‘You don’t have anybody who Juan/(s)he/I could help.’

Analogous behavior is observed for other OC-MEC languages with possible variation concerning the mood that exceptionally steps in to license the wh-subjects. Thus, Hebrew, lacking the subjunctive mood, makes use of the future tense form, (19a), while Italian, though in possession of the subjunctive mood, makes use of the present indicative, (19b). (19)

a. b.

yeš l-a mi še-yišmor al ha-yeladim exists to-her who rel-guard.fut on the-children ‘She has somebody who could take care of the children.’ Maria ha già chi le cura Maria has already who her.cl take.care.of.3sg i bambini. the children ‘Maria already has somebody who could take care of her children.’

Details aside, the empirical pattern entailed by the PRO-wh generalization is clear and cross-linguistically robust and as such calls for an explanation.

Radek Šimík

139

4. Capturing the PRO-wh generalization: the core idea The central question posed by the PRO-wh generalization is: What makes OC PRO and wh-subjects behave on a par and differently than all other types of subjects? The answer that I put forth in this paper is simple: Both OC PRO and wh-words map to logical lambda-operators (in their derived positions) - a type of mapping generally assumed to be excluded for other types of DPs, including referential (RE) and quantificational expressions (QE), but also non-obligatory controlled PRO and other types of pronominal variables. The LF-semantics mapping proposed here is in (20). The first column contains a schematic LF syntax of some embedded (clausal) category XP, containing/being introduced by a particular DP, the second one a simplified logical formula that XP maps to, and the last one the domain of objects that the formula is a member of. (20)

a. b. c. d. e.

||[XP wh-DP … ]|| ||[XP OC PRO … ]|| ||[XP QE … ]|| ||[XP … RE … ]|| ||[XP … NOC PRO/pro … ]||

= Ȝx[P(x)] = Ȝx[P(x)] = ȍx[Q(x) *P(x)] = P(c) = P(x)

 D¢e,t²  D¢e,t²  Dt  Dt  Dt

The idea is that wh-words (20a) and OC PRO (20b) undergo operator movement to the edge of some XP, serving to lambda abstract over a variable (their trace). They literally map to a lambda-operator and hence have no semantic type. The variable they bind is restricted by the descriptive content of the wh-word in the first case and unrestricted in the case of OC PRO.11 The resulting formula is of a functional type - denoting a function from individuals to truth-values/propositions. Quantificational expressions (20c), while also undergoing operator movement (quantifier raising), do not leave their syntactic sister intact, but rather take it as their argument, yielding a truth-value/proposition. They are of type ¢et,t², as standardly assumed. (Replace ȍ by a any quantificational determiner and * by the particular relation between propositions that it entails; e.g., if ȍ is every (), then * is the implication relation o; Q is the restriction of the quantifier and P is its scope.) Referential expressions (20d) and various pronominals (20e) are entity-type expressions and thus can be interpreted 11

The precise manner in which the wh-word contributes the variable restriction is immaterial here. It could either be done via a run-of-the-mill presupposition or by construing their trace as definite descriptions (see e.g. Rullmann and Beck 1998, Sauerland 1998, Johnson in press).

140

The Property Analysis of Control Constituents

in situ. The only difference between the two is that REs denote constants (c), while pronominals denote (free) variables (x). In both cases, the whole XP in which they appear is of a truth-value/propositional type.12 It follows from (20) that wh-clauses (before the application of further operators, e.g. the iota-operator in free relative clauses) and OC constituents are fundamentally different from other embedded clauses. The difference is formulated in the standard property vs. proposition dichotomy. Now, suppose that control predicates always select for property-type expressions, as proposed by Chierchia (1984) (but already assumed in Montague 1973, Bach 1979, and others). Then, in conjunction with (20), we have a handle for understanding why OC PRO and wh-subjects are the only acceptable subjects in MECs, in particular those of the “control type” (see section 3.1). Other types of subjects simply fail to deliver the typetheoretical semantics appropriate for the MEC-selecting predicate. Before I turn to a detailed analysis of control MECs, it should be noted that (20a), i.e. the particular syntax-semantics mapping of wh-clauses, is by no means the only possible one. This logical-lambda construal of whexpressions was utilized by Groenendijk and Stokhof (1984) in their dissertation on questions and also by Heim and Kratzer (1998) in their formal semantics textbook. Perhaps the strongest argument supporting this position comes from the use of wh-expressions as relative operators in headed relative clauses, where virtually no other analysis is well imaginable. Yet, other proposals abound for other uses of wh-expressions, particularly wh-questions and free relative clauses. The analysis that is closest in spirit to the logical-lambda analysis is the one of Caponigro (2003), who assigns wh-words the type of identity functions. Also under his approach, wh-constituents end up denoting properties and to the extent that it is empirically equivalent, it could also be used here.13 There are three other widely used approaches to the semantics of wh-expressions. One treats them as existential quantifiers, whether ordinary ones (Karttunen 1977, May 1977) or dynamic ones (Haida 2007). Another takes them to be simply Heimian individual variables (Berman 1991, Beck 2006) and yet another uses a set-of-individuals construal (Hamblin 1973, Kratzer and Shimoyama 2002). In none of these approaches can a typetheoretical difference between wh-clauses/control constituents on the one 12

Some complements of attitude predicates arguably constitute an exception to (20e), as they denote properties rather than propositions (see Chierchia 1989). In that case, the pronominal functions as a sort of resumptive pronoun, being lambdabound at the edge of the complement. 13 See Šimík (2011:125ff.) for one empirical reason to adopt the logical-lambda approach rather than Caponigro’s identity-function approach.

Radek Šimík

141

hand and all other constituents on the other be naturally postulated: all denote propositions (or sets of propositions in Hamblin semantics). Therefore, to the extent that the present argument is valid, it also provides an argument for a property-construal of wh-clauses and, by transitivity, for a syncategorematic logical-lambda treatment of wh-words (in their derived positions). In summary, I argued that it is possible to draw a type-theoretical line between wh-clauses and obligatory control constituents on the one hand and all other types of embedded clauses on the other. This is the case if the former denote properties and the latter propositions. These analyses have been proposed on independent grounds, for all types of constructions involved, though rarely (if ever) in this particular conjunction. In the next section, I turn to a particular implementation of the property analysis of control MECs.

5. A detailed analysis of control MECs and of the PRO-wh generalization The general idea laid out in the previous section is that MECs, at the level of derivation at which they are selected by the control predicate, must denote properties. This ensures that the only possible constituents that can be “fed” into the relevant predicate are constituents with an OC PRO or a wh-word in their left periphery, thus accounting for the PRO-wh generalization introduced in section 3.2. The particular implementation of this general idea, however, is not as straightforward as one could wish. The main issues we face concern (i) what I call the two-gap problem and (ii) the identity of the very control predicate. The first issue is illustrated in the schematic pseudo-English example (21). Under the present analysis, run-of-the-mill control MECs have two operators at their edge - the wh-operator and the OC PRO. Since both have the semantics of a logical lambda operator, the resulting expression is of type¢e,et², rather than¢e,t², as envisioned in section 4. (21)

a. b.

Ii have [MEC with whom PROi to speak] pseudo-English ||with whom PRO to speak|| = ȜxȜy[can.speak.with’(y,x)]

A simple-minded modification of the proposal in favor of the ¢e,et² type will not address the problem fully, however. The reason is that for multiple-wh MECs, i.e. MECs with more than one wh-word in their left periphery, the type would have to be modified again.

The Property Analysis of Control Constituents

142

(22)

a. b.

Ii have [MEC with whom about what PROi to speak]. ||with whom about what PRO ȜxȜy[can.speak.with.about’(z,y,x)]

pseudo-English to

speak||

=

The issue is therefore more general and will inevitably include the problem of distinguishing wh-words from PRO. The problem of the control predicate itself has two parts. Firstly, the question is why only stative predicates like ‘be’ and ‘have’ exhibit OC, while dynamic ones such as ‘find’, ‘buy’, etc. are NOC (see section 3.1).14 The second question is how exactly the control argument is introduced, esp. in such an impoverished argument structure as the existential ‘be’ presumably has, and what the structural relation is between this argument and the OC PRO. In what follows I propose a solution in terms of a covert applicative predicate being hosted/licensed by the overt predicate that selects the MEC, be it ‘be’/‘have’ or the dynamic predicates ‘find’, ‘buy’, etc. The applicative head functions as a control predicate, having the properties envisioned in section 4. Non-subject wh-words move only after the applicative predicate is “discharged”, finishing the derivation of a MEC, which is then selected by a predicate of existence (‘be’, ‘have’, or a covert predicate in the complex structure of dynamic predicates). After introducing this analysis in section 5.2 and demonstrating how it accounts for both problems mentioned above, I will turn to the cases where PRO is “replaced” by a wh-subject. These will be tackled in section 5.3. But for the start, let us provide a baseline analysis of MECs, forgetting about their control property for a while.

5.1. The baseline analysis of MECs In Šimík (2011), I argued that all MECs (irrespective of the typology in section 3.1) are selected by one and the same predicate, modulo some variation which is irrelevant for the present purposes. In this paper, I will follow the proposal in its essentials and refer the reader to the thesis to

14

An anonymous reviewer is wondering whether infinitival questions embedded under predicates like ‘find/figure out’ exhibit OC or NOC. At least in Czech, these exhibit NOC (and hence behave in a similar fashion as dynamic MEC-selecting predicates). However, I am somewhat skeptical to providing a unified analysis of (non-)control in MECs and embedded infinitival questions, as these two construction types are different in many respects (see Šimík 2011) for discussion.

Radek Šimík

143

resolve potential unclarities.15 The predicate that selects MECs, call it BE, characterizes a relation between some individual and a (minimal) situation/event in which this individual “exists” in a broad sense (is present, available, visible, etc.). BE can either be spelled out as ‘be’ or ‘have’, depending on the language and possibly a number of other factors, or it can be spelled out together with more complex predicates (I assume lexicalization of non-terminal nodes, see e.g. Ramchand 2008, Caha 2009, Starke 2011), such as ‘find’ or ‘buy’, essentially playing the role of these predicates’ result state. The simplified lexical entry for BE, i.e. its syntax, semantics, and phonology is given in (23) (in a form of a ¢syn; sem; phon² triple). This predicate can be phonologically realized as the existential ‘be’. In the case of (24), the lexical entry for the process of finding, there is no phonological realization available (marked by the emptyset in the third element of the triple). The morpheme ‘find’, by assumption, corresponds to the complex predicate FIND+BE, whose lexical entry is provided in (25). It expresses the process of finding (in which some individual x is active) resulting in the state of existence of some y (see also Beck and Johnson 2004).16 (23) (24) (25)

BE = ¢[VP xDP [V’State …]]; e,x[exist’(e) & ș(e) = x]; be/have² FIND = ¢[VP xDP [V’Process …]]; e,x[find’(e) & ș(e) = x]; ‡² FIND+BE = ¢[VP xDP [V’ Process [VP yDP [V’State …]]]]; e,e’,e”,x,y[find’(e’) & ș(e’) = x & exist’(e”) & ș(e”) = y & e = e’ + e”]; find²

Notice that there are argument placeholders (xDP and yDP) in the syntactic entry and corresponding existentially quantified variables in the semantic entry. This is a simplification reflecting the fact that the actual syntactic/semantic values of these placeholders are irrelevant for the conventional association between the three modules encoded in the lexical entry; in the actual syntax and semantics, these placeholders are replaced by actual DPs with whichever semantics they happen to have. The three dots within the (most embedded) VP hint at the possibility to “extend” the 15

In virtue of simplicity, I conflate worlds and events into a single type of situations in this paper, construing events as minimal situations (see Kratzer 2008). 16 The semantic format and the syntax-semantics mapping is inspired by Ramchand’s (2008) constructivist approach to event semantics. Here, I introduce an insignificant simplification by replacing Ramchand’s asymmetric causal “leads-to” relation between subevents (o) by a simple Link (1983)-style operator (+), assuming that the asymmetry between subevents can be read off directly from the syntax.

The Property Analysis of Control Constituents

144

predicate by inserting more structure into its complement. This possibility is what underlies the creation of complex predicates such as FIND+BE and, as we will see shortly, also the possibility to “incorporate” MECs into the event structure of BE. If the MEC is present, it is inserted into the complement of BE and the whole structure characterizes a complex situation of a state of existence of some object making it possible to “do” something with that object. With the notable exception of Spanish “headed” MECs (see Šimík 2011:260ff. for discussion), the external argument of BE remains implicit in MECs and is existentially quantified over. Consider the pseudo-example in (26a) and its structural description in (26b). (26)

a. b.

There is who to visit. ‘There is somebody who one can visit.’ [VP‡i [V’ BE [XP whoi PRO to visit]]]

pseudo-English

The syntax-semantics mapping of (26) is spelled out in detail in (27). The MEC, as shown in (27b), denotes a property concept (type ¢s,et²), a set of situation-individual pairs. Notice that the wh-movement in MECs corresponds simply to lambda-binding, in accordance with the assumptions introduced in section 4. (The intensional construal of the property semantics, used here to capture the modal nature of MECs, does not affect the general point.) BE takes this MEC as its internal argument, while its external argument (the individual whose existence is predicated) is “backgrounded” (removed from the syntax) and existentially quantified over.17 This existentially quantified variable is also “fed” into the open individual argument position of the MEC, created by wh-movement. Notice that BE also introduces the modal existential quantifier, ranging over a set of situations circumstantially accessible from e’. The primary circumstance restricting the quantifier is the most salient one, namely the existence of x. The truth-conditions are derived by feeding (27b) into (27a), the result of which is shown in (27c). The sentence is true iff some individual x is in the state of existence (in the broad sense), characterizing the situation e’, and there is some circumstantially accessible situation e”, in which somebody visits x. (27)

17

a.

||BE|| = ȜQȜee’,x[exist’(e’) & ș(e’) = x & e”A(e’)[Q(e”)(x) & e = e’+e”]]

Alternatively, the position is filled with a phonologically empty property-type nominal, preserving the existential quantification over the individual variable it introduces. See Šimík (2011) for discussion.

Radek Šimík

b. c.

145

||who PRO to visit||= ȜeȜx[visit’(e) & Ag(e) = y & Th(e) = x] ||(26)|| = ||BE||(||who PRO to visit||) = e,e’,x,y[exist’(e’) & ș(e’) = x & e”A(e’)[visit’(e”) & Ag(e”) = y & Th(e”) = x & e = e’+e”]]

There are two provisos to take into account, both included for expository reasons at this point. Firstly, the situation variable introduced by BE is existentially closed, whereas normally it would either be bound by a higher predicate (such as FIND) or by an aspect or tense head. Secondly, the embedded subject is treated essentially as a NOC PRO and represented as an existentially closed variable. A proper analysis of this PRO is what I turn to now.

5.2. Introducing the control predicate Having laid out the basic syntactic and semantic analysis of MECs, let us now move on to the analysis of control. The core proposal made here is that the control predicate is not BE itself, but rather a separate atomic predicate - call it FOR. This predicate is a kind of low applicative head, expressing the state of profiting/benefiting from some event (namely the event that appears in its complement). The implementation of this idea is fully incorporated into the presently assumed system of assembling complex predicates from atomic event predicates. In particular, the idea is that the MEC-embedding predicate BE always comes coupled with the abstract predicate FOR (similarly as the process FIND always comes with the result state BE). Compare the lexical entry in (23) above with the amended one in (28).18 (28)

18

BE+FOR = ¢[VP xDP [V’ State1 [VP yDP [V’State2 …]]]]; e,e’,e”,x,y[exist’(e’) & ș(e’) = x & profit’(e”) & ș(e’’) = y & e = e’+e”]; be/have²

FOR is analyzed as denoting a subevent in the complex event structure of MECselecting predicates. An anonymous reviewer wonders whether there is any empirical evidence for this type of analysis and, if not, whether a potentially simpler analysis, using a non-eventive applicative, is not sufficient. At present, I cannot think of any empirical evidence supporting the subevent analysis - it remains motivated purely conceptually and theory-internally. The reviewer’s suggestion is therefore much appreciated; yet, a potential reformulation along the suggested lines will have to be postponed for another occasion.

The Property Analysis of Control Constituents

146

In order to illustrate how (28) works, let us revisit our simple example (26), this time modified in such a way that the selecting predicate can host an overt controller such as Mary in (29a). In (29b) we can see the proposed structural description of (29a). Notice that Mary is basegenerated in the argument position of the applicative predicate FOR and subsequently gets formally licensed in the matrix TP area.19 (29)

a. b.

Mary has who to visit. pseudo-English ‘Mary has somebody who she can visit.’ [TP Mary1 [VP1‡j [V’ BE [VP2 who2/j [VP2 t1/i [V’ FOR [XP PROi to visit t2]]]]]]]

The derivation of the truth-conditions of (29) is given in (30). Concentrate first on the semantics of the applicative predicate FOR in (30a). Notice that its semantic profile is essentially one of a control predicate in property theories of control (see section 2), modulo event semantics, in that it takes two arguments, a property (Q) and an individual (x) and attributes the property to that individual. In this particular example, FOR takes the control constituent XP as its internal argument, while such constituents, as proposed in section 4, have the same semantics as wh-clauses (cf. (27b) above and (30d) below), namely one of a property concept, where PRO corresponds to a lambda-operator binding the subject variable (the introduction of the lambda-binding could, but need not be achieved by operator/PRO-movement; cf. the discussion in section 2). The external argument is Mary. The semantics of the complete VP is in (30c) and characterizes a situation in which Mary profits from visiting somebody. The resulting constituent is targeted by wh-movement, which corresponds to lambda-binding of the trace.20 The rest of the derivation proceeds just like in (27) - BE (defined in (27a)) takes the wh-clause as its argument and the result (after existentially closing the event variable) is true iff some individual x is in the state of existence (in the broad sense), characterizing 19

In (29b) as well as the structural descriptions to come I use the label V (verb) in a somewhat loose fashion - essentially as a syntactic correlate of any subevent (a cover label of Ramchand’s 2008 init, proc, res, and possibly other types of syntactic categories within the decomposed verb syntax). 20 I adopt a considerably freer theory of wh-movement than usually assumed. In particular, there is no CP or any other projection to host the target of the whmovement, there are no wh-features to check, wh-movement is only regulated by locality and by interface/interpretability conditions. See Šimík (2011:Ch5), but also Pancheva (in press) and Pancheva and Tomaszewicz (2011), for an empirical justification of such a theory.

Radek Šimík

147

the situation e’, and there is some circumstantially accessible situation e”, in which Mary profits from visiting x. (30)

a. b. c. d. e.

||FOR|| = ȜQȜxȜee’, e”[profit’(e’) & Exp(e’) = x & Th(e’) = e”& Q(e”)(x) & e = e’+e”] ||PRO to visit|| = ȜeȜy[visit’(e) & Ag(e) = y & Th(e) = x] ||Mary FOR PRO to visit|| = [||FOR||(||PRO to visit||)](||Mary||) = Ȝee’, e”[profit’(e’) & Exp(e’) = mary & Th(e’) = e” & visit’(e”) & Ag(e”) = mary & Th(e”) = x & e = e’+e”] ||who Mary FOR PRO to visit|| = ȜeȜxe’, e”[profit’(e’) & Exp(e’) = mary & Th(e’) = e”& visit’(e”) & Ag(e”) = mary & Th(e”) = x&e = e’+e”] ||(29)|| = ||BE||(||who Mary FOR PRO to visit||) = e,e’,x[exist’(e’) & ș(e’) = x & e”A(e’), e”’,e””[profit’(e”’) & Exp(e”’) = mary & Th(e”’) = e”” & visit’(e””) & Ag(e””) = mary & Th(e””) = x & e = e’+e”& e” = e”’+e””]]

This analysis is supported by a number of arguments. I would like to mention three of them, drawing from Russian facts. Let us come back to the structural description of control MECs under the present analysis, the general schema of which is repeated below. (31)

[TP … [VP1‡j [V’ BE [VP2 wh2/j [VP2 subji [V’ FOR [XP PROi … t2]]]]]]]

I have assumed that the matrix subject, i.e. the controller of the embedded PRO, is generated in a low position and is licensed by A-movement to the TP domain. In Russian, the TP domain of the MEC-embedding predicate est’ ‘be’ does not possess the right features to license the subject, in other words, the matrix predicate is impersonal. Yet, Russian MEC subjects can be formally licensed in their base position - by dative marking - and the movement that they undergo is a non-feature driven satisfaction of the EPP requirement (see Bailyn 2004 and the references therein): (32)

{Emu / *on} est’ s kem ostavit’ he.dat / he.nom be.imprs with whom leave.inf detej. children ‘He has somebody with whom he can leave the children.’

The Property Analysis of Control Constituents

148

The EPP requirement on T in Russian, however, can also be satisfied by other constituents than the logical subject and in such cases, the subject is predicted to stay in situ. Interestingly, we expect the subject to follow the wh-word in these cases, as the wh-word adjoins to the VP that hosts it. The following examples come from Livitz (2010). Notice that the subject mne ‘me:dat’ in (33a) cannot just be an additional subject in the infinitival, since adding another dative-marked subject results in ungrammaticality, as witnessed by (33b). 21 (33)

a. b.

Zdes’ est’ þto mne here be.imprs what me ‘I have something here that I can wear.’ *Mne est’ þto tebe me.dat be.imprs what you.dat ‘I have something that you can wear.’

nadet’. wear.inf nadet’. wear

Another supporting argument comes from the fact that the dative controller must be animate (or even human), as shown in (34). (34)

{Kole / #Vetru} zdes’ neþego razrušat’. Kolja.dat wind.dat here neg.what destroy.inf ‘Kolja / the wind has nothing more to destroy here.’

Notice that while this is expected under the present analysis where the dative originates as the argument of FOR and bears benefactive semantics, it remains unexplained under the assumption that the dative is simply an embedded subject, which undergoes (optional) raising to the matrix (along the lines of raising MECs, see 3.1, and as argued e.g. by Babby 2000). Finally, the presently assumed structural description of control MECs could also be held responsible for the impossibility to front (e.g. topicalize) the MEC, along with its wh-word, which is demonstrated in (35a). This is because if the wh-clause fronts, the FOR-part of the complex 21

As also observed by Livitz (2010), the dative subject can be accompanied by a prepositional genitive subject - the canonical expression of possessor in Russian: (i) U menja est’ þto tebe nadet’. at me.gen be.imprs what you.dat wear.inf ‘I have something that you can wear.’ This is not problematic for the present analysis, since the dative controller does not have possessor but rather benefactive semantics/syntax and as such does not block the presence of independent possessors, which can presumably be introduced by enriching the event and argument structure of the existence predicate.

Radek Šimík

149

BE+FOR predicate must move along and the adjacency of BE and FOR, required for the lexicalization of the predicate is disrupted. Crucially, fronting of the infinitival MEC is not ruled out per se, as long as the whword is stranded (presumably along with the FOR predicate). Notice also that fronting of wh-interrogatives, (35c), analogous to (35a), is perfectly grammatical (given the right context). (35)

a. b. c.

*[MEC

ýemu pouþit’sja]1 est’ t1. what learn.inf be.imprs ‘There is something that you can learn.’ [XP Pouþit’sja]1 est’ þemu t1. learn.inf be.imprs what ‘There is something that you can learn.’ [Q ýemu pouþit’sja]1 ja neznaju t1 what learn.inf I neg.know.1sg ‘I don’t know what to learn.’

Before I turn to the analysis of MECs with wh-subjects, I would like to show how the problems pointed out in the introduction to this section can be addressed under the present analysis. The first problem, dubbed the “two-gap problem” was that the embedded predicate must somehow be able to tell apart the two operators at the edge of the MEC: the wh-word and the PRO. The solution offered by the present analysis lies in the decomposition of the matrix predicate into the existence predicate BE and the applicative FOR. The structurally lower FOR selects for the control constituent (whose edge only contains the PRO-operator) and at a subsequent point of the derivation, i.e. after the control relation has been established, BE selects for the wh-clause, introducing the predicate relation between the wh-clause and the (phonologically empty) object whose existence is predicated.22 The empirical part of the second issue is summarized by the pattern in (36). The question was why only stative predicates (‘be’/‘have’) but not dynamic ones (‘find’, ‘buy’, etc.) behave as OC predicates. (36)

a.

Tebei est’ þem PROi/*j pisat’? you.dat be.imprs what.instr write.inf ‘Do you have anything that you/*I/*one can write with?’

22 The case of multiple-wh MECs requires a special treatment. See Šimík (2011:232ff.) for discussion.

The Property Analysis of Control Constituents

150

b.

Tyi našel þem PROi/j pisat’? you.nom found what.instr write.inf ‘Did you find anything that you/I/one can write with?

I believe that the present analysis offers a relatively elegant solution which need not resort to ad hoc stipulations. Consider the following two structural descriptions, corresponding to the sentences above (notice that now ‘find’ corresponds to the complex predicate FIND+BE+FOR): (37)

a. b.

[VP1‡j [V’ BE [VP2 þem2/j [VP2 tebei [V’ FOR [XP PROi pisat’ t2]]]]]] [VP1 tyk [V’ FIND [VP2‡j [V’ BE [VP3 þem2/j [VP2 argk/i [V’ FOR [XP PROk/i pisat’ t2]]]]]]]]

What would have to happen in order for ty ‘you’ in the matrix of (36b) to obligatorily control the embedded PRO? Basically, the argument of FOR, marked as arg in (37b), would have to raise to the argument of FIND. In other words, arg would have to be a trace after raising of ty ‘you’. This, however, would be a movement out of a theta-position into another thetaposition. While it has been proposed that such movement should be allowed (see Hornstein 2001), there seems to be no place/need for such an operation in the present system of control, where obligatory control is established in semantics rather than in syntax (e.g. by A-movement). I conclude that the implicit argument arg in (37b) is a non-obligatorily controlled PRO (semantically an existentially closed variable). Notice that this analysis can directly be applied to cases with no overt controller, such as (38). In these cases, the controller is simply implicit - represented as a NOC PRO.23 (38)

Est’ kuda [VP PROarb/i [V’ FOR [XP PROi be.imprs where.to ‘There is a place where one can go.’

idti]]]. go.inf

In summary, I proposed an analysis of the control predicate as being a part of the MEC-embedding predicate, rather than being the predicate itself. 23 In fact, an analogous analysis can be applied to the class of non-control MECs. In such MECs, the overt subject of ‘be’/‘have’ would not originate in the specifier of FOR (which would simply host a NOC PRO), but rather higher in the structure, similarly to the argument of FIND. (Such subjects would be parallel to the possessor u menja ‘at me’ in the Russian example in footnote 19.)

Radek Šimík

151

Though the analysis comes at the cost of postulating covert structure, it has some welcome consequences, both syntactic and semantic.

5.3. The case of wh-subjects So far, the present analysis accounts for the contrast between (39a) and (39b) (see the examples in section 3.2). As was discussed in section 4 and developed in more detail in the previous subsections, (39b) is ruled out for type-clash reasons: the control constituent is of a propositional rather than a property type. It is now time to say more about (39c) and thus complete the explanation of the PRO-wh generalization. (39)

a. b. c.

