Linguistics in the Netherlands 1990 9783110849998, 9783110130362

Citation preview

LINGUISTICS IN THE NETHERLANDS 1990

AVT PUBLICATIONS AVT Publications is a series sponsored by the Algemene Vereniging voor Taalwetenschap (Linguistic Society of the Netherlands). In addition to the annual publication of Linguistics in the Netherlands further publications, resulting from other activities promoted by the Society, may appear in this series.

Still available Linguistics in the Netherlands 1 9 7 7 - 1 9 7 9 Wim Zonneveld and Fred Weerman (eds) Dfl. 3 0 , — / U S $15.00 Language

Development

Peter Jordens and Josine A. Lalleman Dfl. 6 0 , — / U S $30.00 Linguistics in the Netherlands 1989 Hans Bennis and Ans van Kemenade (eds) Dfl. 4 2 , 5 0 / U S $22.00

LINGUISTICS IN THE NETHERLANDS 1990

Reineke Bok-Bennema Peter Coopmans (eds)

1990

FORIS PUBLICATIONS Dordrecht - Holland/Providence Rl - U.S.A.

Published

by:

Foris Publications Holland P.O. Box 509 3300 A M Dordrecht, the Netherlands Distributor

for the U.S.A.

and

Canada:

Foris Publications U.S.A., Inc. P.O. Box 5904 Providence Rl 02903 U.S.A. Sole distributor

for

Japan:

Sanseido Book Store, Ltd. 1-1, Kanda-jimbocho-cho Chiyoda-ku Tokyo 101, Japan

ISBN 90 6765 492 2 © 1990 Algemene Vereniging voor Taalwetenschap seer.: A. van Kemenade, Department of Linguistics, Free University, P.O. Box 7161, 1007 MC Amsterdam No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission from the copyright owner. Printed in the Netherlands by ICG Printing, Dordrecht.

Contents

Preface

vii

List of contributors

ix

Hans Bennis & Pirn Wehrmann On the Categorial Status of Present Participles

1

Norbert Corver Left Branch Extractions out of PP

13

Marcel den Dikken The Structure of English Complex Particle Constructions

23

Jan Don Prefixed Nouns in Dutch

33

Dicky Gilbers On Predicting the Relative Markedness of Vowels

43

Janet Grijzenhout Modern Icelandic Foot Formation

53

Marjan Grootveld Marcus-Parsing and Coordination: A Mismatch

63

Wim de Haas Evidence for a Noncyclic End Rule in English Liliane Haegeman The of Motional de Goan in West Flemish Helen deSyntax HoopObjects & Henriette Swart Indefinite

71 81 91

Sietze Looyenga On the Internal Structure of Nominal Infinitives

101

Anneke Neijt Prosodic Structures and Phonetic Findings - The Case of Equally Stressed Adjectives

113

Hub Prust & Remko Scha VP Ellipsis Induces Clausal Parallelism

123

Willebrord Sluyters The Derivation of Geminate Blockage Effects

133

Ron van Zonneveld Syntactic Development in the Universal Grammar Perspective

143

Joost Zwarts Some Parallels between DP and IP

155

Preface

This volume contains a selection of the papers presented at the twentyfirst annual meeting of the Linguistic Society, held in Utrecht on January 20, 1990. The aim of the annual meeting is to provide members of the society with an opportunity to report on their work in progress. At this year's meeting fifty-two papers were presented. The sixteen papers contained in this volume present an overview of research in different fields of linguistics in the Netherlands. As in previous years, the authors have submitted their articles in camera-ready form. We are grateful to them for their cooperation. We also thank a number of colleagues who helped us referee the papers submitted for this volume. May 1990

Reineke Bok-Bennema Peter Coopmans

List of contributors

H. Bennis University of Leiden, Department of Dutch, P.O. Box 9515, 2300 RA Leiden N. Corver University of Brabant, Department of Language and Literature, P.O. Box 90153, 5000 LE Tilburg M. den Dikken University of Leiden, Department of English, P.O. Box 9515, 2300 RA Leiden J. Don University of Utrecht, Research Institute for Language and Speech, Trans 10, 3512 JK Utrecht D. Gilbers University of Groningen, Department of Dutch, Oude Kijk in't Jatstraat 26, 9712 EK Groningen J. Grijzenhout University of Utrecht, Department of English, Trans 10,3512 JK Utrecht M. Grootveld University of Leiden, Department of Dutch, P.O. Box 9515, 2300 RA Leiden W. de Haas Institute of Dutch Lexicology/University of Utrecht, Research Institute for Language and Speech, Trans 10, 3512 JK Utrecht L. Haegeman University of Geneva, Department of English, 2 Bid des Philosophes, CH-1205 Geneva, Switzerland H. de Hoop University of Groningen, Department of General Linguistics, Oude Kijk i n ' t Jatstraat 26, 9712 EK Groningen 5. Looyenga University of Groningen, Department of Dutch/Department of General Linguistics, Oude Kijk i n ' t Jatstraat 26, 9712 EK Groningen A. Neijt University of Leiden, Department of Dutch, P.O. Box 9515, 2300 RA Leiden

H. Prüst University of Amsterdam, Department of Computational Linguistics, Spuistraat 134, 1012 VB Amsterdam R. Scha University of Amsterdam, Department of Computational Linguistics, Spuistraat 134* 1012 VB Amsterdam W. Sluyters University of Nijmegen, Department of French, Erasmusplein 1, 6525 GT Nijmegen H. de Swart University of Groningen, Department of Romance Languages, Oude Kijk i n ' t Jatstraat 26, 9712 EK Groningen P. Wehrmann University of Leiden, Department of Dutch, P.O. Box 9515, 2300 RA Leiden R. van Zonneveld University of Groningen, Department of Dutch, Oude Kijk in't Jatstraat 26, 9712 EK Groningen J. Zwarts University of Utrecht, Research Institute for Language and Speech, Trans 10, 3512 JK Utrecht

Hans Bennis & Pim

Wehrmann

On the Categorial Status of Present Participles

1. Introduction In the literature the categorial status of V+ing in English is generally discussed with reference to contexts such as in (1) and (2). (1) (2)

a. b. a. b.

the (*veiy) flying Dutchman the (*veiy) sleeping beauty a (very) amusing stoiy a (very) amazing performance

It has been argued by many scholars that the participles in (1) are verbal, while the participles in (2) are either verbal or adjectival (cf. Fabb 1984, Brekke 1988, Milsark 1988). The crucial argument for an adjectival status of the participle in (2) is the possibility of intensification by modifiers such as very. In a recent article, Borer (1990) argues against this analysis. She claims that the possibility of adding very to a participle is not related to the adjectival vs. verbal status of the participle, but correlates with the possibility to add very much to the corresponding verbal projection. This is shown in (3). (3)

a. "The girl slept very much b. This story amazes me very much

She thus relates the unacceptability of very in (1) to the unacceptability of very much in (3a). In this way she is able to analyse the present participles in (1) and (2) as truly adjectival. The differential behaviour of the participles in (1) and (2) is discussed in Brekke (1988). He claims that only present participles of psychological verbs with a non-subject experiencer are able to appear as adjectives. He formulates this restriction in his Experiencer Constraint. This constraint stipulates that only present participles that have a non-subject experiencer may undergo a categorial change from V to A. Borer (1990) takes the distinction between experiencer verbs and other verbs as a general property, unrelated to the categorial status of the present participle. In this paper we shall demonstrate that Dutch provides very clear evidence that a distinction between adjectival and verbal present participles in attributive position is necessary and that something like the Experiencer Constraint is required to determine which verbs allow an adjectival present participle. 2. The distribution of present participles in Dutch A first indication that the present participles of experiencer verbs differ quite substantially from those of non-experiencer verbs is formed by the distribution. There is a striking difference between present participles in adjunct position and those in complement position. Most grammars of Dutch take the view that present participles occur in adjunct position only, i.e. in attributive or adverbial position. To the extent that

2 present participles are able to appear in complement position, they are taken to be lexicalized adjectives (cf. Paardekooper 1984, Van Gestel 1986). 2.1 Present Participles in adjunct position There are two positions where present participles can appear rather freely. These positions are both adjunct positions. Moreover, these are the two adjunct positions in which adjectives appear i.e. the attributive position within NP and the adverbial position within VP. Some examples are given in (4) and (5). (4)

(5)

a. de vliegende Hollander the flying Dutchman b. de zingende detective the singing detective a. Hij zit huilend in de kerk He sits crying in the chapel b. Ik dans zingend in de regen I dance singing in the rain

With the exception of verbs that do not allow a referential subject such as weather verbs, all verbs may appear in these constructions. 2.2 Present Participles in complement position The observations in 2.1 suggest that present participles are adjectives. We thus expect them to appear in complement positions in which adjectives occur. There are at least two constructions in which we find adjectives as complements: adjectives may have a raised subject as in the complement of raising/copula verbs, such as zijn (to be), or adjectives may appear with a lexical subject in the complements of verbs such as vinden (to consider). At first sight it appears to be the case that present participles do not appear as the head of a projection in complement position. (6)

a. Jan is zielig/*huilend John is pitiful/crying b. Iedereen vindt Doornroosje rustig/*slapend Everybody considers the Sleeping Beauty quiet/sleeping

However, present participles such as opvallend (striking) and verheugend (gratifying) do appear in predicative position, as is illustrated in (7): (7)

a. Dat argument is opvallend That argument is striking b. Zijn medewerking is verheugend His cooperation is gratifying

To account for these counterexamples to his claim that present participles appear in adjunct position only, Van Gestel argues that in sentences such as in (7) "the participle has clearly become a lexicalized adjective". This formulation suggests that there is no systematic relationship between the verbs verheugen (enjoy) and opvallen (strike) and the present participles in (7). We do not agree with this view. When we look at the paradigm of present participles occurring in predicative position, there is at least one

3 class of verbs which allow their present participles to appear in predicative position very productively, i.e. psych-verbs. This is illustrated in (8): (8)

Die opmerking is verontrustend/beledigend/onthutsend/opwindend/... That remark is alarming/insulting/bewildering/exciting/...

We are clearly missing a generalization if we analyse these present participles as lexicalized adjectives. As is to be expected, the same class of present participles appears with a lexical subject in the complement of verbs such as vinden. This is shown in (9): (9)

a. Ik vind zijn gedrag opvallend I consider his behaviour striking b. Ik vind dat gebaar ontroerend I consider that gesture touching

The generalization appears to be that only present participles derived from psychological verbs occur in predicative or complement position. 1,2

3. Verbal properties of present participles Barring a few lexically determined cases, adjectives do not have NPs as direct objects. This is taken to be a consequence of the fact that adjectives are not able to assign Case. If present participles are always adjectives, we would not expect them to appear with a lexical NP-object. However, present participles in adjunct position can appear with lexical objects quite easily, as is shown in (10). (10) a. een aardappels etende schilder a potatoes eating painter b. het mij Russisch lerende meisje the me Russian learning girl The data in (10) seem to indicate that the present participle is at least internally verbal. Only the presence of a Case assigning verb is able to account for the occurrence of the objects in (10). Another distinction between adjectives and verbs is that verbs take Small Clause complements very easily, whereas Small Clauses are not allowed as complements of adjectives. Whatever the reason for this difference, if present participles are adjectives we do not expect them to appear with a Small Clause complement. In (11) we find evidence that they take subcategorized or resultative Small Clauses just like verbs. (11) a. het hem ongelukkig makende voorstel the him unhappy making proposal b. de zijn bord leeg etende jongen the his plate empty eating boy The non-occurrence of similar data in English may now receive two potential explanations. If Borer's analysis is correct, the impossibility of (10) and (11) in English would have to follow from the difference in categorial status: V+ing is adjectival in English and V+end is verbal in Dutch. However, there is no reason why a difference of this type should occur between these languages. A more interesting approach is available. Dutch is an OV-language and English a VO-language. In both languages the realization of attributive projections is subject to something like the Head Final Filter

4 (cf.Williams 1982). From this it follows that a VO-translation of the NPs in (10) and (11) is ruled out. 4. The realization of the experiencer argument In the case of psychological verbs with an internal experiencer this argument is generally expressed as an NP with oblique Case. In Dutch adjectives with an experiencer argument do not allow this argument to be expressed in an NP. This argument is either implicit or expressed in a PP headed by the preposition voor (for).3 In a verbal projection an experiencer argument cannot be expressed in such a PP. This is illustrated in (12) and (13): (12) a. Haar gedrag ontroert mij/ *voor mij Her behaviour touches me / for me b. Die opmerking irriteert mij / *voor mij That remark irritates me / for me (13) a. Haar gedrag is *mij / voor mij pijnlijk Her behaviour is me / for me painful b. Die opgave is *mij / voor mij moeilijk That task is me / for me difficult It is interesting to observe that in the case of present participles the experiencer argument can be realized in an NP, as is illustrated in (14). (14) a. de mij ontroerende opmerking the me touching remark b. het mij opvallende meisje the me striking girl This again points to a verbal character of the present participle. The experiencer can also be realized in a voor-PP, as is demonstrated in (15). (15) a. de voor mij ontroerende opmerking the for me touching remark b. het voor mij opvallende meisje the for me striking girl This particular realization of the experiencer argument points to an adjectival status of the present participle. These facts indicate that present participles of experiencer verbs are either verbal or adjectival. In section 2 we saw that there is a contrast between experiencer verbs and non-experiencer verbs in the occurrence of present participles in complement position. It is significant that in complement position the realization of the experiencer argument is fixed. Only a voor-PP is possible, as in (16). The experiencer cannot be realized in an NP. We thus find a contrast between (14) and (17). (16) a. Die opmerking is voor mij ontroerend That remark is for me touching b. Ik vind dat gedrag voor ons verheugend I consider that behaviour for us gratifying (17) a. *Die opmerking is mij ontroerend That remark is me touching

5 b. *Ik vind dat gedrag ons verheugend I consider that behaviour us gratifying The contrast between (14) and (17) indicates that only adjectival present participles appear in complement position in Dutch. We thus may explain the non-occurrence of present participles of non-experiencer verbs in complement position as being a consequence of the impossibility for these present participles to become adjectival. In order to do so, we need something like Brekke's Experiencer Constraint. In this paper we are not concerned with the exact formulation of this condition. What is relevant to our discussion is that the differential categorial status of the present participle is important not only with respect to the internal structure of the present participle projection, but also with respect to the distribution of this projection. 5. Other arguments for a dual status of the present participle Our discussion so far has discovered three types of present participles: a) always verbal: V+end of non-experiencer verbs (only adjunct position); b) always adjectival: V+end of experiencer verbs in complement position; c) verbal/adjectival: V+end of experiencer verbs in adjunct position. In this section we shall present a number of arguments that further support this partitioning of present participles. 5.1 Stress shift In some cases there are two possible stress patterns on the present participle of a experiencer verb. We illustrate this in (18): (18) a. de 6pvallende/opv611ende jongen the striking boy b. de iingstaanjagende/angstaanj&gende situatie the frightening situation The stress pattern of derived and compound adjectives in Dutch is the subject of ongoing discussion (cf.Schultink 1977, Backhuys 1989, and references cited there). For our purposes, it is sufficient to state that when a complex verbal expression has initial stress, the corresponding adjectival expression bears its stress on the syllable preceding the adjectival suffix, cf.dpklappen (to fold back) vs. optd&pbaar (fold back-able). Given our analysis we now expect the following: - present participles of non-psychological verbs do not exhibit the stress pattern of complex adjectives; - present participles of psychological verbs in complement position exhibit the stress pattern of complex adjectives; - present participles of psychological verbs with a NP-experiencer do not exhibit the stress pattern of complex adjectives; - present participles of psychological verbs with a PP-experiencer exhibit the stress pattern of complex adjectives. This is indeed what we find, as is illustrated in (19). (19) a. de opkomende/*opk6mende zon the rising sun

6 b. die jongen is *6pvallend/opv&llend that boy is striking c. de mij 6pvallende/*opv&llende jongen the me striking boy d. de voor mij *6pvallende/opv£llende jongen the for me striking boy 5.2 Adverbial modification Although the argument concerning adverbial modification is not very strong for English, as Borer has pointed out, the situation in Dutch is different in several respects. Translating very by zeer we observe that the class of present participles that allow adverbial modification by zeer is derived from verbs that allow the same adverbial modifier. Once again this is the class of experiencer verbs. In this respect Dutch zeer is similar to Hebrew me'od, as discussed by Borer. This is illustrated in (20): (20) a. Dit verheugt me zeer This enjoys me very b. *Doornroosje slaapt zeer The sleeping beauty sleeps very c. een zeer verheugende mededeling a very gratifying announcement d. *een zeer slapend meisje a veiy sleeping girl If we translate very by heel the situation is quite different. Heel only modifies adjectives. This leads to a different pattern, as is illustrated in (21): (21) a. *Dit verheugt me heel b. *Doornroosje slaapt heel c. een heel verheugende mededeling d. *een heel slapend meisje It appears that heel is not like Hebrew me'od, but more like English very. Just as in English, the adverbial variant heel erg (very much) can only be attached to psychological verbs.4 However, capitalizing on the discussion in previous sections we can now test whether the thematic status of the verb or the categorial status of the present participle is the relevant factor for modification by heel. If the adjectival status is relevant, we expect the following situations: -if the experiencer of the present participle is an NP heel cannot be attached, while zeer and heel erg can; - if the present participle has stress on its initial syllable, heel cannot be attached, while zeer and heel erg can. Although the data are rather subtle, there is no doubt that there is a contrast in the expected direction. These contrasts can only point to a difference in categorial status. (22) a. een mij zeer/heel erg/*heel verheugende mededeling a me very gratifying announcement b. een zeer/heel erg/*heel 6pvallende jongen a very striking boy

7 5.3 Comparative formation The formation of comparatives again constitutes a rather strong argument in favour of a dual categorial status of the present participle of psychological verbs. Borer dismisses more and less as potential arguments for adjectival vs. verbal status of the present participle. We think that her argument is incorrect. She argues that since practically all verbs allow more and less the possible modification by more/less is a test for the existence of verbs and not for the existence of adjectives. The fact that only present participles of experiencer verbs allow modification by more/less thus appears to point in the wrong direction. Another approach seems far more likely. Both adjectives and verbs may be modified by more/less. In English, adverbial modifiers of adjectives are generated to the left of the adjective. However, adverbial modifiers of verbs are generated in the head-initial VP, i.e. to the right of V. In attributive position, the Head Final Filter thus allows only adjectival modifiers and not verbal ones. This explains the contrast in (23). (23) a. *a more sleeping beauty b. a less interesting story This analysis is corroborated by the fact that the Dutch equivalents of more/less may appear in front of present participles of non-experiencer verbs. This is what we expect given the OV-nature of Dutch. (24) a. een meer (dan Jan) lezende jongen a more (than John) reading boy b. een minder (dan Jan) slapend individu a less (than John) sleeping individual More interesting is the fact that only present participles of experiencer verbs allow comparative formation by affixation of the comparative morpheme -er, as is illustrated in (25): (25) a. een nog aangrijpend-ere / beangstigend-ere ontwikkeling a still moving-er / frightening-er development b. *een nog lezend-ere / slapend-ere jongen a still reading-er / sleeping-er boy The rule of -er-affixation is restricted to adjectives and a specific subclass of present participles. If the adjectival status is indeed the relevant factor, we again are able to make several predictions: - if the compared present participle has an experiencer argument realized, it must be a voor-PP; - if the present participle allows two stress patterns, the addition of the comparative morpheme requires an adjectival stress pattern. These predictions are confirmed, as is illustrated in (26): (26) a. een voor mij / *mij nog aangrijpendere gebeurtenis a for me / me still moving-er happening b. een nog *angstaanjagender / angstaanjSgender geluid a still frightening-er sound

8 5.4 On-prefîxation The data on on-prefîxation (un-) are exactly parallel to the data on er-suffixation. As in English (cf.Wasow 1977 a.o.), the possibility of on-prefixation in Dutch points at an adjectival status. Without comment we shall give the relevant facts in (27) and (28), which are entirely parallel to (25) and (26). (27) a. een on-bevredigende / on-opvallende gebeurtenis a un-satisfying / un-striking happening b. *een on-werkende / on-slapende jongen a un-working / un-sleeping boy (28) a. een voor mij / *mij onopvallende afloop a for me / me unstriking end b. een *on6pvallende / onopvâllende jongen a unstriking boy 6. Conclusion In this article we have shown that the categorial status of present participles in attributive position is not uniform. If the verbal stem is a psychological verb with a non-subject experiencer (cf.note 1) the internal structure of the projection may be either verbal or adjectival. In the case of other verbs the present participle is internally verbal. We have developed a number of tests which make the categorial status visible by the application of a number of syntactic, morphological and phonological processes. This analysis is directly in conflict with the approach taken in Borer (1990). As far as we can see there is no alternative analysis for the Dutch facts which takes the present participle to be an adjective uniformly. Apart from very general theoretical objections, Borer presents no arguments against a dual status of the present participle in attributive position. It thus follows that we should analyse these instances of English V+ing in the same way as we do Dutch V+end. All this constitutes a first step in the direction of an explanation of the cross-categorial behaviour of the present participle. Much more can and needs to be said. Milsark (1988) observes that present participles may have any value for the features aN,|JV in English. It is interesting to observe that the distribution of V+end in Dutch is very restricted in comparison to V+ing in English. Dutch has no progressive -end and no nominal gerundives. The fact that Dutch only allows the class of (semi-)adjectival present participles discussed here calls for an explanation. Moreover, given that the distribution of V+end forms a subset of the distribution of V+ing, an analysis of the Dutch present participles must be properly included in the analysis of English gerunds. At the end of this article we list three important questions with respect to the present participles discussed here: a) Why are verbal present participles allowed in the typically non-verbal, attributive position? b) What prevents verbal present participles from appearing in complement position? c) What prevents non-experiencer verbs from becoming adjectival present participles? We hope to discuss these questions in future work.

9 Notes 1. In fact, we have to be somewhat more precise. As in other languages (cf. for example Belletti & Rizzi 1988), the class of psychological verbs in Dutch is not a uniform class. If we take the presence of an experiencer argument to be the decisive criterion for a psychological verb, there are three classes of psych-verbs. An important distinction is the way in which the experiencer is projected in the structure. The experiencer can either be realized as an external argument or as an internal argument. In this way we can distinguish verbs of the type vrezen (fear, 'the temere-class') from verbs of the type beangstigen (frighten). The class of verbs with an internal experiencer can again be divided into verbs which are ergative in all respects (bevallen (please), opvallen (strike), 'the piacere-class') and verbs which appear to be somewhere in between ergative and unergative verbs (ontroeren (touch), verheugen (gratify), 'the preoccupare-class'). These classes differ in the selection of their perfect auxiliary and the way in which past participles may be realized in attributive position. For our present purposes the distinction between different types of psych-verbs with an internal experiencer is not relevant, since they show a similar behaviour in the present participle construction. However, verbs with an external experiencer do not appear in predicative position. They behave just like non-psychological verbs. This is shown in (i): (i)

a. de (mij) bewonderende menigte the me admiring crowd b. de (niemand) vrezende soldaat the nobody fearing soldier c. *De menigte is (mij) bewonderend The crowd is me admiring d. *De soldaat is (niemand) vrezend The soldier is nobody fearing

In addition to present participles of psychological verbs, Dutch has another class of present participles that occur in complement position quite freely. Some examples are given in (ii). (ii)

a. Zijn gedrag is angstaanjagend His behaviour is fear-giving (frightening) b. Deze muziek is rustgevend This music is peace-giving c. Dat schouwspel was adembenemend That spectacle was breath-taking

These complex predicates are traditionally analyzed as being derived by a rule of Synthetic Compound Formation. Alternatively, they could be argued to be the result of syntactic noun incorporation. Whatever the correct approach may be, these predicates are derived by a very productive rule, and the result is clearly an adjectival expression. Interestingly, these expressions show characteristics of psychological predicates, as do their verbal counterparts. For instance, the indirect object of the verbal counterpart of (iib), Deze muziek geeft mij rust (This music gives me rest), is best characterized as a type of experiencer argument. This corroborates the idea that something like the Experiencer Constraint is involved in the formation of adjectival present participles. 2. In complement and predicative position, we also find present participles of ergative verbs. Some examples are given in (i).