Johni has with what PROi/*j to clean it. *Johni has with what for Maryj to clean it. Johni has whoj to clean it.

pseudo-English

The analysis and the background assumptions so far entail that in (39c) the control predicate FOR selects the wh-clause, containing the wh-subject: (40)

[VP1‡j [V’ BE [VP2 Johni [V’ FOR [XP who2/j t2 to clean it]]]]]

Notice that the XP in (40) is of the right semantic type, i.e. ¢s,et², which means it can be selected by FOR and the semantic computation can proceed further (unlike in a case like (39b)). Yet, there are two rather serious problems. Firstly, given the semantics of FOR, John should control who and the sentence should be true just in case John is actually about to do the cleaning. This is obviously wrong. Secondly, BE has no access to the wh-word (and hence to the variable bound by it), so even if the typeclash that now arises between BE and VP2 could be fixed, the whole sentence would mean that there is some individual and that John can do some cleaning. This is also quite wrong. I propose that these semantic issues can be solved by a coercion of the meaning of FOR. The goal is (i) to avoid establishing the control relation between the argument of FOR and the variable bound by the wh-word and (ii) to let the meaning of the wh-word “percolate” one step higher, in order to make it accessible to BE. The coerced meaning of FOR (cf. (30a) above) is given in (41).24

24

The proposed coercion mechanism is closer in spirit to that of Klein and Sag’s (1985) than to the (perhaps) more standard one of Pustejovsky’s (1993). In

152

(41)

The Property Analysis of Control Constituents

||FOR’|| = ȜQȜxȜyȜee’,e”[profit’(e’) & Exp(e’) = x & Th(e’) = e”& Q(e”)(y) & e = e’+e”]

The lambda prefix Ȝx introduces the benefactive argument (John in (40)), while Ȝy corresponds to the percolating wh-word and at the same time fills the entity-type gap in the complement of FOR’ (namely Q). Importantly, the semantic type of Q remains unaffected by the coercion, still allowing us to capture the PRO-wh generalization. Now, with FOR’ replacing FOR, the sentence in (40) has the right truth conditions: it is true iff there is some individual and it is possible that John profits from that individual cleaning something. There are three outstanding questions to be addressed. Firstly, what triggers the coercion of FOR into FOR’? Secondly, how exactly is the whsubject formally licensed, given that it appears in the syntactic environment of OC PRO? Thirdly, what prevents non-subject wh-words from moving to this lower position, utilizing FOR’ instead of FOR, in effect allowing for non-PRO/wh subjects to occur, and ultimately destroying the PRO-wh generalization? In the rest of this subsection I provide a single tentative answer for all these three questions. It is clear that the structure immediately below FOR/FOR’ is intimately associated with formal licensing of subjects. In the run-of-themill cases, this structure (so far designated as XP) will be some sort of a defective TP, capable of licensing PRO (e.g. by checking its Case, cf. Chomsky 1995, Martin 2001). It is obvious that something must be different when the wh-subject is involved. There are two basic cases to distinguish. The first is represented by Spanish, which replaces the infinitive by the subjunctive if (and only if) a wh-subject is present. This seems like a last resort act utilized just to case-license an overt subject. (Notice, once more, that no other overt subject can parasite on this act, as it will not yield an interpretable LF.) If the subjunctive structurally differs from the infinitive, then it is not surprising that the two are also selected by a slightly different predicate. In other words, there is a structural cue for FOR to coerce into FOR’. The second case is represented by Russian. Russian does not rely on a structural change in the complement of FOR, rather, it seems to locate the change in FOR itself - by “shifting” it into an exceptional case-marking (ECM) predicate. The effect of this shift is that the wh-subject in Russian is not nominative-marked (as presumably in Spanish), but rather dative-marked. An interesting consequence of this particular, the coercion does not affect the argument (as in Pustejovsky) but rather the verb/function (as in Klein and Sag).

Radek Šimík

153

shift from an ordinary FOR to an ECM FOR’ is that Russian, as opposed to Spanish, cannot license an overt external argument of FOR’ (recall that BE is impersonal in Russian). The examples in (42) show the relevant contrast between Spanish and Russian.25 (42)

a. b.

tiene quién le escriba. Juan no J. neg has who him write.sbj.3sg ‘Juan has nobody who could write him.’ (*Maše) zdes’ est’ komu rabotat’. M.dat here be who.dat work.inf ‘Maša has/there is somebody who can work.’

Now let us have a look at the structural descriptions of the two sentences above. In Spanish, the wh-word fronts to the edge of the complement of FOR’ and it is the complement itself (presumably the TP/finiteness related structure) that formally licenses this subject. As a result, FOR’ can host its own overt argument, which in turn gets formally licensed in the matrix TP. In Russian, on the other hand, no subjunctive in MECs is available. The structure remains identical and the wh-subject at its edge gets exceptionally dative-marked by FOR’. Because the predicate has already discharged its case-licensing capacity and because there is no caselicensing functional structure in the matrix, its external argument must necessarily be covert and in no need of formal licensing, i.e. presumably a NOC PRO. (43)

a. b.

25

Spanish [TP Juan1[VP1‡j [V’ BE [VP2 t1 [V’ FOR’ [TP[fin]quién2/jT t2 le escriba]]]]]] Russian [VP1‡i [V’ BE [VP2 PROarb [V’FOR’ [XPkomu1/i t1 rabotat’]]]]]

An anonymous reviewer is wondering whether we know that the dative on the wh-word komu ‘who’ in (42b) is not assigned by some structural-case assigner associated with the embedded infinitive mood, as argued e.g. by Moore & Perlmutter (2000), Sigurdsson (2002), and, specifically for Russian MECs, by Babby (2000). While this is in principle possible, such an analysis would leave the fact in (42b) unexplained. If the dative on komu ‘who’ is assigned independently of the dative on non-wh-subjects, then it is unclear why the two cannot co-occur. The hypothesis (advocated by Babby) that all dative subjects in MECs are structural and assigned by the infinitival seems to be refuted by the datapoint in (34). The grammaticality of (42a) in turn suggests that the ban on the co-occurrence of two subjects in Russian MECs cannot possibly be only semantic.

154

The Property Analysis of Control Constituents

Notice that this proposal about the structure of control MECs with whsubjects establishes a very close relationship between the “shifted” structure/form of the MEC and/or the predicate it is selected by and the “shifted” semantics of this predicate, i.e. FOR’. Does the proposal shed any light on the third problem, namely why other than subject wh-words cannot front to the complement of FOR’? I believe it does. Notice that the coercion of FOR into FOR’ is now closely tied to subject-licensing, whether by a structural change of its complement as in Spanish or by a change in the “direction” of case-marking as in Russian. In either case, FOR’ is only motivated if the edge of its complement contains an expression that also must be licensed at that edge. Now, in all other cases, i.e. in non-subject wh-words, this is not the case, the wh-word is always licensed independently and hence, there is no motivation for FOR to shift.

6. Conclusion This paper is a defense of the property analysis of control constituents, as opposed to the more standard proposition analysis. To the extent that the property analysis entails a semantic treatment of control, this paper also supports the latter. The argument comes from a novel observation from the domain of modal existential wh-constructions, which, being replicated fairly steadily across typologically very different languages, gives rise to a generalization, which I called the PRO-wh generalization. This generalization entails a very close relationship between obligatorily controlled PRO and wh-expressions. I argued for a particular explanation of this generalization, which is based on the conjunction of three, previously independently made assumptions: (i) a property-type construal of control constituents (ii) a property-type construal of wh-clauses, and (iii) a strict type-theoretical construal of control predicates. The last assumption was discussed at some length in subsection 5.3, where it was argued that a control predicate retains its type-theoretical construal even if it loses its control abilities due to a structurally motivated semantic coercion.

Acknowledgements This paper is based on parts of my dissertation. I would like to thank Mark de Vries and Jan Koster for their guidance through the writing process. I also received valuable feedback from Gisbert Fanselow, Luis Vicente, and three anonymous reviewers. Finally, I am grateful for the

Radek Šimík

155

comments that I received from the audiences of SinFonIJA 3 in Novi Sad and FDSL 8.5 in Brno.

References Aoun, J. and R. Clark. 1985. On non-overt operators. Southern California Occasional Papers in Linguistics. University of Southern California, 17-36. Babby, L. 1998. Subject control as direct predication: Evidence from Russian. In Formal Approaches to Slavic Linguistics 6: The Connecticut Meeting 1997, edited by Ž. Boškoviü, S. Franks and W. Snyder, 17-37. Ann Arbor, MI: Michigan Slavic Publications. —. 2000. Infinitival existential sentences in Russian: A case of syntactic suppletion. In Formal Approaches to Slavic Linguistics 8: The Philadelphia Meeting 1999, edited by T. H. King and I. A. Sekerina, 121. Ann Arbor, MI: Michigan Slavic Publications. Bach, E. 1979. Control in Montague grammar. Linguistic Inquiry 10, 515531. Bach, E. and B. Partee. 1980. Anaphora and semantic structure. In Papers from the Parasession on Pronouns and Anaphora, edited by J. Kreiman and A. Ojeda, 1-28. Chicago: Chicago Linguistic Society. Bailyn, J. F. 2004. Generalized inversion. Natural Language & Linguistic Theory 22, 1-49. Beck, S. 2006. Intervention effects follow from focus interpretation. Natural Language Semantics 14, 1-56. Beck, S. and K. Johnson. 2004. Double objects again. Linguistic Inquiry 35, 97-124. Berman, S. 1991. On the semantics and logical form of wh-clauses. Doctoral dissertation. University of Massachusetts, Amherst. Boškoviü, Ž. 1996. Selection and the categorial status of infinitive complements. Natural Language & Linguistic Theory 14, 269-304. Bresnan, J. 1978. A realistic transformational grammar. In Linguistic theory and psychological reality, edited by M. Halle, J. Bresnan and G. Miller, 1-59. Cambridge, MA: MIT Press. Caha, P. 2009. The nanosyntax of case. Doctoral dissertation. University of Tromso. Caponigro, I. 2003. Free not to ask: On the semantics of free relatives and wh-words cross-linguistically. Doctoral dissertation. University of California, Los Angeles.

156

The Property Analysis of Control Constituents

Ceplová, M. 2007. Infinitives under ‘have’/‘be’ in Czech. In Czech in generative grammar, edited by M. Doþekal, P. Karlík and J. Zmrzlíková, 31-45. München: LINCOM. Chierchia, G. 1984. Topics in the syntax and semantics of infinitives and gerunds. Doctoral dissertation. University of Massachusetts, Amherst. —. 1989. Anaphora and attitudes de se. In Semantics and contextual expression, edited by R. Bartsch, J. van Benthem and P. van Emde Boas, 1-32. Dordrecht: Foris. Chomsky, N. 1980. On binding. Linguistic Inquiry 11, 1-46. —. 1981. Lectures on government and binding. Dordrecht: Foris. —. 1995. The minimalist program. Cambridge, MA: MIT Press. Clark, R. 1990. Thematic theory in syntax and semantics. London: Croom Helm. Culicover, P. W. and W. Wilkins. 1986. Control, PRO, and the projection principle. Language 62, 120-153. Farkas, D. 1988. On obligatory control. Linguistics and Philosophy 11, 2758. Groenendijk, J. and M. Stokhof. 1984. Studies on the semantics of questions and the pragmatics of answers. Doctoral dissertation. University of Amsterdam. Haida, A. 2007. The indefiniteness and focusing of question wh-words. Doctoral dissertation. Humboldt University, Berlin. Hamblin, C. 1973. Questions in Montague English. Foundations of Language 10, 41-53. Heim, I. and A. Kratzer. 1998. Semantics in generative grammar. Oxford: Blackwell. Hornstein, N. 1999. Movement and control. Linguistic Inquiry 30, 69-96. —. 2001. Move! A minimalist theory of construal. Oxford: Blackwell. Johnson, K. (in press). Towards deriving differences in how whmovement and QR are pronounced. Lingua. Jones, C. 1991. Purpose clauses: Syntax, thematics, and semantics of English purpose constructions. Dordrecht: Kluwer. Karttunen, L. 1977. The syntax and semantics of questions. Linguistics and Philosophy 1, 3-44. Klein, E. and I. Sag. 1985. Type-driven translation. Linguistics and Philosophy 8, 163-201. Koster, J. 1984. On binding and control. Linguistic Inquiry 15, 417-459. —. 1987. Domains and dynasties: The radical autonomy of syntax. Dordrecht: Foris. Kratzer, A. 2008. Situations in natural language semantics. In The Stanford encyclopedia of philosophy, edited by Edward N. Zalta.

Radek Šimík

157

(Spring 2008 edition). . Kratzer, A. and J. Shimoyama. 2002. Indeterminate pronouns: The view from Japanese. In Proceedings of the Third Tokyo Conference on Psycholinguistics, edited by Y. Otsu, 1-25. Tokyo: Hituzi Syobo. Landau, I. 2000. Elements of control: Structure and meaning in infinitival constructions. Dordrecht: Kluwer. —. 2003. Movement out of control. Linguistic Inquiry 34, 471-498. —. 2004. The scale of finiteness and the calculus of control. Natural Language & Linguistic Theory 22, 811-877. Larson, R. 1991. Promise and the theory of control. Linguistic Inquiry 22, 103-139. Link, G. 1983. The logical analysis of plural and mass terms: A lattice theoretical approach. In Meaning, use, and interpretation of language, edited by R. Bäuerle, C. Schwarze and A. von Stechow, 302-323. Berlin: Mouton de Gruyter. Livitz, I. 2010. Distinguishing existentials: Modal possessive constructions in Russian. Presented at Formal Approaches to Slavic Linguistics 19. Martin, R. A. 2001. Null case and the distribution of PRO. Linguistic Inquiry 32, 141-166. May, R. 1977. The grammar of quantification. Doctoral dissertation. MIT, Cambridge, MA. Montague, R. 1973. The proper treatment of quantification in ordinary English. In Approaches to natural language, edited by J. Hintikka, J. Moravcsik and P. Suppes, 221-242. Dordrecht: Reidel. Moore, J. and D. M. Perlmutter. 2000. What does it take to be a dative subject?. Natural Language & Linguistic Theory 18, 373-416. Pancheva, R. 2010. More students attended FASL than ConSOLE. In Formal Approaches to Slavic Linguistics 18: The Cornell Meeting 2009, edited by W. Browne et al., 382-399. Ann Arbor, MI: Michigan Slavic Publications. Pancheva, R. and B. Tomaszewicz. 2011. Experimental evidence for the syntax of phrasal comparatives in Polish. In University of Pennsylvania Working Papers in Linguistics 17: Selected papers from PLC 34, edited by L. Friedman. Penn Linguistics Club, Philadelphia, PA. . Pustejovsky, J. 1993. Type coercion and lexical selection. In Semantics and the lexicon, edited by J. Pustejovsky, 73-96. Dordrecht: Kluwer. Ramchand, G. 2008. Verb meaning and the lexicon: A first phase syntax. Cambridge: Cambridge University Press.

158

The Property Analysis of Control Constituents

Rosenbaum, P. 1967. The grammar of English predicate complement sentential complementation. Cambridge, MA: MIT Press. Rullmann, H. and S. Beck. 1998. Presupposition projection and the interpretation of which-phrases. In SALT 8: Proceedings from the 8th Conference on Semantics and Linguistic Theory, edited by D. Strolovitch and A. Lawson, 215-232. Ithaca, NY: CLC Publications. Šimík, R. 2011. Modal existential wh-constructions. Doctoral dissertation. University of Groningen. Sauerland, U. 1998. The meaning of chains. Doctoral dissertation. MIT, Cambridge, MA. Sigurdsson, H. Á. 2002. To be an oblique subject: Russian vs. Icelandic. Natural Language & Linguistic Theory 20, 691-724. Starke, M. 2011. Towards elegant parameters: Language variation reduces to the size of lexically stored trees. Manuscript. University of Tromso. . Stephenson, T. 2010. Control in centered worlds. Journal of Semantics 27, 409-436. Williams, E. 1980. Predication. Linguistic Inquiry 11, 203-238. Wurmbrand, S. 2002. Syntactic vs. semantic control. In Studies in comparative Germanic syntax: Proceedings from the 15th Workshop on Comparative Germanic Syntax, edited by J.-W Zwart and W. Abraham, 93-127. Amsterdam: John Benjamins.

NONFINAL INFORMATION FOCUS IN SERBIAN: SOME EXPERIMENTAL DATA1 SABINA HALUPKA-REŠETAR

Outline The paper challenges the standard minimalist claim that PF has to pronounce the highest member of a nontrivial chain. The analysis put forward in Halupka-Rešetar 2009, which rests, on the one hand, on viewing [+f] as a lexical feature, and on the other hand, on a version of Boškoviü’s (2008) Pronounce Lower Copy, allowing the pronunciation of a lower copy if doing so does not violate any phonological constraints is put to a test by examining the prosodic properties of senteces uttered by native speakers of Serbian, with the goal of corroborating the assumption that non-final information focus needs to be prosodically marked.

1. The problem The properties of the two types of focus, information focus and identificational focus, which were previously often conflated in the literature, are listed in É.Kiss (1998) as follows: o o

1

identificational focus expresses exhaustive information, whereas information focus merely marks the nonpresupposed nature of the information it carries; certain types of constituents, universal quantifiers, also-phrases and even-phrases, for example, cannot function as identificational focus; but the type of constituents that can function as information focus is not restricted;

Research on the paper was funded by Languages and Cultures in Time and Space, project no. 178002 of the Ministry of Education and Science of the Republic of Serbia.

Nonfinal Information Focus in Serbian: Some Experimental Data

160

o o o o

the identificational focus takes scope; information focus does not; the identificational focus is moved to the specifier of a functional projection; information focus involves no movement; the identificational focus is always coextensive with an XP available for operator movement; information focus can be either smaller or larger; the identificational focus can be iterated, but information focus can project.

Although these conclusions are drawn on the basis of Hungarian and English data, they are attested in other languages as well. Still, some of these claims might sound somewhat rigid to a native speaker of Serbian, a language in which almost any word order variation produces a grammatical sentence (as in some other Slavic languages, e.g. Russian). Specifically, the claim made above about information focus not involving any movement is not in line with the fact that a question like (1) can be answered (at least) by (2a) and (2b): (1) (2)

Ko je video Mariju? who.Nom Aux.Cl.3sg seen.M Mary.Acc ‘Who saw Mary?’ a. MARKO je video Mariju. Mark.Nom Aux.Cl.3sg seen.M Mary.Acc ‘Mark saw Mary.’ b. Mariju je video MARKO.

If information focus does not involve movement, how, and more importantly, why do we get the word orders in (2)? An earlier study of word order variations in Serbian (Halupka-Rešetar 2008) showed that there is even a tendency among native speakers of this language to place the information focus sentence initially (2a) rather than sentence-finally (2b), at least in cases when the subject is the information focus. Rather than assuming that information focus does involve movement, this observation led to the hypothesis that there might be a subtype of information focus, for which the term nonfinal information focus (Halupka-Rešetar 2009) is suggested. This type of focus is not realized in sentence final position, the expected position for new information, which follows the presupposed part of the sentence and carries sentential stress, assigned to it by the NSR (Chomsky & Halle 1968; Cinque 1993; Zubizarreta 1998), which is, at least in Serbian, responsible for licensing

Sabina Halupka-Rešetar

161

information focus, Stjepanoviü (1999) claims. So, in neutral interpretation, focus follows the presupposed material, the order of which is free but restricted to the clause (defocalized phrase scrambling in Stjepanoviü 1999), as illustrated by the ungrammaticality of the following examples: (3)

a.

b.

(4)

a.

b.

Ko si tvrdio da who.Nom Aux.Cl.2sg claimed.Masc that voli Mariju? love.Pres.3sg Mary.Acc ‘Who did you claim loves Mary?’ *Mariju sam tvrdio da Mary.Acc Aux.Cl.1sg claimed.Masc that voli Petar. love.Pres.3sg Peter.Nom ‘I claimed that Peter loves Mary.’ (‘Mary, I claimed, Peter loves.’) Koga si tvrdio da Petar who.Acc Aux.Cl.2sg claimed.Masc that Peter voli? love.Pres.3sg ‘Who did you claim that Peter loves?’ *Petar sam tvrdio da Peter.Nom Aux.Cl.1sg claimed.Masc that voli Mariju. love.Pres.3sg Mary.Acc ‘I claimed that Peter loves Mary.’ (‘Peter, I claimed, loves Mary.’)

So, if the NSR licenses (final) information focus, the question arises what licenses non-final information focus and how we account for it. Namely, in (2b) above, since information focusing does not involve movement to any particular structural position, it is clear that it is not the highest copy of the subject, auxiliary clitic or verb that is pronounced or else the word order would be as in (2a). One possibility is to challenge the standard minimalist claim that PF has to pronounce the highest member of a nontrivial chain - this alone is not a novel idea (see Bobaljik 2002; Boškoviü 2001, 2002, 2004; Pesetsky 1997; Stjepanoviü 1999, 2004, 2007 among others). It has been noted by numerous authors that there are constructions which can be successfully analysed using a version of the copy theory of movement under which movement occurs in the syntax, creating a chain of copies of the moved element, but where PF decides which copy to pronounce. Boškoviü (2001 et seq.) has explored this idea

162

Nonfinal Information Focus in Serbian: Some Experimental Data

in discussing Serbian second-position clitics and has claimed that since clitics cannot occur at the edge of an intonational phrase, in cases when the highest copy of a clitic occurs at the periphery of a phrase, a lower copy of it will be pronounced. Similarly, the third person singular auxiliary clitic je is generated lower than the object clitics, but then it moves above them. However, due to a PF constraint, it has to be pronounced last in the clitic cluster so the lower copy of it will be pronounced: (5) (6)

jei [Agrio dative clitic [Agrdo accusative clitic [VP/AuxP jei ...]]] a. *Dala mi ga. je gave.Part.F Aux.Cl.3sg me.Dat.Cl he.Acc.Cl ‘She gave it to me.’ b. Dala mi ga je.

Wurmbrand & Bobaljik (2005) give an account of Dutch inversion of a verb selecting an infinitival complement and the head of its complement in terms of lower copy pronunciation, claiming that extraposition of the infinitival complement is always optional as far as syntax is concerned. So extraposition creates a chain of copies and it is PF, not syntax proper, that decides which copy receives phonetic instantiation. While it is the default to pronounce the higher copy, nothing prevents PF from pronouncing a lower copy unless doing so interferes with some other phonological constraint. Stjepanoviü (1999 et seq.) has argued the free word order phenomenon observable in SerboCroatian is the result of how the copy deletion mechanism works at PF and has shown that although a subject has to move in overt syntax, a lower copy of it may be pronounced, yielding (7b) in addition to the expected (7a). However, the lower copy receives phonetic instantiation only if the C-NSR,2 which assigns nuclear stress to the most deeply embedded constituent and which in SC applies prior to copy deletion (or at the same time), assigns stress to it, thereby identifying it as the (information) focus of the sentence: (7)

a. b.

2

Petar þita knjigu. Peter.Nom read.3sg.Pres book.Acc ‘Peter is reading a/the book.’ Knjigu þita Petar. book.Acc read.3sg.Pres Peter.Nom ‘Peter is reading a/the book.’

C-NSR (constituent-driven NSR): Given two nodes Ci and Cj, that are metrical sisters, the one lower in the syntactic asymmetric c-command ordering is more prominent (Zubizarreta 1998).

Sabina Halupka-Rešetar

163

The present paper is an extension of the analysis put forward in HalupkaRešetar 2009, which rests, on the one hand, on viewing [+if] as a type of a lexical feature, and on the other hand, on a version of Boškoviü’s (2008) Pronounce Lower Copy, which allows the pronunciation of a lower copy if doing so does not violate any phonological constraints (e.g. the obligatory second position of clitics in Serbian). It is obvious that word forms are generally not lexically specified for being either [–if] or [+if] – whether a word form becomes the information focus of a sentence is clearly not encoded in the lexicon in principle, because if it were, we could have either (8a) or (8b), but not both, contrary to fact (Breul 2004: 42): (8)

a. b.

[Beans][–if] I don’t like [Beans][+if] I don’t like

It is, however, conceivable that the lexicon provides for word forms getting into the numeration supplied with both [–if] and [+if] (or [±if]), with subsequent elimination of one of them. So, lexical items do not enter the numeration specified for [–if] or [+if], but they are specified so at the outcome of the numeration. It is perfectly conceivable that lexical items become specified for [±if] in the numeration if we provide for a mechanism which projects a [±if]-feature on a word form to phrases (for it is phrases that are information focus expressions, not word forms). Assume, therefore (along with Breul 2004) that a (certain type of) pitch accent on a word form is a direct formal reflex of the [+foc]-feature of a focus expression and that some version of the intonational theory of focus projection provides us with the required mechanism of [±foc]-projection from word forms to phrases. This intonational theory of focus projection will not be pursued in this paper, but the assumption is that, depending on the discourse or the communicative intention of the speaker, one or another constituent is chosen to be the information focus – to have [+if] as an interpretable lexical feature (just as the speaker chooses the items from the lexicon, so too he chooses the IF, the only one element which can not be deleted from the answer to the (context) question; cf. German and the fact that any element which is inherently [top] can check the uTop feature of C).3 Support for this idea comes from the fact that any element of a sentence may be the IF of it in the right context (even the functional ones), regardless of its feature composition. Non-final (non-canonical) position 3

That the element in SpecCP is always [top] is not an uncontroversial issue, though, as some suggest this is where expletive elements are/may be merged. See the discussion in Holmberg & Nikanne (2002), Richards & Biberauer (2005), Thráinsson (2007), among others.

164

Nonfinal Information Focus in Serbian: Some Experimental Data

of the IF will be shown to be marked prosodically by being assigned more prominence than final IF (which is, Stjepanoviü claims, assigned nuclear stress by the NSR). To summarize the approach, [+if] is a feature added to a lexical item after it enters the numeration. This feature percolates and at the LF interface, any constituent bearing the feature [+if] can be interpreted as the information focus of the sentence. At the PF interface, a copy of the terminal marked [+if] must carry main stress. Note that in the unmarked case of a sentence with a transitive verb, [+if] on the object enables the object, the VP or even the whole TP to be interpreted as the IF of the sentence. This analysis was challenged using a production test in which the prosodic properties of senteces uttered by native speakers of Serbian were examined, with the goal of corroborating the assumption that non-final information focus needs to be prosodically marked.

2. The experiment Before presenting the details of the experiment two issues need to be pointed out. The first one concerns the status of wh-questions as a means of testing information focus. While the part of the sentence that replaces the wh-element of the context question is standardly taken to be the new information focus, this test has its drawbacks. Namely, in a wider context in which the set of entities from which the answer is selected is already established in the discourse (assume, for example, a situation in which a mother asks her two sons, Marko and Jovan, the question in (1), repeated below) or when the element that answers the wh-word in the question is not in line with the (assumed) expectations of the speaker (e.g. the same example in the context of the first speaker expecting that the answer to the question is Jovan, not Marko), the element replacing the wh-word bears a contrastive interpretation and has a different discourse status than it has in the (intended) context of (1) asking for new information: (1) (2)

Ko je video Mariju? who.Nom Aux.Cl.3sg seen.M Mary.Acc ‘Who saw Mary?’ a. MARKO je video Mariju. Mark.Nom Aux.Cl.3sg seen.M Mary.Acc ‘Mark saw Mary.’

Sabina Halupka-Rešetar

165

It is exactly in order to avoid such a contrastive interpretation that the context was limited to one sentence (the context-question) in the experiment. The second issue regards the fact that the two answers in (2) (and more generally, answers with a final focus and a non-final focus) have the same truth conditions. Following Szabolcsi’s (1981) views, we shall claim that structural focus has an exhaustive feature. This feature is associated with identificational rather than with information focus. Thus, if in the context of question (1), both answers in (2) may be uttered even if Marko was not the only person who saw Mary, i.e. if the set of people who saw Mary is not exhausted by uttering (2a,b) and (2a,b) do not contradict (2c) below, we are indeed faced with an information focus and not identificational focus. (2)

c.

MARKO I JOVAN Mark.Nom and John.Nom Mariju. Mary.Acc ‘Mark and John saw Mary.’

su Aux.Cl.3pl

videli seen.M.Pl

As neither (2a) nor (2b) contradict (2c), we may conclude that the examples analyzed in the paper indeed cover examples of (two types of) new information focus. The experiment that was conducted with 18 native speakers of Serbian in the isolation booth of the Faculty of Philosophy. There were two sections of the test, organized on slides. In both sections, the focus of each utterance was controlled by a preceding question. After a brief oral explanation, in the first section, subjects were presented with a number of context questions and for each question, a set of lexical items they needed to use to form the answer to the question, e.g. (9)

Ko je knjigu who.Acc Aux.Cl.3sg book.Acc ‘Who bought the book?’

kupio?4 bought.M knjigu kupio

je

Jovan

4

The test also included questions with non-fronted objects (e.g. Ko je kupio knjigu?). However, no significant differences were noted in the answers relative to the position of the object in the question.

166

Nonfinal Information Focus in Serbian: Some Experimental Data

Only the answers were to be read out loud. The aim of this part of the experiment was to test how likely native speakers of Serbian are to leave the information focus in sentence final position. Out of the 18 subjects, 3 knew exactly what was being tested and this was obvious from the answers they gave: in the vast majority of the answers in this part of the experiment, they placed the new information in sentence final position and in most cases they took significantly more time to answer a question than the other subjects. (This is not to suggest that the use of final focus correlates with any kind of linguistic expertise or taking more time to answer but rather that the three subjects in question spent (more) time on constructing answers with the ’expected’ final focus.)5 The remaining 15 subjects used non-final information focus mostly in subjectquestions (like (9) above), less frequently with other constituents. The following table sums up the results in terms of the use of the type (final or non-final) and syntactic category of information focus. (Note that sentence 5 asked about the whole TP therefore having a non-final IF was impossible but it was still included in the test): Table 1. Distribution of final vs. non-final information focus relative to syntactic category sentence no. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.

synt. category of IF S VP O S+V TP V O VP S S AdvT AdvDir

final IF 4 16 17 9 5 17 18 4 3 6 17

non-final 11 9 13 12 14 12 1

As the above results show, subjects preferred to use non-final information focus in questions which required new information about the subject (9), the verb (10) and a temporal adjunct (11), whereas they used final 5

The subjects flipped the slides at their own pace which resulted in a considerable difference in the mean time required to do the experiment: 188 seconds for the three expert subjects as opposed to 123 seconds for the other 15 subjects.