10 (i)

a. De man is stervend The man is dying b. Het schip is zinkend The ship is sinking c. De prijzen zijn stijgend The prices are rising

Ergative verbs have the theme argument as their subjects. Likewise, those psychological verbs that allow their present participle to appear in predicative position (cf. note 1) have been argued to have theme subjects at S-structure. This might lead us to the generalization that having a theme subject is the relevant property for a present participle to appear in predicative position. However, there is evidence that the cases in (i) are to be distinguished from constructions with psychological verbs. - First, there is an aspectual difference. Whereas present participles derived from psychological verbs denote stative properties, present participles derived from ergative verbs do not. For instance, the sentences in (i) can adequately be paraphrased with the aan het V-en construction, the Dutch counterpart of the English progressive: (ii)

Het schip is zinkend / aan het zinken The ship is sinking / on the sink

- Second, only the present participle of an ergative verb may appear with an inflectional ending (-e) in predicative position, as is shown in (iii): (iii) a. Deze situatie is opvallend(*e) / verheugend(*e) This situation is striking / gratifying b. De temperatuur is stijgend(e) / dalend(e) The temperature is rising / going down The optional e-ending is found with non-ergative verbs in adjunct position as well, as in (iv): (iv) a. Dansend(e) ging Jan naar zijn werk Dancing went John to his work b. Jan zat lezend(e) op het toilet John sat reading in the bathroom These facts suggest that in contrast to present participles of psychological verbs, present participles of ergative verbs are always verbal. Other evidence points into the same direction. - Present participles of ergative verbs can be modified by expressions that are generally present in verbal projections only: (v)

a. De prijzen zijn sterk stijgend The prices are strongly rising b. Het schip was snel zinkend The ship was fast sinking

- In contrast to present participles of psychological verbs, present participles of ergative verbs cannot undergo comparative formation or on- (un-) prefixation: (vi) a. "Het onzinkende schip The unsinking ship

11 b. *De stijgender prijzen The more rising prices 3. The possibility of implicit arguments in adjectival projections can be derived from the control properties in adjectival structures (cf.Roeper 1987, Bennis & Hoekstra 1989). 4. There exists yet another way to translate very. Erg has a distribution which differs from both zeer and heel. It modifies not only adjectives and psychological verbs, but also non-psychological verbs such as snurken (to snore) which are semantically compatible with the addition of an intensifier (cf. Jan snurkt erg/*heel/*zeer (John snores very)). References BACKHUYS, K.J 1989 Adjectival Compounds in Dutch, in: Bennis & van Kemenade (eds) Linguistics in the Netherlands 1989, Dordrecht: Foris Publications, 1-10 BELLETTI, A. & L.RIZZI 1988 Psych verbs and ©-theory, in: Natural Language and Liguistic Theory 6, 291-352 BENNIS, H. & T.HOEKSTRA 1989 PRO and the Binding Theory, in: Bennis & van Kemenade (eds) Linguistics in the Netherlands 1989, Dordrecht: Foris Publications, 10-23 BORER, H. 1990 V+ing: It walks like an Adjective, It Talks like an Adjective, in: Linguistic Inquiry 21, 95-103 BREKKE, M. 1988 The Experiencer Constraint, in: Linguistic Inquiry 19, 169-180 FABB, N. 1984 Syntactic Affixation, diss. MIT, Cambridge MA GESTEL, F.van 1986 X-bar Grammar: Attribution and Predication in Dutch, Dordrecht: Foris Publications MILS ARK, G.L. 1988

Singl-ing, in: Linguistic Inquiry 19, 611-634

PAARDEKOOPER, P. 1984 Beknopte ABN-syntaxis, 7e druk, Eindhoven ROEPER, T. 1987 Implicit Arguments and the Head-Complement Relation, in: Linguistic Inquiry 18, 267-310 SCHULTINK, H. 1977 Over de accentuering van afgeleide woorden in het Nederlands, in: H.Heestermans (ed.) Opstellen door vrienden en vakgenoten aangeboden aan dr. C.H-A.Kruyskamp, 's-Gravenhage, 180-188 WASOW, T. 1977 Transformations and the Lexicon, in: AAkmajian, P.Culicover & T.Wasow (eds) Formal Syntax, New York: Academic Press WILLIAMS, E. 1982 Another Argument that Passive is Transformational, in: Linguistic Inquiry 13, 160-163

Norbert Corver

Left Branch Extractions out of PP

1. Introduction* This paper presents an analysis of the extractability of PP-internal left branch modifiers in Dutch and English within the Barriers framework as proposed in Chomsky (1986). The accessibility of these modifiers to movement operations is exemplified in (1) and (2) for Dutch and English, respectively (see also Zwarts (1978) a.o.). la

Jan wist niet [[hoe diep]i het lijk [t4 onder de grond] lag] John knew not how deep the body under the ground lay

b

Jan wist niet [[hoeveel meter^ het lijk [t4 onder de grond] lag] John knew not how-many meter the body under the ground lay

2a b

How fart did he go [tt into the woods]? How many feett did the monster raise his head [t4 above the water]?

Besides removal of the left branch modifier to [Spec,CP], movement of the entire PP (involving pied piping) to this landing site is permitted as well: 3a b

Jan wist niet [hoe diep onder de grond] het lijk lag Jan wist niet [hoeveel meter onder de grond] het lijk lag

4a b

[How far into the woods] did he go? [How many feet above the water] did the monster raise his head?

In order to establish that sub-extract ion of a left branch modifier out of a containing PP has taken place in the derivation of the sentences in (1) and (2), I will first discuss the internal syntax of strings like hoe diep/hoeveel meter onder de grond in section 2. Next, I will account for the movement behavior of the left branch modifiers at issue in terms of the government and bounding theory as formulated in the Barriers system. In section 3, I will finally address the question whether the accessibility of left branch PPmodifiers is a property of P-stranding languages, or whether it is also found in non-P-stranding languages. 2. Internal Structure In order to establish that the sentences in (1) and (2) really involve subextraction of a left branch modifier out of a PP, it is necessary to address a number of questions: First of all, is the modifier really extracted out of the PP? One could propose, for example, that the VP diep/ 2 meter onder de prond lag in a sentence like (5) can be assigned two underlying representations, one in which the modifier is internal to the PP (as in (6a)), and one in which it occurs in a PP-external position attached to V' (as in (6b)), although still modifying the PP from that position.1

14 5

Ik denk [dat het lijk [ v p diep/2 meter onder de grond lag]] I think that the body deep/2 meter under the ground lay

6a b

• • • (c? • • • (IF • • • [v? IFF diep/2 meter [ p . onder de grond]] lag]]] ... [ C p • • • [IF • • • U p [diep/2 meter] [ p p onder de grond] lag]]]

Of course, if the left branch modifier need not be part of the PP in strings such as (5) , then the sentences in (1) do not necessarily involve subextraction out of PP. It could be argued that an underlying structure as in (6b) is at the basis of these examples. The sentences in (3), on the other hand, in which the entire PP has been moved into the [Spec,CP] of the embedded clause, would have a representation like (6a) underlyingly. So, if one wants to argue that the derivation of the sentences in (1) involves subextraction of a left branch modifier out of PP, evidence against an underlying representation like (6b) must be presented first. In particular, it must be shown that it is incorrect to analyze a string like diep/2 meter onder de prond as a nonconstituent (as in (6b)). After having established the constituenthood of a string like diep onder de grond. a second question concerning the internal structure of this string should be addressed, namely: Which element is the head of this constituent? The structural analyses given in the above-mentioned examples already show that I will consider this string a PP headed by the P onder. in which de frond occupies the complement position and diep a left branch adjunct-position. However, in order to be sure that the sentences in (l)-(2) involve subextraction of a left branch modifier out of a PP, it must be shown that any alternative structural analysis of this string is inappropriate. An alternative which comes to mind is one in which the string diep onder de grond is considered an adjective phrase which is headed by the adjective diep and in which onder de grond functions as a sort of post-adjectival prepositional modifier. On the basis of a number of syntactic arguments I will exclude this alternative structural analysis. A third question which will be raised in this section is the following: What hierarchical position does the left branch adjunct occupy within the PP? Is it a sister1of P' or is it base-adjoined to PP? Turning to the first question, I will present some arguments in favor of analyzing strings like diep onder de grond and 2 meter onder de grond as one constituent rather than two separate constituents hanging from V' . That diep and 2 meter are part of the PP, and are not for example hanging from V' (i.e. a position external to PP), is shown by a number of facts. Of course, the sentences in (3) and (4), where the interrogative phrase has been moved into the [Spec,CP] position, already show that the strings at issue can form a constituent. However, the possibility of moving these strings into [Spec,CP] does not show that they must form one. The Dutch PP-extraposition facts in (7) suggest that the modifier must be part of the PP. 7a

Jan heeft het lijk diep onder de grond gevonden John has the body deep under the ground found b * Jan heeft het lijk diep tt gevonden onder de grondt c Jan heeft het lijk tt gevonden diep onder de grondt

15 If the P and its complement are moved rightwards to a postverbal position, the modifier has to move along. Under an analysis in which the string diep onder de grond can be analyzed as two separate maximal projections hanging from V' , it is not clear what rules out structure (7b): It would simply involve extraposition of a PP to a postverbal position. Notice also that if one adopts an underlying structure like (6b), it is not clear what rules out the ill-formedness of the following topicalization constructions, since these could simply be analyzed as constructions in which a PP is fronted to [Spec,CP]: 8 *

Onder de prondt heeft Jan het lijk diep t± gestopt Under the ground has John the body deep

9 *

put

Into the woodst he went deep tj

Of course, these sentences are acceptable if the modifier is absent. 10 11

Onder de woods grond he heeft Jan het lijk gestopt Into the went

Under an analysis in which the modifiers in (7)-(9) occupy a position within the PP that is a sister of P' , the ungrammaticality of these sentences may follow from a number of factors: First of all, movement of P' is not in accordance with the stipulation that only maximal projections and X-zero categories can undergo movement (see Chomsky (1986)). Furthermore, it would violate the structure preservingness constraint on substitution operations in the topicalization construction, since a P' (i.e. a non-maximal category) is moved into [Spec,CP], a position which only allows maximal projections. Finally, if the sentences in (7)-(9) involve extraction of a P'-constituent, the Subjacency Condition and the ECP will always be violated within the Barriers system: P' (a non-maximal projection) cannot escape the potential barrierhood of the dominating maximal projections VP and IP (the latter by inheritance). The barrierhood of these maximal projections cannot be voided via adjunction, given the structure preservingness requirement on adjunction operations: a P' cannot adjoin to a maximal projection. So, removal of the P' to the left periphery of the clause crosses two L-barriers (i.e. non-L-marked maximal projections): VP and IP. Consequently, the ECP is violated since the P'-trace is not antecedent governed. Another argument showing that the left branch modifier must be part of the PP comes from the following coordination facts: 12a* ..dat Jo zijn duim diep [zowel in cijn mond als in jouw mond] stak ..that Joe his thumb deep both in my mouth and in your mouth stuck b* Joe stuck his finger deep [both into Sue's mouth and into my mouth] If the modifiers (dlep/deepl in these sentences are base-generated outside of the PP, then it is not clear what blocks emphatic conjunction of the two PPs involved in these sentences. Notice that absence of the adjectival modifiers yields well-formed sentences: 13a b

..dat Jo zijn duim zowel in mijn mond als in jouw mond stak Joe stuck his finger both into Sue's mouth and into my mouth

16 If, however, the adjectival modifiers in (12) are part of the PP, then the ungrammaticality of these strings can be accounted for in the following way: Under an analysis in which the adjectival modifier is attached to P' , the emphatic coordinating conjunctions in (12) conjoin two P'-constituents. This is not in accordance with Neijt's (1979) generalization that emphatic coordinating conjunctions can only conjoin maximal projection categories. Given the considerations above, it seems fair to conclude that a string like (5) should be assigned structure (6a) rather than (6b). Let us now turn to the second question concerning the internal structure of the strings at issue: Which element heads constituents like dlep onder de prond and 2 meter onder de grond? I will assume that the preposition (onder) is the head of this constituent (PP) and that diep/2meter are left branch modifiers within the PP. So, I argue against an Internal structure of these strings as given in (14), where onder de grond functions as a right branch modifier within NP and AP. b

(HP 2 meter [ F P onder de grond]] ( A F diep [ p p onder de grond]]

A first argument in favor of analyzing these sequences as PPs and not as noun phrases or adjective phrases containing a PP-modifier comes from subcategorization. Consider, for example, the following sentence: 15

Jan woont [2 meter/diep onder de grond] John lives 2 meter/deep under the ground

If the string between brackets would have the noun meter or the adjective dlep as its head, with onder de \grond as some sort of a right branch modifier within the maximal projection of this head, then we would expect that the noun and the adjective could also appear alone, i.e. without the PP-modifier. The following sentences, however, show that neither the noun phase 2 meter nor the adjective phrase dlep can occur as a complement subcategorized for by the verb: 16a* Jan woont 2 meter b*?Jan woont diep The following sentence shows complement of the verb wonen. 17

that

the

PP onder

de grond can occur as

Jan woont [onder de grond]

On the basis of these facts, it is clear that structural analysis in (14) is incorrect, and that the correct analysis is one in which the phrase ¿ meter/diep is a left branch modifier within the subcategorized PP. A second argument against the structures in (14) comes from the extraposition example in (7c), where the string dlep onder de grond has been extraposed to a postverbal position. In Dutch, only CP and PP can be moved to a postverbal position. So, if in the sequence diep onder de prond the adjective diep were the head of this phrase, then we would expect that this (adjective) phrase could not occur in extraposed position. (7c), however, shows that dlep onder de grond can occur in extraposed position, which suggests that this string is a PP rather than an AP. My third argument against structures as in (14) is based on the fact from English that fronting of a sequence like deep into the woods may trigger

17 inversion of a nonpronominal subject wich a verb that is in the simple past or present in English. This inversion can occur with fronted directional PPs, but never with fronted APs (cf. Emonds (1976)): 18a* Fast ran this gnu! b

Into the woods ran a herd of gnus!

Consider now the following well-formed example: 19 [Deep into the woods] ran a herd of gnus! The grammaticality of this sentence suggests that a sequence like deep into the woods is a PP containing an adjectival modifier, rather than an adjective phrase that contains a prepositional modifier. My last argument in favor of a PP-analysis is the fact that the strings at issue can appear in focus position in so-called cleft constructions. As is well-known, NPs and PPs can occur in this position, but not APs: 20a It was [under the ground] that I found the body b * It was [carefully] that John stuck his finger into his mouth T h e grammaticality of the following sentence suggests that the string in brackets should be analyzed as a PP containing an adjectival modifier, and not as an AP containing a PP-modifier: 21

It was [deep under the ground] that I found the body

From these arguments I conclude that strings like 2 meter/diep onder de grond and their English equivalents should be considered PPs in which dlep/2 meter f u n c t i o n as left branch modifiers. This brings us to the final question concerning the internal structure of these strings: What hierarchical position do these left branch modifiers occupy within the PP? As I have already hinted at above, I assume that the left branch modifier is a sister of P' (as in (22a)). So I will not adopt a structure in which the modifier is base-adjoined to PP (as in (22b)). 22a b

[ V F [ F P diep/2 meter [ ? . onder de grond]] lag] tvp (pp diep/2 meter [ F ? onder de grond]] lag]

If structure (22b) were adopted, one would expect that rules operating on maximal projections can apply to both PPs, since they are both maximal. It turns out, however, that the string onder de prond and diep onder de grond behave differently with respect to various syntactic processes. Topicalization and extraposition, for example, can only apply to the string including the m o d i f i e r (i.e. the higher PP-segment in (22b))). This is clear from the examples given in (7)-(9). If diep onder de prond were assigned a structure like (22b), one would expect that the lower PP-segment onder de prond could also undergo topicalization or extraposition. Movement of this segment would be in accordance with the stipulation that only maximal and X-zero categories can be moved. Furthermore, substituting the lower PP-segment for [Spec,CP] does not violate the structure preservingness requirement on substitution operations. Finally, extraction of the lower PP-segment out of the complement-PP does not violate the ECP or the Subjacency Condition. The higher PP-segment is not an Lib a r r i e r , since it is L - m a r k e d by the v e r b , because the barrierhood of dominating non-L-marked maximal categories (e.g. VP) can be circumvented by

18 adjoining the PP-segment to these categories. Notice that this adjunction operation is in accordance with the structure preservingness requirement on adjunction operations. Thus, the ungrammatical status of the sentences (7b), (8) and (9) is unexplained under structure (22b). Of course, under structure (22a) their ill-formed status directly follows, as I have already pointed out above in my discussion of the examples (7)-(9). Another potential argument against structure (22b) comes from the coordination constructions in (11). Under structure (22b), the string diep zowel in ailn mond als in 1 ouw mond could be assigned the following structure: 23

[ F p [diep] [zowel [ p p in mijn mond] als [ p p in jouw mond]]]

So, the lower PP-segment consists of a coordinated structure in which the emphatic conjunctions zowel...als conjoin two PPs (i.e. two maximal projections). Notice, that the ill-formedness of the sentences in (12) cannot be explained in terms of Neijt's generalization that only maximal categories can be coordinated by emphatic conjunctions, since the PP-conjuncts are maximal in (23). As I have already pointed out above, however, the ungrammaticality of the sentences in (12) can be accounted for under structure (22a), because in that case the emphatic conjunctions conjoin two P'-constituents (i.e. nonmaximal categories) , which is not allowed. 2. The extractability of left branch modifiers out of PP Having established the internal structure of strings such as diep/2 meter onder de prond- let us consider the question how the extractability of left branch modifiers can be accounted for within a Barriers theory. In the sentences in (1) and (2), a left branch modifier has been reordered out of a PP-complement of the verb. The PP-complement is L-marked by the verb, and therefore is not a blocking category (BC), nor an L-barrier. Via intermediate adjunction to VP, the extracted modifier can reach the [Spec,CP] position without violating the ECP or the Subjacency Condition. Schematically: 24

[ c p hoe diepi [ I P ... [ V P t'j [ V P [ p p t4 [ p . onder de grond]]]]]]

The Subjacency Condition is not violated, since no L-barrier intervenes between the traces of the fronted modifier. The ECP is satisfied since there are no intervening L-barriers or M(inimality) barriers which include the trace but exclude its antecedent.2,3 PP is not a Minimality-barrier excluding the mo d'ifier -trace since it does not contain a head c - commanding the trace. The lower VP-segment contains the trace, a maximal projection containing the trace and a head c-commanding the trace. However, it does not exclude the antecedent trace which is adjoined to VP. Hence, the adjunct-trace contained within the PP is properly governed. The left branch extractions discussed so far all involve removal of a modifier out of a complement-PP. The question now arises whether left branch elements can be reordered out of adjunct-PPs as well. As is well-known, it is not possible to extract modifiers from within adjunct clauses introduced by prepositions. Extraction of an argument from within the same configuration yields a much better result. This is illustrated by the following sentences (cf. also Chomsky (1986)):

19 25a?? Which girl did John faint after he had kissed t A ? b* How passionatelyt did John faint after he had kissed the girl t t ? In the a-sentence, a direct object noun phrase is extracted from within an adjunct-PP, yielding a weak subjacency violation because of the the Lbarrierhood of this adjunct-PP. The ECP is not violated since the initial trace of the moved direct object is antecedent governed by the intermediate trace which is adjoined to VP. The b-sentence, however, is strongly out since, besides weakly violating the Subjacency Condition, it also violates the ECP. The adjunct-PP, which is an L-barrier, blocks antecedent government of the intermediate adjunct-trace which occupies the [Spec,CP]-position of the embedded clause by the next highest potential antecedent governor, i.e. the intermediate trace that is adjoined to the VP of the matrix clause.4 Consider now the following sentences in which a left branch modifier is reordered out of an adjunct-PP: 26a* Hoeveel meterj ontplofte het vliegtuig [t4 boven de stad]? How many meters exploded the airplane above the city b* Hoe diepj ontmoette Jan Marie [tt onder de grond]]? How deep met John Mary under the ground 27a* How many kilometersj did the plane lose a wing [tt above the city]? b* How many metersi does this animal feed its young [tt under the ground]? These sentences are strongly out, since besides weakly violating the Subjacency Condition they violate the ECP. Suppose the PP is embedded inside the VP, then the first adjunction site for the measure phrase is the VP. The intervening adjunct-PP creates an L-barrier blocking antecedent government of the PP-internal trace, and therefore triggers an ECP-violation. Note that the L-barrlerhood of the adjunct-PP cannot be voided via adjunction under the assumption that adjunction to argument type categories is not permitted. The following pairs of examples from Dutch and English clearly show that there is a contrast between argument and adjunct extractions from within adjunct PPs: 28a? De stad waart het vliegtuig een vleugel [t4 boven] verloren had, was The city where the airplane a wing above lost had, was Washington Washington b* Die 3 kilometer die4 het vliegtuig een vleugel [tt boven de stad] Those 3 kilometer that the airplane a wing above the city verloor, waren voldoende om de mensen te evacueren lost, were enough for the people to evacuate 29a? Which treej does this animal feed its young [under t 4 ]? b* How many metersj does this animal feed its young [t4 under the ground]? In (28a) and (29a), a so-called R-pronoun (waar) and a NP have been extracted out of the PP respectively, yielding a preposition-stranding construction. Following Van Riemsdijk (1978), I will assume that P-stranding languages, as opposed to non-P-stranding languages, have a specifier position which can function as an escape hatch for complement extractions. So, the R-pronoun waar

20 and the NP-complement which tree leave the adjunct-PP via the [Spec.PP] position. This means that extraction of the complement out of the adjunct-PP does not yield an ECP-violation, because the initial trace of the moved complement is antecedent governed at S-structure by the intermediate trace in [Spec.PP]. The weak unacceptability of the a-examples is due to a weak violation of the Subjacency Condition: extraction of the complement crosses the L-barrier PP. Consider next the b-sentences, in which a left branch adjunct has been moved out of an adjunct-PP. Clearly, these sentences are worse than the a-examples. Besides Weakly violating the Subjacency Condition, these sentences violate the ECP. The initial trace of the fronted left branch modifier is not properly (i.e. antecedent) governed by the nearest potential antecedent (the intermediate trace adjoined to VP), since the adjunct-PP is an L-barrier excluding the intermediate antecedent-trace. 3. Modifier extraction out of PP in non-P-stranding languages So far, we have seen that left branch modifier extractions out of (complement) PPs are permitted in languages like Dutch and English. As we have seen, these languages also allow P-stranding. The question arises whether there is any relation between the possibility of having P-stranding in a language and the accessibility of left branch modifiers to movement operations. One might hypothesize, for example, that the extractability of modifiers from within PPs is dependent on the possibility of using the [Spec.PP] position as an escape hatch for movement. In that case, one would predict that languages which do not allow removal of the complement of the preposition because of the absence of a PP-internal escape hatch, do not permit extraction of a modifier either. It turns out now that extraction of left branch modifiers out of PPs is permitted in non-P-stranding languages as well. Consider, for example, the following sentences from German, French, Italian, Rumanian and Polish, respectively. 30a

Wieviel Metert hing das Bild [ti über dem Schrank]? How-many meters hang the picture above the cupboard

b(?) [A combien de km's]1 est-ce que Jean habite [tt au Nord de Paris]? By how-many of kilometers John lives to the-North of Paris c

[A quanti 'metri]j Gianni gettò la palla [tj sopra il cesto]? By how-many meters John threw the ball above the basket

d(?) La clt^ metrij a ingropat cadavrul [tt sub pamlnt]? By how-many meters (he) has burried corpse-the under ground e

Jak gt^bokoj schowal zloto [tj pod ziemia]? How deep (he) hid gold under ground

Although the speakers I consulted often prefer the pied piped variants of these sentences, they consider the left branch extractions at issue (fairly) acceptable. They all agree that sentences involving left branch modifier extraction are much better than those in which a complement of a preposition is fronted. This asymmetry between complement-extraction on the one hand and left branch modifier extraction on the other follows from the ECP. Under the assumption that non-P-stranding languages do not have a [Spec.PP] position

21 which can function as an escape hatch for complement extractions, the extracted complement must move in one swoop to the nearest landing site, e.g. a position adjoined to VP. Direct extraction out of PP yields a Minimality violation: PP is a H-barrier including the trace, but excluding the potential antecedent of the trace. It is a M-barrier since it contains: (i) the trace, (ii) a maximal projection containing the trace (viz. PP itself), and (iii) a head c-commanding the trace. Extraction of the left branch modifiers in (30) does not yield a Minimality v i o l a t i o n , s i n c e PP does not count as a M-barrier excluding the nearest potential antecedent of the trace. PP is not a M-barrler for the trace in the l e f t b r a n c h m o d i f i e r p o s i t i o n , b e c a u s e it d o e s n o t contain a head ccommanding the trace.