Sabina Halupka-Rešetar

167

information focus with equal frequency in cases when the question was about the S+V complex (12): (10)

(11) (12)

Šta je Jovan uradio s what.Acc Aux.Cl.3sg Jovan.Nom done.M with knjigom? book.Inst ‘What did Jovan do with the book?’ Kada Marko ide po platu? when Marko.Nom go.3sg.M.Pres for salary.Acc ‘When is Marko going to pick up his salary?’ Šta üe biti s kuüom? what.Nom will.3sg be.Inf with the.house.Instr ‘What will happen to the house?’

The sentences used here were not minimal pairs of final and non-final information focus, so no definitive conclusions could be reached regarding the difference between the two subtypes. Unlike temporal adjuncts, the preferred type of which proved to be the non-final information focus, the subjects in the experiment were more prone to use new information directional adjuncts in final position, just like they did with objects. This tendency might correlate with the fact that the canonical position of both objects and directional adjuncts is clausefinal, while temporal adjuncts are canonically in a clause-internal position. The second part of the experiment consisted of pairs of questions and answers. On each slide there were two context questions, one followed by an answer with non-final information focus and the other answered by a sentence with final information focus. The lexical elements in the questions were not identical so as not to reveal the goal of the experiment, but the syntactic category of the new information was the same. The subjects were instructed to read the questions silently and read out loud only the answers. Although wh-questions can in principle be used to elicit contrastive focus answers as well as new information focus answers, as pointed out above, the lack of a wider context which could/would have suggested any contrast was taken to ensure that the subjects would interpret and read the answers as containing new information focus rather than contrastive focus elements. An example of the task in question is given in (13) and (14) below:

168

(13)

Nonfinal Information Focus in Serbian: Some Experimental Data

a.

b.

(14)

a. b.

Koga je Crvenkapa who.Acc Aux.Cl.3sg Little-Red-Riding-Hood.Nom posetila? visited.F ‘Who did Little Red Riding Hood visit?’ Baku je grandma.Acc Aux.Cl.3sg Crvenkapa posetila. Little-Red-Riding-Hood.Nom visited.F ‘Grandma she visited.’ (... and possibly others, too.) Šta Marija kupuje na pijaci? what.Acc Mary.Nom buy.3sg.Pres on market.Dat ‘What does Mary (regularly) buy at the market place?’ Marija kupuje povrüe na Mary.Nom buy.3sg.Pres vegetable.Acc on pijaci. market.Dat ‘Mary buys vegetables at the market place.’ (... and possibly other things, too.)

The recordings were analyzed using time waveforms in Sound Forge 9.0 in an attempt to determine whether in the cases where non-final information focus was used this element was indeed prosodically more prominent, i.e. whether it had an audio signal amplitude peak level higher than the final information focus used in the answer to the next question.6 Due to obvious facts, one can not talk about an average signal level as each subject had a different signal level. Therefore, no quantitative analysis of the data will be presented, only a qualitative analysis which shows that the audio signal level (measured at the peak) is higher with non-final information focus than with final information focus. The figures below represent the waveforms for sentences (13b) and (14b) as pronounced by one of the subjects:

6

An anonymous reviewer notes that the unfortunate consequence of the production test aimed at testing prosodic distinctions between the two types of focus is that the potential distinction in intensity is obliterated by sentence prosody (both intensity and pitch decrease at the end of the sentence, irrespective of the position of sentence stress). While the author acknowledges this fact, the difference in intensity between final and non-final information foci, clearly visible on the waveforms of sentences containing them, is too significant to be obliterated, although it is probably considerably reduced.

Sabina Halupka-Rešetar

169

Figure 1. Waveforms for non-final (left) and final (right) direct object information focus (grey area)

Note that in (13b) the direct object information focus ‘baku’ appears sentence-initially, while in (14b) the direct object information focus ‘povrüe’ is in sentence-final position. The audio signal level peak is at 0.4dB in (13b) compared to -8.7dB for (14b). A similar (more or less significant) difference with respect to the waveforms and pitch peaks of sentences with non-final IF as opposed to those with final IF can be observed in the case of all other pairs of sentences (except in the statistically irrelevant 4 cases out of the 90 pairs of sentences).

3. The analysis Next we turn to a possible explanation of the licensing of non-final IF in order to predict when and where it is expected occur. Serbian-type languages, which use stress to mark focus, differ from English-type languages with respect to what they do in syntax, and what at PF. In Serbian, there is overt movement to functional projections in syntax (the direct object moves to SpecAgrOdP, the indirect object moves to SpecAgrOiP, the verb to T/PredP, the subject to SpecAgrSP, etc.) and the mechanism of stress assignment cooperates with copy deletion. Namely, the proposal is that the NSR and the [+if] feature will do for most cases of final information focus (except, e.g. when the subject is the IF), but even if we allow PF not to pronounce the highest member of a nontrivial chain we still need the Shift stress PF rule already mentioned which requires a terminal bearing the [+if] to carry main stress, where the amount of stress assigned by this rule is larger than the stress assigned by the (default) NSR to the final IF. This, coupled with the obligatory feature percolation of [+if] up to the root node and the LF rule of interpreting any constituent bearing the [+if] feature as the IF of the utterance could, on the other hand, overgenerate, so we need an economy condition along the lines of Neeleman & SzendrĘi (2004):

170

(15)

Nonfinal Information Focus in Serbian: Some Experimental Data

a. b.

Minimize the number of prosodic peaks (given the targeted interpretation) Minimize stress shift (given the number of prosodic peaks)

While the former economy condition rests on the usual assumption that in any given sentence only one element is selected from the focus set as the actual focus (although nothing in the theory forces this), the latter one aims at having the stress assigned by the NSR rather than having a shifted stress (which is the case with marked stress position such as the subject position in English). So, in answer to the question in (10), repeated here as (16a), either (16b) or (16c) will do: (16)

a.

b. c.

Šta je Jovan uradio s what.Acc Aux.Cl.3sg Jovan.Nom done.M with knjigom? book.Inst ‘What did Jovan do with the book?’ Jovan je knjigu KUPIO. John.Nom Aux.Cl.3sg book.Acc bought.Masc ‘John BOUGHT the book.’ Jovan je KUPIO knjigu.

While in (16b) main stress is assigned by the (default) NSR to the verb in final position and its lowest copy is pronounced, in (16c) the verb receives phonetic instantiation by the Shift stress rule, making it unnecessary for the NSR to assign nuclear stress, too (by the economy condition above) and the highest copy of the verb is pronounced.7 The bracketing representations below show the derivation of (16b-c): (17)

a. b.

[AgrSP Jovan [AgrS je [TP Jovan kupio je [AgrDOP knjigu kupio je [VP je [VP Jovan kupio knjigu]]]] [AgrSP Jovan [AgrS je [TP Jovan kupio je [AgrDOP knjigu kupio je [VP je [VP Jovan kupio knjigu]]]]

The sketch of the theory presented here hopefully gets round the Inclusiveness principle, possibly the biggest drawback of all feature-based theories of focusing, by assuming that no features are added to the lexical 7

Note that it would be odd to claim that a copy other than the highest or the lowest in a nontrivial chain can be pronounced as this would imply that PF can count.

Sabina Halupka-Rešetar

171

elements as a result of the application of syntactic operations but that the lexicon provides for word forms getting into the numeration supplied with both [–if] and [+if], with subsequent elimination of one of them. So, lexical items do not enter the numeration specified for [–if] or [+if], but they are specified so at the outcome of the numeration. In English-type languages, on the other hand, in which there is no morphological case marking on nouns (and therefore nominals have to occur in particular structural positions to reflect their grammatical function) and no V-to-T movement of lexical verbs, IF has to be pronounced in the structural position which determines its syntactic function. Hence the difference between the following Serbian-English pairs of sentences: (18) (19)

(Who punished Peter?) a. MARY punished Peter. b. *Peter punished MARY. (Ko je kaznio Petra?) who.Nom Aux.Cl.3sg punished.Part.Masc Peter.Acc (‘Who punished Peter?’) a. MARIJA je kaznila Petra. Mary.Nom Aux.Cl.3sg punished.Part.F Peter.Acc ‘MARY punished Peter.’ b. Petra je kaznila MARIJA.

Crosslinguistic variation also predicts the existence of languages in which the highest copy of each element has to be pronounced. English is such a language – this is why the all-focus answer to (20a) is not grammatical, unlike its Serbian counterpart in (21). The theme argument is merged as the complement of the unaccusative verb, i.e. in postverbal position, from where it consequently raises to SpecTP. In English, the highest copy of this argument has to be pronounced, while Serbian also allows the pronunciation of the lower copy of Jovan (20b): (20) (21)

a. b. a. b.

What happened? *Blinked John. Šta se desilo? what.Nom se.Cl happen.Past ‘What happened?’ Trepnuo je Jovan. blinked.Part.Masc Aux.Cl.3sg John.Nom ‘John blinked.’

172

Nonfinal Information Focus in Serbian: Some Experimental Data

c.

Jovan je John.Nom Aux.Cl.3sg ‘John blinked.’

trepnuo. blinked.Part.Masc

4. Conclusions The analysis based on the results obtained in the experiment predicts that the non-final type of information focus can occur cross-linguistically in languages in which word order is relatively free (e.g. Serbian), but not in languages in which word order is governed by rules of grammar (e.g. English). Of course, the issue of scrambling and what exactly it implies is also controversial, since the term is often used for expository convenience to refer to a broad range of different pheomena and has so far been elusive to a consensus definition. However, if scrambling, very generally, refers to the possibilty of rearranging the order of arguments within a clause, a neat correlation becomes obvious: languages in which nouns are morphologically marked for case and V-to-T movement is not restricted to auxiliaries are exactly those that allow both scrambling and the pronunciation of copies other than the highest one in a nontrivial chain. This might suggest that these two phenomena could be actually collapsed into one in the sense that the type of word order variations we observe in languages like Serbian is actually not a phenomenon distinct from other types of A’-movement (specifically focusing and topicalization), but at most a combinaiton of these. In this paper, a small step was made in favour of analyzing one instance of word order variation (scrambling) in terms of the possibility of pronouncing lower copies of particular elements, specifically those which carry the [+if] feature. However, more evidence is needed to make such a bold claim.

References Bobaljik, J. 2002. Ⱥ-chains at the PF-interface: Copies and ‘covert’ movement. Natural Language and Linguistic Theory 20/2, 97-267. Boškoviü, Ž. 2001. On the Nature of the Syntax-Phonology Interface: Cliticization and Related Phenomena. Amsterdam: Elsevier. —. 2002. On certain differences between Bulgarian and Serbo-Croatian C(P). Balkanistica 15, 35-48. —. 2004. Topicalization, focalization, lexical insertion, and scrambling. Linguistic Inquiry 35/4, 613-638. —. 2008. On the Clausal and NP-structure of Serbo-Croatian. In Formal Approaches to Slavic Linguistics: The Toronto Meeting, 2006, edited

Sabina Halupka-Rešetar

173

by R. Compton et al., 42-75. Ann Arbor, MI: Michigan Slavic Publications. Breul, C. 2004. Focus Structure in Generative Grammar. Amsterdam, Philadelphia: John Benjamins. Chomsky, N. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. —. 2000. Minimalist Inquiries: The Framework. In Step by Step: Essays on Minimalist Syntax in Honor of Howard Lasnik, edited by D. Michaels and J. Uriagereka, 89- 155. Cambridge, Mass.: MIT Press. —. 2001. Derivation by Phase. In Ken Hale: A Life in Language, edited by M. Kenstowicz, 1-52. Cambridge, Mass.: MIT Press. Chomsky, N. and M. Halle. 1968. The sound pattern of English. New York: Harper and Row. Cinque, G. 1993. A Null Theory of Phrase and Compound Stress. Linguistic Inquiry 24/2, 239-297. É. Kiss, K. 1998. Identificational focus versus information focus. Language 76, 891-920. —. 2010. Structural focus and exhaustivity. In Information Structure: Theoretical, Typological, and Experimental Perspectives, edited by M. Zimmermann and C. Féry, 64-88. Oxford: Oxford University Press. Halupka-Rešetar, S. 2008. Položaj informacionog fokusa i red reþi u srpskom jeziku. Godišnjak Filozofskog fakulteta u Novom Sadu XXXIII/2, 311-321. —. 2009. The structural articulation of focus in English and Serbian: a comparative analysis of the structural and the generative approach. Doctoral dissertation. Faculty of Philosophy, University of Novi Sad. Holmberg, A. and U. Nikanne. 2002. Expletives, Subjects and Topics in Finnish. In Subjects, expletives, and the EPP, edited by P. Svenonius, 71-106. Oxford: Oxford University Press. Neeleman, A. and K. SzendrĘi. 2004. Superman Sentences. Linguistic Inquiry 35, 149-159. Pesetsky, D. 1997. Optimality theory and syntax: Movement and pronunciation. In Optimality theory: An overview, edited by D. Archangeli and D. T. Langedoen, 134-170. Malden, Mass.: Blackwell. Richards, M. and T. Biberauer. 2005. Explaining Expl. In The function of function words and functional categories, edited by M. den Dikken and C. Tortora, 115-153. Amsterdam, Philadelphia: John Benjamins. Szabolcsi, A. l981. The Semantics of Topic-Focus Articulation. In Formal Methods in the Study of Language, edited by J. Groenendijk et al., 513540. Amsterdam: Matematisch Centrum.

174

Nonfinal Information Focus in Serbian: Some Experimental Data

Stjepanoviü, S. 1999. What do Second Position Cliticization, Scrambling, and Multiple Wh-fronting Have in Common. Doctoral dissertation. University of Connecticut, Storrs. —. 2004. A word order paradox resolved by copy deletion at PF. In Linguistic Variation Yearbook 2003, edited by P. Pica, 139-177. Amsterdam, Philadelphia: John Benjamins. —. 2007. Free word order and copy theory of movement. In The Copy Theory of Movement, edited by N. Corver and J. Nunes, 219-248. Amsterdam, Philadelphia: John Benjamins. Thráinsson, H. 2007. The Syntax of Icelandic. Cambridge: Cambridge University Press. Wurmbrand, S. and J. Bobaljik. 2005. Adjacency, PF, and extraposition. In Organizing Grammar: Linguistic Studies in Honor of Henk van Riemsdijk, edited by H. Broekhuis et al., 679-688. Berlin, New York: Mouton de Gruyter. Zubizarreta, M.-L. 1998. Prosody, Focus, and Word Order. Cambridge, Mass.: MIT Press.

CLITIC PLACEMENT AND THE PROPERTIES OF THE INTONATIONAL PHRASE IN SERBIAN MAJA MARKOVIû AND TANJA MILIûEV

Outline In this paper we examine the properties of the Intonational Phrase (IP) in Serbian, with respect to the generalization that second-position clitics in Serbian are sensitive to the I-boundary, which forces them to be spelled out in lower position. We try to establish what set of properties is required for a constituent to be realized as IP, and consequently delay clitics. We show that in addition to specific phonological properties, IPs are distinguished by semantic/pragmatic properties as well. It is their distinct interpretation that sometimes allows phonological features to be somewhat obliterated in the production/perception of I-boundaries.

1. Introduction In this paper we deal with the properties of the Intonational Phrase (IP) in Serbian, in the light of clitic placement. Our discussion has two aims. First, we wish to tackle the issue of the inventory and the properties of prosodic constituents in Serbian. Second, we try to account for the “exceptional” cases of clitic positioning, where the second-position (2P) effect is violated. Those cases include clitics found immediately after ‘pauses’ (i.e., I-boundary), and the so-called delayed clitics in the apparent absence of the prosodic condition. We will argue that clitics in Serbian are indeed a perfect tool for detecting IP constituents. Whenever delayed or clitic-third orders arise, there is an extra IP present blocking clitic attachment. The optionality of prosodic phrasing of a constituent is only apparent, since elements prosodically phrased as IPs are always associated with distinct semantics. We suggest that the strong correlation with interpretation, coupled with the positional cues from clitic placements allows some of the phonological Iboundary features to be at least partly obliterated.

176 Clitic Placement and the Properties of the Intonational Phrase in Serbian

The paper is organized as follows. In Section 2 we present the details of clitic placement in Serbian, and the two possible accounts of the 2P effect. Section 3 discusses the cases of unexpected clitic delay, on the one hand, and the unexpected 2P effect on the other. In Section 4 we give phonological properties of these ‘ill-behaved’ constituents, based on an experimental study, and try to establish the set of relevant features distinguishing IP in Serbian as a prosodic constituent. Section 5 addresses the issue of dual prosodic phrasing of the same syntactic constituent, specifically focusing on the behaviour of appositives and ‘light’ topics with respect to clitic placement, and the meaning differences that arise with clitic-second and clitic-third orders.

2. SC clitics and the second- position effect Clitics in Serbian (and Croatian; henceforth, SC)1 are famous for exhibiting the second position effect (2P), which can be generalized as the requirement that clitics occur in the second position in the clause. The element providing a host for the clitic can be either the first syntactic constituent/a phrase (1P), or the first phonological word (1W), as illustrated in (1) and (2), respectively (in all the examples clitics will be given in bold face). (1)

a. b.

Ovu knjigu sam kupila u this book AUX bought in ‘This book I bought in Subotica.’ Ovu sam knjigu kupila u Subotici

Subotici. Subotica

(1P) (1W)

The sensitivity to seemingly two different types of constituent - prosodic and syntactic - has given rise to a lot of discussion on the role of prosody and syntax in providing a host for clitics. On purely phonological accounts (Halpern 1995; Radanoviü-Kociü 1988, 1996), it is phonology that provides a host for clitics. Clitics must attach/encliticize to the first phonological (phonologically independent, stress-bearing) word. When there is no such host available, the operation of Prosodic Inversion (Halpern 1995), or PF movement of the first 1

We will use the term Serbo-Croatian in discussing facts and views generally accepted in the literature. We restrict ourselves to Serbian when presenting our own findings. We hypothesize that the same situation obtains in Croatian as well, but we do not want to commit ourselves to making any such claims at this point of research.

Maja Markoviü and Tanja Miliüev

177

available phonological word has to take place to license the clitic, which is in sentence initial position in the syntax. (2)

a. b.

sam ovu knjigu kupila u Subotici Ovu sam knjigu kupila u Subotici

(syntax) (PI)

Syntactic approaches, on the other hand, emphasize that phonological word first (1W) orders are possible only when the syntax provides a suitable host. In other words, they correlate with the possibility of syntactic movement, specifically, the Left-branch extraction (cf. Boškoviü 2001, 2004; Progovac 1996; Wilder & ûavar 1994). The strongest evidence that syntactically immobile elements can never be hosts to clitics comes from the fact that prepositions, even when stressed, and thus phonologically appropriate for PI (as is the case with the preposition prema ‘toward’ in (3)), can never hosts clitics because there is no syntactic movement available for them (they cannot be extracted). (3)

*Prema su Mileni hodali toward AUX Milena walked ‘They walked toward Milena.’ (Progovac 1996; Boškoviü 2001, 2004)

Although syntax is responsible for clitic placement, phonology still plays an important role. Namely, it filters out cliticization after appositives (4), parentheticals (5), heavy constituents and (certain) sentence initial topics (be they heavy (6), or light (7)). Such elements regularly fail to host the clitic, which then has to surface in a lower (clause-third) position.2 (4)

2

Majina tetka, profesorica latinskog, objasnila Maja's aunt professor of-Latin explained ablativ mi je me AUX ablative ‘Maja's aunt, a teacher of Latin, explained ablative to me.’

Clitic-delay with non-heavy topics is largely optional, and often considered to be influenced by the (prosodic) weight of the constituent. More details are provided in Section 6.

178 Clitic Placement and the Properties of the Intonational Phrase in Serbian

(5)

(6)

(7)

Slavna balerina, kako saznaje Politka, famous ballerina as finds Politika predstaviüe nam novom ulogom se introduce us REFL with-new role ‘The famous ballerina, as ‘Politika’ finds out, will introduce herself to us with a new role’ a. Pre nekoliko godina palo mi je before several years fell me AUX na pamet da odem u Španiju on mind to go to Spain ‘A few years ago it occured to me I could go to Spain.’ b. Dve osobe iz Bara juþe je two persons from Bar yesterday AUX privela beogradska policija arrested Belgrade police ‘The Belgrade police arrested two people from Bar yesterday.’ c. Glumica Eva Ras otvorila nam je vrata actress Eva Ras opened us AUX door svog doma u Beogradu of-her home in Belgrade ‘The actress Eva Ras welcomed us to her home in Belgrade.’ a.

b.

Lica ne razaznaje im faces not distinguish them ‘He doesn’t distinguish their faces.’ (Radanoviü-Kociü 1988:107) zadaje puno problema Danima im for-days gives them lots of-problems ‘For days he's been giving them lots of problems.’

There is general agreement that the reason why these particular constituents cause delayed clitic placement is that, presumably, they project their own Intonational Phrase, set off by pauses.3 Intonational Phrases typically correspond to the root clause in the syntactic representation (Emonds 1970), but it has been long noted that constructions such as parentheticals, appositives, appositive relative clauses, heavy constituents, tag questions, 3

Schütze (1994:52) warns that the pause in this context should not be understood as actual phonetic silence. This relaxed use of the term is also assumed here.

Maja Markoviü and Tanja Miliüev

179

and certain preposed elements form a separate intonational phrase (cf. Nespor and Vogel 1986; Selkirk 1984, 1986, a.o.). The sensitivity to the presence of an I-boundary has lead to a slight modification of the domain for clitic placement. Rather than being second in their clause, clitics must occur in the second position of their intonational phrase (cf. Boškoviü 2001, 2004, Halpern 1992; Radanoviü-Kociü 1988; Schütze 1994, a. o.). (8)

*([XP] )IP ([clitic ….])IP

The generalization in (8) makes a clear prediction that all late/clitic-third orders involve a constituent in the first position that has prosodic properties of an IP. However, as elegant the account in terms of prosodic conditions is, it faces some complications that need to be addressed. First of all, not all clitic-third orders seem to involve two separate IPs. Second, not all constituents that are somehow by default associated with an Iboundary trigger delayed clitic placement. We discuss these problematic cases in the following section.

3. ‘Exceptional’ cases The first complication raised by the phonological account of the 2P and non-2P effect is the fact that not all clitic-delaying constituents seem to share the same prosodic properties. Namely, the boundary following fronted constituents, especially the non-heavy ones, is generally perceived as not being the same as the one associated with parentheticals and appositives (cf. ýamdžiü and Hudson 2006:324 for a similar observation regarding heavy and non-heavy topics). In addition, the boundary (#) often seems optional. (9)

a.

b.

Tu plavu haljinu (#) obukla that blue dress wore nekoliko puta several times ‘That blue dress she wore several times.’ Lica (#) ne razaznaje im faces not distinguish them ‘He doesn’t distinguish their faces.’

je AUX

If clitic-third orders can arise even when there are no clear indicators of an I-boundary, this poses a serious problem for (8). Moreover, the potential lack of I-boundary features further complicates the issue of cliticization

180 Clitic Placement and the Properties of the Intonational Phrase in Serbian

after fronted topics, given that the alternation between clitic-second and clitic-third orders in this context is already often perceived as being quite free. The second complication is that clitics can be found immediately after I-boundaries. More precisely, elements supposed to project an IP can sometimes serve as clitic hosts. Thus, it is not uncommon to find cases of cliticization after heavy constituents. This happens most often with the third person singular auxiliary clitic je (cf. the discussion in Schütze 1994:82), although other clitics are possible there as well. Interestingly, clitic-second orders are especially common, even preferred, with singleword predicates (10).4

4

It seems that the presence of other VP material (an adjunct or an object) makes clitic-third orders not only more easily available, but actually preferred to cliticsecond orders. (i)

a.

b.

c.

Problemi o kojima üemo danas razgovariti problems about which shall today discuss veoma neprijatni. su nam very AUX us unpleasant ‘The problems that we shall discuss today are very unpleasant to us.’ Markov brat koji živi u Beogradu Marko's brother who lives in Belgrade ženi u nedelju. se marry REFL on Sunday 'Marko's brother who lives in Belgrade is getting married on Sunday' Markov brat koji živi u Beogradu Marko's brother who lives in Belgrade kupio lep poklon. nam je bought us AUX nice present 'Marko's brother who lives in Belgrade has bought us a nice present'

It should be noted that the restriction on clitic-second orders with single-word predicates should not be understood as a kind of clitic's sensitivity to the clausefinal position (in which it would be found if the predicate were to serve as the host). Rather, what seems to be the problem is the obligatory narrow focus interpretation that the single-word predicate, or more precisely, the verb, gets in this position in the presence of an overt subject, which, of course, may and may not be the desired reading of the sentence. Why this interpretation is obliterated or relativized in cases such as those in (i) is the question we leave for further research.

Maja Markoviü and Tanja Miliüev

(10)

a.

b.

181

Problemi o kojima üemo danas problems about which shall today razgovarati neprijatni su nam discuss AUX us unpleasant ‘The problems that we shall discuss today are unpleasant to us.’ Markov brat koji živi u Beogradu Marko’s brother who lives in Belgrade ženi/ stigao/ se je REFL marry/ AUX arrived/ pomogao nam je us AUX helped ‘Marko's brother, who lives in Belgrade, is getting married/has arrived/helped us.’

While the variability of clitic placement with fronted and heavy constituents has been noted, and to some effect discussed in the literature, the fact that clitic-second order is possible with appositives as well has somehow passed unnoticed. This order is illustrated in (11). (11)

Majina tetka, profesorica latinskog mi je Maja's aunt professor of-Latin me AUX objasnila ablativ explained ablative ‘Maja's aunt, a teacher of Latin, explained ablative to me’

Parentheticals, on the other hand, are only highly marginally acceptable with 2P clitics (12).5 (12)

?*Slavna balerina, kako sazanaje ‘Politika’, famous ballerina as finds Politika predstaviti novom ulogom üe nam se will us REFL introduce with-new role ‘The famous ballerina, as ‘Politika’ finds out, will introduce herself to us with a new role’

Clitic attachment to heavy constituents and appositives clearly challenges the phonological account of the 2P effect. There are essentially two ways 5

Parentheticals show an interesting variability in clitic placement before and after them (cf. Boškoviü 2000, Franks 1998). This probably depends on the adjunction site of the parenthetical.

182 Clitic Placement and the Properties of the Intonational Phrase in Serbian

to reconcile the ‘clitic after pauses’ cases with the prosodic condition. One possibility is to assume that the properties demarcating the I-boundary (‘pause’) in those cases are not the same as those of the well-behaved Iboundary which trigger delayed clitic placement. This line of reasoning is found in Schütze's (1994) discussion of cliticization after heavy constituents. He argues for the existence of two prosodic phrasing algorithms, which insert ‘pauses’ at different places in the derivation, one before clitic placement (prosodic projection), and the other after clitic attachment (readjustment), ‘under certain conditions’.6 ‘Pauses’ found after heavy constituents would then result from the latter process. The other avenue would be to assume that clitics found after otherwise delaying elements are not prosodic clitics. The gradual loss of clitichood has been suggested for the auxiliary clitic je, based on its rather idiosyncratic behaviour, clearly distinguishing it from other auxiliary clitics (cf. Boškoviü 2001, 2004; Schütze 1994; Tomiü 1996). Conceivably, other clitics as well could be at the beginning of such a process, starting to lose their prosodic sensitivity. Clitics after heavy constituents and appositives would be such ex-clitics.7 None of these approaches, however, offers an account of clitic delay in the absence of ‘pauses’ (or, more precisely, where (some of) the relevant Iboundary features are missing), as seems to be the case with non-heavy topics. If indeed some of the boundary features are absent, clitic-delaying topics are a different (‘smaller’) prosodic constituent (cf. Zec & Inkelas 1990 and the claim that (heavy) topics can be phonological phrases, rather than IPs necessarily). With the introduction of this type of variability of prosodic factors, phonology can no longer play the crucial role in the 2P effect, and some other motivation needs to be sought. On the other hand, a way to maintain the correlation between the prosodic condition as stated in (8) and the 2P effect/clitic-delay would be to assume that I-boundary does not have a fixed set of phonetic/phonological features; thus, elements of the same prosodic constituency (IP) could be perceived differently, in certain cases. Hoping that some exact phonetic and phonological evidence would shed some light on the issue, we have conducted a production study of 6 Schütze does not want to commit himself to the IP status of heavy constituents in those cases, but he strongly suggests that 'pauses' normally do not demarcate phiphrases. 7 The fact that clitics are still impossible in the sentence initial position clearly presents a problem for this idea (cf. Boškoviü 2004:67). Of course, one can always assume that the change proceeds gradually, targeting only specific contexts at first, before it is completed and generalized to all relevant contexts.

Maja Markoviü and Tanja Miliüev

183

delayed clitic placement and placement after pauses. The results and the discussion are presented in the following section.

4. Phonetic cues signalling the presence of constituent boundaries: results of the experiment In order to analyze the presence of IP boundaries, we conducted a production experiment in which we recorded three speakers reading sentences with clitic-second and clitic-third orders. The reading material contained all types of constructions given in the examples (4-7). The recorded material was then analyzed in order to find out whether the observed constructions contained the phonological cues commonly associated with IP boundaries. The set of phonological features usually assumed to signal prosodic constituent boundaries include: a) temporal disjunctures or pauses periods of actual silence; b) the presence of boundary tones - in the Serbian language L%, signalling finality, and H%, signalling continuation; c) pitch resets - upward shifts of F0, which are generally assumed to indicate a beginning of a new IP; d) preboundary lengthening – the lengthening of final elements in a prosodic unit (Godjevac 2000, 2005; Ladd 1996; Selkirk 2005; Wang & Hirschberg 1995). Another indicator of a prosodic unit boundary may be the absence of phonological changes, such as assimilations.8 In analyzing the recorded material for this study, we focused on the first three features (temporal disjunctures, presence of boundary tones and pitch resets).9 Although the phonetic realization of the utterances may not be altogether conclusive, certain general tendencies regarding the relation 8

This inventory of prosodic unit boundary cues may not be exhaustive, as there seem to be other signals indicating a boundary. Namely, an ongoing research indicates that the timing of F0 peaks may play a significant role in the perception of boundaries in Serbian (Markoviü & Miliüev 2011). 9 The results of the analysis of potential voicing assimilation in the recorded corpus are rather inconclusive. In several cases where some of the boundary cues were clearly present (except for temporary disjuncture), there was also (unexpected) noticeable voicing assimilation. In some of the cases where it would be expected, it was missing. Voicing assimilation generally exhibited a rather irregular pattern of occurrence. The only obvious correlation found was that it is excluded by temporal disjuncture (pause). If a pause is missing, assimilation may occur even at points of undeniable boundary signals. We believe that the presence of assimilation as an IP boundary signal should be more thoroughly researched, and that is why we did not include it in this study.

184 Clitic Placement and the Properties of the Intonational Phrase in Serbian

between clitic placement and syntactic constituents are noticeable throughout the corpus analyzed. In the examples of clitic-delaying appositives, all three speakers divided the appositive from the rest of the sentence by a considerable break, which included all of the above mentioned IP boundary signals – pause, IP boundary tone and pitch reset. In the examples where the appositive was immediately followed by a clitic, most of the boundary signals were suspended, and the speakers seemed to adjoin the clitic phonologically to the appositive. This can be seen from the wave forms and F0 curves of sentences in (13), produced by the same speaker (Fig.1): (13)

a.

b.

Marija, profesorica latinskog, objasnila mi je Marija professor of-Latin explained me AUX ablativ ablative ‘Marija, professor of Latin explained ablative to me’ Marija, profesorica latinskog, mi je objasnila Marija professor of-Latin me AUX explained ablativ ablative ‘Marija, professor of Latin, explained ablative to me’

Figure 1. Clitic 3rd and clitic 2nd position after an appositive.

Maja Markoviü and Tanja Miliüev

185

The marginally acceptable parenthetical construction analyzed is given in (14). (14)

*?Uzgajivaþi malina, kako izveštava Politka, raspberry growers as reports Politika odbili predlog vlade su AUX refused proposal of-government ‘Raspberry growers, as Politika reports, have refused the proposal of the government.’

Interestingly, although the clitic immediately follows the parenthetical clause, all three speakers retain a significant break, including a silent period (marked in the picture), even though we might expect the break to be called off due to the following clitic. Figure 2. Clitic following a parenthetical.

In the case of fronted elements, boundary signals are not as conclusive. Fig. 3 illustrates the pair of sentences in (15). (15)

a.

U tom predivnom odmaralitštu na rivijeri in that beautiful resort on coast nije smetalo nam ništa us nothing NEG-AUX bother ‘In that beautiful coastal resort nothing was unpleasant to us.’

186 Clitic Placement and the Properties of the Intonational Phrase in Serbian

b.

U tom predivnom odmaralitštu na rivijeri in that beautiful resort on coast ništa nije smetalo nam nothing us NEG-AUX bother ‘In that beautiful coastal resort nothing was unpleasant to us.’

Figure 3. Clitic after a fronted element.

The only difference between the two sentences is in a somewhat more significant rise of the post-topic portion of the sentence with the delayed clitic (‘ništa nam nije smetalo’), as the only signal of an IP boundary. All other boundary features are missing – there is no pause, the reset is absent, and there is no lengthening of the preboundary element preceding the clitic (the example on the right). In the case of light topicalized elements, the break features are also rather obliterated in clitic-delay sentences. In the sentence in (16) there is a clear IP boundary rising tone at the end of the topcalized element Sapiru (Fig. 4). (16)

Sapiru jezik instrumenat je to-Sapir language is instrument ‘For Sapir language is an instrument’

Maja Markoviü and Tanja Miliüev

187

Figure 4. Clitic after a light topicalized element.

In the case of a heavy constituent, with 2P clitics and single-word predicates, as in (17), most of the boundary cues are also absent: the heavy subject shows a downstepping tendency, and the predicate, beginning with a clitic, continues the same trend. (17)

o kojima üemo danas razgovariti Problemi problems about which shall today discuss neprijatni nam su AUX to-us unpleasant ‘The problems that we shall discuss today are unpleasant to us.’

Figure 5. Clitic after a heavy constituent

188 Clitic Placement and the Properties of the Intonational Phrase in Serbian

5. Prosodic phrasing and correlations with meaning The production study has provided us with several important insights. First of all, it has shown that constituents regularly exhibit more Iboundary signals with delayed clitics than with 2P clitics. Second, it has revealed that some constituents are more flexible than others. Some allow complete elimination of boundary features (appositives and heavy constituents with single-word predicates), and clearly indicate the possibility of dual prosodic phrasing. In contrast to these are parentheticals, which always retain the I-boundary features, irrespective of the position of the clitic. Preposed (non-heavy) topics, in most cases, never exhibit a full set of Iboundary features, but still show phonological differences in clitic-second and clitic-third orders. Even though constituents in clitic-second and clitic-third orders do not always exhibit a sharp distinction in their boundary features, we will claim that clitics indeed distinguish between different prosodic constituents – they can cliticize only to prosodic constituents lower than IP. In other words, clitics can detect an I-boundary, even when not all its defining phonological features are present. An argument in favour of the view that constituents immediately followed by clitics are not necessarily IPs, and conversely, that cliticdelaying constituents always are IPs, comes from the fact that the presence or absence of I-boundary features correlates with slight changes of meaning. The strongest support that different prosodic phrasing brings about interpretative differences comes from appositives.

5.1. Integrated vs. supplemental appositives Appositives are generally assumed to fall into two classes: integrated and supplemental (Huddleston and Pullum 2002; Potts 2003). Just like with relative clauses, the difference between the two types is based on semantics: integrated appositives are restrictive, and supplemental are specifying. The distinction is also supported by prosodic differences. While supplemental appositives are always associated with the so-called comma intonation (i.e., IP edges), this is not the case with integrated appositives. We argue that the integrated vs. supplemental distinction is also what underlies clitic-second and clitic-third orders with appositives. Specifically, the absence of I-boundary transforms a supplemental appositive into an integrated appositive, which is, in turn, always interpreted as restrictive. With no I-boundary present, clitics are free to occupy the second position.

Maja Markoviü and Tanja Miliüev

189

Consider (18), a slightly modified version of (11). The appositive in (18a) restricts the meaning of the antecedent, while the relation between the antecedent and the appositive in (18b) is that of specification. In other words, (18a) entails that Maja could have more aunts who are not necessarily teachers of Latin, whereas in (18b), this reading is absent. (18)

a.

b.

Majine tetke, profesorice latinskog Maja's aunts professors of-Latin objasnile ablativ mi su AUX me explained ablative ‘Maja's aunts, professors of Latin explained ablative to me’ Majine tetke, profesorice latinskog Maja's aunts professors of-Latin objasnile ablativ su mi explained AUX me ablative ‘Maja's aunts, professors of Latin, explained ablative to me’

That the appositives that clitics can attach to can only be integrated appositives is also evident from the fact that clitic-second orders are impossible with appositives whose antecedent has a unique referent, as illustrated in (19) (cf. de Vries 2002 for the discussion of restrictive relative clauses with such antecedents). (19)

a.

b.

# Ja, tvoja mama, obeüala sam ti I your mother AUX you promised sladoled ice cream ‘I, your mother, promised you an ice cream.’ # Maja Markoviü, fonolog objasnila mi je Maja Markoviü phonologist me AUX explained Praat Praat ‘Maja Markoviü, the phonologist, has explained ‘Praat’ to me.’

It should be noted that integrated appositives and restrictive relative clauses are not completely banned with pronouns and names. Integrated appositives with personal pronouns are possible as long as they do not lead to vacuous quantification, i.e., as long as the notion of a set is available, allowing the relation of restriction.

190 Clitic Placement and the Properties of the Intonational Phrase in Serbian

(20)

a.

b.

Ja tvoja prijateljica ti sve verujem. I your friend you all believe you Ja tvoj šef imam par sumnji. I your boss have couple of-doubts ‘I (as) your friend believe you, but I (as) your boss have some doubts’ Mi Banaüani najbolje razumemo ih we Banatians them best understand ‘We, the people from Banat, understand them best’ (as opposed to people from other regions)

By contrast, in (19), especially (19a), it is difficult to see how in the given context the property contained in the appositive could be applied restrictively. That clitic-delaying appositives are indeed supplemental is supported by fact that they can never be used restrictively (21a). Since the supplemental appositive cannot restrict the value of the antecedent but only assert that their denotation has a certain property, it is difficult to interpret the antecedent in (21a), Ivan Ivanoviü, as having different denotations. When comma intonation (I-boundary) is absent, this interpretation, as expected, becomes available (21b). (21)

a.

b.

# Ivan Ivanoviü, reper, dosta mi se Ivan Ivanoviü rapper quite me REFL dopada. Ivan Ivanoviü, voditelj, nije mi appeal Ivan Ivanoviü host not-is me tako simpatiþan so likeable ‘Ivan Ivanoviü, the rapper, quite appeals to me. Ivan Ivanoviü, the TV host, is not so likeable (to me)’ Ivan Ivanoviü, reper mi se dosta dopada. Ivan Ivanoviü, voditelj mi nije tako simpatiþan

Based on the truth-conditional distinctions between the two types of appositives and their strong correlation with specific prosodic features, it can be concluded that the absence of relevant I-boundary features with non-clitic delaying appositives is not coincidental, nor does it arise as a rescue strategy to accommodate the clitic's prosodic sensitivity. We are clearly dealing with a different type of prosodic constituent with its own semantics.

Maja Markoviü and Tanja Miliüev

191

That manipulating boundary properties necessarily leads to different prosodic phrasing, and consequently the loss of the associated meaning is also supported by parentheticals’ strong resistance to cliticization. In the production study the speakers regularly hesitated when required to pronounce a sentence with the clitic immediately after the parenthetical. No other constituent has been noted to represent such a problem. To use the appositive terminology, it is difficult to turn a supplemental parenthetical into an ‘integrated’ one. In theory, parentheticals can be pronounced in such a way as to accommodate a clitic. However, unmotivated loss of I-boundary features does not lead to felicitous results.

5.2. [+Comma] feature (Potts 2003; Selkirk 2005) The facts observed regarding appositives and parentheticals are in line with Pott's (2003) analysis of supplements (appositives, appositive relative clauses, parentheticals) as realizations of the syntactic [+Comma] feature. This feature has distinct realizations in both semantics and phonology. In the semantics, [+Comma] marked constituent gets the conventional implicature (CI) meaning.10 Potts accepts Grice's (1975: 44-45) original definition of CIs as ‘logically and compositionally independent of what is “said (in the favored sense)”, i.e., independent of the at-issue entailments’. Unlike ‘at-issue entailments’ (the regular asserted content), CIs can be viewed as ‘secondary entailments’, whose function is to guide the discourse in a particular direction or to help the hearer to better understand why the at-issue content is important at that stage. CIs also differ from presuppositions in that they are always non-backgrounded, i.e. discourse new. In the phonology, +Comma feature yields a prosodic constituent with the so-called ‘comma intonation’ (IP, in other words). In addition to parentheticals and appositives, as prototypical supplemental material, Potts (2003, Chapter 3) shows that there is good reason to extend the list of CIs to include a variety of left-peripheral elements, such as speaker-oriented adverbs (‘luckily’), topic/subjectoriented adverbs (‘thoughtfully’), utterance-modifiers (‘frankly’), which are generally associated with comma-intonation. The CommaP status of these constituents stems from their interpretation as the ‘upper layer’ of ‘discourse structure’ - the layer which permits us to talk about and impose conditions on utterances. All CI items can be then understood as 10

The term 'implicature' might be somewhat confusing, since CI's are actually entailments (as opposed to Grice's conversational implicatures (see Potts 2003, Section 1.4.1 for details on the distinction).

192 Clitic Placement and the Properties of the Intonational Phrase in Serbian

comments upon an asserted core, ‘providing a means for a bit of editorializing on the part of speakers’ (Potts 2003:206). Selkirk (2005) adopts Pott's proposal that the syntactic grounding of IP is Comma Phrase. She points out that the root sentence, a default syntactic association of IP, naturally falls into this category since it is ‘performed as a separate speech act, and its meaning is independent of the at-issue entailments of any other Root sentence and of any other Comma Phrase’ (Selkirk 2005: 16). She also considers left-peripheral adjuncts such as ifclauses, as for-topic clauses and left-dislocated phrases to be projections of [+Comma], ‘with their own conventional implicatures and their own potential for being performed as (hidden) speech acts’. In Serbian, left-dislocated phrases (22) and left-peripheral adjunct clauses (23) obligatorily delay clitics. (22)

a.

b. (23)

a. b. c.

d.

Marija, juþe njoj smo se Mary-LFD yesterday AUX REFL her-RP predstavili introduced ‘Mary, we introduced ourselves to her yesterday’ *Marija, smo se njoj juþe predstavili Ako voliš biljke, redovno ih if you-love plants regularly them ‘If you love plants, you must water them regularly’ *Ako voliš biljke, ih redovno zalivaj Kad se vratila, napisala mu when RFL returned wrotten him pismo letter ‘Kad se vratila mu je napisala pismo.’ *Kad se vratila, mu je napisala pismo

zalivaj water je AUX

Of Pott's supplemental adverbials, only speaker-oriented adverbs and utterance-modifiers regularly disallow encliticization in Serbian.11 The 11

Subject-oriented adverbs do not require clitic-delay (ia). Also, clitic-delay seems obligatory only with a subset of speaker-oriented adverbs, since some of them, such as nesumnjivo ‘undoubtedly’ or oþigledno 'evidently' (ib) allow 2P clitics. Most likely, only those speaker-oriented adverbs which express the speaker's attitude towards the act of assertion (as opposed the speaker's attitude towards the truth of the proposition; cf. Bellert 1977) disallow 2P clitics. We leave the matter for further research.

Maja Markoviü and Tanja Miliüev

193

lack of I-boundary/comma intonation allowing 2P clitics always comes with the change of meaning, and these adverbs are regularly interpreted as VP modifying (preposing yields a narrow focus interpretation). (24)

a. b.

(25)

a. b.

Sreüom, dobro ga poznajem by-luck well him I-know ‘Luckily, I know him well’ Sreüom dobro poznajem ga by-luck him well I-know ‘It is by luck that I know him well’ Iskreno to zameram mu honestly that him hold-against ‘Honestly/Frankly, I hold it against him’ Iskreno zameram mu to honestly him that hold-against ‘It is in an honest way that I hold it against him’

The issue of [+Comma] as the syntactic grounding of IPs is somewhat complicated by the fact that supplements, CIs, and independent speech acts do not cover the whole inventory of elements that can be prosodically realized as IP. This issue is discussed in Selkirk 2005. She points out that certain preposed elements can be optionally realized as IPs (Selkirk 2005:23-24). A preposed PP, as on the fourth of July in (26), can have dual phrasing, that of an IP (26a), or a lower prosodic consitutent, which she identifies as Major Phrase (MaP) (26b). (26)

a. b.

(i.)

a. b.

(On the fouH*rth of July H*L-H%)IP, (we'll have a paraH*de and fiH*reworksL-L%)IP ((On the fouH*rth of July H*L-)MaP, (we'll have a paraH*de and fiH*reworksL-L%)MaP) IP

Pametno poneo ga je wisely him AUX taken 'It was wise of him to have taken it with him' Nesumnjivo/Oþigledno ga poneo je undoubtedly/evidently him AUX taken

sa with

sobom self

sa with

sobom self

194 Clitic Placement and the Properties of the Intonational Phrase in Serbian

The difference in prosodic phrasing should be accompanied by (subtle) semantic differences, and it s not immediately clear what they might be.12 Recall that preposed elements in Serbian, especially the light ones, also only optionally delay clitic placement. In addition to preposed adverbials, preposed arguments also give rise to clitic-third orders. For preposed arguments, CommaP status is even more difficult to establish, since they are clearly an integrated part of the root sentence, and their meaning cannot be the same as that of supplements or ‘discourse’ adverbials. These and some other problems which arise with sentence initial topics in Serbian are discussed in the following section.

6. Sentence initial topics – optional IP status In addition to the lack of a straightforward relation between CommaP and topicalization in terms of interpretation, Serbian preposed elements pose an additional problem - they exhibit the least I-boundary signals. This further weakens the motivation for their IP status. Yet, they can delay clitics. In what follows we try to examine what factors might be at play distinguishing clitic-delaying preposed topics as a unique constituent.

6.1. The heaviness factor on clitic delay The generalization commonly found in the literature regarding sentence initial elements in SC is that only heavy elements obligatorily delay clitics, while the non-heavy ones only optionally do that. What counts as a heavy constituent, however, is not entirely straightforward. Zec & Inkelas (1990) claim that the heaviness requirement is prosodic in nature. The contrast between the sentences in (a) and (b) in (27) and (28) indicates a topicalized element needs to consist of two prosodic words, i.e., be prosodically branching. Note that syntactic complexity is

12

The differences between preposed and non-preposed adverbials have been discussed in the literature (cf. Kuno 1975 and Reinhart 1983 for the analysis of preposed adverbials as 'scene-setters' or topics, creating new discourse segments), as well as the different prosodic status of different types of preposed adverbs (specifically, the impossibility of 'neutral' intonation with certain preposed adverbs), but, at least to our knowledge, the consequences of the optionality of prosodic phrasing of one and the same type of a preposed adverbial have not yet received an account.

Maja Markoviü and Tanja Miliüev

195

unlikely to play a crucial role here since in both cases we have proper names.13 (27)

a. b.

(28)

a. b.

Petar Petroviü voleo je Mariju Petar Petroviü loved AUX Marija ‘Petar Petroviü loved Marija’ *Petar voleo je Mariju Petar loved AUX Mary ‘Petar loved Marija.’ U Rio de Ženeiru ostali dve godine su in Rio de Janeiro stayed AUX two years ‘In Rio de Janeiro, they stayed two years’ *U Riju ostali dve godine su in Rio stayed AUX two years ‘In Rio, they stayed two years’ (Zec & Inkelas 1990:373–375)

Radanoviü-Kociü (1988), however, claims that only single-word subjects are impossible to delay clitics, while adjuncts and objects need not obey the prosodic branchingness requirement, and can optionally give rise to clitic-third orders even when they consist of a single prosodic word (29). Truly ‘heavy’, i.e., obligatorily clitic delaying constituents are those consisting of three or more words. (29)

a. b.

Lica ne razaznaje im faces not distinguishes them ‘He doesn't distinguish their faces’ Sapiru jezik instrumenat je to-Sapir language is instrument ‘For Sapir language is an instrument’ (Radanoviü-Kociü 1988:106-107)

We thus arrive at a rather puzzling generalization: an element needs to be ‘truly heavy’ (in terms of both prosodic structure and length) to disallow 13

Zec & Inkelas (1990:375) argue that the lack of syntactic branchingness is particularly evident with proper names such as Rio de Ženeiro, where only the second part, Ženerio can be case marked (u Rio de Ženeiru vs. u Riju de Ženeiru 'in Rio de Janeiro'). The first part, Rio, can bear case when used in isolation (u Riju 'in Rio'). Hence, even though syntactically this NP acts as a single unit, prosodically it is complex, thus satisfying the core requirement for delaying clitics.

196 Clitic Placement and the Properties of the Intonational Phrase in Serbian

clitic attachment; ‘light’ elements can optionally delay clitics, unless they are subjects. This is summarized in (30). (30)

a) b) c)

obligatory clitic delay: truly heavy elements optional clitic delay: ‘light’ adjuncts prosodically branching subjects no clitic delay: light subjects

and

objects;

We, however, claim that the ban on single-word subjects is too strong. We will show that under the right pragmatic conditions, all light (or nonheavy) elements can delay clitics.

6.2. The interpretation Assuming again that clitic-delay arises in the presence of an additional I-boundary, and that comma intonation has semantic correlates, the optionality in the clitic position should also come with certain interpretative differences. It has been noted in the literature that non(truly) heavy clitic-delaying constructions must satisfy certain information structural/semantic requirements. Browne (1975), for instance, argues that subjects with clitic-third orders are possible only when they introduce a new topic of conversation. (31)

Moja djevojka vidjela je juþer medvjeda. my girlfriend seen AUX yesterday bear ‘My girlfriend saw a bear yesterday.’ ‘You know my girlfriend? She saw a bear yesterday.’ (Browne 1975:120)

As for other sentence initial constituents, Browne proposes that an element must be emphatically or contrastively stressed to be able to delay a clitic. (32)

a. b.

*U školi uþiti bez knjige. üe in school study will without book ‘In school he’ll study without a book.’ U ŠKOLI uþiti üe bez knjige. (as opposed to kod kuüe ‘at home') ‘In school he’ll study without a book.’ (Browne 1975: 112)

Maja Markoviü and Tanja Miliüev

197

The same contrast relation seems to be required for light subjects. In order for the non-branching subject ‘Mubarak’ to be able to delay the clitic, it has to be contrasted with another subject (‘Gadafi’). (33)

Mubarak Mubarak je AUX ‘Mubarak it)’

predao vlast (Gadafi je given AUX power Gaddaffi rešio da je saþuva) decided to it keep stepped down from power. (Gaddafi decided to keep

Franks (1998:28) assumes that clitic-delaying clause initial elements bear contrastive focus, which is set off prosodically by an I-boundary (indicated by #). (34)

#Javili su nam da# #prije announced AUX us that ago nekoliko dana# #na toj liniji# several days on that line #voz kasnio tri sata# je train AUX late three hours ‘They announced that, several days ago, on that line, the train was 3 hours late.’

However, the relation between contrastive focus and prosodic phrasing is not so straightforward. Contrastive focus per se, of course, does not need to be set off prosodically. Nor do contrastively focused sentence initial elements block cliticization (35).14 (35)

PETRA mu predstavi (i nikog drugog/a ne Pavla) Peter him introduce ‘You should introduce Peter to him (and no one else/and not Paul)’

Rather, to get multiple IPs there need to be multiple contrastive foci (Selkirk 2003). This is the line of reasoning we will pursue as well. The most likely interpretation of the clitic-delaying topics is that of contrastive topic (cf. Büring 1997, 2003). A contrastive topic is licensed by the presence of another constituent that invokes an alternative, i.e., a focus. 14

See Smiljaniü 2004 for the phonological representation of second-position clitics after narrowly focused elements.

198 Clitic Placement and the Properties of the Intonational Phrase in Serbian

This kind of topic is prosodically distinguished by a special kind of a complex pitch accent, the so-called ‘B-Accent’ (cf. Jackendoff 1972), L*H followed by a L-H% boundary. Significantly for our discussion, contrastive topics are generally assumed to project their own IPs (cf. Selkirk 2005; Wagner 2010, a.o.; for a slightly different view see Féry 1993, where a contrastive topic-focus construction is assigned one IP, consisting of two ‘intermediate phrases’, one for contrastive topic, one for the rest of the sentence containing the focused part). Pragmatically, contrastive topics can be interpreted as contrastive aboutness topic which, just like focus, indicates alternatives, i.e., alternative aboutness topics (Büring 1997; Krifka 2008) (36)

A: B:

What did you buy on 59th street? On fifty NINTH street I bought SHOES. (On 58th street I bought gloves)

By providing an answer with a contrastive topic, the speaker also anticipates answering an alternative question that substitutes the CT element (about the purchase in another street). Another typical use is in answers to multiple wh-questions or pair-list questions (Büring 2003; van Hook 2003; Wagner 2010, a.o.) (37)

A: B:

Who ate what? FREDCT ate the BEANSFOC and MARYCT ate the SPINACHFOC.

Clitic-delaying topics seem to fit these contexts. They can be used in answers to pair-list questions (38). (38)

A:

B:

Koliko ste gde ostali (tokom howAUX where stayed (during long vašeg putovanja po Južnoj Americi)? your trip to South America) ‘How long did you stay where (on your trip to South America)?’ U Riju ostali dve nedelje. smo in Rio we-stayed AUX two weeks (U Karakasu pet.) in Caracas five ‘In Rio we stayed two weeks. (In Caracas five)’

Maja Markoviü and Tanja Miliüev

199

Both the sentence initial element and the focused element in the body of the sentence induce alternatives – the topicalized locational PP u Riju ‘in Rio’ has to be contrasted with another comparable location – u Karakasu ‘in Caracas’, and the focused NP dve nedelje ‘two weeks’ requires the introduction of another temporal interval. Due to the fact that CT implies the assertion of one or more related propositions, sentences containing contrastive topics cannot be used in isolation. Indeed, clitic-delaying topics do not seem appropriate as answers to single wh-questions or multiple wh-questions with single-pair reading.15 (39)

A: B:

Koliko ste ostali u Riju? Koliko ste gde ostali? How long did you stay in Rio? How long did you stay where? #U Riju ostali smo dve nedelje.

That the focused part in the sentence needs to be contrastive in order to license a clitic-delaying contrastive topic can be verified by (34). Even though the second PP ‘in Caracas’ can be assumed to be contrastively related to the PP ‘in Rio’, without a contrastively focused element, it cannot be a contrastive topic, and the clitic-third orders in those cases are infelicitous. (40)

U Riju ostali smo dve nedelje. in Rio stayed AUX two weeks #U Karakasu sreli smo neke prijatelje. in Caracas met AUX some friends ‘In Rio we stayed two weeks. In Caracas we met some friends’

Topic initial sentences with clitic-second orders can also be used as answers to multiple wh-questions. The main point of difference, however, seems to be that, unlike contrastive topic-focus sentences, these can be used as single-pair answers. In other words, they do not necessarily indicate the presence of alternatives, i.e., they are not instances of contrastive focus.

15

The interpretation of multiple wh-questions in Serbian is subject to some controversy. For some speakers, all multiple wh-questions require a pair-list reading, while for some pair-list reading is available only in the absence of Superiority effects, i.e. with wh-fronting, as opposed to wh-movement (cf. Boškoviü 2003). According to our judgements, both types of multiple questions can have single-pair reading.

200 Clitic Placement and the Properties of the Intonational Phrase in Serbian

Finally, let us remark that we are aware of the possibility that not all preposed arguments triggering clitic delay match the interpretation we propose here, which is mainly based on truly light (one-word) constituents. However, we expect that both the notions of contrast and focus will prove to be relevant for the ‘special’ semantics requiring the accompanying prosodic IP phrasing. At this point, we also have little to say about the minimal requirement on I-boundary signals found with clitic-delaying topics, as opposed to other clitic-delaying structures. Whether this is common to all leftperipheral IPs, needs yet to be investigated. Size or phonological weight might be an important factor (only light topics have been investigated), but we do not believe it to be the crucial one.

6.3. Clitic-delaying subjects As noted earlier, clitic-delaying subjects do not need to be contrastive topics, but rather introduce new topics of conversation. Without a detailed study of the pragmatic properties of these subjects, it is difficult for us to properly evaluate the claims found in the literature. But if we assume that ‘new topic of conversation’ is indeed the interpretation obtained in cliticthird orders, the semantics motivating separate IP phrasing would probably have to be sought in the interpretation associated with the socalled Hanging Topics, since they are commonly used to introduce new discourse material, or create new discourse segments. The syntactic construction associated with hanging topics is leftdislocation, specifically Hanging Topic Left Dislocation (HTLD). Leftdislocation is usually distinguished from topicalization by the presence of a resumptive element in the body of the sentence.16 In order to account for the absence of an overt resumptive in cliticdelaying constructions of this type, we can speculate that sentences with HTLD subjects are instances of Clitic Left-Dislocation, where the resumptive pronoun is realized in the form of clitic. Since there are no overt nominative clitic forms, the resumptive element would be realized as a null clitic, or pro.

16 Some authors, it seems, do not require a resumptive pronoun for the identification of a constituent as left-dislocated, but rather rely on the presence of an I-boundary/comma intonation and clitic-delay (see, for instance, Boškoviü 2003; Halupka-Rešetar 2009) It is unclear if the assumption is supported by syntactic and interpretative facts as well.