4.

Conclusion

In this paper I have investigated the internal syntax of PPs containing left b r a n c h modifiers and the accessibility of these left branch modifiers to movement operations. It was shown that modifier-extraction out of adjunct-PPs is much worse than removal out of complement-PPs. This asymmetry was accounted for in terms of the ECP: Adjunct-PPs, as opposed to complement-PPs, are Lbarriers and therefore block antecedent government of the PP-Internal trace of the fronted modifier. It was further shown that left branch modifiers are also accessible to movement operations in non-P-stranding languages, which shows that it is incorrect to conclude from the absence of P-stranding in a certain language that PPs are absolute islands to movement operations in that language.

Notes *

I am grateful to Henk van Riemsdijk for his comments on an earlier version of this paper. I am also indebted to the following people for d i s c u s s i o n a n d n a t i v e j u d g e m e n t s : D e n i s Delficto, Viviane Deprez, Peter Kipka, Hans-Peter Kolb, Carole Paradis, Donca Steriade, Craig Thiersch, Hanna Ualinska and Raffafilla Zanuttini. I also want to thank the reviewers for their comments.

1.

That a phrase X can enter into a modification relation vith a phrase Y in spite of the fact that they do not form a constituent, is shown, for example, by the following sentence from Dutch in which the second adjunct-PP (in de Tulnstraat) specifies as it were the location expressed by the other adjunct-PP (in Tilbury): (i)

Ik heb haar [in Tilburg] toen [in de Tuinstraat] ontmoet I have her in Tilburg then in the Gardenstreet met

2.

'Minimality' is defined as follows: £ is a M-barrier for ¡2 if £ includes fe, jJ (an X° c-commander of fe), g (a maximal projection not necessarily distinct from a) (Chomsky (1986: class lectures Fall)).

3.

Following Lasnik & Saito's (1984) theory of proper government, I will assume that all traces of a moved adjunct must be properly (i.e. antecedent) governed in LF and that only the initial trace of a moved argument must be properly governed in S-structure.

22 4.

I will assume that adjunction is not permitted to argument type categories. Since PP can count as an argument (i.e. it can get theta-marked in D-structure), it does not qualify as a target of adjunction (see also Sportiche (1988)).

references Chomsky, N. 1986 Barriers. Cambridge, Massachusetts: MIT-Press Emonds, J. 1976 A Transformational Approach to English avntax. New York: Academic Press Lasnik, H. & M. Saito 1984 O n the Nature of Proper Government, in Linguistic Inquiry 3, 269320 Neijt, A. 1979 Gapping. Dordrecht: Foris Riemsdijk, H. 1978 A Case Study in Syntactic Markedness. Dordrecht: Foris Sportiche, D. 1988 Conditions on Silent Categories, ms., UCLA Zwarts, F. 1978 Extractie uit Prepositionele Uoordgroepen in het Nederlands, in Berkel A. e.a. (eds) Proeven van Neerlandistiek. 303-399

van

Marcel den Dikken

The Structure of English Complex Particle Constructions 0. Introduction* The well-known word-order alternation in the English verb-particle construction illustrated in (1) has given rise to a rich variety of competing analyses in the Government-Binding framework, even if we restrict ourselves to those analyses that start off on the premise that the particle in these constructions is the predicate of a Small Clause (SC) in the complement of the verb. (1)

a. They looked the information up b. They looked up the information

Thus, Kayne (1985) opts for an analysis of the pair in (1) according to which the variant in (lb) is derived from the underlying word order in (la) through extraposition of the subject NP of the particle-headed SC, as illustrated in (2a). Though his arguments for doing so are not particularly strong, Kayne explicitly rejects an alternative approach to the alternation in (1) which holds that (lb) is obtained by moving the particle into the verb, this movement operation being an instance of head incorporation in the sense of Baker (1988). This analysis is given in (2b). Both approaches take (la) to represent the base-generated word order pattern, (lb) being derived from it. Taking the opposite tack, however, one could suggest a third structure, given in (2c), according to which the particle is ergative and hence selects only an internal argument. The word order in (la) is then obtained through NP movement of the complement of the particle into the ©' subject position of the particle-headed SC. NP movement is optional, since the verb's structural Case may also be transferred to the object of the particle in situ, through reanafysis (cosuperscripting) of the verb and the particle.2 (2)

a. They looked [sc the information [pp up]] I I b. They [v looked „J [sc the information [pp up]] I I c. They looked [sc [NW JJ [pp UP the information!! I I

Although the analysis of particles as ergative prepositions, on which (2c) is based, is not a novelty (cf. esp. Gu6ron 1986; 1989; also cf. Taraldsen 1983), it has not exactly managed to carry the day. Nonetheless this approach has a number of significant advantages, one of them being that it offers an insightful account of the properties of so-called complex particle constructions in English, which will be the topic of this article. L The English Complex Particle Construction In English we can distinguish several types of complex particle constructions, representative examples of which are given in (3): (3)

a. b. c. d. e.

They made John out a liar They put the books down on the shelf They sent a schedule out to the stockholders They sent the stockholders out a schedule They fixed the unemployed up with a job In this paper I shall confine myself to a discussion of the examples in (3a-c). To present an analysis of the very intriguing and complex properties of the constructions in (3d,e) would lead me far beyond the space limitations imposed on this article. What makes these complex particle constructions so difficult to analyse is the fact that, on a SC analysis of particle constructions, these constructions will have to contain two SCs in the complement of the verb, both of which must in addition occupy A-positions. To take the case of (3a), this construction contains both SCs found in the examples in (4a,b) at the same time, one of these SCs (John a liar) moreover not being a constituent at S-structure in (3a). (4)

a. They called [sc John a liar] b. They found [sc John out]

24 On a strictly binary branching analysis, and on the assumption that the two SCs in (3a) are indeed generated in A-positions (and hence are not adjuncts), there are precisely two conceivable D-structures for (3a): (5) (6)

They made [sci [sc2 John a liar] [ P P out]] They made [sci [NW e] [PP out [502 John a liar]]]

The structure in (S), which is adopted by Kayne (1985), postulates the SC John a liar in the subject position of the unergative particle out, while this SC is generated in the complement of the unaccusative particle out in (6). Which of these analyses should we choose? In order to establish this, let us first of all have a closer look at the properties of the English make out a liar construction. As Kayne (1985) points out, the construction in (3a) exhibits a very rigid word-order pattern within VP: the particle may only be positioned in between John and the predicative NP a liar (cf. (7)).3 Moreover, the ungrammatically of (8) bears out that the make out a liar construction does not allow the predicative NP to be WH-extracted. In this respect this construction contrasts with a semantically parallel example that lacks a particle, such as the impeccable What kind of a liar did they call John?. (7)

a. They made John out a liar b. 'They made out John a liar c. 'They made John a liar out (8) 'What kind of a liar did they make John out? It is interesting to note that once we add the infinitival copula to be to the example in (3a), so that we arrive at the sentence in (9a), verb-adjacent particle placement and WH-extraction of the predicative NP yield grammatical results, as shown in (9b) and (10). (9)

a. They made John out to be a liar b. They made out John to be a liar c. *They made John to be a liar out (10) What kind of a liar did they make John out to be? These two differences between (7)/(8) and (9)/(10) will turn out to be crucial in determining which of the structures in (5) and (6) is correct. As mentioned above, Kayne (1985) reconciles the underlying constituency of the SC John a liar with its surface discontinuity by generating this SC in the subject position of the particle-headed SC (cf. (5)). In the mapping of D-structure onto S-structure, the predicate of the subject-SC will subsequently undergo movement into an adjunction position to the particle-headed SC, as is illustrated in (11): (11) They made [sci [sci [sc2 John '¡] [pp out]] a liarj] I I Case Theory is held responsible for the obligatory extraposition of the inner-SC predicate. The subject NP of SC2 must receive Case, but is separated from the Case assigner, V, by two SC boundaries. One of these boundaries must hcnce be removed. On a definition of a barrier in terms of L-containment, extraposition of the predicate of SC2 will in effect render the inner SC transparent, so that the Case which is assigned by the matrix verb can now reach the inner-SC subject. This explains the obligatoriness of the movement operation displayed in (11), and at the same time the ungrammaticality of (7c) as a violation of the Case Filter. Kayne's account of the ungrammaticality of verb-adjacent particle placement (cf. (7b)) is heavily dependent on his analysis of the word-order alternation in the verb-particle construction, which I represented in (2a), above. On this analysis the word order in (7b) could only be derived from the underlying structure in (5) through rightward movement of the entire inner SC. That, however, would leave unaffected the violation of the Case Filter, since the barrierhood of SC2 is not voided, its predicate not being removed from the SC. Notice, though, that on a particle-incorporation approach to the alternation in (1), reflected by (2b), it is not so clear why (7b) should be ill-formed if we base ourselves on the structure in (5). The obligatory adjunction of the predicate of SC2 to SCI, in conjunction with a specific definition of a variable, furthermore provides Kayne with an explanation for the fact that the presence of the particle in the make out a liar construction renders WH-extraction of the predicate of SC2 impossible, as is shown by (8). If we assume that the predicate of SC2 is WH-moved to SpecCP from the adjunction position to SCI, problems arise since the trace in the (A'-) adjunction position to SCI does not qualify as a variable, since variables by definition stand in A-positions. The initial trace in the D-structure position of the inner-SC predicate cannot function as the necessary variable either, since a variable must be locally operator-bound. The local binder of the trace in (11) is the intermediate trace in the adjunction position to SCI. That, however, is not an operator. SpecCP does contain an operator; that, however, does not locally bind the most deeply embedded trace. The mandatory adjunction of the inner-SC predicate to SCI hence rules (8) ungrammatical, on Kayne's assumptions concerning operator-variable constructions.

25 Notice, however, that in the current Barriers theory (Chomsky 1986), in which successive cyclic WH-movement via intermediate adjunction to a maximal projection (notably, VP) is found all over the place, Kayne's view of variables cannot be maintained, so that his account of the ungrammaticality of (8) should be abandoned. This account has a number of problematic aspects anyway. In the first place the A'-movement operation displayed in (11) is potentially problematic in view of the Empty Category Principle (ECP) (or Kayne's 1984 Connectedness Condition): extraction takes place out of a left b r a n d . Kayne circumvents this problem by adjoining the predicate of SC2 to SCI, so that the moved predicate can function as a local antecedent for its trace. Secondly, the fact that Kayne's analysis crucially involves extraposition of a SC predicate may raise some eyebrows. While the existence in English of pairs such as (12a), which I analysed in Den Dikken (1987) in terms of optional extraposition of the predicate of the complement-SC (onto the wagon) around the VPadjoined SC (headed by green) (cf. the structure in (13)), indicates that English is not as strict as Dutch, which forbids extraposition of SC predicates altogether,4 the fact that not all constructions of the type in (12a) are grammatical (cf. (12c), taken from Emonds 1976:109) suggests that even in English extraposition of SC predicates is subject to constraints: (12) a. They loaded the hay (onto the wagon) green (onto the wagon) b. They consider Sue (an ugly woman) nude (an ugly woman) c. They painted the house (red) unsanded (*red) (13) [vp IVP load (sci the hay onto the wagon] [502 PRO green]] „. i i The grammatically of (12b) shows that apparently, predicative NPs may undergo extraposition in English. Thai is in perfect accordance with Kayne's (1985) analysis of the make out a liar construction. The fact, however, that extraposition of predicative APs is impossible (cf. (12c)) while on Kayne's analysis it is apparently obligatory in the complex particle construction in (14) (from Jackendoff 1977:67), renders an extraposition approach to complex particle constructions, based on the structure in (5), highly problematic: (14) a. They painted the barn up red b.*They painted up the barn red c. *They painted the barn red up Kayne's analysis of the make out a liar construction also fails empirically since it cannot adequately differentiate between (7) and the minimally different example in (9), whose particle-placement and WHextraction properties differ significantly from those of (7) (cf. (7b) vi. (9b) and (8) vi. (10)). The contrast between (8) and (10) is not mentioned by Kayne, and it is unclear how a solution for it based on his analysis would read. Kayne (1985:114) does notice the contrast between (7b) and (9b), and suggests by way of a solution that the latter example could be derived from the D-structure in which John to be a liar finds itself in the subject position of the particle-headed SC by first of all adjoining the entire ECM infinitival John to be a liar to the particle-headed SC, after which to be liar is then moved rightward yet another time (in order to render the extraposed infinitival clause transparent for Case assignment by V to its subject, John): (15) ... make [[e]j out] [John

[e]v]j

[to be a liar]k

As Kayne (1985:fn. 26) himself also acknowledges, however, there is no a priori reason why the predicate of the extraposed inner SC (a liar) in the ungrammatical example in (7b) should not be allowed to also move rightward a second time, making the inner-SC subject accessible to Case assignment by V in the same way as in (15). Apparently, an analysis of the type in (15) must be unavailable in this case, but the theory gives us no clue to establishing why this should be so. We may conclude, then, that there is ample reason to reject Kayne's analysis based on the structure in (5) in favour of the analysis in (6). Apart from being the only option left to us after the abandonment of (5), the structure in (6) also gains considerable initial appeal from a consideration of the example in (16), which differs from (7) and (9) in that the SC/infinitival clause in these sentences has been exchanged for a finite CP. What is striking about the example in (16) is that this finite clause obligatorily follows the particle without there being a dummy argument (it) in the position preceding the particle. In this respect (16) contrasts sharply with an example such as (17), in which the presence of it is compulsory: (16) They made (*it) out that John is a liar (17) They made *(it) indisputable that John is a liar Vanden Wyngaerd (1989:16) concludes from this contrast that the finite clause in (16) is base-generated as the complement of the particle, which must hence be ergative. On this analysis (16) features no CP extraposition, which explains the absence of it. In (17), on the other hand, the SC predicate is unergative, and CP extraposition leads to the obligatory presence of the dummy pronoun. The contrast in (16)/(17) hence motivates the analysis of out as an ergative preposition, and at the same time gives us an argument in favour of the analysis of the make out a liar construction in (6), which is also proposed by Vanden Wyngaerd (1989). In what follows, I shall develop this analysis in full detail, and extend it to the constructions in (3b,c).

26 Let us first of all return to the examples in (7)-(10). In the D-structure of the make out a liar construction the subject of SC2, John, is generated in a non-Case-marked position: the particle is ergative and hence, given Burzio's generalisation, unable to assign Case to its complement. In order to avoid a violation of the Case filter (or the Visibility Condition), we should hence move John into the ©' subject position of the particleheaded SC in the mapping of D-structure onto S-structure. In that position, John can receive Case from the matrix verb in the familiar fashion. This derivation is illustrated in (18): (18) They made [sci [NPO- J o h n j [ P P out [sc2 tj a liar]]] The obligatoriness of NP movement explains the rigid word-order pattern exhibited by the make out a liar construction. The Case-dependent NP John cannot remain in situ in the subject position of SC2, not even if the verb and the particle are reanalysed (as in (lb), given the ergative-P analysis in (2c), above). The reason for this is that, as I shall motivate shortly, SC2 is a barrier, so that the Case assigned by the verb will always be prevented from reaching John. SC2 is a barrier since it is not L-marked. Although it receives a ©-role from its governor, the particle, I shall assume that the particle, just like INFL, is not sufficiently lexical to be able to fiinction as an L-marker. The Case filter hence forces NP movement of John into the specifier position of SCI. 5 This straightforwardly accounts for the ungrammaticality of (7b). We now also explain the ill-formedness of the example in (7c), since the current Barriers framework does not provide us with any tools that could conceivably make the particle surface at the right periphery of the clause. Rightward particle movement is clearly blocked by Structure Preservation, heads not being allowed to adjoin to maximal projections. Moreover, movement of the predicate of SC2 (which, by Spec-Head agreement, is coindexed with its subject) into an adjunction position to the left of the particle, apart from being unmotivated, would have the undesirable result of making the NP-trace of the moved subject of SC2 locally A'-bound by the PP-adjoined predicate of SC2. Before turning to the ungrammaticality of the WH-question in (8), I would first of all like to account for the influence that to be exerts on the possibilities of particle placement, i.e. the difference in grammaticality between (7b) and (9b). On the assumption that the range of possible D-structures is strongly curtailed by a principle of the type in (19), and given that the thematic roles assigned to SC2 in (7) and the infinitival IP in (9) are identical, the conclusion is forced upon us that the D-structure of (9) must be completely analogous in all relevant respects to that of (7). This leads to the D-stucture in (20). (19) The Uniformity of Theta Assignment Hypothesis (UTAH) Identical thematic relationships between items are represented by identical structural relationships between those items at the level of D-structure (Baker 1988) (20) They made [sci [NW e ] [pp out [ip John to be a liar]]] Just as in the case of (6) this structure will yield an ill-formed output unless it is manipulated in some way. Here as in (6), NP movement of John is a way of avoiding a Case filter violation. Hence (9a) is grammatical. Apparently, however, the structure in (20) also allows the subject of the infinitival clause to remain in situ, in view of the grammaticality of (9b). This means that in this case reanalysis of the verb and the particle renders it possible for the Case assigned by V to reach the subject of the particle's complement. This in turn implies that the complement of the particle apparently is not a barrier in (20). Nonetheless, out does not Lmark its complement in (6) and (20) alike. There must hence be some other way in which the barrierhood of the particle's complement can be voided in this case. Broadening our horizon to include the prepositional particle constructions in (3b,c), we see that something similar must be going on in these constructions: as the b-examples in (21) and (22) indicate, verb-adjacent particle placement is possible here as well: (21) a. They b. They c. 'They (22) a. They b. They c. 'They

put the books down on the shelf put down the books on the shelf put the books on the shelf down sent a schedule out to the stockholders sent out a schedule to the stockholders sent a schedule to the stockholders out

On the null-hypothesis that the underlying structure of these examples is analogous to that of the make out (to be) a liar construction, the D-structure of the prepositional complex particle construction in (21) looks as in (23), with a SC in the complement of the particle: (23) They put [sci

INPS'

e] [pp down [sea the books on the shelf]]]

Just as in the structures given above, the particle is unable to L-mark its complement, so that we should expect Case assignment to the subject of SC2 in situ to be impossible. The opposite is true, however, (21/22b) being perfectly acceptable.

27 The task we are facing now is to lift the barrierhood of the particle's complement in (9), (21) and (22), but not in (7). Adhering to the particle's inability to function as an L-marker will straightforwardly yield us the barrierhood of SC2 in the structure in (6), as required. We could now prevent the SC in the complement of the particle in (23) from becoming a barrier by capitalising on the notion of a segment, which also figures prominently in Chomsky's (1986) definitions of a barrier. In the prepositional complex particle construction the predicate of the SC complement of the particle is categorially identical to the predicate of the higher SC, of which the particle is the head. The substructure in (24) hence forms an integral part of the structure of the examples in (21) and (22): (24) ... [ PP Prt ... [ PP ...]] This structure can be looked upon as a base-generated instantiation of the multi-segment structure that is obtained through adjunction (cf. also Chomsky 1986:76 for another implementation of this idea in a different context). In this structure the lower PP is a segment of the entire, multi-segment PP. Since the outermost segment of this PP is L-marked by the verb, and since~as Chomsky (1986) and more recently also Lasnik & Saito (forthc.) have argued--L-marking percolates down to the head of a projection, we are led to the desirable conclusion that the complement of the particle in the structure in (23) is not a barrier since it is Lmarked. What we see, then, is that just in case the syntactic category of the particle's complement is identical to that of the particle itself, the lower SC is not a barrier, so that transmission of the Case assigned by the verb to the subject of SC2 in (23) is possible (provided, of course, that the verb and the particle are reanalysed). By making crucial use of the concept of a segment we are hence in a position to account for the grammatically of (21/22b) without at the same time~wrongly~predicting that (7b) should be grammatical as well: in the structure in (6) the two SC predicates are not categorially identical, so the possibility of interpreting the structure in terms of a multi-segment constituent as in (24) never comes into play. This analysis of the contrast between (7b) on the one hand, and (21/22b) on the other, can be directly transposed to the example in (9b). The only thing that we need assume in addition to what we have assumed already is that the English infinitival marker to, which is homophonous to the preposition to to begin with, has the categorial status of a preposition. On this plausible assumption, which is also made in e.g. Reuland (1983) and Gudron & Hoekstra (1988), the IP complement of the particle in (20) is identical in all relevant respects to a PP, so that (20) is another instance of the substructure in (24). The prima facie puzzling contrast between (7b) on the one hand, and the b-examples in (9), (21) and (22) on the other can hence be shown to be accountable for in an insightful fashion on the basis of our analysis of complex particle constructions. Apart from the problems posed by the different particle placement possibilities found in the various English complex particle constructions, for which we have now furnished an explanation, the examples in (7)-(10) also confront us with a second problem: why does WH-extraction of the predicative NP yield an ungrammatical result in (8), while (10), and also a sentence as closely related to (8) as What kind of a liar did they call John?, are perfectly acceptable? A consideration of the prepositional complex particle constructions shows that the contrast between (8) and (10) is not an isolated fact: the examples in (25) and (26) are further illustrations of the fact that WHextraction of the constituent immediately following the particle is consistently impossible, while subextraction from this constituent does lead to a well-formed result: (25) a. *On which shelf did they put the books down? b.'To whom did they send a schedule out? (26) a. Which shelf did they put the books down on? b. Who did they send a schedule out to? This empirical generalisation can be given formal substance by once again making use of our earlier assumption that particles are not full-fledged lexical categories. This lexical defectiveness of particles can now be held responsible for the fact that extraction of the predicate of SC2 in the structures in (6) and (23) results in ungrammatically. As the particle is not a lexical governor, the trace left by WH-extraction in (8) and (25) is not lexically governed. This results in a violation of the Empty Category Principle (ECP), on the assumption that lexical government is a conditio sine qua non, in any event for cases of extraction from an A-position. The ungrammaticality of the examples in (8) and (25) can hence be viewed as a consequence of the ECP, while the fact that (10) and the examples in (26) are well-formed follows from the fact that there the ECP's demands are met, the WH-trace being lexically governed. Apart from subextraction from PP, as in (26), there are two other ways in which the ungrammatical examples in (25) can be rescued. One is to WH-extract the particle and the transitive PP together, as in (27) (adapted from Sag 1982 and Stowell 1981:339): (27) a. Down on which shelf did they put the books? b. Out to whom did they send a schedule?

28 Notice that (27) lends further support to the analysis of complex particle constructions in (6), in which Prt and PP form a constituent which does not also include the subject of SCI (PP in (23)). On an analysis & la Kayne (198S), no such constituent can be identified since the only projection that comprises both the particle and the transitive PP in the post-adjunction structure in (11) (i.e. the upper SCI) also necessarily includes the subject NP of SC2 (the books, a schedule), which is not preposed in (27). The one remaining way to salvage the examples in (25) emerges from a comparison of these sentences with the constructions in (28):6 (28) a. On which shelf did they put down the books? b. To whom did they send out a schedule? What these sentences show is that once the particle is placed in verb-adjacent position, WH-extraction of the predicate of SC2 yields a grammatical output. The crucial difference between the examples in (25) and those in (28) is the fact that in the latter reanalysis of the verb and the particle has taken place, which allows the verb's Case to be transmitted to the subject of SC2 in situ. My conception of reanalysis is in essence identical to that of Baker (1988), who looks upon this phenomenon as an abstract instance of the general rule of head incorporation, expressing it in terms of (»indexation (cosuperscripting) of the two heads, which is interpreted as being equivalent to the coindexing induced by head movement. Being an instantiation of incorporation, reanalysis of the verb and the particle hence displays all the properties of head incorporation. In particular, it will feature the effects of Baker's (1988) Government Transparency Corollary, given in (29): (29) The Government Transparency Corollary A lexical category which has an item incorporated into it governs everything which the incorporated item governed in its original structural position (Baker 1988) For the case at hand, the Government Transparency Corollary entails that as a consequence of V-Prt reanalysis the matrix verb will come to govern everything that the particle governed at D-structure in (28). In other words, the trace of the WH-moved PP in these examples ends up being governed by the matrix verb, which, in contradistinction to the particle, is a lexical governor. As a result of reanalysis the WH-extractions in (28) hence satisfy the ECP, which explains the well-formedness of these examples. Since nothing in principle prevents us from assuming that in the make out a liar construction, too, reanalysis of the verb and the particle may take place, what we should now account for is why even if the particle is in verb-adjacent position the example in (8) does not improve: (30)*What kind of a liar did they make out John?