Maja Markoviü and Tanja Miliüev

(41)

[LFD

Moja my

devojka]IP [RootClause proRSP girlfriendNOM juþe medveda]IP yesterday bear ‘My girlfriend, she saw a bear yesterday’

201

videla seen

je AUX

Clearly, much further research is needed before we can establish with certainty both the syntactic and the pragmatic status of these cliticdelaying subjects. At this point it seems clear that much more than the notion ‘introduction of new discourse participants’ is required for these subjects to fit the interpretation of CommaP constituents.17

7. Conclusion Having examined the phonological properties of the ‘exceptional’ cases of clitic placement, namely, 2P clitics after constituents which normally project IPs, and delayed clitics after non-heavy constituents, which are normally prosodically realized below IP, we have shown that cliticization is possible only when the I-boundary features are erased, while clitic-delay is licit when minimal indication of an I-boundary is present. We have also established that the change in prosodic phrasing brings about a change in meaning. We find that Pott's (2003) analysis of the strong correlation between phonology/prosody (‘comma intonation’) and semantics (‘CI meanings’) in terms of Comma Phrase can well capture most of the exceptions to the 2P effect with clitics in Serbian, and is certainly worth pursuing in order to account for those cases which, at this point, do not seem to straightforwardly fit into the proposal. The influence of contrastive focus-marking on the variation of prosodic phrasing (Selkirk 2005) has also been confirmed. Based on our analysis of clitic-delaying light topics as contrastive topic-focus constructions, the generalization that multiple contrastive foci require multiple IPs seem to emerge. While we have not dealt with how the factor of length/heaviness actually influences the change in prosodic phrasing, we have established that with truly heavy constituent, the I-boundary is quite freely ignored 17 The clitic-delaying subjects seem to be most closely related to Pott's upper layer of discourse structures, much like discourse adverbials. What kind of speakeroriented comment on to the assertion core this is and why in this particular case 'discourse newness' needs to be introduced by special means is, of course, the crucial issue.

202 Clitic Placement and the Properties of the Intonational Phrase in Serbian

when the interpretation of other elements (the verb, specifically) in the sentence is affected by the spell-out of the clitic in a lower position.

Acknowledgements The research presented in the paper is supported by the technological project “The Development of Dialogue Systems for Serbian and Other South Slavic Languages” (TR32035) of the Ministry of Science and Technological Development of the Republic of Serbia.

References Bellert, I. 1977. On Semantic and Distributional Properties of Sentential Adverbs. Linguistic Inquiry 8, 337-351. Boškoviü, Ž. 2001. On the Nature of the Syntax-Phonology Interface: Cliticization and Related Phenomena. Amsterdam: Elsevier. —. 2003. On the interpretation of multiple questions. Linguistic variation Yearbook 1, 1-15. —. 2004. Clitic placement in South Slavic. Journal of Slavic Linguistics 12, 37-90. Browne, W. 1975. Serbo-Croatian enclitics for English-speaking learners. In Contrastive analysis of English and Serbo-Croatian, Vol. 1, edited by R. Filipoviü, 105-134. Zagreb: Institute of Linguistics, Faculty of Philosophy, University of Zagreb. Büring, D. 1997. The Meaning of Topic and Focus - The 59th St. Bridge Accent. London: Routledge. —. 2003. On D-trees, beans and B-accents. Linguistics and Philosophy 26, 511-545. ýamdžiü, A. and R. Hudson. 2007. Serbo-Croat-Bosnian clitics and Word Grammar. Research in Language 5, 5-50. ûavar, D. and C. Wilder. 1994. Clitic third in Croatian. Eurotyp Working Papers, Grammatical Models Selection, 19-61. Emonds, J. 1970. Root and Structure-Preserving Transformations. Doctoral dissertation. Department of Linguistics, MIT. Franks, S. 1998. Clitics in Slavic. Position paper for Comparative Slavic Morphosyntax workshop, Bloomington, Indiana, June 1998. Féry, C. 1993. German Intonational Patterns. Tübingen: Niemeyer. Godjevac, S. 2000. Intonation, Word Order and Focus Projection in Serbo-Croatian. Dissertations in Linguistics, Ohio State University, Columbus.

Maja Markoviü and Tanja Miliüev

203

—. 2006. Transcribing Serbo-Croatian intonation. In Prosodic Typology -

The Phonology of Intonation and Phrasing, edited by S. Jun, 146-171. Oxford: Oxford University Press. Halpern, A. 1995. On the placement and morphology of clitics. Stanford: CSLI Publications. Halupka-Rešetar, S. 2009. Strukturna artikulacija fokusa u engleskom i srpskom jeziku: uporedna analiza strukturalnog i generativnog pristupa. Doctoral dissertation. University of Novi Sad. Hook, van H. 2003. The rise in the rise–fall contour: does it evoke a contrastive topic or a contrastive focus?. Linguistics 41/3, 515–563. Huddleston, R. and G. K. Pullum. 2002. The Cambridge Grammar of the English Language. Cambridge, UK: Cambridge University Press. Jackendoff, R. 1972. Semantic Interpretation in Generative Grammar. Cambridge, MA: MIT Press. Krifka, M. 2008. Basic notions of information structure. Acta Linguistica Hungarica 55, 243-276. Kuno, S. 1972. Functional sentence perspective: A case study from Japanese and English. Linguistic Inquiry 3, 269–336. Ladd, D. R. 1996. Intonational Phonology. Cambridge: Cambridge University Press. Markoviü, M. and T. Miliüev. 2011. I-boundary features and the role of F0 peak timing. Ms. University of Novi Sad. Nespor, M. and I. Vogel. 1986. Prosodic Phonology. Dordrecht: Foris. Potts, C. 2003. The Logic of Conventional Implicatures. Doctoral dissertation. Department of Linguistics, University of California at Santa Cruz. Progovac, Lj. 1996. Clitics in Serbian/Croatian: Comp as the second position. In Approaching second: Second position clitics and related phenomena, edited by A. Halpern and A. Zwicky, 411-428. Stanford: CSLI Publications. Radanoviü-Kociü, V. 1988. The grammar of Serbo-Croatian clitics: A synchronic and diachronic perspective. Doctoral dissertation. University of Illinois, Urbana-Champaign. —. 1996. Serbo-Croatian clitics: A prosodic approach. In Approaching second: Second position clitics and related phenomena, edited by A. Halpern and A. Zwicky, 429-445. Stanford: CSLI Publications. Reinhart, T. 1981. Pragmatics and linguistics: An analysis of sentence topics. Philosophica 27/1, 53–94. Schütze, C. 1994. Serbo-Croatian Second Position Clitics Placement and the Phonology-Syntax Interface. MIT Working Papers in Linguistics 21, 373-473.

204 Clitic Placement and the Properties of the Intonational Phrase in Serbian

Selkirk, E. 1984. Phonology and Syntax: The Relation between Sound and Structure. Cambridge, Mass.: MIT Press. —. 1986. On derived domains in sentence phonology. Phonology 3, 371405. —. 2005. Comments on the Intonational Phrasing in English. In Prosodies, edited by S. Frota, M. Vigario and M. J. Freitas, 11-58. Berlin: Mouton de Gruyter. Smiljaniü R. 2004. Lexical, pragmatic, and positional effects on prosody in two dialects of Croatian and Serbian: an acoustic study. London: Routledge. Tomiü, O. Mišeska. 1996. The Balkan Slavic clitics. Natural Language and Linguistic Theory 14, 811-872. Vries, M. de. 2002. The syntax of relativization. Utrecht: LOT. Wagner, M. 2010. Contrastive topics decomposed. Ms. Cornell University. (to appear in Semantics & Pragmatics). Wang, M. Q. and J. Hirschberg. 1992. Automatic Classification of Intonational Phrase Boundaries. Computer Speech and Language 6, 175-196. Zec, D. and S. Inkelas. 1990. Prosodically constrained syntax. In The phonology-syntax connection, edited by S. Inkelas and D. Zec, 365– 378. Chicago: University of Chicago Press.

PART TWO. NATURAL LANGUAGE PROCESSING

NAMED ENTITY RECOGNITION IN THE SYSTEM FOR INFORMATION EXTRACTION STAŠA VUJIýIû STANKOVIû

Outline One of the crucial issues of natural language processing is named entity recognition. In this paper, we discuss the general problem of named entity recognition in a system for information extraction, and the problem of named entity recognition for the Serbian language in particular. We propose a solution to the problem based on language resources developed for the Serbian language, implying the usage of the GATE1 and the Unitex2 systems.

1. Introduction At the beginning of this paper, the areas of interest and the main viewpoints of this research are to be defined. We concentrate on the issues in information extraction and named entity recognition, and the systems of interests for better understanding of the approach presented here. The general field of interest for this research is information extraction in general and named entity recognition in particular. The goal of information extraction is to automatically extract information in a predefined structure, from the text documents that are not necessarily given in a structured form. One of the tasks of information extraction is named entity recognition. The main issue of named entity recognition is to identify proper names in text documents and to classify them in some of the predefined categories such as: ENAMEX category (persons, organizations, and locations), TIMEX category (temporal expressions), or NUMEX category (numeric expressions), in the spirit of Chinchor (1998). 1 2

GATE homepage: http://gate.ac.uk/. Unitex homepage: http://www-igm.univ-mlv.fr/~unitex/.

Staša Vujiþiü Stankoviü

207

For example, the first sentence from the most translated Verne’s Novel “Around the World in Eighty Days,“ with marked entities is given below in (1). Persons are bolded, Locations are underlined, and Dates are normalized. (1)

Mr. Phileas Fogg lived, in 1872, at No. 7, Saville Row, Burlington Gardens, the house in which Sheridan died in 1814. Mr. Phileas Fogg lived, in 1872, at No. 7, Saville Row, Burlington Gardens, the house in which Sheridan died in 1814.

The main task of the system for information extraction would be to perform this kind of identifying and marking automatically. To solve the problem of named entity recognition in both English and Western European languages, a series of techniques that give good results were developed, while there is still plenty of room for development for the Slavic, typically morphologically rich languages. Up to now, the development of electronic morphological dictionary based on finite automata and development of algorithms based on statistical methods were the approaches used for resolving this issue in the Serbian language. This paper deals with the problem of identification of named entities (Sekine & Ranchhod 2009) in Serbian texts, using the GATE (cf. Cunningham et al. 2011) and the Unitex (cf. Paumier 2008) system. The organization of the paper is as follows. Section 2 provides the description of the general problem of information extraction, and especially the problem of information extraction for the Serbian language. Section 3 introduces two systems for language processing, GATE and Unitex, followed by a review of some possible approaches to solve the problem and the models of possible solutions that could be proposed. The essence of the proposed solution is described in Section 4, while its evaluation is given in Section 5. Finally, conclusion and future work are presented in Section 6.

2. Description of the problem In this section, we focus on the following issues of importance for the understanding of the basic orientation of this research: (a) Problem

208

Named Entity Recognition in the System for Information Extraction

definition, which is to define and explain the focus of this research, (b) Elaboration of problem importance (why is it important to focus on this problem), and (c) An assessment of the problem development trends (discussing why will the importance of this problem grow over time). Problems of information extraction in general, and its essential component named entities recognition in particular, are extremely complex in the texts of contemporary Serbian language and stem from several sources rooted in the fundaments of the Serbian language. Some of the problems are represented below: o

o

o

o

3

The inflectional and derivational morphologies of the Serbian language are very rich. For example, Serbian language has 7 cases and 3 values for numbers - singular, plural, and paukal3 (according to Stanojþiü & Popoviü 2002). Additionally, for instance, proper names may have a number of derived forms, adjectives or adverbs. In the texts in Serbian, both Cyrillic and Latin alphabets are approximately equally used. This represents the problem because the transliteration is not unified. For example, the toponym New York is represented in the Latin alphabet as Njujork (or even NJujork), but in the Cyrillic alphabet it could be represented as ɇʁɭʁɨɪɤ or ȵɭʁɨɪɤ. Moreover, the orthography of Serbian does not give precise definition how to treat the foreign names that inevitably occur in texts, such as the names of organizations, people, and so on. For example, the Microsoft company takes the forms Mikrosoft and Majkrosoft. Compounds represent a great problem, because each compound, or some part of it, can be inflected. For example, in the compound Novi Sad, both parts are inflected: Novom Sadu.

Serbian Grammar dictates that the numeral one agrees with a noun in gender, number, and case, numerals greater then or equal to five do not inflect and they agree with a noun in the genitive plural, while numerals two, three and four either inflect, and agree with a noun, or they don’t inflect. If the latter is the case, they agree with a noun in so-called “paukal”, a special kind of grammatical number used with selected small values (2, 3 and 4). Not all grammarians agree with that; some of them argue that in the case of these small values, the noun agrees in genitive singular since, in most cases, the forms are the same. However, electronic dictionaries for Serbian use the notion of “paukal” since it better explains a numerous agreement phenomena that are not within the scope of this paper (as analyzed in Krstev & Vitas 2007).

Staša Vujiþiü Stankoviü

209

These characteristics, typical for Slavic languages, are analyzed in Przepiórkowski (2007), and make them extremely difficult for automated processing. Serbian language, as a representative of the South Slavic languages, represents a challenge for researchers. The main problem of this research is to design a system for information extraction for the Serbian language. The importance of the problem will grow over time because the usage of natural language processing tools tends to grow over time, and it will be crucial to apply the best possible resources and algorithms on the growing amount of documents collected from the process of digitalizing traditionally collected materials and the emerging amount of information from the Web, for the reasons of performing different analyses. The importance of the problem will grow due to the fact that infrastructures for natural language processing are applicable to a number of applications, which are either traditionally important (like language engineering, search engines, teaching tools, etc.) or emerging (like online machine translation services, speech recognition systems, semantic web applications, etc.).

3. Proposed solutions and their criticism from the view point or goal defined in the introduction In order to solve the presented problem, we have used two systems for natural language processing, Unitex and GATE, and various resources, which have been developed by natural language processing group at the Faculty of Mathematics, University of Belgrade, and adapted for the purpose of this research.

3.1. Review of the Unitex system and lexical resources Unitex is a corpus processing system, based on automata-oriented technology. It has been in development since 2001 by Sébastien Paumier at the Institut Gaspard-Monge, University of Paris-Est Marne-la-Vallée. It is an open source software, designed for applying electronic dictionaries and grammars to texts. Unitex provides processing in multiple languages with more complex queries than regular expressions over characters. It has been used for many teaching and research projects. Unitex system is distributed with the electronic dictionaries of the type DELA (Dictionnaires Electroniques du LADL or LADL electronic dictionaries). The DELA format for morphological dictionaries has been developed within the Laboratoire d'Automatique Documentaire et

210

Named Entity Recognition in the System for Information Extraction

Linguistique under the guidance of Maurice Gross. This format satisfies key features suitable for automatic text transformation - contains the information that makes it possible to resolve problems of segmentation, as well as, morphological, syntactic, and semantic processing of text. Electronic dictionary in the DELA format is text file where each line represents a word entry and describes the inflected form of the word, the lemma of the word and some grammatical, semantic, and inflectional information. The system of the DELA morphological dictionaries is presented in Figure 1. It consists of the dictionaries of simple words, noninflected forms, called DELAS (simple forms DELA), the dictionaries of inflected forms DELAF (DELA of inflected forms). On the other side, it contains the dictionaries of compounds DELAC (DELA of compound forms DELA) and the dictionaries of inflected compound forms DELACF (DELA of compound inflected forms). Figure 1. The LADL electronic dictionaries.

An example entry from the English dictionary of simple words is shown in (1), where books is the inflected form of the entry, which is mandatory, book is the canonical form (lemma) of the entry, N+Conc is the sequence of grammatical and semantic information (N denotes a noun, and Conc denotes that this noun is a concrete object), p is an inflectional code which indicates that the noun is plural. More on DELA format can be found in Courtois & Silberztein (1990). (2)

books,book.N+CONC:P

The morphological electronic dictionaries in the DELA format for Serbian have been developed for years by the natural language processing group at the Faculty of Mathematics, University of Belgrade. According to Krstev (2008), the present size of the Serbian morphological dictionary of simple

Staša Vujiþiü Stankoviü

211

words, DELAS, contains 121,000 lemmas. Close to 87,000 simple lemmas belong to general lexica, while the remaining 34,000 lemmas represent various kinds of simple proper names. Dictionary of inflected forms, DELAF, contains approximately 1,450,000 different lexical words. The size of DELAC and DELACF dictionaries are approximately 3,000 and 54,000 lemmas, respectively. An entry from the Serbian dictionary of simple word forms with appropriate grammatical and semantic code (similar to (2)) is shown in (3), where the word form stanovao is the singular (S) masculine gender (M) of the active past participle (G) of the verb (V) stanovati ‘to dwel’ that is imperfective (Imperf), intransitive (It), and Ireflexive (Iref). (3)

stanovao,stanovati.V+IMPERF+IT+IREF:GMS

Besides these morphological dictionaries, numerous finite-state transducers were developed in Unitex for Serbian language. The goal is to use them for solving the problems of the morphological analysis and compound words by recognizing them in the proper way and annotate them if necessary with common XML tags such as ENAMEX, TIMEX, and NUMEX. An example of the finite-state transducer graph that recognizes temporal expressions, for instance sequences of characters of the form 01.10.2010. or 01-X-2010., and annotate them with TIMEX tags is shown in the Figure 2. Figure 2. Finite-state transducer graph for extracting and annotating temporal expressions.

We can conclude that Unitex system provides a possibility of working at the levels of morphology, lexicon, and syntax, and supports various kinds of electronic dictionaries and finite-state transducers, which represents a good bases for processing texts in the Serbian language.

212

Named Entity Recognition in the System for Information Extraction

3.2. Review of the GATE system The GATE (General Architecture for Text Engineering) is an architecture, a framework, and a development environment for natural language processing applications. It is an open-source free software, in development by the University of Sheffield natural language processing Group since 1995. The GATE is used for various tasks of natural language processing in multiple languages by a wide range of scientists, researchers, teachers, students, and companies all over the world. The main advantage of the GATE architecture is that everything is organized in terms of components. Every component is designed for a specific task, and there is a clear difference between data storage, data visualization, and the algorithms for natural language processing. This approach provides an opportunity for different types of users to develop data and processing components independently, as well as usage of the same language resources in different graphical environments. Three main types of components are distinguished – Language Resources, Processing Resources and Visual Resources. Language Resources represent linguistic data that involve all kinds of documents, corpora (a list of documents), and ontologies. Processing Resources are algorithmic resources, processing components suitable for deployment and adaption for new languages or domains. Visual Resources represent graphical user interface that allow the visualization and editing of Language Resources and Processing Resources. Among many other tasks of natural language processing that can be done in the GATE system, the task of utmost importance for this research is information extraction. It is based on various existing processing resources integrated in the ANNIE (A Nearly-New information extraction) System. More on ANNIE can be found in Cunningham et al. (2002). Each one of these resources can be used individually or combined with userprovided new resources. The ANNIE System consists of the following processing resources: tokeniser, sentence splitter, POS (part-of-speech) tagger, gazetteer, semantic tagger and orthomatcher. The tokeniser splits text into tokens. It recognizes many types of words, numbers, symbols, white patterns and punctuation, and frequently can be used on different languages without (significant) modifications. The sentence splitter segments the text into sentences. It is based on a cascade of finite-state transducers and usually does not require major changes to be used in different types of applications or for different languages.

Staša Vujiþiü Stankoviü

213

The POS tagger produces a part-of-speech tag in the form of annotation for each word. The version of POS tagger originally provided with the GATE system is based on the Brill tagger. The POS tagger is a language-dependent resource and one of the main goals of this research is to adapt it to the Serbian language. The gazetteer contains lists of cities, countries, personal names, organizations, etc., as well as, lists of company designators and titles. These lists are used to annotate the occurrence of the list items in the texts. The excerpt from one such list (the list of the cities) is shown in Figure 3. Figure 3. The excerpt from the list of the cities. The lists are plain texts.

The semantic tagger is based on the JAPE (Java Annotations Pattern Engine) language (Cunningham 2002), which describes patterns to match previously produced annotations and new annotations that have to be produced as the results. The JAPE performs finite state processing over annotations based on regular expressions. The JAPE grammar consists of a set of phases. Each phase consists of a set of rules that describes the pattern to be recognized and the action to be taken. The grammar always has two sides: Left Hand Side (LHS) and Right Hand Side (RHS). The LHS of the rule contains the identified annotation pattern that may contain the Kleene regular expression operators. The RHS outlines the action to be taken on the detected pattern and consists of annotation manipulation statements. The phases run sequentially and constitute a cascade of finite state transducers over annotations.

214

Named Entity Recognition in the System for Information Extraction

An example of the JAPE grammar is given in Figure 4. Input specification consists of the name of the phase, set of input annotations to use in this rule, and control options that define the method of rule matching: brill style (when more than one rule matches the same region of the document, they are all fired), applet style (only one rule can be fired for the same region of text), or first style (a rule fires for the first match that is found and no other matching attempts are made for that region of the document). Macro has the same form as left hand side and matches the word or number tokens. The LHS corresponds to the pattern that matches the email address, while the RHS represents appropriate action. Figure 4. The JAPE grammar for specifying email address.

The orthomatcher performs coreference, identification of the relations between named entities found by the semantic tagger. It improves results of previously annotations, by producing new annotations based on relations between named entities. As mentioned above, GATE has a graphical environment that provides simple way for user to manage the documents and results of different processing. For illustration, Figure 5 shows the GATE graphical user interface with the result of loading the example document with Vern’s sentence, and processing it with the ANNIE components. Document content is presented in the window’s central position. On the left side is presented the tree with loaded applications, language resources and processing resources. On the right side are shown founded and highlighted annotation sets like Person, Location and Date in different colors, and

Staša Vujiþiü Stankoviü

215

below the text is presented annotation list with features for every annotation. With all above in mind, we can conclude that GATE works using domain-specific list lookup components (gazetteer lists) and JAPE grammars (finite state transducers over annotations) without sophisticated grammatical information and gives excellent results in tasks of named entity recognition for various languages that are supported in GATE. Therefore, he goal is to develop such resources for the Serbian language.

3.3. Presenting the models of the possible solutions The output from the GATE system, in solving the problem of named entity recognition, is a text with tagged named entities. As we mentioned before, the basic process of tagging is done through several phases: 1) identification of tokens and segmentation (tokeniser and sentence splitter) required for the tagger; 2) association with the appropriate parts of speech (POS) required for phase 3); and 3) tagging the named entities identified by the list (gazetteer) and JAPE grammars. Figure 5. The GATE graphical user interface. Resource panel with all loaded applications, language resources, and processing resources, is shown on the left side; on the right side are shown the founded and marked annotation sets like Person, Location, and Date in different colors, and below the text is presented the annotation list with features for every annotation.

216

Named Entity Recognition in the System for Information Extraction

The application of the GATE on texts in Serbian does not provide equally good results in the individual phases: i.e., while phase 1 (identification of tokens and segmentation), gives satisfying results, the same can not be said for phase 2 (association with the appropriate parts of speech), or for phase 3 (tagging the named entities identified by the list). In fact, the problem of ambiguity which appears in phase 2 remains present even after the recognition of named entities in phase 3. One possible idea would be to use the existing statistical POS taggers adjusted for the Serbian language, but they have not given the results that are accurate enough yet, according to Popoviü (2008). In addition to ambiguity, there is also a problem of numerous morphological forms of named entities that would have to be in the gazetteer lists, which will be discussed in details below. With all above in mind, in order to overcome these problems, we came to the idea to use Unitex resources, electronic dictionaries, and local grammars, and to adapt them so they can be used in the GATE, as gazetteer lists and JAPE grammars, and give the best results for later stages of text processing in the GATE, which are language independent or can be adjusted to the Serbian language (Figure 6). Figure 6 . The main goal of this research: Adapting Unitex resources for GATE.

In the case that the organization of Unitex electronic dictionaries as GATE gazetteer lists is the goal, numerous morphological forms of named entities that would have to be in the lists represent the problem in construction gazetteer lists from Unitex electronic dictionaries in a straightforward way. Such extended lists could be generated on the basis of the electronic dictionary of named entities in the Serbian language, as illustrated in Figure 7, but that solution opens more problems: a) the size of the lists would grow by adding all the types of named entities, and, b) the inability to contain the lists of all named entities that appear in the Serbian texts.

Staša Vujiþiü Stankoviü

217

Figure 7. The emerging size of the lists with all the types of named entities. Impossible to make the lists of all named entities that appear in the Serbian texts.

In the case that the goal is to organize Unitex local grammars as GATE JAPE grammars, conversion represents the problem because of the diverse representations of finite automata in these systems. The approach assumes that two different systems, GATE and Unitex, use finite automata to annotate the text, and differ only in the representation of these automata. As there is a number of graphs for identifying named entities in the Serbian language developed for Unitex, this approach suggests the attempt of creating the equivalent JAPE grammars. In the Unitex system we distinguish two types of representation for local grammars: The graphical representation and appropriate file-format grammar representation. An example graph that corresponds to the JAPE grammar for specifying email address (Figure 4) is given in Figure 8. Graphical representation of the local grammar is presented on the left side and the appropriate file-format grammar representation is presented on the right side of the figure. More on file-format grammar representation one can find in Paumier (2008).The idea was to make automatic conversion from file-format grammar representation into the JAPE grammar. In our experimental work, we concluded that this kind of automatic conversion can not be done in a straightforward way. It will be a challenge for us in our future research work.

218

Named Entity Recognition in the System for Information Extraction

Figure 8. Unitex representations of the example graph.

4. The proposed solution and its essence In the eventually chosen approach, we first use the Unitex resources to produce electronic dictionaries for given documents. Then, based on the produced electronic dictionaries, we generate the list for the POS tagger (cf. example excerpt in Figure 9) and the appropriate gazetteer lists. We modified the POS tagger for the Serbian language, and use it to provide the appropriate part-of-speech information for documents, consulting the generated list. The GATE semantic tagger then uses word categories and produced gazetteer lists to annotate named entities. This ensures the possibility to load the text with appropriate lists into the GATE system, to clearly spot the detected named entities and categories, and then use other tools of the GATE system.

Staša Vujiþiü Stankoviü

219

Figure 9. The excerpt from the list generated for the POS tagger from the DELAF dictionary.

Example sentences from the Verne’s Novel “Around the World in Eighty Days,“ in the Serbian language, with marked entities, are given below in (4). Persons are bolded, Locations are underlined, and Dates are normalized. (4)

Godine 1872, u kuüi broj 7 u Ulici Sevil-rou Barlington Gardenz, u kojoj je 1816. godine umro Šeridan stanovao je gospodin Fileas Fog, jedan od najþudnovatijih i najzapaženijih þlanova londonskog Reform-kluba, iako je izgledalo da se on trudi da ne uþini ništa što bi moglo na njega privuüi pažnju. Mr. Phileas Fogg lived, in 1872, at No. 7, Saville Row, Burlington Gardens, the house in which Sheridan died in 1816. He was one of the most noticeable members of the Reform Club, though he seemed always to avoid attracting attention.’ Nije bilo knjižice ni knjiga koje bi gospodinu Fileasu Fogu bile suvišne, jer mu je Reform-klub stavljao na raspolaganje dve knjižice, jednu posveüenu lepoj književnosti, drugu pravu i politici. ‘There was no study, nor were there books, which would have been quite useless to Mr. Fogg; for at the Reform two libraries, one of general literature and the other of law and politics, were at his service.’

The above is what has to be automatically marked by the system. The first step is to do preprocessing of the document in Unitex and to get the appropriate DELAF dictionary as a result. The next step is to load

220

Named Entity Recognition in the System for Information Extraction

the lists generated from this dictionary for the POS tagger and gazetteer in the GATE system. After processing, document in the GATE named entities are annotated. Example sentences after processing in the GATE are shown in Figure 10. Person, Date, and Location are highlighted, and the excerpt from the annotation list is emphasized. As we can see, Sevil-rou and Gardenz are not recognized as Locations, and this is primarily because they are not in the Unitex electronic dictionaries. But, we can see in the first sentence that GATE annotated gospodin Fileas Fog ‘Mr. Phileas Fogg’ as Person in nominative, and as well in the second sentence the system annotated gospodinu Fileasu Fogu ‘Mr. Phileas Fogg’ as Person in dative, because it has gazetteer lists provided from Unitex electronic dictionaries so it can work with words in different cases and tenses. Figure 10. Example sentences in the Serbian language, with annotated named entities in GATE.

The question of named entity recognition in this case is a question of coverage of proper names with Unitex electronic dictionaries, as well as coverage of other types of entities with local grammars.

5. Evaluation For the purpose of evaluation, we used the most frequent and basic MUC evaluation measures of precision, recall and F-measure (Chinchor 1992).