(cf. (8))

In my analysis of the various English complex particle constructions one assumption plays a key role: the idea that a particle is not a full-fledged lexical category. This assumption has two important implications. In the first place the particle's lexical defectiveness causes the complement of the particle not to be L^marked. Hence it is a barrier (unless it is categorially identical to the particle). Secondly, the non-lexical status of the particle implies that the particle is not a lexical governor, so that extraction of a constituent governed by the particle is not allowed. In the examples in (25) and (28) we only had to reckon with the second implication, the complement of the particle being exempt from barrierhood by virtue of its categorial identity with the particle. In (8) and (30), by contrast, there is no such categorial identity. As a consequence the complement of the particle will always be a barrier by lack of L-marking. Reanalysis of particle and verb will not change this. While incorporation of a lexical element into a non-lexical 3-marker will turn the latter into an L-marker (as e.g. in the case of V-to-I movement, which causes INFL to L-mark its VP complement), the opposite situation does not occur. That is, incorporation of a non-lexical ©-marker into a lexical element will never turn the former into an L-marker. In other words, even after reanalysis of V and the particle SC2 will not be Lmarked. Being a barrier as a result, this SC will hence block transmission of the verb's structural Case to the subject of SC2, so that (30), just like (7b), violates the Case filter. Moreover, and perhaps even more significantly, the SC forms a barrier for proper government of the WH-trace, so that (30) also violates the ECP. The contrast between the examples in (28) on the one hand, and the ungrammatical construction in (30) on the other is hence straightforwardly explained on the present analysis.9

2. Concluding Remarks In this article I have shown that a consistent implementation of the SC analysis leaves precisely two potential analyses of English complex particle constructions. On the first approach, taken by Kayne (1985), the subject position of the particle-headed SC is occupied by another SC, whose predicate is obligatorily extraposed in the mapping of D-structure onto S-structure. I have shown that this analysis, apart from suffering from some nontrivial theoretical problems, fails to make the desired distinction between the examples in (7)/(8) and (9)/(10). The alternative approach, which postulates a SC in the complement of an ergative particle, manages to explain the properties of the English complex particle construction with the aid of some rather minimal assumptions.10

29 The analysis of complex particle constructions proposed in this article lends important support to the Small Clause analysis. The various particle placements and the possibilities of WH-extraction of the constituent following the particle can-as I have shown in this article-be straightforwardly explained if it is assumed that the particle is an independent syntactic head projecting a SC. An analysis that generates particles as subparts of complex verbal categories has no insightful perspectives to offer in this respect. The SC analysis of verbparticle constructions is thus strongly motivated. Moreover, our analysis of complex particle constructions supports the view that locative and directional PPs (as found in (3b,c)) are SC predicates, as is also argued extensively in Hoekstra & Mulder (1990). Finally, the analysis of complex particle constructions presented here lends interesting support to the crucial assumption that particles are ergative prepositions, which is also very desirable in a variety of other respects (cf. Guiron 1989; Gu6ron & Hoekstra 1990 for discussion). Notes *

The research leading up to this article was part of the Leiden University research programme 'Structural Properties of Language and Language Use', financed by the Dutch Ministry of Education. Earlier versions of this paper were presented at the Utrecht University researchers meeting, 13 October 1989, at the first Leiden Conference for Junior Linguists, 16 November 1989, and at the annual conference of the Linguistic Society of the Netherlands, 20 January 1990. I thank the audiences present on those occasions, as well as Hans Bennis, Frits Beukema, René Mulder, Guido Vanden Wyngaerd, Jan-Wouter Zwart and the editors of this volume for their comments and suggestions. The present form of the paper owes much to various stimulating discussions with Teun Hoekstra. The usual disclaimers apply.

1. Here and elsewhere in the paper, the label 'SC' is used purely for expository purposes. Since it is of no consequence to the issues raised here, 1 do not take a stand on the categorial status and internal structure of small clauses (cf. Den Dikken 1987 for relevant discussion). 2.

Reanalysis as intended here does not involve restructuring, but rather is similar to the mechanism adopted by e.g. Hornstein & Weinberg (1981) to enable a preposition (in their case the head of a prepositional object, in the present case an ergative preposition) to transmit the verb's structural Case to its complement. More will be said about the precise nature of V-Prt reanalysis further below.

3.

While I have given (7b) a full asterisk, Kayne assigns it the judgement *?. The fact that this example is apparently somewhat less bad than the totally unacceptable (7c) is not explained by Kayne, however. Since it is of no concern to me either, I have chosen not to differentiate between (7b) and (7c). (Incidentally, the same slight contrast is noted by Jackendoff 1977:76 for the examples in (14b,c), below.)

4.

This is shown by the well-known fact that predicative PPs (i.e., prepositional SC predicates) fail to undergo PP-over-V in Dutch: (i) a.

dat Jan het boek op de plank zette that John the book on the shelf put 'that John put the book on the shelf b. *dat Jan het boek zette op de plank

5.

One may ask how this NP movement is licit, given that a barrier (SC2) is crossed. Notice that this issue is not peculiar to the construction at hand. On the plausible assumption that subjects are generated VPinternally, NP movement of the subject NP into the specifier of IP, which is forced by the Case filter just as in (18), likewise crosses a barrier, given that English VP is never L-marked and that adjunction to VP in order to void VP's barrierhood is not allowed since that would result in illicit movement. A similar situation furthermore arises in raising and passive constructions (unless we are willing to adopt Chomsky's 1986 extended chains proposal). Apparently, then, NP movement is not blocked by the presence of a barrier between the extraction and landing sites of the movement operation. This could be captured by assuming that NP movement is not subject to the ECP but to some other constraint, possibly Condition A of the Binding Theory, which, as the reader will verify, is satisfied in all of the above cases (but which, it should be noted, will be unable to account for the strong deviance of such 'super-raising1 constructions as *John seems that it is certain to win, pleonastics normally only giving rise to "weak" binding violations; cf. Chomsky 1986:18n). Notice that Lasnik & Saito's (forthc.) recent adaptation of the Barriers theory also opens up the possibility that the ECP does not apply to NP movement. I cannot follow them, however, in requiring that not a single barrier separate NP-trace and its antecedent.

30 With respect to the prepositional complex particle construction, there is a remarkably close correspondence between English and (some dialects of) Norwegian. Thus, Alma Naess (p.c.) informs me that in the example in (i), verb-adjacent particle placement is allowed, while clause-final particle placement is not (cf. (21)), and that for many speakers A'-extraction of the predicate of SC2 is not allowed unless the particle is in verb-adjacent position, subextraction from the predicate of SC2 as well as movement of the particle and the transitive PP together being perfect for all speakers (cf. (ii) to the a-examples in (25)-(28)): (i) a.

Han satte katten ned pi gulvet he put the-cat down on the-ground b. Han satte ned katten p& gulvet c. *Han satte katten pS gulvet ned (ii) a. *P& gulvet satte han katten ned b. PS gulvet satte han ned katten c. Ned pS gulvet satte han katten d. Hva satte han katten ned pi? what put he the-cat down on The same parallelism is found between the English examples in (22) and the b-sentences in (25)-(28) and their Norwegian counterparts in (iv)-(v) (thanks to Arnfinn Vonen for constructing and judging these sentences): (iv) a. b. c. (v) a. b. c. d.

De sendte m^teprogrammet ut til aksjonaerene they sent the-schedule out to the-stockholders De sendte ut mtfteprogrammet til aksjonaerene *De sendte mtfteprogrammet til aksjonaerene ut *Til aksjonaerene sendte de mtfteprogrammet ut Til aksjonaerene sendte de ut miteprogrammet Ut til aksjonaerene sendte de mdteprogrammet Hvem sendte de m^teprogrammet ut til? who sent they the-schedule out to

It may now be asked why English should feature only abstract particle incorporation, or, more concretely, why the ungrammatical word order in (7b) couldn't now be derived by NP-moving John into the subject position of SCI in (6) (so that a Case violation is averted) and by applying actual incorporation to the particle, adjoining it to the verb, as shown in (i): (i) They [ v made outj] [ SC i [ N W John,] [ PP tj [sc2 t; a liar]]] In order to rule out this derivation of (7b) we can suggest that right-adjunction of the particle to the verb, as in (i), creates a complex verbal category that is morphologically ill-formed, since it violates Williams' (1981) Right-hand Head Rule, which defines as the head of a complex X° category the right-hand member of the word. If this suggestion is viable, we now face the problem of accounting for why (ii), with the particle /¿/(-adjoined to the verb (so that the Right-hand Head Rule is complied with), is still ungrammatical: (ii) *They out-made John a liar That particles do not left-adjoin to V in English, as opposed to e.g. Dutch, might be related to an independent difference between these two languages: the fact that English is VO while Dutch is OV, so that in the former, in contrast to the latter, the particle is not adjacent to the verb in the syntactic structure, which might be held responsible for the impossibility of actual incorporation. Notice, though, that even apart from the fact that this suggestion is difficult to formalise, its empirical validity is rendered dubious by the fact that in Norwegian, a VO language, particles can be found to the left of V in passives, as is shown in (iii) (adapted from Afarli 1984:13). While I leave this issue open, it is worth pointing out that, to the extent that the passive examples in (iii) can be argued to have been derived through actual incorporation of the particle into the verb, these sentences lend support to our claim that the word-order alternation in the active examples in (iii) (and in the English verb-particle construction in general) is not obtained through actual particle incorporation, but rather through abstract incorporation (i.e. V-Prt reanalysis, or cosuperscripting). (iii) a. b. c.

Vi sparka (ut) hunden (ut) we kicked out the-dog out Vi klipte (av) trSden (av) we cut off the-thread off Vi kjtfrte (bort) avfallet (bort) we drove away the-waste away

/ / / / / /

Hunden vart utsparka the-dog was out-kicked TrSden vart avklipt the-thread was off-cut Avfallet vart bortkjtfrt The-waste was away-driven

31 8. Interestingly, double SC constructions of the type in (12a) behave exactly as expected when a particle is added. Robin Smith (p.c.) gives the following judgements of the examples in (i): (i) a. They b. They c. They d. 'They

carried carried carried carried

up the hay the hay up up the hay the hay up

onto the wagon green onto the wagon green green onto the wagon green onto the wagon

Our analysis correctly ensures that PP-extraposition across the VP-adjoined adjectival SC is possible just in case the particle is reanalysed with the verb, so that V can lexically govern PP's trace. The facts in (i) hence at once support the extraposition approach to the word-order alternation in double SC constructions proposed in Den Dilcken (1987) and the account of WH-extraction in complex particle constructions presented here. Notice that the grammaticality of (ic) is problematic for Kayne's (1985) approach to complex particle constructions: from the adjunction position to SCI in (11), above, the inner-SC predicate (PP) should not be allowed to move any further, given Kayne's views on operator-variable relations. Thus the paradigm in (i) also offers an additional argument against Kayne's analysis. 9. Notice that the fact that the particle's SC complement no longer L-contains its predicate is not sufficient to lift the SC's barrierhood (contra Kayne 1985). Lack of L-marking will continue to ensure that the SC is a barrier. 10. On Kayne's analysis, as on ours, recursion of particles is structurally possible, Kayne's approach in principle allowing for infinitely many subject-SCs and the present proposal being compatible with a series of ergative particles all selecting particle-headed SCs. As Gu6ron (1989) has pointed out, however, the ergative-P analysis of particle constructions allows for a principled way in which this unwanted recursion of particles can be prevented. Gu6ron argues that the particle, being a perfective predicate, cannot initiate a new tense domain and hence cannot select another particle-headed SC, which on Gudron's assumptions constitutes such a new tense domain (for details I refer to the original paper). References ÂFARLI, T. 1984 Norwegian Verb Particle Constructions as Causative Constructions, Working Papers in Scandinavian Syntax 11. University of Trondheim BAKER, M. 1988 Incorporation: A Theory of Grammatical Function Changing, Chicago: The University of Chicago Press CHOMSKY, N. 1986 Barriers, Cambridge, MA: MIT Press DIKKEN, M. DEN 1987 Secundaire predicate en de analyse van small clauses, in GLOT 10. 1-28 EMONDS, J. 1976 A Transformational Approach to English Syntax, New York: Academic Press GUÉRON, J. 1986 Clause Union and the Verb-Particle Construction in English, paper delivered at NELS 16; ms., Université de Paris 8 1989 Particles, Prepositions, and Verbs, ms., Université de Paris 8; published in Mascaró, J. & M. Nespor (eds.) (1990), Grammar in Prog-ess, Dordrecht: Foris GUÉRON, J. & T: HOEKSTRA 1988 Les chaînes-T et les verbes auxiliaires, in Lexique 7. 61-85 1990 Tense, Particles Causatives, abstract Time Conference, MIT HOEKSTRA, T. & R.andMULDER 1990 Unergatives as Copular Verbs: Locational and Existential Predication, in The Linguistic Review 7:1. 1-79 HORNSTEIN, N. & A. WEINBERG 1981 Case Theory and Preposition Stranding, in Linguistic Inquiry 12:1. 55-91 JACKENDOFF, R. 1977 X'-Syntax, Cambridge, MA: MIT Press

32 KAYNE, R. 1984 Connectedness and Binary Branching, Dordrecht: Foris 1985 Principles of Particle Constructions, in Gudron, J., H.-G. Obenauer & J.-Y. Pollock (eds.), Grammatical Representation. Dordrecht: Foris. 101-40 LASNIK, H. & M. SAITO forthc. Move Alpha: Conditions on Its Application and Output, Cambridge, MA: MIT Press REULAND, E. 1983 Government and the Search for Auxes: A Case Study in Cross-linguistic Category Identification, in Heny, F. & B. Richards (eds.), Linguistic Categories; Auxiliaries and Related Puzzles (Vol. I). Dordrecht: Reidel. 99-168 SAG, I. 1982 Coordination, Extraction, and Generalized Phrase Structure Grammar, in Linguistic Inquiry 13. 329-36 STOWELL, T. 1981 Origins of Phrase Structure, diss., MIT TARALDSEN, T. 1983 Parametric Variation in Phrase Structure, diss.. University of Tromsd WILLIAMS, E. 1981 On the Notions 'Lexically Related' and 'Head of a Word', in Linguistic Inquiry 12:2. 245-74 WYNGAERD, G. VANDEN 1989 Raising-to-Object in English and Dutch, Dutch Working Papers in English Language and Linguistics 14

Jan Don

Prefixed Nouns in Dutch 0. Introduction* Since Trommelen & Zonneveld (1986) it seems well-established that Dutch morphology is sensitive to the Right-hand Head Rule (henceforth: RHR) as originally formulated by Williams (1981). The present paper deals with some puzzling facts with respect to this rather strict principle of morphology, concerning prefixed nouns in Dutch. Consider the following data: (1)

a. ver-raady be-roep v ont-zety ver-dervy

'to betray' 'to appeal to' 'to relieve' 'to pervert'

het het het het

ver-raad be-roep ont-zet ver-derf

'to end' 'to summon1 'to throw in' 'to print' 'to cross'

de de de de de

af-loop op-roep in-gooi uit-draai over-steek

'the treason' 'the profession' 'the relief 'the ruin'

b. [prt-Mlv af-loopv op-roepv in-gooiy uit-draaiy over-steeky

c. de af-stand het over-gewicht de uit-hoek het voor-deel

'the 'the 'the 'the

'the end' 'the summons' 'the inthrow (lit.)' 'the print-out' 'the crossing'

distance' overweight' remote place' advantage'

Nouns prefixed with be-, ver-, and ont- all take the definite article het (la). As is well-known, prepositions may function as 'particles' in derived verbs, and nouns related to these verbs all take the definite article de (lb). Finally, if these particles function as true prepositions in compound nouns, the RHR provides the definite article (lc). The nouns in (la) are clearly a problem for the assumption that Dutch is RHR-sensitive. Trommelen & Zonneveld try to cope with these apparent counterexamples by assuming that the RHR is overruled in these cases by redundancy-rule (2): (2) All existing prefixed nouns from any source are neuters.

34 Although empirically correct, (2) is not a very satisfying answer to the problem just sketched. It immediately returns the question in a different guise: why should (2) hold for the grammar of Dutch? The data in (la) arc also problematic form another perspective. In Don (1990b) I have argued that noun-forming conversion in Dutch is a zero-itf/Fixation process that produces nouns that take the definite article de. So, a form like de loopK ('the run') has a structure as in (3), in which the zero-suffix determines the gender of the noun, in keeping with the RHR. (3) N

Although naturally we need not immediately accept this proposal, it is a striking fact even from the observational angle that the prefixed nouns in (la) which are related to phonologically identical verbs are neuters, i.e. they take the definite article het. First, we will try to provide an answer to these questions. This part of the paper (section 1) is a summary of Don (1990a). The solution offered there, however, raises new questions which will be the main topic of this paper. The solution in section 1 only explains why the data in (la) exist. It does not explain why grammatically equivalent forms taking the article de do not exist. This will lead us to the hypothesis that the zero-affix is restricted to monosyllabic bases. This hypothesis seems to be contradicted by the data in (lb), but we will show that this problem can be elegantly solved. The paper is organized as follows. In section 1 we will give an answer to the question why (2) holds for the grammar of Dutch. Section 2 tries to cope with the problem of the nonexistence of the de verraad-typt nouns, and section 3 deals with the data in (lc). Section 4 summarizes the main conclusions.

1. Ge-deletion In Don (1990a) we have proposed the following answer to the question why (2) holds for the grammar of Dutch. Schultink (1973) observes that, whereas Dutch past participles are normally formed with the prefix ge- and with the suffix -d or -r (depending on the voicequality of the final consonant), the past participles of verbs prefixed with be-, ver- and ontlack the ge- prefix. Schultink ascribes this to a rule of ge-deletion, which, somewhat simplified, states that ge- is deleted before a stressless prefix. Thus, a form like bewerkt (past participle of the verb be-werky) is underlyingly ge-be-werk-t. If we combine this analysis with the observation that ge- normally derives nouns which all take the article het, then we might suppose that this prefix is underlyingly present in the data in (la) and is deleted by the rule Schultink proposed for the past participle-data. Thus, a form like het beraad underlyingly looks like het ge-be-raad. This analysis explains the fact that the data in (la) are all neuters: ge-, which is the head in these constructions clearly makes neuters. (4) illustrates this:

35 (4) het het het het het

gebed gebod gebruik gedrag geduld

prayer 'command' 'usage' 'behaviour' 'patience'

het het het het het

geheim gehoor gelag geloof gemis

'secret' 'audience' 'score' •belief 'lack'

Notice that under this analysis of the data in (la) we do not have to make a special statement in the grammar to determine the gender of the nouns in (la). However, these data are still not precisely what one would expect to find in a RHR-sensitive language: ge- as a prefix determines category and gender. Ge- must be marked as an exceptional head, like e.g. the prefix en- in English. For further details of this analysis, we refer the reader to Don (1990a).

2. Noun-forming conversion In Don (1990b) we have argued that Dutch boasts a process of zero-affixation that derives nouns from verbs. (5) provides some examples: (5) V loop val trap strijd trek bouw

'to walk' 'to fall' 'to kick' 'to fight' 'to migrate' 'to build'

N de loop de val de trap de strijd de trek de bouw

'the walk' 'the fall' 'the kick' 'the fight' 'the migration' 'the construction'

This list could be extended almost indefinitely. Given the fact that this is a quite productive process, we may again ask why besides the data in (la), we do not encounter such forms taking the article de. In other words, why is it that prefixed verbs may undergo nounformation with the prefix ge-, but not with the zero-affix of the data in (5). Next to valv, we have de valN and het gevalN, but next to vervaly, we only encounter het vervalN, but crucially not de vervalN. We would like to offer the following hypothesis. Let us assume that the zero-affix in the data in (5) requires a monosyllabic stem. Several otherwise puzzling data fall out straightforwardly. First, the so-called 'Romance' verbs in Dutch, which are formed with the suffix -eer or its variant -iseer, do not undergo this type of zero-affixation. (6) gives some examples: (6) argumenteer publiceer informeer

'to argue' 'to publicize' 'to inform'

de argumenteer de publiceer de informeer

'the argue'(lit.) 'the publicize' (lit.) 'the inform' (lit.)

36 It is generally agreed in the literature ( Booij (1977), Van Beurden (1987), Trommelen & Zonneveld (1989)) that -eer, like all other 'Romance' morphology in Dutch is a level l-suffix. From this it follows that we can not derive the ungrammatically of the nouns in (6) by ordering the affixation with the zero-affix before the affixation of -eer, simply because there is no earlier level in Dutch morphology than the level at which -eer attaches. The ungrammaticality has to follow from some other constraint. Our hypothesis is thus confirmed by these data. Secondly, it has been pointed out by Trommelen (1989) that Dutch underived verbs are generally monosyllabic1. However, there is a small class of exceptions which do not allow noun-forming conversion by means of the zero-affix. (7) illustrates this: (7) bakkelei plavei neurie ravot

'to tussle' 'to pave' 'to hum' 'to romp'

*

de bakkelei de plavei de neurie de ravot

'the tussle'(lit.) 'the pave'(lit.) 'the hum'(lit.) 'the romp'(lit.)

The ungrammaticality of the nouns in (7) immediately follows if we assume that the 0-affix only attaches to monosyllabic verbs. We need to add at this point that there are a lot more verbs that do not fit the pattern 'monosyllabic'. All these verbs are related to a phonologically identical noun. Clearly, these verbs contradict both the generalization of Trommelen and the constraint on O-affixation, we are advocating here. But we agree with Trommelen that these verbs are derived from the nouns and not the other way around. E.g. oliev ('to oil') is related to a noun de olieN ('the oil'), but it does not seem strange to assume that the verb is denominal here. One might object that this point is unfalsifiable. If a non-'monosyllabic' verb does not have a related noun, this is taken as evidence for the constraint on O-affixation. And, if there is a related noun, it is assumed that the verb is derived. However, we think that denominal verbs are independently characterised by the fact that they are never strongly inflected. In other words, the constraint on the zero-affixation that we propose here, can be falsified by showing that there are strong verbs with related nouns that do notfit-the 'monosyllabic' pattern. To our knowledge such data do not exist. At this point we might conclude that our hypothesis receives considerable empirical support. The reason that nouns of the type de verraad fail to exist, is now explained by the fact that the zero-affix which could derive such forms can not attach to other than 'monosyllabic' stems. Notice that such a phonological condition on affixation is not so strange as one might be inclined to think at first sight. In English, it is a well-established fact that the comparative suffix -er only attaches to monosyllabic and bisyllabic stems if the second syllable is light. The noun-forming zero-affix appears to obey a similar restriction. Although this restriction seems to be quite well motivated for Dutch, it clearly encounters abundant counterevidence from the data in (lb). Verbs like af-loopw are not 'monosyllabic', but seem to undergo noun-forming conversion quite easily. In the next section we will address this problem.

37 3. Bracketing Paradoxes Before we turn to the problem as it emerges from the discussion in section 2, we will first have a closer look at the data in (lb), repeated here as (8a), and some new data (8b). (8) a. de af-loop de op-roep de in-gooi de uit-draai de over-steek

'the end' 'the summons' 'the inthrow' (lit.) 'the print-out' 'the crossing'

4c de in-burger de in-blik de af-room *de op-hoog de uit-diep

'the naturalization' 'the in-can' (lit.) 'the skimming' 'the hightening' 'the deepening'

The verbs in (8) are all so-called 'separable' verbs. They can be 'separated' in syntax by e.g. verb-second. (9)

a.

b.

Ik zei dat het leger hem oproept. 'I said that the army him oncalls' (lit.) I said that he gets drafted into the army Het leger roept hem op. The army calls him on' (lit.)