Staša Vujiþiü Stankoviü

221

Precision is the fraction of the correctly identified items from all identified items, and the recall is the fraction of the items that actually were identified from all the items that should have been identified. The Fmeasure is a weighted average of the precision and recall. The origin of the de¿nition of the F-measure is presented in van Rijsbergen (1979). The evaluation metrics are defined in (5) below, using the notions presented in Table 1. Table 1. The notions needed to define evaluation metrics. The correct items are the items in the texts annotated by a native speaker that are considered to be “gold standard” against which to compare our system. Identified by the system not identified by the system

(5)

correct items tp (true positives) fn (false negatives)

not correct items fp (false positives) tn (true negatives)

Precision: Recall: Fmeasure: Default balanced Fmeasure:

We used the modified GATE system to annotate the corpus with the texts from Verne’s Novel “Around the World in Eighty Days“. These texts were also annotated by a native speaker. Then we used automated scoring tools in the GATE system to evaluate the results of the system. The results were 78% precision, 61% recall and an F-measure of 68%. This experiment of the named entity recognition from the texts in Serbian language is in early stages and because of that, the currently available results are limited to the Unitex system and its resources.

6. Conclusion and future work In this paper, the GATE system upgrade was presented. Enhancement was made in order to improve the results of its application on the texts in

222

Named Entity Recognition in the System for Information Extraction

the Serbian language. Namely, in the GATE system a number of tools was developed (for instance for Romance and Germanic languages). These resources applied on the texts in the corresponding languages show remarkable results. As the problem of the information extraction, and in particular the problem of named entity recognition is crucial in the processing of texts, it is of interest to set a good foundation for the usage of benefits that GATE could provide for the Serbian language that has different morphology and language nature from Romance or Germanic languages. It turned out that the use of the tokeniser and sentence splitter that already exist in GATE are satisfactory, and can be reused directly, while the POS tagger and gazetteer lists have to be adapted for the Serbian language. This is achieved by using resources that were previously developed for the Serbian language and the Unitex system. The solution of the problem presented in this paper is very important because better results could be achieved, new tools for processing texts in Serbian could be developed, and further improvement could be done to the existing ones, which have proved poor and inadequate for usage on the texts in Serbian. Moreover, the defects that were observed on existing solutions could be eliminated and therefore comprehensive improvement of increasingly popular area of information extraction and especially named entity recognition from the texts in the Serbian language could be done. Development of the tools for processing natural languages has a growing importance because in all aspects of human life the central part are increasing amount of information incorporated into the texts in electronic form, which are necessary to process in a fast and efficient manner.

Acknowledgments The author is thankful to professor Duško Vitas of the Faculty of Mathematics, University of Belgrade, Serbia, for valuable comments and suggestions and natural language processing group at the Faculty of Mathematics, University of Belgrade, for all resources provided.

References Chinchor, N. 1992. MUC-4 Evaluation Metrics. Proceedings of the Fourth Message Understanding Conference. Association for Computational Linguistics, Stroudsburg, PA, 22-29.

Staša Vujiþiü Stankoviü

223

—. 1998. Named Entity Task Definition. Proceedings of the 7th Message Understanding Conference (MUC-7). Fairfax, Virginia. . Courtois, B. and M. Silberztein. 1990. Dictionnaires électroniques du français. Langue française 87. Paris: Larousse. Cunningham, H. 2002. GATE, a general architecture for text engineering. Computers and the Humanities 36/2. Springer, Netherlands, 223-254. Cunningham, H. et al. 2002. GATE: A Framework and Graphical Development Environment for Robust NLP Tools and Applications. Proceedings of the 40th Anniversary Meeting of the Association for Computational Linguistics (ACL'02). Philadelphia, July 2002. . Cunningham, H. et al. 2011. Text Processing with GATE (Version 6). < http://tinyurl.com/gatebook>. Krstev, C. 2008. Processing of Serbian - Automata, Texts and Electronic dictionaries. Belgrade: Faculty of Philology, University of Belgrade. Krstev, C. and D. Vitas. 2007. The Treatment of Numerals in Text Processing. Proceedings of 3nd Language & Technology Conference, edited by Z. Vetulani, 418-422. Poznan: IMPRESJA Widawnictwa Elektroniczne S.A. Paumier, S. 2008. Unitex 2.1 User Manual. . Popoviü, Z. 2008. Evaluacija programa za obeležavanje (etiketiranje) teksta na srpskom jeziku. Beograd: Matematiþki fakultet, Univerzitet u Beogradu. Przepiórkowski, A. 2007. Slavonic information extraction and Partial Parsing. Balto-Slavonic Natural Language Processing, Association for Computational Linguistics (ACL'07). Prague: Association for Computational Linguistics, 1–10. Sekine, S. and E. Ranchhod. 2009. Named Entities: Recognition, Classification and Use. Amsterdam, Philadelphia: John Benjamins Publishing Company. Stanojþiü, Z. and Lj. Popoviü. 2002. Gramatika srpskoga jezika. Beograd: Zavod za udžbenike i nastavna sredstva. Rijsbergen, C. v. 1979. Information Retrieval. London: Butterworths.

PROSODY PREDICTION IN SPEECH SYNTHESIS BASED ON REGRESSION TREES MILAN SEýUJSKI, DARKO PEKAR, DRAGAN KNEŽEVIû AND VLADIMIR SVRKOTA

Outline The paper examines the possibility of automatic prediction of prosodic parameters, namely, f0 contour and phonetic segment durations, in the context of text-to-speech synthesis in the Serbian language. The prosodic parameters are predicted using regression trees trained on a speech corpus containing approximately 4 hours of previously recorded and annotated speech. The speech corpus was annotated with both phonological and phonetic markers, as well as markers related to prosody (lexical pitch accents, prosodic phrase boundaries and prosodic prominence). The results, in terms of values of objective measures, are comparable with those reported for other languages.

1. Introduction The paper presents the method for automatic generation of f0 contour and phonetic segment durations for the purposes of speech synthesis. The method is used within the speech synthesizer AlfaNumTTS, the most sophisticated speech synthesis system for Serbian language to date, developed at the Faculty of Technical Sciences, University of Novi Sad, in cooperation with the company “AlfaNum – Speech Technologies Ltd.” from Novi Sad (Seþujski et al. 2007). The method is based on regression trees trained on a studio recorded single speaker corpus of Serbian. The corpus has been annotated for phonemic identity as well as a number of prosodic events such as pitch accents, phrase breaks and prosodic prominence. The concept of regression trees (Breiman et al., 1984) has so far been successfully used to treat this problem for a variety of languages (cf. e.g. Dusterhoff et al. 1999 and Ni et al. 2009 for f0 prediction and Romportl & Kala 2007, Wagner & Klessa 2010, Lee & Oh 1999, Öztürk

Milan Seþujski, Darko Pekar, Dragan Kneževiü and Vladimir Svrkota

225

& Çilo÷lu 2006 and Lazaridis et al. 2007 for phonetic segment duration prediction). Being a language with not many speech or text corpora, and one of the few pitch-accent languages of Europe, Serbian has proven to be especially challenging from the point of view of the development of spoken language technologies. At the current stage of the research we use the traditional description of the intonational phonology of Serbian through four distinct accent types. It should, however, be noted that the approach using regression trees can also be used in conjunction with the representation of Serbian accents as tonal sequences, as suggested in linguistic literature (see e.g. Inkelas & Zec 1988 and Goÿevac 2005 for two somewhat different variants of the same approach). The remainder of the paper is organized as follows. Section 2 will outline several points related to the prosody of the Serbian language, relevant for its formal treatment within the process of speech synthesis. The following section will give a brief overview of the methodology of classification and regression trees with particular emphasis on its use in natural language processing. Section 4 will give a description of the AlfaNumTTS system as well as speech and language resources it relies on. The speech corpus used for the experiment as well as the conventions used in its annotation will also be described in detail in this section. Section 5 will present the system for automatic prosody prediction for Serbian based on regression trees, using the traditional paradigm of four distinct lexical pitch accents. This section will also include the results of subjective and objective evaluation of the quality of synthesized speech carrying automatically generated prosody. The concluding section will discuss the results and give an outline of the future work.

2. Remarks on the prosody of the Serbian language Serbian is one of the few Slavic languages with a pitch accent, which is a linguistic term of convenience for a variety of restricted tone systems that use variations in pitch to give prominence to a syllable or mora within a word. Unlike languages which use pitch for pragmatic highlighting, Serbian assigns pitch accent at the level of lexicon and uses it to differentiate between word meanings or values of morphological categories. Traditional grammars of Serbian define pitch accent through four distinct accent types, which involve a rise or fall in pitch associated to either long or short vowels, with optional post-accent lengths (see e.g. Iviü & Lehiste 1986). The pitch accent system of Serbian represents a challenge for the development of speech technologies (primarily text-to-

226

Prosody Prediction in Speech Synthesis Based on Regression Trees

speech synthesis and automatic speech recognition) in Serbian, for several reasons. Firstly, the task of part-of-speech (POS) tagging, one of the crucial tasks of high-level synthesis which is also closely related to text accentuation, has to be performed with high accuracy, since not only the position but also the type of the accent of each word have to be determined prior to speech signal synthesis. On the other hand, the study or prediction of the acoustic realisation of accent types in various prosodic contexts are more complex and require significantly more data in case of multiple accent types. On the phrasal level, traditional analyses of Serbian prosody identify three different intonational contours, corresponding to (1) declarative utterances, (2) prosodic questions (with syntax of a declarative statement) and (3) the vocative chant, while in Goÿevac (2005) two additional ones have been introduced: (4) signalling of continuation and (5) yes-no question. Based on these assumptions, and according to the ToBI standard (see Beckman & Hirschberg 1994), two phrase accents (Øí and LHí) and three intonational phrase boundary tones (H%, L% and HL%) have been introduced in order to describe the observed intonational contours. Most recent analyses of Serbian prosody recognize the phonological word and the intonational phrase as the prosodic constituents in Serbian. While the pitch range of each phonological word in an intonational phrase is generally smaller than the one of the preceding word, whenever a phrase contains a relatively large number of phonological words (in Goÿevac 2005 it is reported to be more than 5), the down-stepping trend is interrupted by the adjustment of the pitch range to a higher level in order to allow the speaker to sustain the remaining text within the phrase. While there is some freedom to the choice of the placement of this adjustment, it often coincides with boundaries between major syntactic constituents. The abovementioned analyses, however, fail to interpret this result as the evidence of a hierarchical prosodic structure within an intonational phrase. The system of prosodic annotation presented in Section 4 attempts to account for the variable extent of pitch range adjustment by introducing multiple levels of prosodic phrase boundaries.

3. Regression trees and their use in natural language processing The classification and regression tree (CART) technique, introduced in Breiman et al. 1984, is based on creating tree structures based on the samples in the training set (training phase) and using them for prediction of the behaviour of an unknown parameter whose value is known for all

Milan Seþujski, Darko Pekar, Dragan Kneževiü and Vladimir Svrkota

227

samples in the training set. Having been trained, the tree is able to predict an unknown quantity in a sample that does not correspond to any of the samples in the training set, but is drawn from a population with the same statistical distribution. While classification trees are concerned with assigning one of a finite number of classes to a previously unseen sample, regression trees aim at the estimation of a continuous quantity related to the sample. Owing to their practicality, both classification and regression trees have been used in many scientific fields, in cases when it is necessary to give an estimation of a quantity that is known to be dependent on a number of factors in a highly non-linear and non-homogeneous fashion. One of the areas where CART techniques are extensively used is natural signal processing. Classification trees can be used for a number of tasks related to decision-making, such as phonetic transcription (see Klatt & Shipman 1982), POS tagging (see Orphanos et al. 1999), syntacticprosodic parsing (Wang and Hirschberg 1991) or language modelling in general (see Heeman 1999). Regression trees, on the other hand, have been widely used for tasks such as automatic prediction of prosodic features of synthesized speech. By estimating the duration of phonetic segments in particular contexts using regression trees (as is the case with any other regression-based prediction) one avoids the need for explicit mathematical modelling of the influence of relevant linguistic factors, which is known to be an extremely complex problem (see van Santen et al. 1990; Campbell & Isard 1991; van Santen 1992). Besides actual values of segment durations, the method also offers insight into the complex way that particular linguistic factors interact when affecting the duration of speech segment. The basic principles of regression-tree based prediction can be explained on an (oversimplified) example related to the prediction of phoneme durations. Under the assumption that an independent tree is to be developed for each phoneme, we can consider its development for one particular phoneme (e.g. phoneme /a/) for which there are N instances in the training corpus. The first and the most important step is the identification of the feature set that will be used in the experiment. This (and only this) step requires expert knowledge, and the remainder of the procedure is entirely automatic and requires no human supervision. The feature set contains the features that are considered to be relevant to phonetic segment duration – in this case, for the sake of simplicity, we may assume that the relevant features are the stress (values: stressed and unstressed, applicable only to vowels), the position in relation to the syllable (onset, nucleus, coda) and the position in relation to intonational

228

Prosody Prediction in Speech Synthesis Based on Regression Trees

phrase boundary (measured in the number of syllables). In the training speech corpus the values of all these features, as well as actual phone durations, are known. Such a corpus is, thus, the basis for prediction of the duration of other instances of the same phoneme, based on the known values of their features. The development of the tree proceeds as follows. All possible yes/no questions based on feature values (such as “Is the phone stressed?”, “Is the distance from phrase boundary more than 3?”) are considered for all phone instances in the training corpus. A question divides the N instances of the phoneme in question (the “root node” of the tree) into two groups (“child nodes”) based on the answer, and each of the questions generally does it in a different way. The most relevant question is the one that reduces the total diversity within each of the two child nodes in comparison to the diversity within the initial root node. In other words, among the initial N phones there are phone instances of various durations, and a question is considered relevant if it divides the initial group of N phones in such a way that in one of the two subgroups there are predominantly “long” phone instances and in the other there are predominantly “short” ones. In this case, the diversity within each of the two subgroups is low, and the total diversity is lower than the diversity within the initial group (for different mathematical measures of diversity and other details see Breiman et al. 1984). In this example the notion of diversity is related to phone duration because it is phone duration that the tree is intended to predict. At this point the root node is split according to the answer to the most relevant question (e.g. “Is the phone stressed?”) into two child nodes, and the procedure is repeated recursively for each of the child nodes, eventually leading to the development of an entire tree. In such a tree, each of its terminal nodes (“leaves”) has one specific value of duration assigned to it, calculated as the average value of the duration of the phone instances associated with that node (for the sake of brevity, suffice it to say that in the regression tree terminal nodes can contain more than one phone instances, i.e. the trees are usually not fully branched). While the development of the tree is a complex and computationally demanding procedure, its use as a predictor is quite simple. A phone with known feature values and unknown duration is passed through the tree, from the root node to one of the leaves, following a path based on the answers to the questions that the phone encounters on its way, and the predicted duration is the duration assigned to the node in which the phone has ended. For example, if it is a stressed /a/, and the question related to the root node was “Is the phone stressed?”, the phone will descend from the root node into the child node containing stressed instances of /a/, where it

Milan Seþujski, Darko Pekar, Dragan Kneževiü and Vladimir Svrkota

229

will encounter another question, etc.

4. Overview of the AlfaNumTTS system and the speech corpus This section will give a brief description of the AlfaNumTTS speech synthesizer and the speech and language resources it is based on, with special emphasis on the speech corpus used for the training of the automatic prosody generation module and the conventions used in corpus annotation. Most tasks related to high-level synthesis within AlfaNumTTS, including phonetization, part-of-speech (POS) tagging (with the assignment of lexical pitch accent) and the detection of prosodic events such as prosodic phrase boundaries and prominence from text, are carried out by expert systems. Namely, POS tagging relies on a morphological dictionary containing 100,000 lemmas (3.9 million lexical forms), and lexical disambiguation is based on a combination of hand-written grammar rules and transformation rules automatically inferred from a corpus of 200,000 words of previously POS-tagged text (Seþujski 2009). As to the detection of prosodic phrase boundaries and prosodic prominence from text (assignment of specific prosodic markers to POS-tagged text), the current version of the system relies only on hand-written rules. Due to the extreme complexity of the general problem of prediction of prosodic prominence or deaccentuation from POS-tagged text, the system is restricted to assigning prominence markers only in the most obvious cases. Because of a certain degree of freedom in the choice of positions of both prosodic boundaries and prominence/deaccentuation, no objective assessment of the accuracy of the prosodic marker assignment module is possible, but its positive impact on the naturalness of synthesized speech has been confirmed through subjective evaluation (Staniü Molcer 2010). However, for the purposes of this study prosodic markers have been assigned manually, since it was our intention to evaluate the accuracy of prosody generation at the acoustic level only. Automatic prediction of prosodic features of speech at the acoustic level (f0 targets and phonetic segment durations) is based on regression trees trained on a large speech corpus, which is used for low-level synthesis as well. The corpus contains 4875 sentences (approximately 4 hours of speech) uttered by a single female voice talent, a professional radio announcer using the Ekavian standard pronunciation of Serbian. General intonation in the corpus ranged from neutral to moderately expressive, and the effort was made to keep the speech rate approximately

230

Prosody Prediction in Speech Synthesis Based on Regression Trees

constant. The corpus was recorded in a sound-proof studio and sampled at 44 kHz. The corpus has been annotated with both phonological and phonetic markers (phonemic identity and specific information related to the manner of articulation) as well as markers related to prosody (lexical pitch accents, prosodic phrase boundaries and prosodic prominence). In cases of phones whose articulation consists of more than one phonetically distinct phase (such as occlusion and explosion of stops or vocalic and non-vocalic segments of the vibrant R), annotation was carried out on a sub-phonemic level. Phonemic/phonetic annotation was carried out semi-automatically (AlfaNumASR speech recognition system, Deliü et al. 2010, was used for time-alignment of phone labels, and the results have been inspected and corrected manually), while prosodic annotation was entirely manual. In both cases manual annotation was carried out using AlfaNum SpeechLabel software. Prosodic annotation included the following: o marking of the accented vowel of phonological words (in terms of traditional pitch accents) as well as post-accent lengths; o marking of prosodic phrase boundaries and intonational phrase boundary tones, as proposed in Goÿevac 2005 (no instance of HL%, the “calling contour”, was encountered in the corpus); o marking of prosodic prominence of particular phonological words (two markers: F+, signalling positive prosodic prominence, and Fí, signalling deaccentuation of words representing ‘given’ information or function words originally accented). Each word in the speech corpus was also lemmatized and POS tagged, but most of this information is not used in prosody prediction. The inventory of prosodic labels also includes markers reserved for lexical accents, prosodic phrase boundaries and prominence that the labeller was unsure of, preventing debatable phoneme/prosodeme instances from being used for training.

5. Automatic prosody generation F0 target points and phonetic segment durations, the principal prosodic features necessary for synthesis of natural-sounding speech, are generated using independent regression trees. The corpus of 4 hours of recorded speech provides approximately 517,000 phoneme instances for training.

Milan Seþujski, Darko Pekar, Dragan Kneževiü and Vladimir Svrkota

231

5.1. Experiment In the absence of a specific acoustic model of intonation for Serbian, f0 targets are predicted directly. The prosody generation module predicts the values of f0 at 1/6, 3/6 and 5/6 of the predicted duration of each voiced phone, first time derivatives of f0 at the same locations, as well as the difference in f0 between mid-points of successive vowels. At no point is the entire f0 curve generated, and the interpolation between f0 targets is in fact postponed to the low-level synthesis stage, where segments from the speech corpus are selected so as to fit the f0 targets with minimum need for postprocessing. The prediction of phone durations is based on an independent regression tree operating on a set of features largely overlapping with the one used for f0 target prediction. As the size of the corpus allows the use of an independent regression tree for each phoneme, the phoneme identity is determined within the root node of the duration predictor. The standard approach of predicting z-scores instead of actual phone durations is used (Black & Taylor 1997). The z-score represents a number of standard deviations from the mean, and the training based on zscores requires the calculation of the mean duration for each phoneme and the z-score for each phoneme instance in the corpus. After the prediction of z-scores for a previously unseen utterance, actual durations are calculated as: phone_dur = mean + (z-score * standard_deviation) An overview of the most important features used in the experiment is given in Table 1. The phrase accents proposed in (Goÿevac 2005) were not used in the experiment, as too few instances of HLí were encountered in the corpus, which contains quite little highly expressive speech. Practically all utterances in it carry the Øí accent, typical of declarative phrases, whquestions and non-emphatic yes/no questions.

5.2. Results Prosody features for 20 utterances withheld from the training data were generated and the resulting synthesized speech was subject to objective and perceptual evaluation. The performance of the prosody predictor was evaluated by objective measures RMSE (root mean square error), MAE (mean absolute error) and CC (correlation coefficient), predominantly used in standard literature. The objective measures were calculated by the comparison of the actual values of prosodic parameters from the speech

232

Prosody Prediction in Speech Synthesis Based on Regression Trees

corpus and the values present in the synthesized speech. It should be noted here that the state of the art in the field of computational treatment of prosody does not effectively take into account the fact that there is no single correct prosody for a sentence, either in terms of segment durations or in terms of the intonational contour. For that reason, the results of objective measures, when reported, are at best accompanied with some kind of perceptual evaluation of the model. The perceptual evaluation was carried out through listening tests, where 20 listeners (native speakers) rated the TTS system performance in terms of naturalness of synthesized speech on a scale from 1 (unnatural, robotic speech) to 5 (speech with apparently natural prosodic features). The results of objective and perceptual evaluation are given in Table 2. Table 1. Feature set covered by CART questions Feature Values General prediction features lexical stress stressed, unstressed phrase weak, medium, strong, very strong, end of utterance boundary focus F+ (prosodic prominence), Fí (deaccentuation), none position position in relation to syllable (onset, nucleus, coda) and to consonant/vowel clusters, distance to stressed syllable/word boundary/phrase boundary, measured in the number of phones/morae/syllables/feet other number of syllables in word, number of syllables/words in foot/intonational phrase, syllable structure, function/content word i.p.b. tones H%, L% Features specific to duration prediction identity the identity of a particular phone (43 values, including occlusions and releases of stops, several allophones) manner of art. vowel, fricative, lateral, stop,... place of art. bilabial, alveolar, palatal, velar,... voicing voiced, unvoiced other the same features for the preceding and the following context; the match or mismatch in feature values between the phone and its preceding/following context Features specific to f0 prediction accent accent of the vowel/phonological word (SF, SR, LF, LR, post-accent length, none) position distance to specific accent measured in the number of syllables, accent of the previous/next foot

Milan Seþujski, Darko Pekar, Dragan Kneževiü and Vladimir Svrkota

233

Table 2. Results of subjective and objective measurements of the quality of synthesized speech in terms of naturalness

f0 durations

RMSE 18.6 Hz 15.8 ms

Objective evaluation MAE 16.3 Hz 12.0 ms

CC 0.61 0.91

Subjective evaluation MOS score 3.94

As to the prediction of phonetic segment durations, it can be noted that the results of the objective evaluation are somewhat better than those reported in most recent studies (cf. e.g. Öztürk & Çilo÷lu 2006 for an informative overview), which may be explained by the following factors: o o o o

o

the size of the corpus (in general, the more training data the better the results, and the corpus of 4 hours is generally considered “large”) the absence of highly expressive content from the corpus the concerted effort to sustain a constant speech rate throughout the recording sessions the meticulous procedure of manual inspection of the alignment of phone labels, with precisely defined criteria for label positioning (the same speech corpus is used for actual low-level speech synthesis within AlfaNumTTS) the use of markers intended for explicit prevention of the use of phones of apparently abnormal durations (due e.g. to speaker hesitation) in training; the markers were also used to exclude occasional bursts of particularly fast speech from training

Unlike the case of durations, the results of the objective evaluation of f0 prediction do not compare favourably with state-of-the-art (cf. e.g. Dusterhoff et al. 1999 and Ni et al. 2009), but this fact is of less importance in view of the fact that the results of subjective evaluation show that the synthesized speech is generally considered “very good” on the MOS scale, having been rated by the listeners as 3.94 out of 5. There is another reason which may explain the inferiority of the results of f0 prediction in comparison to most results reported for other languages so far, especially stress-accent languages. At this point one should remember that the corpus, which is considered large from the point of view of training for segment duration prediction, may not be large at all from the point of view of the prediction of intonational contours, for the

234

Prosody Prediction in Speech Synthesis Based on Regression Trees

reasons mentioned in Section 2. For example, in case of a simple stressaccent language, the acoustic realisation of its stress accent usually depends on a number of well-known factors (to a satisfactory degree), and can be successfully predicted in case the values of these factors are known, using a corpus of a certain size. For the sake of comparison, one can consider a pitch-accent language in which the acoustic realisation of each of its pitch accent types depends on mostly the same factors as in the previous example and the actual type of the pitch accent. Approximately speaking, the amount of training data needed for successful prediction of the acoustic realisation of a stress accent can be assumed to be equal to the amount of training data needed for just as successful prediction of the acoustic realisation of any one of the pitch accent types of a pitch-accent language. This simple reasoning shows that the corpus required for successful prediction of the acoustic realisation of Serbian accents can be expected to be several times larger than the corpus required for the same level of accuracy in a stress-accent language.

6. Conclusions and future work The paper presents an integrated approach to fully automatic generation of prosodic features of synthesized speech in the Serbian language, a pitch-accent language considered challenging from the point of view of speech technologies. Both f0 targets and phonetic segment durations have been predicted using data-driven regression trees, and the results obtained are comparable to those reported in literature for other languages. It is interesting to note that the results confirm well-known theoretical principles, such as phrase-final lengthening or the inverse dependency of phone duration from the number of syllables/words in the intonational phrase. However, an important but often underestimated factor that may have a negative influence on prosody prediction accuracy is the speech rate. Namely, it is impossible for the speaker to sustain a constant speech rate regardless of the concentrated effort, especially if the corpus recording is carried out in multiple sessions. This is particularly harmful in view of the fact that speech rate variability affects both segment durations and f0 target realization in a non-linear fashion. To the best of our knowledge, this issue has not been addressed in the related research in a systematic way, but it is, nevertheless, easily identified as the source of a considerable amount of variability in the leaves of the regression trees. It is for this reason that our future research will be oriented on robust estimation of speech rate and its incorporation into the regression-tree based model.

Milan Seþujski, Darko Pekar, Dragan Kneževiü and Vladimir Svrkota

235

Acknowledgements The research presented in the paper is supported by the technological project “The Development of Dialogue Systems for Serbian and Other South Slavic Languages” (TR32035) of the Ministry of Science and Technological Development of the Republic of Serbia.

References Beckman, M. and J. Hirschberg. 1994. The ToBI Annotation Conventions. Ms. Ohio State University, OH. Black, A. and P. Taylor. 1997. The Festival Speech Synthesis System: System Documentation. Technical Report HCRC/TR-83. Human Communciation Research Centre, University of Edinburgh , Scotland, UK. . Breiman, L. et al. 1984. Classification and Regression Trees. Boca Raton, FL: Chapman&Hall, CRC. Campbell, N. and S. Isard. 1991. Segment Durations in a Syllable Frame. Journal of Phonetics 19. Elsevier, 37–47. Deliü, V. et al. 2010. Speech Technologies for Serbian and Kindred South Slavic Languages. In Advances in Speech Recognition, edited by N. Shabtai, 141-164. InTech, Rijeka. Dusterhoff, K. E., A. W. Black and P. Taylor. 1999. Using Decision Trees within the Tilt Intonation Model to Predict F0 Contours. Proceedings of EUROSPEECH’99. Budapest, Hungary, 1627-1630. Goÿevac, S. 2005. Transcribing Serbo-Croatian Intonation. In Prosodic Typology: The Phonology of Intonation and Phrasing, edited by S. A. Jun, 146-171. Oxford: Oxford University Press. Heeman, P. A. 1999. POS Tags and Decision Trees for Language Modeling. Proceedings of Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora. College Park, MD, 29-137. Inkelas, S. and D. Zec. 1988. Serbo-Croatian Pitch Accent: the Interaction of Tone, Stress, and Intonation. Language 64/2, 227-248. Iviü, P. and I. Lehiste. 1986. Word and Sentence Prosody in SerboCroatian. Cambridge, MA: MIT Press. Klatt, D. and D. Shipman. 1982. Letter-to-Phoneme Rules: a Semiautomatic Discovery Procedure. Journal of the Acoustical Society of America 82/3, 737-793.

236

Prosody Prediction in Speech Synthesis Based on Regression Trees

Lazaridis, A. et al. 2007. A CART approach for Duration Modeling of Greek Phonemes. Proceedings of SPECOM 2007. Moscow, Russia, 287–292. Lee, S. and Y.-H. Oh. 1999. Tree-Based Modeling of Prosodic Phrasing and Segmental Duration for Korean TTS Systems. Speech Communications 28/4, 283-300. Ni, J. et al. 2009. CART-Based Modelling of Chinese Tonal Patterns with a Functional Model Tracing the Fundamental Frequency Trajectories. Proceedings of ICASSP 2009. Taipei, Taiwan, 4253-4256. Orphanos, G. et al. 1999. Decision Trees and NLP: A Case Study in POS Tagging. Proceedings of ACAI 1999. . Öztürk, Ö. and T. Çilo÷lu. 2006. Segmental Duration Modeling in Turkish. Proceedings of INTERSPEECH’06. Pittsburgh, PA, 2378-2381. Romportl, J. and J. Kala. 2007. Prosody Modelling in Czech Text-toSpeech Synthesis. Proceedings of 6th ISCA Workshop on Speech Synthesis. Bonn, Germany, 200-205. Santen J. P. H. v. and J. Olive. 1990. The Analysis of Contextual Effects on Vowel Duration. Computer, Speech & Language 4/4. Elsevier, 359–390. Santen, J. P. H. v. 1992. Contextual Effects on Vowel Duration. Speech Communication 11, 513–546. Seþujski, M. 2009. Development of Language Resources for the Serbian Language Required for Part-of-Speech Tagging. Speech and Language: Interdisciplinary Research III. Belgrade, Serbia, 125-139. Seþujski, M. et al. 2007. An Overview of the AlfaNum Text-to-Speech Synthesis System. Proceedings of SPECOM 2007. Moscow, Russia, 37. Staniü Molcer P., V. Deliü and M. Seþujski. 2010. Possibilities of Efficient Evaluation of a TTS System. Proceedings of DOGS 2010. Iriški venac, Serbia, 81-84. Wagner, A. and K. Klessa. 2010. F0 Contour and Segmental Duration Modeling Using Prosodic Features. Proceedings of Speech Prosody 2010. Chicago, IL. Wang, M. Q. and J. Hirschberg. 1991. Predicting Intonational Boundaries Automatically from Text: The ATIS Domain. Proceedings of DARPA Speech and Natural Language Workshop. Pacific Grove, CA, 378-383.

ATTENTION AND LINGUISTIC ENCODING OF MOTION EVENTS IN HUMAN-MACHINE INTERACTION MILAN GNJATOVIû AND VLADO DELIû

Outline This paper introduces a computational model of a schematic system of attention in spatial language. It expands upon the previous work – the focus tree computational model of attentional state in task-oriented human-machine interaction (HMI). This study extends the concept of the focus tree with an insight from the cognitive semantics into a dichotomy between two fundamental cognitive functions in language – of the Figure, and of the Ground. On the conceptual level, the introduced model provides an account on how to organize spatial semantic categories, and how to correlate linguistic and spatial contexts in HMI. We discuss that it supports robust automatic processing of spontaneously formulated users’ commands of different syntactic forms with no explicit syntactic expectations. Finally, the paper introduces an application of the model in the scope of therapeutic HMI with children with receptive language difficulties.

1. Introduction One of the challenges in developing conversational agents is how to address a whole variety of dialogue phenomena that are inherently present in human-machine interaction (HMI). The essence of this research question lies on the fact that utterances spontaneously produced by users do not necessarily fall within a preset system’s grammar. In the domain of HMI, particularly important manifestations of spontaneous spoken language include: varying syntactic complexity of utterances, ellipses, anaphora, exophora (e.g., references to a spatial context), constituent negation (e.g., in self-corrections), etc. Insisting on “well-formed” input would require users to devote additional effort to follow a significantly

238

Attention and Linguistic Encoding of Motion Events

restricted set of interaction rules. However, in interaction scenarios that impose a high cognitive load on the users, or wherein affected user behaviour is expected, this approach would not be well accepted by users. Therefore, there is a need to develop structures and algorithms that support the system’s decision making processes when it is confronted with spontaneously produced user inputs. It is widely accepted that attentional information is of fundamental importance for automatic processing of utterances in HMI (Grosz & Sidner 1986; Moubaiddin & Obeid 2009). This paper introduces a computational model of a schematic system of attention in spatial language. It expands upon the previous work – the focus tree computational model of attentional state in task-oriented HMI, introduced by Gnjatoviü et al. (2011b). This study extends the concept of the focus tree with an insight from the cognitive semantics into a dichotomy between two fundamental cognitive functions in language – of the Figure, and of the Ground. On the conceptual level, the introduced model provides an account on how to organize spatial semantic categories, and how to correlate linguistic and spatial contexts in HMI. To the extent that this model is computationally appropriate, we discuss that it supports robust automatic processing of spontaneously uttered users’ commands of different syntactic forms with no explicit syntactic expectations. Finally, the paper introduces an application of the model in the scope of therapeutic HMI with children with receptive language difficulties. The specification of a prototype dialogue system is inspired by a therapy scenario that includes two participants – a child and a therapist – and a nonverbal spatial context shared between them. The spatial context comprises a set of basic geometric objects that are placed on a table. As the interaction unfolds, the therapist utters various instructions that relate to manipulation of geometric objects (e.g., “show the red square”, “move it to the left”, etc.). The task of the child is to interpret these verbal stimuli and to perform the therapist’s instructions. In our HMI scenario, the user takes the role of the therapist, while the system takes the role of the child. The spatial context is represented by a graphical interface that supports manipulation of graphic objects. The user is allowed to issue commands (in textual form) and to manually manipulate graphic objects. The task of the system is to interpret and perform the user’s commands.

2. The point of departure: the focus tree The focus tree is a computational model of attentional state in taskoriented HMI introduced by Gnjatoviü et al. (2011b). The idea underlying

Milan Gnjatoviü and Vlado Deliü

239

the focus tree was to provide computationally appropriate representation of attentional information that imitates the function of a focus of attention in human perception. Therefore, this conceptualization of attentional state integrates several lines of research: neurocognitive understanding of the focus of attention in working memory (Bledowski et al. 2010; Oberauer & Lange 2009, Oberauer 2002; Cowan 1998), the notion of attention related to the theory of discourse structure in the field of computational linguistics (Grosz and Sidner 1986), investigation of a corpus that comprises recordings of spontaneous speech-based HMI (Gnjatoviü and Rösner 2010). Here, we provide a brief insight into the focus tree (for a more detailed discussion cf. Gnjatoviü et al. 2011b). The focus tree introduces a hierarchical model of attentional information. This is in line with observations made by Li and Tsai (2010). Considering the research question of defining a computationally appropriate structure to represent human knowledge, they discuss that human cognition has generally hierarchical properties, and that, therefore, hierarchical structures (i.e., trees) are appropriate to describe human cognition. However, there was an additional important reason to use a tree structure to represent attentional information. One of the most obvious principles that can be found both in cognitive models of working memory and in the theory of discourse structure is that available entities are distinguished with respect to their access status. A tree structure is used to reflect this principle. Various tree-structured discourse models in HMI have already been reported by different authors (Kirschner 2007; Stede & Schlangen 2004; Jokinen et al. 1998; Moeller 1996; McCoy & Cheng 1991; Hovy & McCoy 1989). However, a distinct feature of the focus tree model introduced by Gnjatoviü et al. (2011b) is that it is conceptualized as a system for simultaneous storage and processing of attentional information. An example of the focus tree that models spatial attention for a restricted spatial context is given in Fig. 1. Each node of the focus tree represents a specific semantic entity bound temporarily to a spatial context. These semantic entities divide in four levels: o o o

Context level – This level contains a root node that relates to the observed spatial context. Object level – This level contains nodes that relate to objects in the given spatial context that can be manipulated. The focus tree in Fig. 1 introduces two objects: the triangle and the square. Action level – This level contains nodes relating to actions that can be performed on a selected object. The focus tree in Fig. 1

Attention and Linguistic Encoding of Motion Events

240

o

specifies that two actions can be performed on the introduced objects: show and move (i.e., translate). Direction level – This level contains nodes that further specify actions on the objects. The focus tree in Fig. 1 specifies that each of the introduced objects can be translated in four directions – leftwards, rightwards, upwards and downwards.

Figure 1. An example of the focus tree.

During the course of interaction, semantic entities represented by the nodes of the focus tree may come into the focus of attention. At every moment of interaction, the current focus of attention is placed on exactly one node. Gnjatoviü et al. (2011b) introduce algorithms for robust processing of spontaneously uttered users’ commands with no explicit syntactic expectations. We briefly illustrate the mapping of users’ commands onto the focus tree for the command: I would like to shift the triangle rightwards. On the surface level, this command has three key-phrases (the triangle, shift, rightwards) that communicate propositional content (cf. the notion of filler introduced by Campbell 2007; 2006). Simultaneously, these phrases express certain semantic entities represented in the focus tree. After this command is processed, the focus of attention is placed on the node marked with grey background in Fig. 2. The criterion for determining a new focus of attention may be formulated as follows: each key-phrase in the given command express a semantic entity assigned either to the node that carries the new focus of attention or to some of its antecedents in the focus tree. Since the focus of attention is placed on a terminal node, it signals that the system has all necessary information to perform a command.

Milan Gnjatoviü and Vlado Deliü

241

Figure 2. Mapping of a command onto the focus tree.

Gnjatoviü et al. (2011b,c; 2009; 2008; 2007) discuss in detail how the focus tree may enable the system to resolve various dialogue phenomena (e.g., syntactically ill-formed utterances, elliptical utterances, anaphora, context-dependent utterances, constituent negation, etc.), and how it can be used to define adaptive dialogue strategies intended to support the user to overcome problems in interaction. However, we argue here that the assumption that – at every moment of interaction – the current focus of attention is placed on exactly one node in the focus tree is too restrictive. For example, this assumption is not appropriate for mapping of commands such as move the triangle under the square. In other words, it is not possible to map such a command onto the focus tree by placing the focus of attention on just one node. In this study, we address this limitation of the focus tree.

3. A step further: dichotomy between the figure and the ground Talmy (2000a:323-336; 2000b:21-146; 1978) provides an account on the systematic relations in language between semantic elements (i.e., within the domain of meaning) and surface (i.e., linguistic) expressions. These relations may be observed in different contexts (e.g., spatial, temporal, causal, etc.), but here we focus primarily on their manifestations in relation to semantic event of motion and location. Therefore, we provide a brief insight into the sketch of a Motion event introduced by

242

Attention and Linguistic Encoding of Motion Events

Talmy. The basic Motion event consists of four semantic entities: Figure, Ground, Path and Motion. Talmy conceptualizes these semantic elements as follows. The Figure is a (conceptually) movable entity whose path, site or orientation is conceived as a variable, while the Ground is a reference entity with respect to which the Figure’s path, site, or orientation is characterized. The Path is the path followed or site occupied by the Figure object with respect to the Ground object, while the Motion refers to the presence per se of motion or locatedness in the event (cf. Talmy 2000a: 312, 2000b:25-27). All these semantic entities can be observed in the sentences given in Fig. 3. Figure 3. Relations between surface expressions and semantic elements of the basic Motion event.

The triangle Figure Place Motion Show Motion

is Motion it Figure the triangle. Figure

above Path under Path

the square Ground the square. Ground

These sentences exemplify linguistic patterns for expression of a semantic event of location (the first and the third sentence) and motion (the second sentence). The noun phrase the triangle and the pronoun it relate to the object that functions as the Figure, and the noun phrase the square relates to the object that functions as the Ground. The verbs express the semantic category of the Motion, and refer to location (to be, to show) and motion (to place). Similarly, the spatial prepositions above and under express the semantic category of the Path, and refer respectively to a site occupied and a path followed by the Figure. It is important to note that a dichotomy between the cognitive functions of the Figure and of the Ground is essentially important for human perception. Thus, in line with our cognitively inspired approach to modelling attentional information in HMI, this dichotomy should be encapsulated in a computational model intended to support linguistic encoding of motion events. Another important remark refers to the partitioning of a spatial scene. In the interaction, the spatial scene may be partitioned in two general ways – a bipartite and a tripartite partitioning (cf. Talmy 2000a:313). In the bipartite partitioning, the Figure object and the Ground object are considered only with respect to their relation to each other, as it is the case in command move the triangle under the square. In the tripartite

Milan Gnjatoviü and Vlado Deliü

243

partitioning, the background is also taken into account as a reference frame. In command move the triangle leftward, the Figure object is considered with respect to the background. As the interaction unfolds, the user may interchangeably adopt a bipartite and a tripartite partitioning of the spatial scene. An example of the interaction domain when a bipartite partitioning may appear as appropriate is solving the Tower of Hanoi puzzle (cf. Fig. 4a). In this puzzle, disks are movable entities, and therefore they function as Figures, while the pegs are reference entities (and not movable), and function as Grounds. A typical command in this domain (with the propositional content fully formulated at the surface level) would be move the second disk on the third peg. However, a corpus-based investigation (Gnjatoviü and Rösner 2010) shows that subjects also utter commands such as rightwards. In the latter command, the subject omits to utter which disk should be moved, since he believes that this information is already known by the system (e.g., from previous interaction). With respect to propositional content, this command is elliptical. But there is one additional difference when compared with the first command – the subject adopts a tripartite partitioning of the spatial scene. The subject instructs the system to move the disk that is in the focus of attention on the next succeeding peg on the right side. This action is defined with respect to the background. To illustrate this, we may change the graphical representation of the puzzle. The underlying idea of this puzzle would not change if we decided to place the pegs one under another (cf. Fig 4b), instead of one next to another (cf. Fig 4a). The first command, move the second disk on the third peg, would communicate the same propositional content as in the first graphical representation. However, it does not hold for the second command, rightwards. In the latter graphical representation, the background (i.e., a reference frame) has changed, and this command might be interpreted as incorrect. Figure 4. A 3-disk version of the Tower of Hanoi puzzle.

a)

b)

244

Attention and Linguistic Encoding of Motion Events

4. A computational model of a schematic system of attention in spatial language This section introduces a computational model of a schematic system of attention in spatial language. It is based on the concept of the focus tree, and primarily intended to address the limitations of the focus tree discussed in the previous section. In order to encapsulate the dichotomy between the Figure and the Ground, the introduced model is a composite of two distinct but interrelated focus trees – one primarily relating to the Figure, and the other primarily relating to the Ground. We refer to them as to the Figure focus tree and the Ground focus tree. Consequently, the focus of attention also contains two components: the Figure focus of attention and the Ground focus of attention. We illustrate the introduced model for a concrete HMI domain. The prototype dialogue system includes a graphical representation of the spatial context shared between the user and the system. The spatial context comprises a set of twenty-four graphic objects (i.e., four basic geometric shapes – square, circle, triangle and rectangle – that appear in six different colours: red, green, blue, white, yellow and black). The user is allowed to manipulate these objects in two ways – by using a mouse to drag them, and by issuing textual commands. The role of the system is to interpret and perform these commands. It may also ask the user (in textual form) to provide additional information. For the purpose of easier representation of the model, we limit the number of graphic objects to two – the black square and the black triangle. The model is represented in Fig. 5. A screen display of the prototype dialogue system is given in Fig. 6. Both these objects may function as the Figure and as the Ground. The user is allowed to instruct the system to show (S) or to move (M) an object. Additional semantic entities that refer to the presence of translation of an object are: upwards, downwards, leftwards, and rightwards (cf. nodes marked with S, M, Ĺ, Ļ, ĸ, ĺ in the Figure focus tree in Fig. 5). Possible sites occupied by the Figure object with respect to the Ground object are: above, under, to the left of, and to the right of (cf. nodes marked with Ĺ, Ļ, ĸ, ĺ in the Ground focus tree in Fig. 5). To formulate a command, the user is allowed to adopt both the bipartite and the tripartite partitioning of the spatial scene. A typical command for the bipartite partitioning is move the triangle under the square, and for the tripartite partitioning – move triangle downwards (i.e., relative to its current position).

Milan Gnjatoviü and Vlado Deliü

245

Figure 5. The simplified model of attentional state.

Figure 6. A screen display of the prototype dialogue system.

To illustrate the implementation of the introduced model within the prototype dialogue system, we give an analysis of a dialogue fragment between the user and the system (cf. Fig. 7). The introduced model supports robust automatic processing of spontaneously formulated users’ commands of different syntactic forms with no explicit syntactic expectations. To achieve this, syntactic analysis of utterances is restricted to keyword spotting (cf. also Moubaiddin & Obeid 2009). To each node in the focus trees, a set of key-phrases that may express the semantic entity represented by the node is assigned. We refer to these key-phrases as to focus stimuli. During the interaction, the system takes as input a textual

246

Attention and Linguistic Encoding of Motion Events

version of the given command, and detects focus stimuli in it. The processing of the user’s commands is discussed in more detail in the rest of this section. Figure 7. A dialogue fragment between the user and the system.

I want you to show the square. User1: Move it down. User2: And now the triangle under the square. User3: (clicks on the square) User4: Shift. User5: Where should I shift the square? System6: Above the triangle. User7: (Pause in interaction) Shift that object down. User8: Which object do you want to shift: the triangle or System9: the square? Not the square. User10: In command User1 (I want you to show the square), the system detects two focus stimuli: show and the square. Observed out of the interaction context, the first focus stimulus (show) may express two semantic entities represented by two nodes marked with S in the Figure focus tree. The second focus stimulus (the square) may express two semantic entities represented by two nodes marked with Ŷ that belong respectively to the Figure focus tree and the Ground focus tree. To interpret this command, the system needs to properly relate these focus stimuli to corresponding semantic entities. Although show comes before the square in the command, the second focus stimulus is more general1, so the system first relates the triangle to a semantic entity. Since the system does not detect spatial prepositions in the command, it concludes that the triangle does not express a semantic entity related to the Ground. Therefore, it relates this focus stimulus to the node Ŷ in the Figure focus tree. As an intermediate step, the temporary Figure focus of attention is placed on this node (cf. Fig. 8). The mapping of the first focus stimulus (show) is now restricted to the sub-tree determined by this node as its root node. This restriction implies that show will be related to the child node (marked with S) of the 1

In our approach, a focus stimulus s1 is considered to be more general than a focus stimulus s2 if s1 expresses a semantic entity that is related to a node that is on a higher level of the focus tree than a node that represents a semantic entity expressed by s2.

Milan Gnjatoviü and Vlado Deliü

247

node that carries the temporary Figure focus of attention. The Figure focus of attention is now placed on that child node. This node is a terminal node, which means that the system has all necessary information to perform a command. In the graphical interface, it marks the square in an appropriate manner. Transition of the Figure focus of attention for command User1 is given in Fig. 8. For this command, the Ground focus of attention remains undefined. Figure 8. Transition of the Figure focus of attention for command User1.

The processing of command User2 (Move it down) illustrates how the system resolves anaphora by taking into account attentional state. It should be noted that, at this point, the system is not aware that there is a reference (it) at the surface level that relates to a graphic object. In this command, the system detects just two focus stimuli: move and down. The first focus stimulus may be related to two nodes marked with M in the Figure focus tree. However, the system’s decision is determined by the current Figure focus of attention and the topology of the Figure focus tree. On the conceptual level, the system relates the focus stimulus move to a node marked with M that is “the closest” to the current Figure focus of attention, and then places the Figure focus of attention on this node. On the implementation level, the system detects the minimal sub-tree that contains both the current Figure focus of attention and a node marked with M (cf. Fig. 9). In general, the second focus stimulus, down, may be related to two nodes marked with Ļ in the Figure focus tree. However, the system relates this focus stimulus to the node Ļ that belongs to the sub-tree determined by the temporary Figure focus of attention as its root node, and then places the Figure focus of attention on it. This node is a terminal node, which means that the system has all necessary information to perform a command. The reference of the observed anaphora is resolved by taking the Figure focus of attention into account. The referenced object (that should be shifted downwards) is represented by an antecedent node of the current Figure focus of attention, i.e., the square.

248

Attention and Linguistic Encoding of Motion Events

We point out that the anaphora was resolved although the system is not even aware of the reference it on the surface level. Such an approach has two advantages. The first advantage is that even if the reference it had not been explicitly specified, the command would have been processed in the same manner. This is important because the user does not necessarily have to explicitly include anaphora, especially when the referenced object is already appropriately marked in the graphical interface. For example, command move down does not contain anaphora, but its propositional content is the same. The second advantage is that it allows a robust processing of users’ commands without taking syntactic information related to anaphora into account. Transition of the Figure focus of attention for command User2 is given in Fig. 9. The Ground focus of attention still remains undefined. Figure. 9. Transition of the Figure focus of attention for command User2.

In the command User3 (And now the triangle under the square) contains three focus stimuli: the triangle, the square and under. Since the spatial preposition under comes immediately before the square, the system concludes that the square expresses the Ground. The Ground focus of attention is first placed on the node Ŷ in the Ground focus tree, and then transited to its child node Ļ (cf. the right part of Fig. 10). Now, when the Ground in determined, the system concludes that the remaining focus stimulus (the triangle) expresses the Figure, and places the Figure focus of attention on the node Ÿ in the Figure focus tree (cf. the left part of Fig. 10). After processing of this command, the Figure focus of attention determines which object should be moved (i.e., the triangle), while the Ground focus of attention determines where it should be moved (i.e., under the square).

Milan Gnjatoviü and Vlado Deliü

249

Figure 10. Transition of the Figure and the Ground foci of attention for command User3.

User4 is a nonverbal command. The user clicks on the square in the graphical interface. The system concludes that the user wants to perform an action on this object. As a respond, it placed the Figure focus of attention on the node Ŷ in the Figure focus tree (cf. the left part of Fig. 11). The command User5 (Shift) is elliptical – it provides just one focus stimulus. Observed out of the interaction context, this command does not specify which object should be shifted. However, similarly as discussed above, the system resolves this ellipsis by taking the current Figure focus of attention and the topology of the Figure focus tree into account. It relates the focus stimulus to the closest node marked with M, and places the Figure focus of attention on it (cf. the left part of Fig. 11). At this point, the system is aware that the square should be shifted, but still does not have information on where it should be shifted. Therefore, according to its dialogue strategy, the system takes initiative and asks the user where to shift the square (System6). In command User5 (Above the triangle), the user provides the requested information. In this command, the spatial preposition above comes immediately before the trianlge, so the system concludes that the trianlge expresses the Ground. The Ground focus of attention is first placed on the node Ÿ in the Ground focus tree, and then transited to its child node Ĺ (cf. the right part of Fig. 11). Now, when the Ground is determined, the system has complete information, and shifts the square above the triangle.

250

Attention and Linguistic Encoding of Motion Events

Figure 11. Transition of the Figure and the Ground foci of attention for command User4, User5 and User7.

Then a pause occurs in the interaction. We recall that the idea underlying our approach is to provide computationally appropriate representation of attentional information that imitates the function of a focus of attention in human perception. Humans have an inherently limited attention span, and when a pause occurs in the interaction, the system assumes that semantic entities that were in the Figure and the Ground foci of attention before the pause will not necessarily be salient after the pause. In other words, the system intentionally “forgets” that the square was in the Figure focus of attention. Consequently, for command User8 (Shift that object down), there are two possible interpretations: shift the triangle down and shift the square down. Two possible transitions of the Figure focus of attention are given in Fig. 12. According to its dialogue strategy, the system takes initiative and asks the user to choose one of these two options (System9) . Constituent negation in command User10 (Not the square) excludes one of these options. Figure 12. Transition of the Figure focus of attention for command User8 and User10.

Milan Gnjatoviü and Vlado Deliü

251

5. Concluding remarks This paper introduced a computational model of a schematic system of attention in spatial language. Although the model is discussed for a concrete HMI scenario, the introduced concepts are designed to be used in a wide range of spatial contexts. Gnjatoviü (2011a) reports a corpus-based investigation on the conceptual appropriateness of the focus tree for updating the focus of spatial attention in speech-based HMI. The corpus contains about fifteen hours of recordings of speech-based HMI produced in a Wizard-of-Oz simulation (cf. Gnjatoviü & Rösner 2010). In the simulation settings, ten subjects (native German speakers) were asked to solve a set of problems in a graphical system. Some of the problems are illustrated in Fig. 13. The conceptual appropriateness of the focus tree was confirmed with respect to its hierarchical organization, and correlation between linguistic and spatial contexts. Figure. 13. Domains of interaction in the NIMITEK corpus.

a) Tangram

b) Grid puzzle

c) Tower of Hanoi puzzle

d) The three jugs problem

e) Filling empty space

f) Classification

The introduced model is by no means intended to be a general solution for modelling spatial information, but rather a conceptual improvement of the existing methodology for modelling spatial attention in HMI. It extends the concept of the focus tree with an insight from the cognitive semantics

252

Attention and Linguistic Encoding of Motion Events

into a dichotomy between two fundamental cognitive functions in language – of the Figure, and of the Ground. On the conceptual level, the introduced model provides an account on how to organize spatial semantic categories, and how to correlate linguistic and spatial contexts in HMI. This is in line with recent findings in neurosciences that spatial attention and internal representations in working memory are closely interrelated (Nobre et al. 2004; Griffin & Nobre 2003; Wheeler & Treisman 2002; Awh & Jonides 2001). In addition, the previous section discussed how the introduced model enables the prototype dialogue system to process spontaneously formulated users’ commands of different syntactic forms with no explicit syntactic expectations. We provided an analysis of the dialogue fragment between the user and the system, and explained in detail how the system processes anaphora, ellipses, constituent negations, nonverbal commands, pauses in the conversation, etc. It is important to note that the observed interaction scenario was deliberately selected for the purpose of illustration, because it does not include additional information related to an underlying task structure (or intentional structure). In general, the system may be aware of specific rules of the task and of projected intentions of the user. This information may additionally support robust processing of the user’s utterances (cf. Gnjatoviü 2011c; 2008). However, in this paper, we deliberately excluded2 such additional information from the interaction scenario, in order to illustrate robust processing of the user’s commands that relies solely on attentional information. Finally, we point out an important research question that falls out of the scope of the reported corpus-based investigation, namely whether the introduced model of a schematic system of attention in spatial language supports language-dependent categorization of spatial relations. In classical physics, space is considered as represented by a continuum. However, in language, we partition space into discrete basic spatial categories (Bowerman 1996:145). For example, to describe a relation between two objects in space, we may choose one of the following spatial semantic categories: containment (i.e., object A is in object B), support and contiguity (i.e., object A is on object B), occlusion (i.e., object A is under object B), etc. On the level of surface expressions in English, each of these semantic categories is expressed by a spatial preposition, e.g., in, on, under, etc. We observe this semantic-to-surface association to emphasize the fact that languages differ in how they categorize spatial 2 A careful reader may have noticed that the prototype dialogue system did not take actual positions of the graphic objects into account when processing commands.

Milan Gnjatoviü and Vlado Deliü

253

relations. This is illustrated for the following three spatial relations in English, German and Serbian (Fig. 14): Figure 14. Examples of spatial relations in English, German and Serbian.

The picture Das Bild ɋɥɢɤɚ

hangs hängt ɜɢɫɢ

on an ɧɚ

the wall. der Wand. ɡɢɞɭ.

The book Das Buch Ʉʃɢɝɚ

lay liegt ɥɟɠɢ

on auf ɧɚ

the table. dem Tisch. sɬɨɥɭ.

In Auf ɇɚ

the picture, dem Bild ɫɥɢɰɢ

there is ist ʁɟ

a man. ein Mann. ɱɨɜɟɤ.

In Serbian, all three spatial relations fall in a single spatial category. Thus, the same spatial preposition (ɧɚ) expresses all three spatial relations. In contrast to this, in English and German, these relations fall in two categories. In English, the first and second spatial relations (expressed with the preposition on) belong to the same spatial category, while in German, the second and third spatial relations (expressed with the preposition auf) belong to the same spatial category (cf. Fig. 15). A more detailed discussion on different categorization of spatial relations is provided by Khetarpal et al. (2009), Bowerman & Choi (2004), Levinson & Meira (2003), Munnich et al. (2001), Talmy (2000a:177-254) and Bowerman (1996). Here, we used these examples of spatial relationships to introduce the question of whether the introduced model supports language-dependent categorization of spatial relationship. Figure 15. Categorization of spatial relations in English, German and Serbian

254

Attention and Linguistic Encoding of Motion Events

We conducted a preliminary investigation on whether the introduced model is appropriate for modelling spatial attention in English, German and Serbian (e.g., the prototype dialogue system introduced in this paper supports interaction in these languages). The results are promising and indicate that the model could support cross-linguistic modelling of spatial information. A more extensive investigation on this issue is planned.

Acknowledgements The presented study is performed as part of the projects “Design of Robots as Assistive Technology for the Treatment of Children with Developmental Disorders” (III44008) and “Development of Dialogue Systems for Serbian and Other South Slavic Languages” (TR32035), funded by the Ministry of Education and Science of the Republic of Serbia. The responsibility for the content of this paper lies with the authors.