These verbs are problematic for an autonomous morphology. Normally, it is assumed that syntactic rules can not refer to material under the X^-level. So, on the basis of the data in (9), one could argue that these verbs are syntactic constructs of some type. But, two arguments contradict this view. First, it is a well known fact that the 'separable' verbs may be input to derivational morphology. (10) gives some data: (10)

aan-raady af-roepv mede-deely op-vraagy

'to advice' 'to call over' 'to inform' 'to ask for'

aanraad-erN afroep-ingN mededeel-ingN opvraag-baarA

'the suggestion' 'the call over' (lit.) 'the information' 'askforable' (lit.)

Second, the semantic properties of the 'separable' verbs are not transparently related to those of the particle and the verb. The non-transparant semantics also occur in the derivations in

38 (10). So, we will assume in this paper that these verbs are morphological constructs, and leave the problem of the 'separability' unaddressed here. From the data in (8) we may conclude that clearly not all 'separable' verbs are suitable input to the noun-forming conversion process. The crucial difference between the verbs in (8a) and (8b) is that in the former the right-hand member is an underived verb, while in the latter it is a noun or an adjective. This shows that we cannot account for the ungrammatically of the data in (8b) by a simple subcategorization-frame on the suffix; we need to have access to information that is one cycle too 'deep' 3 ; such a subcategorization-frame for the zero-affix would violate the atom-condition, or a variant thereof, which restricts morphological operations to immediately adjacent cycles (cf. Allen (1978), Siegel (1978), Williams (1981)). Informally speaking, one would wish to say that the zero-affix is attached to the head of the separable verb and that because of the fact that the heads of the verbs in (8b) are not of the suitable type (not 'monosyllabic' verbs) the constructions are out. In such an account the problem of section 2 would be solved as well: the reason that the phonological restriction on the zero-affix is not violated, is that the affix in (8a) is attached to the head of the verb and not to the verb as a whole. In the remainder of this paper we will try to develop an analysis along these lines. It should be clear that we will be dealing with so-called 'bracketing paradoxes'. Phonologically, one would like to propose a structure as in (1 la), because this would account both for the grammaticality in (8a) and for the ungrammatically of the data in (8b), but semantically the structure in ( l i b ) seems more appropriate. (ll)a.

b. N

op

N

roep

0

op

roep

0

The analysis we want to propose is based on Williams (1981) 4 . We will very briefly illustrate this theory, taking the more or less 'classic' unhappier-pamdox as an example. The paradox runs as follows. The English comparative suffix -er only attaches to monosyllabic stems and bisyllabic ones if the second syllable is 'light'. Hence, the adjective unhappier should be structured as in (12a). However, compositional semantics require the structure of (12b). Unhappier does not mean "not happier", but "more unhappy". (12)a.

b. A

un

A

happy

er

un

happy

er

39 Williams gives up a compositional semantics, and introduces the notion "lexically related", which is defined as follows. (13)

(cf. Williams (1981), p. 261) X can be related to Y if X and Y differ only in a head position or in the nonhead position.

The definition of the head is now commonly known as the RHR. The definition of the nonhead reads as follows: (14)

(cf. Williams (1981), p. 261) Nonhead: the highest left branch of a word.

With this definition of relatedness, Williams is able to relate unhappier (with the single structure (12a)) to unhappy, thus accounting for its semantics. It does not seem impossible that Williams' account can be reformulated in terms of a restructuring operation applying between S-structure and LF, thus retaining a compositional semantics, but nothing hinges on this at present. If we propose such an analysis for the cases in (8a), we get the following picture; (8a) is assigned the structure (15). (15) N

op

roep

0

This structure, although not containing the separable verb as a constituent, can be related to the verb since the zero-affix is a head. Hence, without violating either the phonological condition on the attachment of the zero-affix or its subcategorization-frame, the noun can be "related" to the separable verb. If we treat the data in (8b) in the same way, the following picture emerges. We cannot attach the zero-affix to the separable verb, because of the phonological restriction that it obeys. The other possibility is to attach it to the noun or adjective which is the right-hand member of the separable compound. But then we do not satisfy the subcategorization-frame of the zeroaffix. Notice that we are assuming that the separable verbs in (8b) are 'exocentric' constructs: the left-hand member is a particle, the right-hand member is an adjective or a noun, but the complex is a verb. These then, are structures on a par with well-known examples like [pushv UPP]N- This assumption however is not crucial. If we assume instead, that the right-hand

40 nouns and adjectives are converted (zero-affixed) to verbs prior to the attachment of the particle, the noun-forming zero-derivation will be excluded on independent grounds. The structure would be as in (16): (16) N

in

burger

0

0

(16) requires the stacking of two zero-affixes. Such stacking, however, turns out to be generally disallowed in Dutch. Thus, the /ier-nouns in (17a) can be converted to verbs (17b), which cannot in their turn be converted to de-nouns (17c). (17)

a.

hetdeel hetfeest het boek het huis

'the part' 'the party' 'the book' 'the house'

b.

deel-0 v feest-0 v boek-0 v huis-Oy

c.

*de deel0 v -0 N defeestO v -0 N de boek0 v -0 N de huisOv-0N

Thus, whatever structure is proposed for the separable verbs in (8b), we cannot generate phonologically identical nouns which can be related to them. Having accounted for the data in (8) in this way, again the question crops up why the data in (la) are grammatical only with the /icf-article. Although a phonological restriction prohibits the structure in which the zero-affix is attached to the prefixed verb, we cannot prohibit the structure (18) which at first sight could be related to the verb verraady. (18) V

ver

raad

0

Note however, that the zero-affix in (18) is neither the head nor the non-head, therefore we cannot relate (18) to the verb verraady. Also prefixes like be-, ver-, and ont- determine the category of the complex and consequently, (18) is a verb and not a noun. To put it differently, although we may attach the zero-affix as indicated in (18), this will not help us in converting the verb to a noun, because the attachment of the prefix will render the structure verbal.

41 4. Conclusion In this paper w e have tried to answer several questions concerning the prefixed nouns in (1). First, w e have argued, following D o n (1990a) that nouns o f the type het verraad are derived with the prefix ge-, which is deleted by an independently motivated phonological rule of gedeletion. The non-existence of nouns of the type de verraad

turns out to result from a

phonological restriction on the attachment of the noun-forming zero-affix. This phonological condition is not contradicted by nouns of the type de aanloop.

W e have shown that Williams'

(1981) notion of "lexically related" can cope with these "bracketing paradoxes". This analysis receives independent motivation from the non-existence of nouns of the type "de

inburger.

Notes * The author is supported by the Stichting Taalwetenschap (project nr. 300-164-013) which is funded by The Dutch Organization of Scientific Research. He wishes to thank Lisan van Beurden, Wim de Haas, Eddy Ruys, Mieke Trommelen and Wim Zonneveld for comments and discussion. 1 The term monosyllabic is used here in a theoretical sense and not just observationally. In Kager & Zonneveld (1986) it is argued that bisyllabic forms with a schwa in the rightmost syllable are to be considered as monosyllables phonologically. We adopt their view here and use the term 'monosyllabic' in this theoretical sense. ^ The nouns in (8b) are ungrammatical only as nominalizations of the corresponding 'separable' verbs; they can be formed as straightforward compound nouns (cf. (lc)), but without the semantics of the verb. ^ Notice that this conclusion also holds if we assume that the nouns or adjectives in (8b) have been converted to verbs before the formation of the separable compounds. 4 Limited space prevents us from extensively justifying this choice here. It turns out that an analysis along the lines of Pesetsky (198S) runs into serious problems. Without amendment, Pesetsky's proposal cannot make a distinction between structures like de [ver[[raoJx> in which the prefix functions as a head, and de [af[[loop]\0]v]i, in which af is not a head. If we extend Pesetsky's proposal in such a way that it can deal with this distinction, we think the result is very similar to Williams' relatedness. Possibly an analysis similar to the one to be proposed here is possible in Huybregts' (1986) colinearityframework. We leave this matter for further research.

References ALLEN, M.R. 1978 Morphological

Investigations,

diss. Univ. of Connecticut.

B E U R D E N , L. V A N 1987 "Playing level with Dutch Morphology", in: F. Beukema & P. Coopmans (eds.) Linguistics in the Netherlands 1987, Foris, Dordrecht, pp. 21-30. B O O U , G. E. 1977 Dutch Morphology: A Study of Word Formation in Generative Grammar, Dordrecht, Foris.

42 DON, J. 1990a "Ge-deletion in Dutch Nouns", to appear in P. Coopmans et al. (eds.) OTS-Yearbook 1990. 1990b "Tegen "against any directional rule of conversion"", to appear in De Nieuwe Taalgids. HUYBREGTS, M.A.C. 1986 "Colinear Structures in Morphology", paper presented at the 'Morphology & Modularity-conference, June 1986, Utrecht. KAGER, R. & W. ZONNEVELD 1986 "Schwa, Syllables and Extrametricality in Dutch", in: The Linguistic Review 5, pp.197221.

LIEBER, R. 1980 On the Organization of the Lexicon, diss. MIT. PESETSKY, D. 1985 "Morphology and Logical Form", Linguistic Inquiry 16-2, pp. 193-246. SCHULTINK, H. 1973 "Het prefix ge- in Nederlandse (en Duitse) verleden deelwoorden", De Nieuwe Taalgids 66, pp. 409-418. SIEGEL, D. 1978 "The Adjacency Constraint and the Theory of Morphology" in: NELS VIÜ, Amherst, Mass., pp. 189-197. TROMMELEN, M. 1989 "Lettergreepstructuur en Woordcategorie", De Nieuwe Taalgids 82, pp.64-77. TROMMELEN, M. & W. ZONNEVELD 1986 "Dutch Morphology: Evidence for the Right-hand Head Rule", Linguistic Inquiry 17, pp. 147-169. 1989 Klemtoon en Metrische Fonologie, Coutinho, Muiderberg. WILLIAMS, 1981 "On the E. Notions 'Lexically Related' and 'Head of A word'", Linguistic Inquiry 12, pp. 245-274.

Dicky Gilbers

On Predicting the Relative Markedness of Vowels 0. Introduction* This paper investigates the consequences of the DIGPHON (-Digital Phonology) model for the theory of markedness. This model attempts to provide a highly constrained theory of intrasegmental structure in a way that is phonologically adequate and phonetically realistic. DIGPHON makes heavy use of notions from electronic circuits to model those aspects of the phonological competence of speakers that are nowadays often captured in terms of trees (so-called 'feature geometry', cf. e.g. Clements 1985, McCarthy 1988) or dependency relations defined on atomic units (cf. e.g. Anderson & Ewen 1987, Van der Hulst 1989). The DIGPHON system has two main advantages over these alternatives: (1) it is more restrictive (see Gilbers 1989b), and (2) it offers a way of representing segments which might mirror more closely the neural circuitry responsible for production of speech. The latter, more speculative point, is interesting from the perspective of e.g. Halle (1983), who suggests that "the distinctive features correspond to controls in the central nervous system which are connected in specific ways to the human motor and auditory systems" (op. cit. , p.95). It should be emphasized, however, that the model so far was based primarily on standard phonological data and argumentation and has been tested primarily in traditional phonological tasks such as accounting for vowel inventories of natural languages, phonological processes such as lowering rules, and phonotactic regularities. In this paper, we will explore how the model accounts for observations on relative markedness of vowels. The notion of markedness has played a role in various linguistic theories, beginning with the important work of the Prague Structuralists, and persisting in modern generative phonology (cf. e.g. Kean 1975). This notion has not always been defined in the same way (cf. Basb^ll 1981) and this poses a problem for the construction of a markedness hierarchy of vowels. Often, markedness is connected with the frequency of a particular vowel in vowel systems. For example, according to the sample of languages in Maddieson (1984) /i/ has a greater frequency of occurrence than /e/; so we could qualify /i/ as less marked than /e/. However, one could as well relate markedness to the frequency of occurrence of a vowel in a text of a particular language, or to the phonotactic behavior, 1.e. the distributional capacities, of a vowel In that language. Clements (1989) claims that a theory of phonological features must be able to distinguish between more- and less-favored vowel inventories. But the great phonological interest in first language acquisition also requests that such a theory must account for the order in which vowels are acquired by a language learning child. Not surprisingly, these different definitions of the notion 'markedness' accordingly may give rise to distinct vowel hierarchies. So, even if we base our markedness hierarchy on one criterion, we have to be cautious in proposing a vowel hierarchy of markedness values. Nevertheless, this is exactly what we would like to do in this paper. It turns out that the DIGPHON model provides a natural account of markedness in terms of articulatory complexity. He will propose a markedness hierarchy of vowels, based on a DIGPHON model of the universal vowel. In section 1 an outline of the model will be presented. Furthermore, we will deal with the link between favored vowel inventories and the relative simplicity of our model in this section. Section 2

44 shows a markedness hierarchy which is related to the degrees of complexity of the vowel derivation in our model. Finally, we will propose a measuring unit for the relative markedness values of the vowels in section 3. 1. The DIGPHON universal vowel model In DIGPHON vowels are represented by 0-1 codes. These codes are not arbitrarily chosen, but they represent a certain configuration of modules, the organization of articulatory features, in a universal model of the vowel. In the model we distinguish carrier modules and modulator modules. Three carriers, having the value 1 or 0, respectively standing for 'activated' and 'not activated', constitute a unit that makes it possible to subdivide the vowel triangle into domains of articulation. Because it is impossible to articulate a vowel simultaneously 'front* and 'back', the theoretically possible domains 101 and 111 are not incorporated. (1) a. unit of carriers:

b. domains: 100

/i/

001

no

000 /•/ 010 /«/

on /»/

M

In each domain we want to generate several vowels. For instance, in domain 100 we would not only like to derive /i/, but also /i/. The places of articulation in (lb) can be modified by modulation. In this particular case of /i/ we use a modulator [lowered] which corresponds acoustically with increasing the value of the first formant and articulatorily with a more open lower jaw. (2) F1 modulation:

In a previous version of DIGPHON (Gilbers 1989a) we proposed a modulator [tense], but in the present model we would like to connect the notion of 'markedness' with the structure of the model. Generally speaking, modulated vowels are more marked than non-modulated vowels, /i/ occurs in fewer vowel systems than /!/, so /i/ is more marked than /i/. The modulator [round], which we need to derive /y/ in domain 100, forms an exception to this general markedness principle. For rounded front vowels are marked, but rounded back vowels are unmarked. So in the case of the modulator [round] we cannot maintain the idea that modulated vowels are more marked than un-modulated vowels. However, as distinct from modulators, like [nasal], [lowered] or [breathy voiced], [round] is an underspecified feature in our model. The value of [round] depends on the 3-bit carrier code. In the unmarked case this modulator is automatically activated in the domains 001 and 011, and not activated in the other domains. To express the passive, underspecified character of [round], the module is connected with the unit of carriers in our model supplied with a polarity switch. The model then will be provided with a system of connections in such a way that [round] is automatically activated in the domains 001 and 011 and disconnected from the input in the other domains if the polarity switch is in position A in (3a). If this switch is in position B

45 vowels with the reverse values of [round] will be derived in the domains. Negative polarity of roundness so to speak mirrors the value of roundness in an imaginary vertical axis in the vowel triangle (3b). b. Round polarity:

(3) a. R-reverse:

|A I» +0 o-

H



output

Another passive, underspecified feature in our model is the value of [high/low]. The value of [high/low] also depends on the carrier code. As we have seen in (lb), the possible carrier codes give shape to the vowel triangle in such a way that 100 and 001 imply [high], a closed lower jaw, and 010 implies [low], an open lower jaw. The combinations 110 and 011 correspond with [mid], a half open lower jaw. With these features we can fill the less marked positions in the vowel triangle. However, we would like to propose a phonological feature system that can generate all vowels of any vowel system. Some vowel systems contain vowels that are articulated in the complementary positions of the domains that we can derive so far. In English, for instance, a central, high vowel occurs: /i/; the. vowel system of Russian includes a fronted, low vowel: /g/; and Hungarian holds a back, low vowel /d/. Again, we can make use of a polarity switch to obtain these more marked vowels. As negative polarity of roundness mirrored the vowel triangle in an imaginary vertical axis, negative polarity of [high/low] mirrors the vowel triangle in an imaginary horizontal axis, the e-o axis (4). In the case of negative polarity all height values will receive the complementary value: [high] will be [low]; [low] will be [high]; [lowered] will be [raised] and [raised] will be [lowered]. b. High/low polarity:

(4) a. HL-reverse: a| >| +o o-

C| D| +0 o-

H

H

output

direct contact ' indirect contact > no contact

H +o

oA LîL

/ o o0 O I O O

1 - diode (one-directionality)

/

o o X Y binary awitch

C o

o

o

D

1 E

F Cl

e •o o

0 0 F| C o /: o o H| Ij J

/

o o o X Y Z ternary evltch

o o o o X Y X Y double-poled awitch

e

H

+o oX Y polarity switch

|cal«»d I pl front l|^Hlcencr*llHBMl back I I Hlrev l - T * — p 1 I—' j—'

I I /- o o o H l| J

0 o F| G o o o H| I J

I I o/• o o Hj^lj J

47 The model enables us to derive all vowels according to Maddieson (1984, p.204) (7). It covers all possible supralaryngeal variation. Some of Haddieson's symbols have had to be modified for practical reasons. (7) Segment Index Maddieson (1984, p.204): Central

Front l-R) (+R| High Lowered High

1 1

Higher mid Hid Lower mid

4

y

2

Y

I

Back [+R| [-R|

I-RI I+Rl 1 I

u H

3

u 0

ut Ul -r

domain 100 HLt domain 010 HL- domain 001 HL+

Raised Low Low

i e e

£ e *

5 i 9 3

4

0 *

domain 110

domain 000

7

8

«

*

i

*

a

a

*

*

6

£ o 3

jr V A

domain 011 9

£ D

J a

domain 100 HL- domain 010 HL+ domain 001 HL-

Model (6) may seem to be complicated, but one has to realize that this model stands for a non-actual vowel system, i.e. for the abstract universal vowel system. No language actually exploits all the theoretical options. For the derivation of the vowels of the most common systems in the world we only make use of the carriers. According to Ruhlen (1975), which is based on a sample of 693 languages, the most common vowel pattern contains the vowels /i,eta,o,u/. 27Z of the languages in his sample exhibit this pattern (8). Furthermore, IX of the languages add shwa to these five vowels (8 minus the connection between A and [central]). So for the description of the vowel systems of more than a third part of all natural languages, we only have to use the carriers of our model and the modulator [round] in its underspecified positive polar position. (8) Most common vowel system:

oucpuc Next in the scale of unmarked vowel systems is the seven vowel system (e.g. the Italian system). 6% of the sample add /e/ and /o/ to the five vowels of the system in (8). So we have to add the modulator [lowered] to the model in (8) in order to accomplish the derivation of this seven vowel pattern. With this arrangement of the model we predict the relative markedness value (from now on: M-value) of vowel systems. In the next section we will see in what way the relative M-value of each vowel in isolation is predictable from model (6).

48 2. A markedness hierarchy of vowels In DIGPHON a markedness hierarchy of vowels can be derived from Che universal vowel model based on two principles: (9)

1. Complexity Principle 2. Complement Principle

The Complexity Principle states that the M-value of a particular vowel is a function of the complexity of the connections needed to derive that vowel. The Complement Principle is an elaboration of the notion of 'polarity' as described above. It states a certain tendency to balance: after every pole the antipole follows. It can be seen as a sort of Yin and Yang effect in phonology. For instance, if a front vowel has been derived, there will be a strong tendency to derive a back vowel. Both these principles can strengthen each other, but in case of contradiction the Complement Principle overrules the Complexity Principle. From these two basic principles we can derive a number of related principles: (10) 1. The H-value of a vowel Is In proportion to the nunber of carriers activated for the derivation of that vowel. 2. Symmetrical carrier codes correspond with a lower Itvalue. 3. Left perlpherallty Is less marked than rlghtparlpherallty. 4. Modulation Increases the M-value. 5. Modulated modulation Increases the H-value which Is the result of modulation. 6. Modulation in one-actlvated-carrler*domalns corresponds with a higher M-value than modulation In two-actlvatedcarrler-domalns. (Place modulation balance (PMB)) 7. Independent Input-module connections cause a lower Mvalue than dependent Input-module connections. 6. Negative polarity Is more marked than positive polarity. 9. HL-reverse corresponds with higher M-value than Rreverse. 10. PMB-adjustment leads to a lower M-value than negative polarity.

(10.1) is directly derived from the Complexity Principle. But the question arises whether there is a difference in M-value between /a/, /i/ and /u/ or not. Each of these three vowels is represented by a code that consists of one activated carrier, respectively 010, 100 and 001. Nevertheless, these vowels can be distinct in terms of derivational complexity. We have to use a simple binary switch, F/G in (6), in order to activate the carrier [central], and a ternary switch, H/I/J in (6), in order to attach the value 1 to [front] or [back], the latter switch being more complex than the former. So, we may conclude that the derivation of /a/ is a little less complex than the derivation of /i/ or /u/ (10.2). A slight difference between /i/ and /u/ may be found in the automatical activation of [round] if [back] is activated (10.3). It will be clear that modulation complicates the derivation and thereby increases the M-value (10.4). In the present model modulated modulation is only involved if [feedback on raised] is activated, but when the model will be expanded with laryngeal features and the soft palate feature [nasal] in forthcoming papers, principle (10.5) will play a more important role in defining the M-value of a generated vowel. An example of the Complement Principle overruling the Complexity Principle can be found in (10.6): place modulation balance (PMB). In case of modulation the carrier codes 110 and 011 lead to less marked vowels than the carrier codes 010; 100 and 001. We can motivate the influence of the Complement Principle as

49 follows. After we have taken up the angular points of the triangle from a physical point of view, i.e. after we have utilized the extreme possibilities of articulation in unmodulated domains, we are looking for the largest space to vary articulation in the case of modulated domains. (11)

i •

•« u t e I

t •« o 4 t a

(10.7) concerns different forms of place modulation (PM). To activate [raised] in (6) we have to follow a path DH from the input to the module. H is necessary, but H is depending on D. So, in order to attach the value 1 to [raised] we have to make use of a dependent switch. The opposite holds if we want to activate [lowered]; just an independent switch in position E in (6) will do. So, the generation of [raised] vowels such as /a/ is more complicated than the generation of [lowered] vowels such as /e/ and /o/. He make use of push/polarity switches A/B and M/N in (6) to express the lower Mvalue of positive polarity (10.8). HL-reverse concerns a more complicated change of derivation than R-reverse (10.10). In the case of R-reverse, there is only an alteration at the end of the path of derivation, while HL-reverse concerns a change at the input of the model, it establishes a totally different path of derivation as compared to the derivation path in the case of [high/low] positive polar. Compared to negative polarity PMB-adjustment is the least complicated, no mirroring takes place and the value of modulation is not altered. Within HLpolarity we can change R-polarity, and within R-polarity we can adjust the PMB. Thus, we can predict a hierarchy of increasing M-values, according to the Complement Principle: (12)

a. b. c. d. e. f. g. h. 1. J. k. 1. m. n. o. p.

llL+/R+/no modulation: HL+/R+/no modulation: llL.+/R+/noclulatlon: I1L+/R+/modulation: IIL+/R-/no modulation: 1IL+/R-/"0 modulation: HL+/R-/nodulatlon: HL+/R-/modulation: HL-/R+/no modulation: HL-/R+/no modulation: llL-/R+/modulatlon: llL-/R+/modulatIon: HL-/R-/no modulation: IIL-/R-/no modulation: IIL-/R-/modulatlon: HL-/R-/modulatlon:

1 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1

carrier activated carriers activated carriers activated carrier activated carrier activated carriers activated carriers activated carrier activated carrier activated carriers activated carrlera activated carrier activated carrier activated carriers activated carriers activated carrier activated

(PMB-adjustment) (PHB-adjustment)

(PHB-adjustment) (PHB-adjustment)

(PHB-adjustment) (PHB-adjustment)

(PHB-adjustment) (PMB-adJustment)

(12) corresponds with a hierarchy of preferred paths in the model (6): (13)

1. 2. 3. 4.

HA NA HB NB

(HL+/R4-) (1I1.+/R-) (HL-/R+) (IIL-/R-)

is less marked tlian Is less marked than Is less marked than

The following preference rule hierarchy holds for each of these four polarity parts:

50 (14)

al. bl. 2. 3. cl. 2. dl. 2. el*. 2. 3. £1. Bl.

This

hierarchy

increasing (15)

FIIC CI1C FIC FJC GC JCC ICE JCE CHE FIE FJE GDI II. GDHK of

M-value a.