References Awh, E. and J. Jonides. 2001. Overlapping mechanisms of attention and spatial working memory. Trends in Cognitive Sciences 5/3, 119–126. Bledowski, C., J. Kaiser and B. Rahm. 2010. Basic operations in working memory: contributions from functional imaging studies. Behavioural Brain Research 214/2, 172–9. Bowerman, M. 1996. The origins of children's spatial semantic categories: cognitive vs. linguistic determinants. In Rethinking linguistic relativity, edited by J. Gumperz and S. C. Levinson, 145-176. Cambridge, UK: Cambridge University Press. Bowerman, M. and S. Choi. 2004. Space under construction: Languagespecific spatial categorization in first language acquisition. In Language in mind, edited by D. Gentner and S. Goldin-Meadow, 387– 427. Cambridge, MA: MIT Press. Campbell, N. 2006. On the Structure of Spoken Language. Proceedings of the 3rd International Conference on Speech Prosody 2006. Dresden, Germany. —. 2007. Towards conversational speech synthesis; lessons learned from the expressive speech processing project. Proceedings of the Sixth ISCA Workshop on Speech Synthesis (SSW6). Bonn, Germany, 22–27. Cowan, N. 1988. Evolving conceptions of memory storage, selective attention, and their mutual constraints within the information processing system. Psychological Bulletin 104/2, 163-191.

Milan Gnjatoviü and Vlado Deliü

255

Gnjatoviü, M. 2009. Adaptive Dialogue Management in Human-Machine Interaction. Munich: Verlag Dr. Hut. —. 2011. Updating the Focus of Spatial Attention in Speech-Based Interaction: A Corpus-Based Investigation. Proceedings of the 14th International Conference “Speech and Computer”. Kazan, Russia, 441-446. Gnjatoviü, M. and D. Rösner. 2007. An approach to processing of user’s commands in human-machine interaction. Proceedings of the 3rd Language and Technology Conference (LTC’07). Adam Mickiewicz University, Poznan, Poland, 152-156. —. 2008. Adaptive Dialogue Management in the NIMITEK Prototype System. Proceedings of the 4th IEEE Tutorial and Research Workshop Perception and Interactive Technologies for Speech-Based Systems (PIT’08). Lecture Notes in Computer Science 5078. Berlin, Heidelberg: Springer-Verlag, 14–25. —. 2010. Inducing genuine emotions in simulated speech-based humanmachine interaction: The NIMITEK corpus. IEEE Transactions on Affective Computing 1, 132–144. Gnjatoviü, M., D. Pekar and V. Deliü. 2011. Naturalness, adaptation and cooperativeness in spoken dialogue systems. Toward Autonomous, Adaptive, and Context-Aware Multimodal Interfaces, Theoretical and Practical Issues. Lecture Notes in Computer Science 6456. Berlin, Heidelberg: Springer-Verlag, 298–304. Gnjatoviü, M., M. Janev and V. Deliü. 2011. (in press). Focus Tree: Modeling Attentional Information in Task-Oriented Human-Machine Interaction. Applied Intelligence. Springer, Netherlands. DOI: 10.1007/s10489-011-0329-5. Griffin, I. C. and A. C. Nobre. 2003. Orienting attention to locations in internal representations. Journal of Cognitive Neuroscience 15/8, 1176–1194. Grosz, B. and C. Sidner. 1986. Attention, Intentions, and the Structure of Discourse. Computational Linguistics 12/3, 175–204. Hovy, E. and K. McCoy. 1989. Focusing your RST: a step toward generating coherent, multisentential text. Proceedings of the 11th Annual Conference of the Cognitive Science Society. Ann Arbor, 667674. Jokinen, K., H. Tanaka and A. Yokoo. 1998. Context Management with Topics for Spoken Dialogue Systems. Proceedings of COLING-ACL 1998, 631–637. Khetarpal, N., A. Majid and T. Regier. 2009. Spatial terms reflect nearoptimal spatial categories. Proceedings of the 31st Annual Conference

256

Attention and Linguistic Encoding of Motion Events

of the Cognitive Science Society, edited by N. Taatgen et al, 23962401. Austin, TX: Cognitive Science Society. Kirschner, M. 2007. Applying a Focus Tree Model of Dialogue Context to Interactive Question Answering. Proceedings of the ESSLLI’07 Student Session. Dublin, Ireland, 135–147. Levinson, S. C. and S. Meira. 2003. Natural concepts in the spatial topological domain—adpositional meanings in cross-linguistic perspective: An exercise in semantic typology. Language 79, 485-516. Li, S.-T. and F.-C. Tsai. 2010. Constructing tree-based knowledge structures from text corpus. Applied Intelligence 33/1. Springer, Netherlands, 67–78. McCoy, K. and J. Cheng. 1991. Focus of attention: Constraining what can be said next. In Natural Language Generation in Artificial Intelligence and Computational Linguistics, edited by C. L. Paris, W. R. Swartout and W. C. Moore, 103–124. Norwell, Mass.: Kluwer Academic Publishers. Moeller, J.-U. 1996. Domain Related Focus-Shifting Constraints in Dialogues with Knowledge Based Systems. In Trends in Natural Language Generation: An Artificial Intelligence Perspective, edited by G. Adorni and M. Zock, 188-204. Heidelberg: Springer-Verlag. Munnich, E., B. Landau and B. Dosher. 2001. Spatial language and spatial representation: A cross-linguistic comparison. Cognition 81, 171–208. Moubaiddin, A. and N. Obeid. 2009. Partial information basis for agentbased collaborative dialogue. Applied Intelligence 30/2. Springer, Netherlands, 142–167. Nobre, C. A. et al. 2004. Orienting attention to locations in perceptual versus mental representations. Journal of Cognitive Neuroscience 16/3, 363–373. Oberauer, K. 2002. Access to information in working memory: Exploring the focus of attention. Journal of Experimental Psychology: Learning, Memory, and Cognition 28/3, 411-421. Oberauer, K. and E. B. Lange. 2009. Activation and binding in verbal working memory: A dual-process model for the recognition of nonwords. Cognitive Psychology 58/1, 102-136. Stede, M. and D. Schlangen. 2004. Information-Seeking Chat: Dialogue Management by Topic Structure. Proceedings of the 8th Workshop on Semantics and Pragmatics of Dialogue, CATALOG 04. Barcelona, Spain. . Talmy, L. 1978. Figure and Ground in Complex Sentences. In Universals of Human Language: Syntax 4, edited by J. Greenberg, 625-649. Stanford: Stanford University Press.

Milan Gnjatoviü and Vlado Deliü

257

—. 2000a. Towards Cognitive Semantics, Vol. 1: Concept Structuring

Systems. Cambridge, Mass.: MIT Press. —. 2000b. Towards Cognitive Semantics, Vol. 2: Typology and Process in

Concept Structuring. Cambridge, Mass.: MIT Press. Wheeler, M. E. and A. M. Treisman. 2002. Binding in short-term visual memory. Journal of Experimental Psychology: General 131, 48–64.

CONTRIBUTORS

Joanna Báaszczak is an Associate Professor at the Institute of English, University of Wrocáaw, Poland. She received her PhD in 2000 from the Humboldt-University of Berlin and Habilitation degree in 2007 from the University of Potsdam. She is currently Head of the Centre for General and Comparative Linguistics at the University of Wrocáaw. She is also coordinating the project TRAIT (Temporal Relations at the Interfaces) sponsored by the Fundation for Polish Science. In 2010 she obtained an additional grant for setting up a neurolinguistic laboratory with the EEG equipment and is currently involved in various Event Related Potentials (ERP) experiments. Her scientific interests include but they are not limited to theoretical syntax (in particular minimalist syntax), processing of negative polarity items, negation, indefinites, case, copula constructions, existential and possessive constructions, aspect, tense, modality, (un)ergativity, Information Structure. Željko Boškoviü is Professor of Linguistics at the University of Connecticut. His main professional interests include Syntactic Theory and Comparative Syntax. He has authored two books, The Syntax of Nonfinite Complementation: An Economy Approach (MIT Press, 1997) and On the Nature of the Syntax-Phonology Interface: Cliticization and Related Phenomena (Elsevier, 2001), and has co-edited Minimalist Syntax (with Howard Lasnik, Blackwell, 2007). He has also published a large number of highly influential articles in refereed journals and books. He has held visiting positions throughout Europe and the United States, as well as in Brazil and India. He has supervised over 30 doctoral dissertations. Vlado Deliü, PhD, is Professor, researcher and project manager at the Faculty of Technical Sciences, University of Novi Sad, Serbia, and a visiting professor at the University of Banjaluka, Bosnia and Herzegovina. He has been leading major research projects in the field of speech technologies in Serbia. Currently, he manages the largest (2011-2014) regional research project dedicated to the development of human-machine dialogue systems and speech technologies, bringing together 32 researchers from six scientific research institutions and three companies from Serbia, FYR Macedonia, Slovenia, and Bosnia and Herzegovina.

Selected Papers from SinFonIJA 3

259

Prof. Deliü has (co-)authored several books, two patents, and more than 150 research and technical articles in scientific journals and conference proceedings. His main professional contribution is the establishment and leadership of the most prominent research and development group in the field of speech technologies in South Slavic languages. His work also focuses on innovative applications of speech technologies: voice portals, assistive technologies for the disabled, e-mail readers, advertising monitors, voice dial systems, etc. For his contribution to innovations, Prof. Deliü received several prestigious awards, including the national award “Miša Anastasijeviü” and the “Gold Archimedes” grand prize of the Moscow International Salon of Industrial Property (2008). Milan Gnjatoviü received his PhD degree in computer science from Ottovon-Guericke University Magdeburg, Germany, in 2009. He works as a postdoctoral researcher at the Department of Power, Electronics, and Communications Engineering at the University of Novi Sad, Serbia. He has also worked as a research assistant in the Department of Knowledge Processing and Language Engineering at Otto-von-Guericke University Magdeburg. His research interests include adaptive dialogue management in human-machine interaction, cognitive technical systems, affective computing, and natural language processing. He participated in several interdisciplinary research projects in the field of human-machine interaction, including neurobiologically inspired technical communication systems (the Excellence Programme “Neuroscience” funded by the federal state of Saxony-Anhalt, Germany), cognitive technical systems (the Transregional Collaborative Research Centre SFB/TRR 62 funded by the German Research Foundation), and natural language spoken dialogue systems (the Research Cycle 2011-2014 funded by the Republic of Serbia). He was a member of the European COST 2102 International Scientific Committee. Currently, Milan Gnjatoviü is the leader of the subproject “Multimodal human-machine dialogue” within the interdisciplinary research project “Design of Robots as Assistive Technology for the Treatment of Children with Developmental Disorders“, funded by the Ministry of Education and Science of the Republic of Serbia. Pavel V. Grashchenkov obtained his PhD in 2006 at the Department of Theoretical and Applied Linguistics with the dissertation entitled “Syntax and Typology of Genitive Phrase” and then joined the Language Department of the Institute of Oriental Studies in Moscow. In syntactic theory, his primary areas of interest are the syntax of verbal and nominal projections. P. Grashchenkov’s studies of the internal structure of nominal

260

Contributors

and verbal projections deals with the problems of DP-CP parallelism, inherent vs. structural case distinction, position of structural case assignment, DP and VP cartography and similar phenomena. As a typologist, Pavel Grashchenkov studied Turkic, Daghestanian, Mongol, Iranian, Slavic and some other languages. He regularly takes part in the field research trips of the Department of Theoretical and Applied Linguistics. He has more than 25 papers on syntax and typology, mostly written in Russian. In addition to working as a researcher in the Institute of Oriental Studies and teaching courses at the Department of Theoretical and Applied Linguistics, P. Grashchenkov also works as a computational linguist. Sabina Halupka-Rešetar is an Assistant Professor at the Department of English, Faculty of Philosophy, University of Novi Sad, where she currently teaches undergraduate courses in Minimalist syntax and Information structure, as well as graduate courses in Comparative syntax and the Left periphery. She received her PhD in 2009 at the University of Novi Sad, having defended the thesis entitled The structural articulation of focus in English and Serbian: a comparative analysis of the structural and the generative approach. Her main scientific interest lies in generative syntax, information structure, the syntax-discourse interface and comparative (primarily English-Serbian-Hungarian) linguistics, with special interest in the issue of (the realization of) contrastive focus and whconstructions. She has participated in various linguistics conferences and has published over thirty papers at home and abroad. Patrycja JabáoĔska is an Assistant Professor at the Institute of English, University of Wrocáaw, Poland. She received her PhD in 2007 from the University of Tromsø. She is currently involved in the project TRAIT (Temporal Relations at the Interfaces) sponsored by the Fundation for Polish Science. As part of the project she designs and runs electrophysiological experiments using the Event Related Potentials technique. Her research interests include: nanosyntax, division of labour between syntax and morphology, argument structure issues (e.g. passives, causatives, reflexives, etc.), argument coding (nominative/accusative versus ergative/absolutive systems), implementation of scales in formal grammar (especially Silverstein’s hierarchy), degrees of finiteness, lexical categories, and neurolinguistics.

Selected Papers from SinFonIJA 3

261

Dorota Klimek-Jankowska is an Assistant Professor at the Institute of English Studies, University of Wrocáaw, Poland. She studied linguistics at the University of Wrocáaw and at Utrecht University. She obtained her PhD in 2008 at Wrocáaw University. Currently, she is involved in the project TRAIT (Temporal Relations at the Interfaces) sponsored by the Fundation for Polish Science. In her subproject she focuses on the interaction of Tense, Aspect and Modality in the expression of future time reference and in the expression of actuality entailment in Slavic. Her main areas of interest are formal semantics, syntax-semantics interface, semantic variation (macro- and micro-parameters) and the application of experimental methods in the study of semantics and pragmatics of natural language. Dragan Kneževiü is a PhD student at the Faculty of Technical Sciences, University of Novi Sad, Serbia, where he currently works as a Research Assistant. He is a member of the Communications and Signal Processing Group. He is engaged as a researcher and developer in Digital Signal Processing and Speech Synthesis. Through the results of his work he contributed to the development of the highest quality speech synthesizers in Serbian and Hebrew. This software is now widely used, particularly by the blind and visually impaired computer users. His current research is related to the prediction of prosodic features. Maja Markoviü, PhD, works as an Assistant Professor at the English Department of the University of Novi Sad, where she teaches courses at BA and MA level. She specialized in English Phonetics and Phonology and Second language phonology. Her main work has been in the field of experimental phonetics, where she investigates the segmental and suprasegmental features of Serbian, the acquisition of the phonological system of English as a second language and the phonological systems of Serbian and English in contrast. Her recent work has been in the field of phonology/syntax interface, with special focus on clitics and their phonological behaviour in Serbian. Maja Markoviü is the editor of the Book of Selected Papers from the 1st International Conference ELALT 2011 - English Studies Today: Views and Voices (together with Dr. Ivana Ĉuriü Paunoviü).

262

Contributors

Franc Lanko Marušiþ got his PhD from Stony Brook University in 2005, when he joined the University of Nova Gorica as an Assistant Professor. He teaches Slovenian syntax, general linguistics, and a language consultancy seminar at undergraduate level and syntax at PhD level. His main areas of interest are Slovenian syntax, comparative Slavic syntax, DP syntax and syntactic theory. He published papers in various journals (including Linguistic Inquiry and Natural Language and Linguistic Theory), gave several talks at various (not only Slavic) conferences and published several papers in refereed conference proceedings (including NELS, FASL, FDSL, ConSOLE etc.). He also co-edited the volume Studies in Formal Slavic Linguistics with selected papers from a conference organized in Nova Gorica. He is one of the initiators of the SinFonIJA conference series. Krzysztof Migdalski is an Assistant Professor at the Institute of English Studies, University of Wrocáaw, Poland. He studied linguistics at the University of Tromsø, then he worked as a research assistant at Tilburg University, where he defended his doctoral dissertation “The Syntax of Compound Tenses in Slavic” in 2006. From 2006 till 2008 he was a postdoctoral researcher at the University of Connecticut, where he carried out a project investigating second position phenomena in syntax. He is interested in the comparative syntax of Balkan and Slavic languages and historical linguistics. He is currently involved in the the project TRAIT (Temporal Relations at the Interfaces) sponsored by the Foundation for Polish Science “Temporality at the Interfaces”, in which he investigates diachronic modifications of the tense systems in Slavic languages. Tanja Miliüev works as a lecturer at the English Department, Univeristy of Novi Sad, where she teaches English syntax, historical linguistics, and a research seminar in theoretical grammar. Her research interests include Old English syntax, Slavic syntax, pronominal syntax, and the syntaxphonology/pragmatics interface. The topic of her PhD thesis (in progress) is the influence of pragmatic/information-structural factors on word order variation in Old English. She has published her work as volume chapters (Grammatical change: Origins, Nature, Outcomes), articles in refereed conference proceedings (including Formal Approaches to the Description of Slavic Languages, English Historical Linguistics 2006) and journal articles (including Acta Linguistica Hungarica, York Papers in Linguistics), and has presented it at various international conferences (e.g. FDSL, York-Holland Symposium, DGfS, SinFonIJA). She was a co-editor of the volume Novi Sad Generative Syntax Workshop 2 Proceedings.

Selected Papers from SinFonIJA 3

263

Darko Pekar, MSc, graduated from the Faculty of Technical Sciences, University of Novi Sad, in 1998, in the field of speech technologies. Until 2003 he was the leading expert at the AlfaNum project, where he obtained wide-ranging experience in the field of speech technologies and their applications, as well as management of scientific and technological projects. In 2003 he became CEO of the AlfaNum company, and in cooperation with the FTS he continues to manage teams working on the development and application of speech technologies. He leads a team of 10 software engineers and accompanying staff. Although he focuses on practical development and providing market-ready ASR&TTS products, he has also published numerous articles and papers in domestic and international journals. In this project he will be mainly involved in the adaptation of ASR and TTS, their optimisation and integration into SDS, as well as management of Axon. Teodora Radeva-Bork currently holds the position of a postdoctoral researcher and lecturer at the University of Potsdam, Institute of Slavic Languages, Germany. Her main research interests are related to syntax, the morphology-syntax and syntax-semantics interfaces, first language acquisition as well as the nature of language from a biolinguistic perspective. In her dissertation at the University of Vienna, Institute of Linguistics, Austria she studied the phenomena of single and double cliticization from the perspective of both adult and child grammar. Although the main focus is the analysis of data from Bulgarian, the work draws cross-linguistic parallels to some of the Romance and Balkan Sprachbund languages. Two unprecedented experimental studies with monolingual Bulgarian children are conducted: elicited production of clitics in pre- and post-verbal position, and elicited comprehension of clitic doubling constructions. Henk van Riemsdijk is Professor Emeritus of Linguistics at the University of Tilburg. His seminal ideas have shaped the development of generative grammar ever since the late 70s. Since his Introduction to the Theory of Grammar (with E. Williams, 1986), he co-authored a number of highly influential books on linguistic theory. Most recently, he was one of the editors of a major reference work The Blackwell Companion to Syntax (with Martin Everaert, 2006). He is co-editor of the Journal of Comparative Germanic Linguistics and is a consulting editor for the Linguistic Review, as well as editor of the series Studies in Generative Grammar since its foundation. Henk van Riemsdijk is the initiator and co-founder of the GLOW society, the most prominent generative linguistics research

264

Contributors

platform in Europe. As one of the world leading authorities in natural language syntax he has published numerous papers in a number of prominent linguistic journals. He was the head of linguistic department at the University of Tilburg and the leader of a well-known Models of Grammar project, which ended in 2008. In the course of his impressive teaching career he has taught in the major linguistic departments worldwide and has influenced the work of many prominent syntacticians today. Milan Seþujski, PhD, is an Assistant Professor at the Faculty of Technical Sciences, University of Novi Sad, Serbia, where he is a member of the Communications and Signal Processing Group. He is engaged as a lecturer in Digital Signal Processing, Digital Filters as well as Acoustics and Audio Engineering. His areas of scientific interest include computational linguistics and human-machine interaction. In 2002 he earned the MSc degree with the thesis “Prosodic elements in text-to-speech synthesis in the Serbian language”. The result of the thesis research is the highest quality speech synthesizer in Serbian, with elements of natural intonation. This software is now widely used, particularly by blind and visually impaired computer users. In 2009 he earned the PhD degree with the thesis “Automatic part-of-speech tagging in the Serbian language”, which resulted in an improvement in the accuracy of automatic POS tagging in Serbian, which is important for many applications of language and speech technologies (text-to-speech synthesis and speech recognition, machine translation, dialogue systems, information retrieval etc.). His current research includes natural language processing as well as mathematical modeling of the prosodic features of speech. Vladimir Svrkota currently works as a researcher and software developer at the company “AlfaNum – speech technologies”, Novi Sad, Serbia. He graduated from the Faculty of Technical Sciences, University of Novi Sad, and has since worked on a number of software development projects. In the company “AlfaNum - speech technologies” he actively participates in the development of text-to-speech systems for Serbian as well as Hebrew. In this work he is engaged at tasks such as prosody modelling, as well as the development of speech application programming interfaces (SAPI) and interfaces suitable for the visually impaired.

Selected Papers from SinFonIJA 3

265

Radek Šimík got his PhD in general linguistics in 2011 at the University of Groningen after completing his undergraduate studies in Czech and English philology at the Palacký University Olomouc. He is currently affiliated to the University of Potsdam, where he works as a postdoctoral researcher in the Collaborative Research Center 632 “Information Structure”. In his work he has mainly concentrated on the syntax, semantics, and information structure of various operator-variable structures such as relative clauses, focus-fronting constructions, wh-questions, and modal existential wh-constructions, on which he wrote his PhD dissertation. The research question that he is generally concerned with is to what extent restrictions that have often been thought of as syntactic can be explained in terms of semantics and pragmatics. Staša Vujiþiü Stankoviü is a PhD Student at the Department of Computer Science and Informatics, University of Belgrade, Faculty of Mathematics, in Serbia. She graduated from the Faculty of Mathematics, University of Belgrade, Serbia, in Computer Science. One of her main interests and the topic she researches as a part of her PhD studies is Natural language processing, and in particular Information extraction in Serbian using the lexical resources for Serbian developed from the Natural Language Processing Group at the Faculty of Mathematics, University of Belgrade, whose member she is. Her current research interests also include Information Retrieval and Web Search, Formal languages and automata theory, Databases, and Semantic web. She has published several conference papers and has given numerous talks at conferences in Europe. Rok Žaucer got a PhD in Linguistics from the University of Ottawa in 2009. He is currently an Assistant Professor in general and Slovenian linguistics at the University of Nova Gorica, Slovenia. His research interests lie in syntax, morphosyntax and event structure within generative grammar. He has worked primarily on the morphosyntax and event structure of Slavic verbal prefixation, the syntax of resultative secondary predication, and the syntax of the noun phrase. He has published his work as journal articles (including Natural Language and Linguistic Theory), conference proceedings articles (including Formal Approaches to Slavic Linguistics) and volume chapters (Morphology and Its Demarcations), and has presented it at various international conferences (e.g. FASL, FDSL, NELS, SinFonIJA). He served as a coeditor of the volume Studies in Formal Slavic Linguistics (Peter Lang) with selected papers from an FDSL conference held in Nova Gorica, Slovenia.

INDEX

A absorption, 17, 22, 23, 24, 27, 28 affix hopping, 3, 4, 5, 6, 8, 9, 10, 11 aspect, viii, 31, 32, 35, 40, 41, 42, 43, 50, 55, 61, 66, 90, 138, 255 imperfective, 31, 34, 42, 43, 46, 48, 206 perfective, viii, 31, 32, 34, 35, 40, 41, 42, 43, 44, 46, 50, 61, 62, 63 Azerbayjani, 55

B Bulgarian, viii, 3, 7, 8, 9, 12, 21, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115, 116, 117, 118, 127, 166, 260

C cartography, 257 Catalan, 3, 103 Chuvash, 60, 62 clause structure, 52, 58, 117 clitics, viii, ix, 7, 9, 10, 13, 99, 101, 102, 105, 108, 114, 116, 117, 118, 155, 157, 169, 170, 171, 172, 173, 175, 176, 181, 182, 183, 186, 187, 188, 189, 190, 191, 195, 196, 197, 198, 258, 260 clitic after pauses, 175 clitic doubling, viii, 99, 100, 102, 103, 105, 107, 108, 111,

112, 113, 115, 116, 117, 118, 260 delayed clitic placement, 172, 173, 175, 176 Czech, viii, 3, 51, 52, 88, 89, 91, 92, 125, 126, 127, 129, 130, 135, 148, 231, 261

D demonstratives, viii, 71, 72, 73, 74, 76, 77, 78, 79, 80, 81, 82, 83, 84, 90, 91, 92 Dutch, 24, 25, 26, 27, 28, 156

F focus tree, x, 233, 234, 235, 236, 237, 240, 242, 243, 244, 245, 247 French, 21, 25, 102, 103, 105 future, viii, 21, 31, 32, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 67, 72, 90, 91, 131, 202, 213, 216, 220, 229, 258 periphrastic future, viii, 31, 32, 34, 35, 36, 37, 38, 39, 41, 42, 43, 44, 45 simple future, 31, 32, 35, 36, 39, 43, 44, 45

G GATE system, 207, 208, 210, 213, 215, 216

Selected Papers from SinFonIJA 3 German, 3, 24, 25, 26, 27, 30, 157, 196, 247, 249, 250, 256 grammatical categories, 58, 60, 64 Greek, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 21, 101, 102, 116, 117, 127, 128, 231

H Hebrew, 3, 12, 21, 68, 130, 131 HMI, x, 233, 234, 235, 238, 240, 247, See Human-machine interaction human-machine interaction, x, 233, 251, 256, 261 Hungarian, 3, 21, 90, 126, 127, 128, 129, 130, 154, 257

I imperatives, x, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13 information extraction, ix, 201, 202, 203, 204, 207, 217, 218 information focus, ix, 153, 154, 155, 157, 158, 159, 160, 161, 162, 163, 166, 167 intonational phrase, 155, 172, 221, 222, 225, 227, 229 Italian, 3, 9, 20, 21, 28, 102, 103, 117, 126, 131

K Kazakh, 56, 57, 59, 62, 65 Kyrgyz, 57, 60, 62, 63, 64, 65, 67

M Macedonian, 3, 7, 8, 9, 100, 101, 105, 106 modality, 255 multiple wh, viii, 11, 15, 16, 17, 20, 28, 30, 192, 193

267

N named entities, 202, 203, 209, 210, 211, 212, 213, 215 negation, 3, 4, 5, 6, 10, 11, 12, 13, 33, 34, 46, 47, 55, 233, 237, 246, 255

P phonology, 8, 136, 170, 171, 176, 185, 195, 198, 220, 258, 259 Polish, viii, 3, 31, 32, 33, 34, 35, 36, 42, 43, 45, 46, 47, 48, 49, 50, 51, 125, 150, 255, 257, 258, 259 Portuguese, 3, 21, 126, 129, 131 prosody, x, 162, 170, 195, 198, 219, 220, 221, 224, 225, 226, 229

R regression trees, ix, x, 219, 220, 222, 224, 225, 229 relative clauses, vii, 15, 17, 21, 22, 28, 133, 172, 182, 183, 185, 261 Romance, 4, 13, 91, 99, 101, 102, 103, 105, 106, 130, 217, 260 Romanian, 3, 21, 101, 103 Russian, 3, 11, 21, 46, 51, 52, 120, 126, 129, 140, 141, 143, 145, 146, 147, 148, 150, 151, 154, 257

S Serbian, ix, x, 73, 92, 127, 153, 154, 155, 157, 158, 159, 160, 163, 165, 166, 167, 169, 170, 177, 186, 188, 193, 195, 196, 197, 198, 201, 202, 203, 204, 205, 206, 208, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 224, 225, 229, 230, 231, 249, 250, 257, 258, 261, 262

268 Slavic, x, 11, 12, 13, 32, 46, 51, 52, 72, 91, 101, 103, 116, 117, 118, 130, 148, 150, 154, 166, 196, 198, 202, 203, 220, 230, 250, 256, 257, 258, 259, 260, 262 Slovak, viii, 51, 75, 88, 89, 91, 125 Slovenian, viii, 3, 31, 32, 46, 47, 48, 49, 50, 71, 72, 73, 74, 75, 76, 77, 78, 82, 83, 84, 85, 87, 88, 89, 90, 92, 125, 258, 262 Spanish, 3, 10, 13, 21, 25, 89, 90, 103, 120, 124, 125, 126, 131, 137, 145, 146, 147 speech synthesis, ix, 219, 220, 228, 250, 261 syntax, viii, 8, 11, 13, 29, 31, 32, 68, 69, 91, 117, 118, 119, 123, 124, 127, 132, 136, 139, 141, 143, 148, 149, 150, 151, 155, 156, 163, 167, 170, 171, 198, 206, 221, 255, 256, 257, 258, 259, 260, 261, 262 syntax-semantics, 120, 133, 137, 258, 260

Index

T tense, viii, 31, 32, 34, 35, 40, 41, 42, 43, 47, 50, 55, 131, 138, 255, 259 Tubalar, 54, 55, 60, 63 Turkic languages, 53, 57, 64 Tyva, 57, 62, 64, 69

U Unitex system, 204, 206, 212, 216, 217

W wh-questions, vii, 17, 23, 29, 30, 117, 133, 158, 161, 193, 262 word order, 68, 104, 109, 112, 113, 114, 115, 117, 154, 155, 156, 166, 168, 259