(carriers o f f ) (1 carrier on «central O , no FH) (1 carrier on , no PM) (1 carrier on , no PH) (2 carriers o n , no PH) (2 carriers o n , no F H ) (2 carriers o n , lndep. FH) (2 carriers o n , lndep. PH) (*: there Is no lndep. PH on < C > ) (1 carrier on , lndep. PH) (1 carrier o n , lndep. PH) (1 carrier on , dep. FH) (1 carrier o n , dep. PH + fb)

preferred as

given

paths in

in

the

model

Is Is Is Is Is Is Is Is Is Is Is Is

less less less less less less less less less less less less

stands

narked narked narked narked narked narked narked narked narked narked narked narked

for

the

than than than than than than than than than than than Chan

vowel

sequence

(15).

HL+/R+

b.

HL+/R-

pref. | code;

rule: fl

pach:

vowel:

1 FCB lrf RHL B dom.mod.pol.

010.Oil.00 J

c.

|[C.DK.Ma| /3/

j

HL-/R+

pref. rule:

code:

2.al

j 000.000.10

j HF.C.NA

/0/

2.bl 2.b2 2.b3

I accld. gap 1 100.000.10 I 001.000.00

1 HC.C.NA I IF.C.NA 1 JF.C.NA

111 /«•/

2.cl 2.c2

I 110.000.10 1 Oil.000.00

| 1C.C.NA I JC.C.NA

l»l /»/

2.dl 2.d2

I 110.100.10 1 011.100.00

I IC.E.NA j JC.E.NA

/«/ /A/

2.

o I I u/ - u

Some words change all their as when a morpheme - u (for the oblique cases of feminine singular weak nouns), -ur (for nominative and accusative feminine plural weak nouns), -um (for dative plural nouns and third person plural verbs), or Sum (for third person plural past tense of weak verbs) is added, other words change one a only. This can be explained if we recognize that u-umlaut is a rule which takes into scope the adjacent syllable. In kalla+Sum the syllable containing u is adjacent to a weak syllable with a which consequently changes to u. This affected syllable follows a strong syllable with a, thus causing a change from a to o as follows: (20)

F /\ o o„ kalla + Sum

—>

Fs F /\ I o oH I kolluSum

In almanak and apparat only one a is transformed in the dative plural case, viz. the one immediately preceding the suffix um. If we assume bounded feet it follows that this a is in a strong position within the foot. In accordance with (19a) it thus changes to 8. As there is no rule which weakens 6 to u in a strong syllable, nor a rule which causes a change of a

59 Note that the parameters for quantity sensitivity, for bounded feet and left dominancy are the same as for Dutch (cf. Trommelen & Zonneveld 1989). 3. Stress Patterns of Icelandic

Derivations

With respect to stress assignment two kinds of suffixes can be distinguished in Icelandic: 1. suffixes that attract secondary stress unto themselves: •auming.ja+ legur

' aumingj a.legur 'miserable looking -sam(ur) 'miskun.nar+ samur •miskunnar.samur 'merciful' -full(ur) 'miskun.nar+ fullur •miskunnar. fullur 'merciful' -heit 'merki.legur+ heit 'merkileg.heit 'arrogance 1 -d6m(ur) 'keisa.ri + d6mur 'keisara.d6mur 'emperorship' -ist(i) 'kontra.bass + isti ' kontrabass.isti 'bassplayer' Arnason mentions in addition to these -leikur, -skapur, anskur, -latur, -iskur. -leg(ur)

2. suffixes that do not affect the stress pattern of the stem: -ugur -6ttur -ingur -andi -ari

'kiia.skit + ugur

' ktia.skitu.gur 'covered with cowshit' 'dSkkblS.rSnd + 6ttur ' d6kkbl6.r5nd6t.tur 'darkblue striped' 'Akra,nes + ingur ' Akur.nesing.ur 'inhabitant of Akranes 'skipu,leggja + andi ' skipu.leggj an.di 'organiser 1 'undir,rita + ari 'undir.ritar.i 'one who signs'

With Arnason (1985a:107) I assume that suffixes that do not affect the stress pattern do not have a (lexical) foot of their own. Suffixes which attract secondary stress unto their initial syllable do not undergo footformation as described in section 2. Presumably they have their own foot structure. This accounts for the fact that secondary stress is assigned to these suffixes even though they are the fourth syllable of a word, which is the syllable to which stress is usually not assigned.

60 I like to add the suffix -ani to the list of suffixes that attract secondary stress. First of all, the fact that it is a lexical foot with a strong and a weak sister explains that Amerikani has secondary stress on the fourth syllable: /'a: me ri .ka nl/. Second, it explains that in this suffix a turns into 5 in dative plural of words like Japani/ JapBnum. 4.

Stress Patterns of Icelandic Compounds

What remains to be explained are the forms below which are apparent violations of the claims that (a) Icelandic is a quantity sensitive language, (b) that u-umlaut changes a to 5 in stressed syllables only, and (c) that Icelandic has an alternating stress pattern: (24al) 'ungbarn.ifi (from ' ung 1 + 'barnifi') 'the infant' (24a2) 'framburfi.ur (from 'bera fram') 'pronunciation' (24bl) 'stresstSsk.u (from 'stress' + 'tosku 1 ) •a businessman's briefcase' (24cl) 'platalma.nak (from 'plat' + 'almanak')'fake almanak' (24c2) 'vinrabbar.bari (from 'vin' + 'rabbarbari 1 ) 'wine rhubarb' These examples have in common that they are compounds with a monosyllabic first component. To account for the stress behaviour of these compounds I assume that "defooting" (cf. Hayes 1982) applies in Icelandic as follows: (25)

F

—>

0/ F. -T® i o

i.e. a foot is "defooted" after a monosyllabic strong foot. The effect is a stress shift rightward under the influence of a stronger stress on the left. Note that "defooting" applies to the metrical trees in (23). This explains why secondary stress is assigned to the third syllable of akSrnum, apSldrum and Japonum. It also accounts for the fact that 'bqrn', 'bur6' and 'tfisk' lose their strength in compounds. They are adjoined to the preceding foot by stray adjunction: F / \ °s ung + barni6 stress + tSsku fram + burSur Ja + p8 num

F i

(26) F

F

;

—> —> —> —>

ung barn i6 stress tfisk u fram bur6 ur Ja p8 num

— — — —

> > > >

F

S H ung barn i6 stress t8sk u fram bur6 ur Ja p5 num

61

In order to account for the presence of o as a result of u umlaut in the second (weak) syllable of (24b), we have to assume that this rule applies at an earlier level. This is supported by Arnason's (1988:4) observation that u-umlaut does not occur at the level of compounding. Consider, for instance, the absence of o in the first syllable of (24a2). According to Arnason (1985a:94), the second members of compounds with a monosyllabic first component can be liable to reduction; stresstaska and framburSur may be pronounced as [strestska] and [frambrSYr]. If this is correct it can be put forward as evidence for "defooting" and subsequent "stray adjunction". As a final example let us consider foot formation, "defooting" and "stray adjunction" for the compounds in (24c): (27) P / \ s w plat + almanak

I i --> plat almanak

—>

(28)

A A" F

vin

F /\ s w s w s w rabbarbari --> vinrabbarbari

F P ¡\ " s \ / \ \ s w w platalmanak

\

F \ A / \ \ s w w s w --> vinrabbarbari s

The above derivations correctly predict secondary stress on the fourth syllables of 'platalmanak 1 and 'vinrabbarbari'. 5. Conclusion In the present paper word stress phenomena are examined for Icelandic. On the basis of the fact that Icelandic clearly displays an alternating stress pattern, I have argued that it has left dominant bounded feet assigned from left to right. This analysis is supported by the phenomena related to the u umlaut rule. The exceptions to the alternating pattern, which induced Arnason (1985a, 1988) to assume unbounded feet, can be explained if we follow Hayes (1982) and assume a rule of "defooting" and the process of "stray syllable adjunction". Note *

I thank Kristj£n Arnason, Inge Marsman, Wim Zonneveld and the editors of this volume for helpful comments.

62 References ANDERSON, STEPHEN R. 1972 Icelandic u-Umlaut and Breaking in a Generative Grammar, in E.S. Firchow et al. (eds) Studies for Einar Haugen, 13-30 1976 On the Conditioning of Icelandic u-Umlaut, in Language Sciences, 26-27 ANDERSON, STEPHEN R. AND GREGORY K. IVERSON 1976 Icelandic u-Umlaut: An Exchange of Views, in Language Sciences, 28-34 ANDERSON, STEPHEN R. 1984 A Metrical Interpretation of Some Traditional Claims about Quantity and Stress, in M. Aronoff and R.T. Oehrle (eds) Language Sound Structure, Cambridge: The MIT Press, 83-106 ARNASON, KRISTJAN 1985a Icelandic Word Stress and Metrical Phonology, in Studia Linguistica 39:93-129 1985b Morphology, Phonology and u-Umlaut in Modern Icelandic, in Edmund Gussmann (ed) Phono Morphology: Studies in the Interactions of Phonology and Morphology, 10-22. 1988 Problems in the Lexical Phonology of Icelandic (Ms) HAYES, BRUCE 1981 A Metrical Phonology of Stress Rules, Doctoral Dissertation, MIT, Cambridge (Ms) 1982 Extrametricality and English Stress, in Linguistic Inquiry 13, 227-276 KIPARSKY, PAUL 1984 On the Lexical Phonology of Icelandic, in Claes-Christian Elert et al. (eds) Nordic Prosody III, 135-164 RÖGNVALDSSON, EIRIKUR 1981 U-hljó6varp og önnur a-ö vixl in nütimaislensku, in Islenskt mài 3, 26-58 SVAVARSDOTTIR, 'ASTA & MARGRET JONSDOTTIR 1988 Islenska fyrir Utlendinga, Reykjavik: Malvisindastofun Hàskóla Islands TROMMELEN, MIEKE S. WIM ZONNEVELD 1989 Klemtoon en Metrische Fonologie, Muiderberg:Coutinho

Marjan Grootveld

Marcus-Parsing and Coordination: A Mismatch 1.

Introduction*

This paper investigates the appropriateness of a class of deterministic parsers, the Marcus-type parsers, for analysing coordination phenomena. After an introduction to determinism and Marcus-parsing in general, implementations are described which pay attention to coordinate structures. The possibilities of handling coordination of complete constituents, backward conjunction reduction and gapping will be discussed in sections 3,4 and 5, respectively. None of the proposals is able to cover the whole range of coordinate structures and therefore their most promising parts are combined in section 6 in order to try to build the ultimate parser, but the desired analyses do not follow. It should be pointed out that none of the accounts considered here pays attention to coordinate constructions consisting of more than two conjuncts. Furthermore, and is the only conjunction starring in examples. It is the prototypical coordinate conjunction, but grammar rules written to handle and are not to be applied blindly to any conjunction, as for instance the distribution of but is more limited. The conclusion in section 7 will be that, even in this restricted empirical domain, coordination inevitably violates Marcus-type principles. To put it differently: the Marcus-parser falls short.

2.

Determinism and Marcus-parsing

It has often been observed that processing natural language sentences by humans proceeds in a fast and efficient way. Most of the time, people are not aware of local ambiguities, numerous though they are: words that belong to different parts of speech, substrings of the sentence that can be assigned more than one analysis etc. Cases in which normal sentence processing is blocked and a semi-conscious error recovery component takes over are exceptional (cf. Milne 1983). These observations have been a challenge to computational linguists, since large grammars tend to slow down the parser. Moreover, most parsers produce several analyses for one input sentence and only by ad hoc regulations can the quantity of the output be reduced. For these reasons, Mitchell Marcus (1980) advocates a deterministic approach to parsing natural language. When local ambiguity confronts a non-deterministic parser, it can try all alternative analyses either in parallel (branches leading to failure eventually die out) or one after the other (decisions that turn out to be wrong are revoked, a method called "backtracking"). In contrast, a parser is called deterministic if it can neither store alternative structures, nor revoke decisions. The absence of backtracking and parallelism in human sentence processing leads Marcus (1980) to implementing a computer model of a deterministic parser, that is, a system building the desired structure without either parallelism or backtracking. Many Marcus-type parsers have been built since. The reader is referred to Milne (1983), Baart & Raaijmakers (1988) and works cited there. The architecture of a Marcus-parser is conspicuous, for there are only two data structures: an Active Node Stack of incomplete constituents, looking for daughters, and a buffer of complete constituents, looking for mothers. Elements in the buffer are either preterminal categories - their order reflecting the order of the words in the sentence to be parsed - or higher categories, in which case they have been "popped" from the stack. The stack grows and shrinks in a strictly last-in first-out fashion. A constituent is pushed onto the stack at its creation and is pushed off it and prefixed to the buffer as soon as it is completed. The element on top of the stack is called the "Current Active Node" (CAN) and this node is part of the limited context the parser can base its actions on. The other accessible constituent in the stack is the most recently created S or NP node, called "Current Cyclic" (CC). In the buffer a limited number of cells is accessible: the parameter "Width of look-ahead window" is set to three, since this setting seems necessary and sufficient (cf. Marcus 1980 and, for a different claim,

64 Milne 1983). The pattern side of each grammar rule must match the contents of the look-ahead, whereas the action side consists of a list of operations. Efficiency is increased by clustering grammar rules according to their use: the grammar consists of rule packets for parsing NP's, VP's, inversion etc. Creation of a new constituent and attachment of the leftmost buffer cell as the rightmost daughter to the CAN are the primitive operations we will need in the remainder of this paper. It is emphasized that when a node has been attached, it cannot receive any daughter node. Determinism is guaranteed by the requirement that all nodes, labels and attachments be permanent and be part of the output structure. With these characteristics of Marcus-parsing in mind, consider how this design might work out for coordination.

3.

Coordination of complete constituents

The most detailed description of a Marcus-implementation that is claimed to handle coordination is Snarr's (1984). She bases her model upon coordination without conjunction reduction: Rick likes Bill and Sue. At first sight, a look-ahead pattern like (1)

[ = ] [ = "and"] [ = ]

and an action specifying conjunction of the first and third node in the buffer is appropriate. Snarr notes, however, that it is possible for two conjoined constituents to contain subconstituents that satisfy the pattern, e.g. Rick likes Bill and Sup. liltp-< Kim The NP nodes dominating Bill and Sue match the first and third cell of the pattern, respectively, eventually causing the parser to fail in recognizing likes Kim. Preventing this by ordering the conjunction rule after the object rule does not solve the problem, for in a coordination of objects, the first NP would be attached as an object, leaving the conjunction and the second NP stranded in the buffer. Instead of writing several ad hoc rules. Snarr introduces four rules, cooperating to analyse most of the coordination cases. Proceeding from left to right, the parser has not analysed the second conjunct when a conjunction appears in the buffer window, so it cannot determine which constituent preceding the conjunction is the left conjunct (except for cases like A and B buy C. where the node dominating A is the only possible left conjunct). Consider the phase at which the material to the left of T in the example has been processed (the tree structure conforms to Snarr 1984:38): (2)

Rick likes Bill! and Sue likes Kim

(3)

St /

\ NPj VPr I / \ I vx np 2 I I I Rick likes Bill Regardless of what is in the buffer, it is possible to define a subset of the current set of nodes in the stack that may be the left conjunct, because the left conjunct must contain the word which immediately precedes the conjunction. For a given conjunction, Snarr forms the set of its RightMost nodes (i.e. nodes which do not have sisters to their right): the RM set. In structure (3) the RM set associated with the conjunction in the first buffer cell consists of NP2, VP^ and S^. Finding the elements of the set is easy: since in Snarr's opinion completed nodes are always dropped into the first buffer cell, RM nodes which have not yet been labelled "RM node" will be recognized and labelled as soon as they appear in the look-ahead. Snarr assumes that the right conjunct must be able to syntactically replace the left conjunct. Therefore, while noting the RM nodes, the parser creates the set of their associated rule packets, in order to parse any constituent which could replace an RM node. The structure up to the conjunction having been created, and is in the first buffer cell and triggers the creation of a CONJP node, and is attached and the set of replacement packets is activated. By removing and from the buffer, the parser can use all three look-ahead cells. This is necessary, because patterns which succeeded during the analysis of the left conjunct, need to succeed in the subsequent conjunct as well.

65 Having parsed a right conjunct, the parser attaches it to CONJP as rightmost daughter. A new coordination rule compares the potential right conjunct for predicate-argument compatibility to the RM nodes of the conjunction. When the left conjunct has been identified, CONJP is dropped into the buffer, the left conjunct is detached from its dominating node and dropped also. In parsing Rick likes Bill and hates Sue, the current buffer state would be [ = < vp, {v,np} > ] [ = < conjp, {"and",vp} > ] [ = < t > ], where the immediate daughters are between braces. t.is a variable that matches everything, as for the time being the contents of the third cell are irrelevant. At this moment a node is created whose type is the same as that of the left conjunct (in this case a VP). The left conjunct and CONJP are attached to it, after which the new node is dropped and attached in turn to the node previously dominating the first conjunct. (4) NP, I I I

/

S1

\

VP Q

VP /

Vt

Rick likes

\ NP 2

CONJP \ I I

Bill and hates Sue

Having given an impression of Snarr's analysis it is time to assess her account. The original architecture of the Marcus-parser has received some extensions: the power to collect RM nodes and to activate rule packets for parsing the second conjunct, a procedure for checking compatibility of conjuncts, and a device detaching structure. Of course, grammar rules invoking these devices will extend the grammar. Evidently, adding rules and the notion of RM nodes is a natural course in the act of enabling the parser to handle coordination. Since coordination, as opposed to subordination, is often characterized by parallelism, the use of "old" replacement packets is a nice implementation of the expectations a conjunction gives rise to. Snarr adds to this the requirement that where rules in the set of replacement packets compete (having the same pattern and the same priority in their respective packets), the rule in the packet associated with the lowest RM node should be favoured. Research into memory limitations in natural language parsing supports this choice (cf. the discussion in Schubert 1986). All these changes seem to be natural extensions, except for detach. This is the most remarkable and dangerous divergence from the original system (Marcus 1980). Snarr states (the exclamation mark is mine and refers to the remark immediately following the quotation): Using this (lowest node] rule, the smallest constituent following the conjunction which can syntactically replace an RM node is attached to the CONJP node, even if it could belong to a larger constituent. If the node is determined not to be the right conjunct then the parser will need the power to detach it from the CONJP node and return it to the buffer (!) but will never have to destroy existing nodes. The parser must have the power to detach a node in any case, so that it can detach the left conjunct before performing the conjunction. It is important to note that giving the parser this power does not allow it to do general backtracking. Any node that is detached by the parser must be dropped into the buffer; thus no nodes are ever destroyed. Furthermore, any node that is detached must reappear intact somewhere else in the tree (Snarr 1984:45). If not backtracking, what is this? In order to find the left conjunct there is no need to attach the assumed right conjunct to CONJP, since the rule that checks for compatibility merely inspects the first buffer cell, whereas it is permitted to inspect three cells. We are forced to conclude that Snarr deliberately takes a chance to do things that have to be undone eventually, which runs counter to the deterministic demands stated in Marcus (1980). Claiming that the possibility to detach is needed in any case is not an impressive defence. It is common practice to limit the power of a natural language processing system as much as possible, so even if something like detach turns out to be necessary, one should try to restrict the situations in which it is used to a minimum The final two sentences of the quotation express obedience to the demand that "all syntactic substructures created by the grammar interpreter for a given input must be output as part of the

66 syntactic structure assigned to that input" (Marcus 1980:12). However, the main quality of the deterministic parser should be that "all syntactic substructures created by the grammar interpreter are permanent" (ibid). Detaching a node from its mother node (cf. V P j in the trees above) is in blatant contradiction of this requirement. Snarr increases the parser's power and she does so in a way that runs counter to at least the spirit of determinism. 4.

Backward conjunction reduction

Snarr assumes that her approach to handle full coordination can be extended to handle backward conjunction reduction (BRC). Her proposal hinges on rightmost nodes again: ^ For a constituent A to be shared by two conjoined constituents C^ and C2, there must be rightmost subnodes, N^ in C} and N2 in C2, to which A could be grammatically attached. (...) This suggests that the parser must split into two separate processes: one attempting to attach a trace of A to a rightmost node in C4 and the other attempting to attach a trace of A to a rightmost node in C2. If both processes succeed, then A is shared by C} and C2 (Snarr 1984:57). Several comments are in order. First, notice that Snarr assumes BCR to reduce one node only. In sentences like Judy gave and Bill sold fa bookl [to Marvl two constituents are left unexpressed in the first conjunct. This may be remedied by replacing "a constituent A" in the quotation with "a string A of constituents" (ignoring for a while the attachment of traces). Second, what strikes me as unattractive is the fact that not just a conjunction, but any string A triggers the collecting of RM nodes. Always creating every possible RM set is expensive and redundant, especially if the sentence turns out to contain no coordination, ellipsis or other "complicating" phenomenon. In case Snarr's model contains a licensing condition on storing information, she does not mention it. As a third point to be noted, Snarr presumes that the parser has access to the conjuncts. Again the proposal in is conflict with Marcus' design, because after labelling a node "rightmost", that node is attached and therefore closed; no daughter node can be attached to it anymore. For that reason both conjuncts are opaque for trace attachment. We cannot but conclude that Snarr's proposal is unfit, because of the rigid early attachment strategy. Church (1982) presents a treatment of backward conjunction reduction in the Marcus-type parser YAP ("Yet Another Parser"). YAP has been designed as a competence model, but unacceptable (although grammatical) input (e.g. multiple centre-embeddings) will cause the parser to fail: YAP simulates time and space limitations. To achieve the finite memory limitation, the active node stack has been provided with a "forgetting" procedure. The implemented counterpart of (mental) forgetting is called closure, which is accomplished by attachment to a higher node. BCR is handled in the following way: When there is a conjunction (e.g. andl in [the first buffer cell] and [the current active node] cannot close, then YAP assumes right node raising. (...) Having detected the deletion, YAP undoes the transformation, inserting an empty noun phrase into the buffer (Church 1982:132). Church scrupulously notes two problems. First, the inserted NP is not bound to an object in the right conjunct, which would lead YAP to accept *I took and vou went. Second, it is only of obligatory elements that the absence is detected; the parser wil not propose an object of ate. in Church's example (489) (5)

I ate ([ n p ]) and you drank everything they brought

Church does not offer a solution to either problem. However, since missing optional elements render a sentence globally ambiguous, the second one seems no real problem for a deterministic parser: for a given grammatical input one expects exactly one output, so global ambiguity should not be recognized. Nevertheless, we are facing a serious problem, which is clear when both BCR and gapping occur. In such a sentence the unexpressed constituent is obligatory. (6) (7)

Bob called Sue and Mary considered Kim a fool Bob called Sue and Mary Kim a fool

67 In these examples, Bob called Sue a fool could be the unreduced version of the first conjunct, a fool being optional in sentence (6), the parser would correctly consider the sentence to be unreduced. In (7) however, the same procedure would fail, since gapping forces a stricter kind of parallelism upon the conjuncts. Examples like these entail that one needs information about the right conjunct to satisfactorily analyse the left conjunct.^ Another and very serious drawback of this approach is the fact that inserting one empty NP might be insufficient and even wrong, since all kinds of elements can be reduced by BCR. The empty categories to be inserted before and are between brackets. (8) (9) (10) (11)

Judy gave (NP, NP) and Bill sold Mary a book Judy gave (NP, PP) and Bill sold a book to Mary Rick bought three (N') and Sue bought five books Jack told (CP) and Jane denied that Sue was pregnant

I hope to have shown that Snarr's approach to BCR fails because of the early attachment strategy. Church is equally unable to handle BCR, since his analysis is based on the wrong assumption that it is possible to predict the type of structure that has been reduced. 5.

Gapping

Church (1982) also sketches a way to treat gapped constructions. Again, the proposed tactics holds only for very simple instances of reduction. When the buffer contains a conjunction followed by two NP's, YAP inserts a copy of the antecedent verb into the buffer. The advantage of this proposal is the availability of subcategorization information while parsing the second conjunct. By copying, so to speak, the antecedent verb, sentences like (12)

*Judy met Bob and Kim Bill a record

can be rejected at the beginning of a record, which I take to be psychologically most plausible and therefore a requirement of a parsing strategy claiming psychological reality. A strong objection concerns the disregard of other kinds of remnants, e.g.: (13) (14)

The boy talked to his boss and [fvjpthe girl] [ppto her supervisor] At our house, we played poker and [ppat her house], [^pbridge]

A comparison of these remnants and their correspondents in the antecedent clause is necessary in order to prevent copying the verbs in sentences like (15) (16)

The boy talked to his boss and [Npthe girl] [ppwith black hair] was interested in it At our house, we played poker and [ppafter the match] [j^pthe bridge] collapsed

Moreover, sentences with three remnants exist as well: Bob met his mother on Sunday and Bill his sister on Monday. A non-parallel third remnant (reluctantly instead of on Mondavi renders the sentenceungrammatical. Since the look-ahead window has no room for the third remnant, it is unclear how Church will prevent YAP from copying the verb if it is not by enlarging the look-ahead. Marcus (1987) presents a very sketchy analysis of gapping. Like Church, he only mentions gapping in the context of two NP's. In (17)

John bought a ball and Mary an umbrella

the connective breaks off the structure after ball. By means of a restart device (not to be discussed in this article) the rest of the sentence is parsed. By completely breaking off the structure preceding and (i.e. by closing the left conjunct) Marcus creates an ordering paradox for sentences showing both backward conjunction reduction and gapping: (18)

Hij gunt de hond en zij de kat {aan de buren / een beter thuis} He allows the dog and she the cat {to the neighbours / a better home}

where gunt is obligatorily bitransitive. To ensure that the constituent that is left unexpressed in the first conjunct is present in what follows and in order to assign it the right label (NP or PP), one has to

68 parse the right conjunct. In other words, if we close the left conjunct, which is a prerequisite in Marcus' treatment of the gapping context, and if we choose to do so by labelling the right number of empty constituents correctly, the right conjunct has to have been analysed already. The right conjunct, however, happens to be the gapping context... This paradox disappears when closure is defined as a default action: "Close, unless another command overrules this one". Situations as described above seem to give rise to "another command". Summarizing the findings of this section: gapping is known to occur in several contexts, a fact which has not been taken into account by either Church (1982) or Marcus (1987). Furthermore, Marcus* approach shows once more that early closure is the wrong strategy in parsing coordination. Church's technique of looking back to the antecedent verb will be challenged in the next section.

6.

The failure of eclecticism

Now that we have seen some separate implementations that are unable to cope with all coordination phenomena, an optimist might hope to construct a perfect parser out of them. I will outline the performance of this eclectic model as benevolently as possible and yet be forced to conclude that perfection cannot be accomplished. The eclectic model would presumably consist of the original Marcus-parser, the devices for creating an RM set and replacement packets, a perfect syntactic knowledge that is reflected in the grammar rules, a procedure for looking back to the earlier occurrence of a gapped verb and, as a means of passing information about reduction sites, a register.3 This mechanism will improve the parsing of BCR, since the BCR context information that is stored in the register will guide the parser in the right conjunct. (See Grootveld (1988) for a description of a data percolating technique.) An instance of BCR that the assumed revised Marcus-parser cannot satisfactorily deal with is (19)

I ate and you drank everything they brought

since Marcus-parsers, when proceeding in a "stand-alone" fashion, are essentially unable to compute more than one analysis for a given sentence. It is for the same reason that the "semi-parsability" of backward reduced sentences like these provides no argument against this model or against Marcusparsing in general, but only illustrates a characteristic that has been known from the beginning: global ambiguity is not recognized.'* Weighing whether one can or cannot accept this ought to be independent of specific linguistic phenomena, so I will end the discussion on these ambiguous BCR sentences here. Because of the fact that gapping is subject to a bounding constraint (compare (20)-(21) to (22), showing VP ellipsis, which allows unbounded distance between antecedent and empty position), Berwick & Weinberg (1984) argue that the local ambiguity that gapping may give rise to is resolved easily, since the disambiguating antecedent verb is structurally close. An example of this ambiguity is presented in sentences (23)-(24), where the parser finds an NP and a PP in the look-ahead, without clues for distinguishing between the possible PP-readings (sister of the verb and of the VP, respectively). (20) (21) (22) (23) (24)

Bob saw Bill and Kim Joe *Bob saw Bill and I don't believe Kim Joe Bob saw Bill and I don't believe Sue said ... that Kim did Kim ran after Sue and ! Joe after Bill Kim arrived after Sue and ! Joe after Bill

Linearly larger distances can be bridged as well: the (linearly!) intervening embedded clause in example (26) contains a saturated verb, which suffices to close the relative clause, yielding the same structure as in (25). (25) (26)

Joe saw Kim and ! Sue Bill Joe saw the man who met Kim and ! Sue Bill

Until now I have assumed, in conformity with the optimistic point of view, that the expanded Marcusparser can analyse full coordination, BCR and gapping. Yet, in the beginning of this section I claimed

69 that this point of view cannot be correct; the argument relies on sentences like (27), which are detrimental to the antecedent copying strategy. (27)

Kim ran after the man who gave Sue and ! Joe after the woman who gave Bill a book

Sentence (27) having been processed up to the exclamation mark, the relative clause who gave Sue is the current cyclic domain, which obviously does not contain the antecedent verb ran. Since closure of the embedded VP is not justified, closure of the relative clause is neither and so ran is invisible. According to the treatment of backward reduction we advocate (preferring late to premature closure), the embedded clause will not be completed before the parser has "scanned" the whole buffer while shifting completed constituents onto the stack. At that moment the stack will be reduced and when the VP gave Sue becomes CAN again, it will receive another daughter (in this case a copy of the NP ¿book). Next, the relative clause can be attached, by which means the matrix verb becomes accessible. Only then can the subcategorization information be copied onto an empty verb in the right conjunct and only then will sentences like * You saw the man who gave Sue and I after the woman who gave Bill a book be rejected. In such a "worst case" the assumed parser recognizes ungrammaticality later than the human parser, which is startled as soon as it notices after. Therefore, the method implemented by Church (1982) and theoretically motivated by Berwick and Weinberg (1984) - i.e. to look back at the antecedent verb - falls short when confronted with the combination of limited stack access and (possibly) multiple empty elements.

7.

Conclusion

Investigating the feasibility of a Marcus-parser that can deal with coordination, including cases with backward conjunction reduction and gapping, leads to the conclusion that none of the implementations described is appropriate and that even an eclectic approach to them will not yield the desired analyses. All shortcomings follow from and concern principles of Marcus-parsing. First, the claim that a Marcusparser simulates a human sentence parser, in that the mechanical parse is blocked if and only if a human parser detects an error, cannot be correct: a human will reject "worst case" sentences like 'You saw the man who gave Sue and I after the woman who gave Bill a book before a Marcus-parser does. Second, the "shift a lot, reduce (initially) a little" approach, although technically unproblematic, runs counter to the idea of a limited look-ahead, because in the model sketched here the shifting is not a goal in itself, but merely a consequence of the need to inspect more of the right context. Third, coordinate structures may show global ambiguity and in that quality they might reinforce the serious objection to deterministic parsers, that maximally one parse is presented. In the light of coordination facts one cannot maintain the principles of Marcus-parsing, so we have to conclude to the inappropriateness of the Marcus-parser.

Notes

*.

This research was supported by the Foundation for Linguistic Research, which is funded by the Netherlands organization for research, NWO. I wish to thank Anneke Neijt for valuable discussions and Marcel den Dikken for correcting my English.

1.

"rightmost" in the quotation is to be understood as rightmost with respect to the start of the allegedly shared constituent, not with respect to the conjunction.

2.

In case the second sentence leads the reader "down the garden path", namely, confuses the reader and forces him to backtrack, a Marcus-type parser ought to fail in parsing it (cf. Marcus 1980).

3.

The reader should notice that the depth of RM node collecting is investigated neither in Snarr's thesis, nor in this paper: the (absolute or relative) stack position of the oldest RM node (probably the deepest node in the stack and eventually the highest node in the tree) is ignored. Because the left context is potentially infinite, the quest for RM nodes is potentially infinite as well and the

70 optimist will have to solve this computational problem of intractability. 4.

However, we can think of a larger scale system, which embeds both a Marcus-parser and a higher level module. The latter, more "conscious", component might restart the parse if it presumes the possibility of an alternative structure. Of course, it will have to control the Marcus-parser in order to prevent it from applying exactly the same grammar rules as before.

References BAART, J. & S. RAAIJMAKERS 1988 'Dutch as a deterministic language*. In: Linguistics in the Netherlands 1988. Ed. by P. Coopmans & A. Hulk. Dordrecht: Foris Publications. (AVT Publications 4) 1 -10 BERWICK, R.C. & A.S. WEINBERG 1984

The grammatical basis of linguistic performance (...). Cambridge (Mass.): The MIT Press

CHURCH, K. 1982 On memory limitations in natural language parsing. Bloomington: Indiana University Linguistics Club GROOTVELD, M J . 1988 De Marcus-parser en de tweede fase: De hoofdzin. MA thesis, Leiden University: Dept. of linguistics MARCUS, M.P. 1980 A theory of syntactic recognition for natural language. Cambridge (Mass.): The MIT Press 1987 'Deterministic parsing and description theory". In: Linguistic theory and computer applications. Ed. by P. Whitelock, M.M. Wood, H.L. Somers, R. Johnson & P. Bennett. London: Academic Press. 69 112 MILNE, R.W. 1983

Resolving lexical ambiguity in a deterministic parser. Ph.D. thesis. Edinburgh

SCHUBERT, L.K. 1986

'Are there preference trade-offs in attachment decisions?'. In: AAAI-86.1,601 - 605

SNARR, V.L. 1984 Theory and parsing of the coordinate conjunction "And". Toronto: University of Toronto. (Technical report CSRI-171).

Wim de Haas

Evidence for a Noncyclic End Rule in English

1. Introduction* In his contribution to the description of English word stress, Kager (1989) showed that stressing and destressing rules are formally identical. He reanalyzed structure-changing destressing rules such as Sonorant Destressing and the Arab Rule as structure-building stress rules, and furthermore showed that the rule of Pre-Stress Destressing is redundant. Assuming that Kager's theory is basically correct, the principal goal of this paper is to show that his analysis of the data in (1) misses an important generalization: (1)

a:

c&nt£en dkyt6na b&ndlna

b:

police c: manipulate banina

sfcgm6ntal mfcntility mdrbific

d:

parental at6mic acidic

The data in (la,b) and (lc, d) are accounted for differently in Kager's theory. The contrast between the initial syllables of the underived words in (la) and (lb) is an automatic consequence of the organization of his theory, while the same contrast between the derived words in (lc) and (Id) is due to a language-specific Prominence-Flattening 'convention'. In section 2 I will show that this additional destressing mechanism is inevitable given Kager's claim that the End Rule is cyclic in English. Both the lopsidedness of the analysis, and the addition of flattening strike me as undesirable. To solve this problem, I will argue that the End Rule is noncyclic, i.e. primarystress assignment takes place after the cyclic word-formation rules and the cyclic phonological rules. It will become apparent that this assumption enables us to account for the data in (1) in a similar vein. The structure of this paper is as follows. In section 21 will outline the relevant aspects of Kager's (1989) account of English word stress; in particular his analysis of the data in (1). In section 3 I will present a reanalysis of these data and claim that the End Rule is noncyclic. Next, I will discuss the cyclic adjectival suffix -ic with a stress condition on its base. I will show that this situation does not necessarily imply cyclicity for the End Rule, but can be accounted for straightforwardly even under the assumption of a noncyclic End Rule.

2. Kager's (1989) Pre-Stress Destressing and Prominence Flattening Kager adopts and develops the so-called bracketed-grids theory proposed by Halle & Vergnaud (1987) and Hayes (1987). In this theory the grid is a hierarchically layered representation consisting of columns of grid marks. The height of each column indicates its

72 stress level, and in this way the subordinative property of stress is expressed. Furthermore, the property that stressed and unstressed syllables alternate in many languages is captured by constructing binary feet (which are indicated by parentheses). The stress of the English word dpaldchicdla is represented as in (2): (2) *

*

* *

line 2 line 1

(* *) (* *) t* *) l i n e 0 a pa la chi co la The English stress system has the following properties: (i) it is quantity-sensitive; (ii) it is bounded; and (iii) it is left dominant (cf. Hayes 1981, Halle & Vergnaud 1987, and Kager 1989). Kager accounts for the first property by assigning to each heavy syllable (i.e. a bimoraic syllable) an inherent stress marker by the Quantity-Sensitivity Rule (3) below. 1 In (2), the penultimate syllable is heavy, and in the first pass through the word a line-1 asterisk is assigned to this syllable: (3)

Quantity-Sensitivity Rule (QS) Assign a level-1 element to a heavy syllable. (Kager 1989:3)

The second and third property are captured by constructing binary left-headed feet starting at the right edge by means of the Light Syllable Stress Rule (4): (4)

Light Syllable Stress Rule (English) Adjoin a line 0 element (i.e. a syllable place-holder) leftward. * line 1 *) lineO * * => (* OCT a a (Kager 1989:106,129)

When QS (3) has assigned a line 1 grid marker to the penultimate syllable, the syllables preceding the penult are gathered into two binary feet, and a third binary foot is erected over the penultimate and ultimate syllable by a subrule of (4), i.e. the rule of Stray Syllable Adjunction which adjoins an extrametrical syllable leftwards (cf. Hayes 1981). Finally, the primary stress on the penultimate syllable results from an End Rule in the sense of Prince (1983). This rule adds a line 2 grid mark to the rightmost metrical stress. The LSSR in (4) adjoins a light syllable as a weak node into a left-dominant foot. To account for the so-called Sonorant-Destressing effects in words such as hdckensack and hilminthoid, Kager postulates a second foot-construction rule, viz. the Closed Syllable Stress Rule in (5), which is ordered after the LSSR (4): 2 (5)

Closed Syllable Stress Rule (CSSR) Assign binary left-dominant feet whose weak node is a syllable which contains a short vowel. (Kager 1989:129)

73 The Sonorant-Destressing facts play a minor role in this paper, and therefore I will only illustrate the effect of (5) on the basis of two sample derivations (underlined syllables are extrametrical here and elsewhere): (6)

a: *

*

*

*

*

*

* * LSSR n.a. CSSR (* *) * ER (* *) * * ha cken sack ha cken sack ha cken sack

b: *

*

*

* * * * LSSR a di ron dack

*

*

*

*

*

(* *) * * CSSR n.a ER (* *) * * a di ron dack a di ron dack

The LSSR cannot apply in (6a) because of the fact that there are not sufficient light syllables available to build a foot. The CSSR on the other hand is free to adjoin the VSon syllable as the weak node of a binary foot. The situation is different in (6b). In this example the LSSR can erect a binary foot over the initial and second syllables. Next, the CSSR is blocked by Prince's (1985) Free Element Condition, which states that rules of primary metrical analysis can only apply to unfooted metrical elements.

2.1. Pre-Stress Destressing In (7a) words are presented in which two light syllables precede the primary stress. In (7b, c, d) on the other hand the primary stressed syllables are preceded by exactly one syllable: (7)

a:

alabima Minnes6ta &riz6na

b:

c&nt6en benzine b&ndina

c:

fcilden diyt6na 6hio

d:

police manipulate banana

The data in (7a) show that stressed and unstressed syllables alternate. When this is impossible, two situations arise: when the pretonic syllable is heavy (7b, c), it bears a secondary stress, otherwise it is stressless (7d). Since Halle's (1973) analysis of English word stress, it has been assumed that the stress rules do not distinguish between (7b-d), i.e. initial pretonic syllables acquire a stress marker or monosyllabic foot regardless of whether they are light or heavy. The absence of a secondary stress in (7d) is usually attributed to a rule of Pre-Stress Destressing which removes a stress over a light syllable if another stress immediately follows, cf. (8): (8) -> * na ba

(* na

*> na

ban da na

Kager's analysis differs from the earlier ones in a significant way: he argues that stress constituents are strictly binary, i.e. nonbranching and ternary branching feet are ruled out. Given this Strict Binarity Hypothesis, Kager is able to remove Pre-Stress Destressing from

74 the grammar, and the contrast between (7b, c) and (7d) follows from the fact that QS (3) will not add a line 1 asterisk to the initial syllable of the examples in (7d), as shown in (9): (9)

a: * * * ba na na

QS

n.a. LSSR

* (* *) ba na na

ER

* (* *) ba na na

b: * * * ban da na

QS

* * * LSSR ban da na

* (* *) ER ban da na

* (* *) ban da na

The conclusion seems to be that Kager's analysis of the Pre-Stress Destressing facts is superior to those of, for instance, Halle (1973), Hayes (1981, 1982) and Halle & Vergnaud (1987), since in his account there is no need for a language-particular destressing rule.

2.2. Prominence Flattening If Kager is correct in hypothesizing that destressing rules can be reanalyzed as stressing rules, one may wonder whether destressing rules are still required. The somewhat disappointing answer given in Kager (1989) is 'yes'. Essentially two types of destressing rules remain, viz. Post-Stress Destressing and Prominence Flattening. The former rule is a refooting rule which changes *a(bràca)(dàbra) (the output of the stressing rules) into (àbra)ca(ddbra), and the latter removes all cyclically assigned line 1 grids over light syllables except for the one dominated by a line 2 grid. I will not go into Kager's account of Post-Stress Destressing here; the interested reader is referred to De Haas (1990) for an alternative analysis of these facts. The destressing rule I want to discuss here is Kager's Prominence Flattening. Before I tum to a discussion of this convention, it is necessary to go into the issue of cyclic stress assignment, and particularly the solution offerred in Kager (1989). Stress assignment appears to be cyclic in English as is illustrated by the data in (10) below. They show that the primary stress of an underived word is preserved as a secondary stress under embedding: (10)

a:

[còndènsjàtion [consideration [stàndardiz]àtion

b:

[còmpensàt]ion [clàssifì]càtion [èmàncipàt]ion

At least in some dialects of English the second syllable of còndènsdtion cannot reduce contrary to the second syllable of còmpensàtion. This contrast corresponds to the different stress pattern of the embedded morphemes condènse and cómpensàte. Under the assumption that the rules of primary metrical analysis apply cyclically, the data in (10) can be accounted for straightforwardly. Kager (1989) accounts for the data in (10) in the following way. With Halle & Vergnaud (1987), he assumes that for each cycle a new stress plane is erected, and within each plane the stress rules are unable to see the effects of their earlier application, and therefore apply as if the input were an underived word. To arrive at a single representation, he proposes a

75 mechanism of plane conflation which is governed by Kiparsky's (1982) version of the Elsewhere Condition in (11): (11)

Elsewhere Condition Rules A, B in the same component apply disjunctively to a form 9 if and only if (i) The structural description of A (the special rule) properly includes the structural description of B (the general rule) (ii) The result of applying A to

Eo [in-the-bath-tub & Eo [sing (j,o) & p]]]

(j,o)

In this formula o is a variable over occasions or spatio-temporal locations. Every stage level predicate introduces such a variable, and it is bound by an existential quantifier. Always is interpreted as a conservative generalized quantifier over dynamic propositions, which means that it will bind the variable p. The standard definitions of Dynamic Montague Grammar make (13) equivalent to the predicate logical formula (14), which shows that binding the dynamic proposition variable p amounts to binding of the occasion variable: (14) Ao [ [in-the-bath-tub (j,o)] -->

[sing (j,o)]]

95 We can replace the proper name by an indefinite NP and quantify over dynamic propositions which contain two indefinites. For instance (15) gets the interpretation in (16a) which is equivalent to (16b). Similarly, (17) is interpreted as in (18): (15) When a student is in the bath-tub he always sings (16) a. Ap [[Ex Eo (student(x) & in-the-bath-tub (x,o) & p] — > Ex Eo [student (x) & in-the-bath-tub (x,o) & Eo (sing (x,o) & p]]} b. Ax Ao [[student (x) & in-the-bath-tub (x,o)] --> sing(x,o)) (17) When John sings a love song, he always sings it well Eo [love(18) a. Ap [Eo Ex [love-song (x) & sing (j,x,o) & p] — > song (x) & sing-well (j,x,o) & p]]] Ax Ao [[love-song (x) & sing (j,x,o)] --> sing-well (j,x,o) b. All indefinites are under the scope of the universal quantifier, which leads to the equivalence of (16a/b) and (18a/b). Chierchia does not treat quantification over indefinite objects as a special case, because predicate logic does not make a crucial difference between the analysis of indefinite NPs in subject and object position. Chierchia's interpretation of quantifying adverbs is then equivalent to the classical DRT approach. The integration of Kratzer's distinction between stage level and individual level predicates in this framework, leads to a constraint on Q-adverbs similar to Kratzer's prohibition on vacuous quantification. In Chierchia's system only existential quantifiers give rise to dynamic propositions. Therefore, he can formulate a rule requiring quantifying adverbs to apply only to dynamic propositions. Let us see now how the problem of indefinite objects which arises in Kratzer's theory is treated within Chierchia's approach. The contrast between the two theories is most obvious when individual level predicates» are concerned, because in that case there is no Davidsonian argument. We have seen that in such a context Kratzer predicts that the indefinite object cannot supply the variable for the quantifier to bind. Only subjects from outside the VP can get bound by the unselective quantifier. This hypothesis provides a natural explanation of the difference between (2) and (19). (2) *When Pedro has a donkey, he beats it (19) When a farmer has a donkey, he beats it On the other hand the same hypothesis makes it impossible to explain the well-formedness of (lc), because the only variable that is available here, is the one provided by the indefinite object, as was pointed out above: (1)

c.

When Mary knows a language, she knows it well.

Chierchia's analysis turns the problem exactly the other way around. There is an existential quantifier present in (lc), so the adverb can bind a dynamic proposition. But the same is true of (2), and yet, no quantification is allowed. More precisely, Chierchia predicts that indefinite objects of individual level predicates in constructions such as (2) and (lc) end up in the restrictive clause. Only existential quantifiers create dynamic propositions, and the only available existential quantifier for the adverb to operate on in sentences such as (2) and (lc) is introduced by the object. This means that Kratzer's approach does not allow enough quantificational relations, which forces her to make a distinction between well-behaved and ill-behaved indefinite objects, whereas Chierchia's theory generates too many interpretations, because he always permits objects to end up in the restrictive clause.

96 5. Comparison, discussion, and conclusion So far, we concentrated mainly on indefinite objects with individual level predicates because in this context the theories of Kratzer and Chierchia make opposite predictions. In this section, however, it will be shown that the problems also involve stage level predicates and subjects. Furthermore we will discuss the possible extensions of both theories in a more general perspective. Since Chierchia (1989) accepts Kratzer's account of the distinction between stage level predicates and individual level predicates as a difference in argument structure, he also inherits some problems related to this analysis. Consider the sentences below, taken from Kratzer (1989). (20) When Sue likes a movie, she recommends it to everyone (21) Uhen Anne appreciates a paper, she tries to really understand it Clearly, the objects in (20) and (21) can only be labeled "ill-behaved" if the predicates in the antecedent of the conditionals are individual level, for stage level predicates always leave open the possibility of quantification over spatio-temporal locations. According to Kratzer, the predicates have to be individual level in the above sentences, due to the ungrammatically of (22) and (23). (22) *When Sue likes "Wings of Desire", she recommends it to everyone (23) *When Anne appreciates this paper, she tries to really understand it In (22) and (23) there is no possible free variable for the quantifier to bind, which gives rise to ill-formedness. But now have a look at the contrast between (24)-(25) and (26)-(27). (24) (25) (26) (27)

When When *When *When

Sue made a movie, she recommended it to everyone Mary built a cottage, she built it well Sue made "Wings of Desire", she recommended it to everyone Mary built the cottage in this village, she built it well

Both Kratzer and Chierchia predict the sentences (26) and (27) to be grammatical. The predicate is stage level, so the adverb can bind the variable for spatiotemporal location. Yet, the sentences are ungrammatical in the intended reading where when has the meaning of "whenever".^ Intuitively, it is clear why we need an indefinite object in this type of sentences: the relation between the verb and its arguments is such that every object can only be associated vith one situation. Movies and cottages can be made only once, so quantification over locations has to correlate with quantifications over objects. Obviously, the distinction between stage level and individual level is irrelevant here. The general picture is that quantification over situations is only possible if one of the arguments is variable. Because individual level predicates denote permanent properties of individuals, they will more often give rise to ill-formedness in quantificational contexts than stage level predicates. Nevertheless, the exact relation between the predicate and its arguments appears to be essential rather than the nature of the predicate itself. This requires a more compositional approach to the construction of events (cf. Verkuyl, 1989). Moreover, just as the stage level/individual level distinction is only important insofar its role in the general picture is concerned, so is the distinction between subjects and objects. Consider once again the difference between (2) and (19). (2) *When Pedro has a donkey, he beats it

97 (19) When a farmer has a donkey, he beats it In Kratzer's theory subjects of individual level predicates are basegenerated outside the VP and hence they always provide the variable needed to avoid a violation of the prohibition against vacuous quantification. Therefore, (19) is well-formed whereas (2) is not. However, Kratzer does not observe that when the object is fixed instead of the subject, the same ill-formedness is obtained as in (2). (28) *Uhen a farmer has Smokey, he beats him. In (28) the subject should provide the necessary variable, but apparently it does not, because the sentence is ill-formed. That means that if the predicate to have is taken as a true individual level predicate, the arguments can only vary simultaneously. If one argument is fixed, the entire situation is fixed, no matter whether it is the object or the subject, and as a consequence, quantification over situations is not possible anymore. It will not come as a surprise by now that even stage level predicates which are "once-only11 with respect to the subject position require indefinite NPs in order to allow quantification. (29) (30) (31) (32)

When Uhen *When *When

a man in India died, his widow usually killed herself an Italian came of age, he usually got married Anil died, his widow usually killed herself Mario came of age, he usually got married

Again, neither Kratzer nor Chierchia can account for the ill-formedness of (31) and (32), since the stage level predicate should in principle supply the variable. It goes without saying that a theory where variables for spatiotemporal location and individual variables are assigned values independently of each other can never account for the observations made in this section. Evidently, this provides us with a strong argument against any approach of the stage level/individual level distinction of the kind Chierchia and Kratzer defend. The observation that Kratzer's and Chierchia's analyses partially share the same inadequacies does not mean that they are in fact equivalent. The real point here is that they present a very different perspective on the problem of indefinites. This difference is basically of a philosophical and methodological nature, but it has important empirical consequences. In the unselective binding approach Kratzer advocates, all indefinites are analyzed as variables. They get an existential reading when they are caught by existential closure of the nuclear scope, and a variable interpretation when they are bound in the restrictive clause of a modal operator or an adverb of quantification. In terms of the classical Generalized Quantifier framework this can be formulated as a "weak" existential interpretation in the nuclear scope, and a "strong" generic reading in the restrictive clause. In this light, Diesing's (1988) proposal can be seen as a syntactic reformulation of Milsark's (1977) observation that only strong quantifiers are compatible with individual level predicates, while stage level predicates allow both classes of quantifiers. The disadvantage of Diesing's approach is that although it is quite attractive for bare plurals and maybe also for a N, it cannot be extended to other indefinites (although Kratzer seems to claim the opposite). Numerals and other weak determiners (such as nany/few) get an existential interpretation when they enter the nuclear scope. But in the subject position of individual level predicates they do not necessarily give rise to generic readings. In this position they usually are interpreted as implicit partitives, witness (33).

98 (33) a. b.

Few cats are ill Few cats are intelligent

(existential or partitive) (partitive only)

The importance of this observation is that strong partitive interpretations cannot be explained in a variable analysis. The NP is interpreted as a Generalized Quantifier, rather than as a variable bound by a sentence operator. This means that Kratzer's approach offers little hope for interesting extensions to the whole class of indefinites. Chierchia also presents a unified analysis of indefinites: in his approach they are always interpreted as dynamic existential quantifiers. Unfortunately, this does not allow him to interpret all scrambling facts correctly. We saw that Kratzer assumes that ill-behaved objects can scramble out of the nuclear scope into the restrictive clause, which gives rise to a generic interpretation. In English this has to be done at the level of logical form, but in languages like German and Dutch, objects can scramble at S-structure with a concomitant difference in interpretation. Kratzer supposes that objects that can scramble at S-structure in German can do so at the level of logical form in English. As for Chierchia, problems arise in explaining the Dutch sentences in (34), which, in Kratzer's formalism, get the representations under (35): (34) a. b. (35) a. b.

... omdat Martin altijd Noorse sokken breit •because Martin always knits Norwegian socks' *... omdat Martin altijd Zweedse tennissterren bewondert 'because Martin always admires Swedish tennisstars' Always i [[location (1) & Ex sock (x) & knit (Martin, x, 1) ] [Norwegian sock (x)]] Always [Ex tennisstar (x) & admire (Martin, x) j [Swedish tennisstar (x)]]

The objects in (34) did not scramble. Instead, they remained in their original position, i.e. adjacent to the verb. As usual, Chierchia would get existential quantifiers in both sentences and thus his rules would generate not only (34a), but also (34b). Kratzer on the other hand correctly predicts (34b) to be ungrammatical. When the object is not scrambled, it is interpreted in the nuclear scope and bound by existential closure of the VP. This makes (34b) an instance of vacuous quantification. Since the predicate in (34a) is stage level, the adverb can bind the Davidsonian argument, and the sentence is fine. In order to make (34b) well-formed, the object has to scramble, which gives rise to a generic reading, witness (36). (36) ...omdat Martin Zweedse tennissterren altijd bewondert 'because Martin always admires Swedish tennisstars' Kratzer's view on free variables and existentially bound indefinites allows her to distinguish between an interpretation of non-scrambled indefinites as weak, existential NPs and scrambled indefinites as variables. This approach allows us to reformulate Chierchia's problem in an obvious way. His unified analysis of indefinite NPs as existential quantifiers does not distinguish between positions of indefinite NPs which create "strong" or "weak" readings. This leads to overgeneration in contexts such as (34). The point now is not to prefer one theory to the other on the basis of the facts given, but on the basis of possible extensions of the analysis. Kratzer cannot go beyond the discussion of bare plurals and a N. Other indefinites cannot be accounted for without serious modifications of her theory. Chierchia's analysis on the other hand is formulated in the framework of Generalized Quantifiers. This means that, in our case, we can appeal to general definitions of the weak/strong distinction in order to

99 accomodate the problems pointed out with respect to (34) and other instances of overgeneration. It is this perspective on interesting empirical and theoretical developments of Chierchia's theory which provides the final argument for us to definitely prefer this approach to Kratzer's analysis. A formalization of our findings will have to await the results of further research. Notes *

This research was supported by the Foundation for Linguistic Research, which is funded by the Dutch organization for research, NWO (grant 300163-032), and the Faculty of Arts of the University of Groningen (VFO LETT 8-24). We thank Ale de Boer for correcting the English.

1.

The terms stage level predicate and individual level predicate are borrowed from Carlson (1977). Individual level predicates quantify over individuals; they denote permanent properties. Stage level predicates, on the other hand, quantify over situations, i.e. spatiotemporal slices of individuals; they denote temporary properties, stages.

2.

We introduce some typographical simplifications which are rather innocent in this context, but which make the formulas easier to read. A correct and complete formalization of the results reported in this section can be found in Groenendijk and Stokhof (1989) and Chierchia (1988, 1989).

3.

Chierchia (1988) presents an interpretation of quantifying adverbs as generalized quantifiers over occasions, while in his (1989) paper they quantify over dynamic propositions. The reasons behind this modification do not concern us here. From now on we will follow the version of Chierchia (1989), because it accepts Kratzer's account of the distinction between stage level and individual level predicates, and this will make it easier to compare both theories.

4.

The sentences are given in the past tense in order to avoid the interference of aspectual problems related to the choice between simple present and progressive.

References BARWISE, J. 1987 Noun Phrases, Generalized Quantifiers and Anaphora, in: Gärdenfors (ed.). Generalized Quantifiers. Dordrecht: Reidel, 1-30 CARLSON, G. 1977

P.

Reference to Kinds in English. New York: Garland

CHIERCHIA, G. 1988 Dynamic Generalized Quantifiers and Donkey Anaphora, in: M. Krifka (ed.). Genericitv in Natural Lanpuape. Tübingen: SNS, 53-84 1989 Anaphora and Dynamic Logic, paper presented at the 7th Amsterdam Colloquium DIESING, M. 1988 Bare Plural Subjects and the Stage/Individual Contrast, in: M. Krifka (ed.). Genericitv in Natural Language. Tübingen: SNS, 107-154

100 GROENENDIJK, J. and M. STOKHOF 1989 Dynamic Montague Grammar, ms. University of Amsterdam HEIM, I. 1982 The Semantics of Definite and Indefinite NPs. diss., University of Massachusetts, Amherst HOOP, H. de, and H. de SWART (to appear) Over Indefiniete Objecten en de Semantiek, in Glot 12.1

Relatie

tussen

Syntaxis

en

KAMP, H. 1981 A Theory of Truth and Semantic Representation, in: J. Groenendijk e.a. (eds.). Formal methods in the study of language. Amsterdam: Mathematical Centre. Reprinted in J. Groenendijk e.a. (eds.)(1984). Truth. Interpretation and Information. Dordrecht: Foris, 1-41 KRATZER, A. 1989 Stage-level and Individual-level Massachusetts, Amherst

Predicates,

ms.

University

of

SCHUBERT, L. and F. PELLETIER 1988 Generically Speaking, or, Using Discourse Representation Theory to Interpret Generics, in: G. Chierchia, e.a. (eds.). Properties. Types and Meaning. Dordrecht: Reidel, 193-268 VERKUYL, H. 1989 Aspectual Classes and Philosophy 12, 39-94

Aspectual

Composition,

Linguistics

and

Sietze Looyenga

On the Internal Structure of Nominal Infinitives 0. Introduction* In this article I present an analysis of Dutch nominal infinitives (NIs) within the theory of Government and Binding (Chomsky 1981, 1986). Nominal infinitives are phrases that have an infinitival verb as a head, but at the same time show the external distribution of noun phrases. Some examples of NIs are given in (l):1 (l)a b c

Fietsen stelen (is spannend) bicycles stealing (is exciting) Het stelen van fietsen (is een misdaad) the stealing of bicycles (is a crime) Het fietsen stelen van Jan (bracht hem in grote problemen) the bicycles stealing of John (took him into big trouble)

These sentences show that NIs, while behaving as noun phrases externally, exhibit both nominal and verbal properties internally. These properties are discussed in a.o. Dik (1985), Van Haaften et al. (1985), Hoekstra & Wehrmann (1985), Reuland (1988) and Van Zonneveld (1990). In order to account for the mixed character of nominal infinitives, it is generally assumed that an NI is a nominal phrase containing a verbal one. That is, an NI is considered a verbal projection that is somehow transposed into a nominal projection. In this article I discuss some of the analyses of NIs that have been proposed heretofore, and I give the outlines of an alternative. Broadly speaking, I follow Reuland (1988) in assuming an empty nominal variable in the sense of Higginbotham as the head of an NI. I will, however, embed this analysis in a DP-like framework. Furthermore, I will argue, contrary to previous analyses (cf. Hoekstra & Wehrmann 1985, Hoekstra 1986, Van Haaften et al. 1985), that NIs with an overt determiner contain a non-lexical (PRO) subject and that NIs without an overt determiner contain a generic determiner. This approach has some advantages, as I will show. 1. The morphosyntactic analysis Previous analyses of NIs share one assumption: that an NI is to be looked upon as a nominal phrase containing a verbal one. According to Hoekstra & Wehrmann (1985) and Hoekstra (1986), the transposition from verbal to nominal is accomplished by the nominal affix -en, which can be attached to the verb at any level in the verbal projection, thereby transposing it into a nominal one. These analyses are based on the Deverbalization Rule of Jackendoff (1977): (2)

X1

->

affix

V1

This morphosyntactic analysis presupposes that affixation can take place in the syntactic derivation. The analysis implies that the verbal and the nominal part of an NI can be separated. Initially, the projection is verbal, but after the affix has been attached to the verb the projection is nominal. This, however, is incompatible with the facts, as noted by Hoekstra (1986, 564):

102 (3)a b

Het the Het the

urenlang hooghouden van de bal for hours holding aloft of the ball met een mesje schillen van de aardappels with a knife peeling of the potatoes

In the NIs in (3), the direct object follows the head of the NI, which means that this head is nominal: in Dutch, the direct object of a noun follows the head, while the direct object of a verb precedes it. So, in these NIs deverbalization must have taken place at the V°-level, in other words: the affix -en must have been attached to the verb at the V°-level. The presence of an adverbial modifier, however, implies that deverbalization must have taken place at the highest level in the V-projection, at least if we adopt the assumption that an adverb such as urenlang ('for hours') is adjoined to the VP. These results of the analysis are incompatible. So, the morphosyntactic analysis is unable to describe NIs adequately. More evidence in favour of this conclusion can be obtained if we look at NIs with a verb that takes two complements, such as geven ('to give') or sturen naar ('to send to'): (4)a b

Het the Het the

aan je moeder geven van bloemen (is een blijk van waardering) to your mother giving of flowers (is a sign of appreciation) naar bed sturen van je kinderen (blijft een hele opgave) to bed sending of your children (remains a tall order)

In (4), the direct object is once more realized in a PP following the head, which means that the affix must be attached to the verb at the V°-level, as pointed out above. The other object of the verb, however, precedes the head, which means that the head has to be verbal at the V' -level and that the affix is not yet attached to the verb at the V°-level. These sentences show once more that the morphosyntactic analysis is unable to describe NIs adequately. 2. Infinitives as hybrids Van Haaften et al. (1985) propose that infinitives bear the features [+N,+V] when they occur in an NI with a determiner, while they bear the features [-N,+V] when they occur in an NI without a determiner. So, an infinitive in an NI with a determiner is able to assign a thematic role both to its left (by virtue of the feature [+V]) and to its right (by virtue of its feature [+N]). This allows us to analyze the van-PP containing the direct object as the (right) sister of the verb in an NI with a determiner. Since the verb does not govern to the right, the preposition van is inserted in order to prevent a violation of the Case-filter. The assumption also enables us to analyze the NIs in (3) and (4), because it doesn't imply that the nominal and the verbal part of an NI are strictly separated. According to this analysis, the structure of the NI in (4a) is as in (5) (cf. Van Haaften et al. 1985):2 [+N,-v]B™

(5) SPEC

[+N,+V]1

geven

van bloemen

If we assume that adverbs are adjoined to a [+N,+V]™M, we can also describe the NIs in (3) adequately. Although descriptively adequate, however, the

103 analysis of V a n H a a f t e n et al. is in essence stipulative. Y e t , as w e will see in the n e x t section, it can b e embedded in an a n a l y s i s that seems to be theoretically m o r e satisfactory.

3. Nominal infinitives a n d the DP-framework In Abney (1987) a n analysis of gerunds is p r e s e n t e d that is v e r y elegant and appealing a t first sight. According to Abney, a n N P is h e a d e d b y a functional element b e l o n g i n g to the category D of determiners. D is to b e l o o k e d upon as some k i n d o f nominal inflection, showing a g r e e m e n t w i t h the h e a d of the thematic category it selects, just like I does. T h e e m p i r i c a l p r o b l e m that m o t i v a t e d the DP-analysis was the d e m a n d for a n adequate a n a l y s i s of gerunds. Since gerunds are g e n e r a l l y considered the English c o u n t e r p a r t s o f NIs, it is obvious t h a t w e s h o u l d try to embed an analysis of N I s in the DP-framework. J u s t like E n g l i s h gerunds, D u t c h NIs can be a n a l y z e d as DPs in w h i c h the determiner selects a V P instead of an NP. G i v e n the a s s u m p t i o n t h a t functional elements such as D a n d I lack descriptive content a n d only regulate or specify the interpretation of their complement, we can say that w h e n a V P is selected b y an element o f the category I, it is u s e d as a p r e d i c a t e and it denotes a property, w h i l e w h e n a V P is selected b y an element o f the category D, it is u s e d as a n argument a n d it refers to the individual correlate o f a property (cf. C h i e r c h i a 1982, 1984). So, the c o n t r i b u t i o n of the functional elements to the interpretation o f a V P is that they determine w h e t h e r it is a predicate or a n argument. A c c o r d i n g to this analysis, the structure of the NI J a n s fietsen stelen ('John's b i c y c l e s s t e a l i n g 1 ) is (6): (6)

DP

DP

V

fietsen

stelen

A t first sight, an a n a l y s i s along these lines is v e r y attractive, because it captures in a rather n a t u r a l w a y the relation b e t w e e n a n NI such as Jans fietsen s t e l e n ('John's b i c y c l e s stealing') and the c o r r e s p o n d i n g sentence Jan steelt f i e t s e n ('John steals bicycles'). In spite of this, the analysis cannot b e maintained, because it is unable to describe the facts correctly. In particular, it is u n a b l e to account for the n o m i n a l p r o p e r t i e s that are e x h i b i t e d b y N I s (and b y gerunds as well). In order to a c c o u n t for these properties, A b n e y a s s u m e s that a g e r u n d contains a n affix -ing that can be attached to v a r i o u s levels in the V - p r o j e c t i o n and t h a t transposes the verbal p r o j e c t i o n into a n o m i n a l one. Abney, too, makes a n appeal to Jackendoff's D e v e r b a l i z a t i o n Rule (cf. (2)). This, however, r e d u c e s h i s analysis to a v a r i a n t o f the m o r p h o s y n t a c t i c analysis discussed in the first section, and makes it inapplicable to NIs. We can try to accommodate Abney's original proposal in s u c h a w a y that its primary attractiveness is maintained and that a n appeal to Jackendoff's d e v e r b a l i z a t i o n schema is unnecessary (cf. also Z w a r t & H o e k s t r a 1989). This revision s h o u l d make a structure such as (7) acceptable:

104 (7) SPEC Jan

Stelen

van netsen

In (7), the infinitive must have the features [+N.+V], which makes this analysis similar to the analysis of Van Haaften et al. (1985). Within a theory that presupposes the existence of functional categories, we can even argue for this feature specification. I will do this in the next section. 4. Functional categories Within a theory that presupposes the existence of functional (or non-lexical) categories, it is generally assumed that there is agreement between a functional head and the head of the thematic (or lexical) projection that it selects, and that the functional head thereby licenses its complement. In spite of all the attention that has been paid recently to all kinds of functional categories and their projections (cf. Fukui 1986, Abney 1987, Chomsky 1988, Pollock 1989), the exact nature of this agreement is not yet completely understood. In the spirit of the proposals mentioned above, we can assume that agreement is to be looked upon as a form of feature sharing. Relevant features seem to be phi-features (of number, gender and case) and categorial features. This suggests that a functional element should be able to license a thematic projection if this element and (the inflectional affix attached to) the head of the thematic projection bear the same phi-features and the same categorial features. As is well-known by now, we have to distinguish two types of inflectional elements: next to verbal inflection, there is also nominal inflection. For example, the genitive affix -es in the German noun phrase (8) is a nominal inflectional affix. This affix is licensed by a nominal functional head, the determiner dieses: (8)

Die Traum dieses Mannes the dream this-GEN man

So, overt case markers in languages such as German are nominal inflectional affixes that agree with a nominal inflectional head. We could, in the spirit of Chomsky (1988), consider the determiner dieses as belonging to the category AGR-G (for AGR-genitive). It seems reasonable to assume that nominal inflectional heads always select NPs, while verbal inflectional heads can only take VPs as their complements. This would be a consequence of the assumption that there is agreement between a functional (i.e. inflectional) head and the head of the thematic projection that it selects, and that this agreement involves (among others) categorial features. This conception of agreement, however, makes it impossible for an element of the nominal category D to select a projection of an element of the verbal category V. So, we are forced either to give up the idea that an NI is a DP in which a VP occurs as the complement of D, or to redefine our notion of agreement. Another reason for reconsidering this notion is that it is uncertain whether functional categories can be characterized in terms of the categorial features [±N] and [±V] (cf. Chomsky 1986). If they cannot, and if we want to maintain the

105 assumption that a functional head can license a thematic projection if this head and the head of the thematic projection show agreement, we are led to the conclusion that this agreement doesn't involve categorial features. Alternatively, we can assume that the agreement holds between the functional head and the (inflectional) affix that is attached to the thematic head. This assumption presupposes that we distinguish thematic categorial features and inflectional categorial features, and that an inflectional affix bears an inflectional categorial feature. If we characterize functional heads as expressions that somehow establish a link between thematic projections, we can assume that these inflectional categorial features indicate how a selected thematic projection has to be linked to another thematic projection. So, we can assume that a projection of a functional head with the feature [+N] is always an argument, while a projection of a functional head which bears the feature [+V] can only be a predicate. This is compatible with the basic idea of the DF-analysis outlined above. So, the affix attached to the thematic head determines by which functional heads a thematic projection can be selected. In order to be able to analyze NIs as DPs in which a VP is the complement of D, we can assume that the infinitival affix -en has the (inflectional) features [+N,+V] , and as a consequence can agree both with a nominal and with a verbal functional head. The assumption that -en has the features [+N,+V] can be related to the fact that infinitives can occur both as a predicate and as an argument. If we reduce this to properties of the affix, we can explain why only infinitives have this property. In order to account for the [+N,+V]status of the infinitive, we need one further assumption: that the verb to which the affix is attached inherits the features of the affix, or that the features of the affix are added to those of the verb. In this way, we can embed the ideas presented in Van Haaften at al. (1985) in an analysis that is theoretically more satisfactory. There is, however, a decisive conceptual argument against this analysis: for principled reasons, it is incompatible with a theory of functional categories. In fact, within such a theory, an affix with the (inflectional) features [+N,+V] is impossible. The feature [+V] reflects the tense properties of a verb, i.e. it indicates whether a verb is finite or non-finite. The feature [+N] , on the other hand, corresponds to phi-features (of number, gender and case). Because a nominal phrase can be assigned different cases, there are different (possibly null) affixes that can be attached to a noun. NIs, too, occur in positions where they receive Nominative, Objective or Oblique case: (9)a b c

Jans fietsen stelen wordt door iedereen veroordeeld John's bicycles stealing is by everyone condemned Wij betreuren het voortdurende stenen gooien van uw zoontje we regret the continuous stones throwing of your son Ik houd niet van het schieten op dieren I like not (of) the shooting on animals

Theoretically speaking, the infinitives in (9) all bear different affixes. So, we cannot assume that the infinitive bears an affix that can be licensed both by a nominal and by a verbal inflectional category. The properties expressed by the feature [+N] and the feature [+V] are incompatible for principled reasons. This means that the analysis according to which a VP with an infinitive as its head can be selected both by a nominal and by a verbal inflectional category is essentially wrong. Rather, these considerations suggest the following canonical structure for an NI: (10)

[ DP [„• D [ IP

[vr •• V+aff1+aff2 ..] I ]]]]

106 So, the infinitives in (9) bear two affixes, a nominal one, aff2 in (10), licensed by the element of the category D, and a verbal one, the (infinitival) affix -eQ, represented as affj and licensed by an element of the category I.3 This can be illustrated with an example from a language with overt case markers, such as German, that has nominal infinitives that resemble Dutch NI's (Abraham 1989): (11)

die Bestätigung des Empfangens dieses Briefes the affirmation of receiving this-GEN letter

Next to the affix -en, the verb in (11) bears another affix, reflecting the case that is assigned to the NI as a whole. This analysis is similar to the analysis presented in Reuland (1988), in which it is assumed that the head of an NI is a variable in the sense of Higginbotham (1983) and that this variable can be combined with an IP that functions as the restriction of the variable. In the next section, I will outline the proposal that NPs contain a variable. 5. The nominal variable in NPs and NIs In Higginbotham (1983, 1985) and also in Reuland (1986, 1988) it is argued that nominal phrases contain a variable or an open place that has to be bound. That is, the canonical structure of NPs is (12): (12)

[„p DET [„. N(x) ]]

In this representation N' is a set expression, while N is a set predicate and x a variable that has to be bound. The determiner acts as the binder of the variable. Higginbotham (1983) argues that (12) is an LF representation, but Higginbotham (1983) gives evidence that suggests that the variable or open place is present at S-structure as well. Besides this, we can also argue that the internal structure of an NP at S-structure and at D-structure must be sufficiently articulated to allow a mapping on the LF-structure (12) (Reuland 1988). So, at D-structure and at S-structure the variable is present too. Reuland (1988) assumes that the variable belongs to a syntactic category that is able to project. In fact, this variable is considered the (empty) nominal head of an NI. Reuland assumes that in Dutch V and I project simultaneously. Furthermore, he argues that the various levels in the V/I-projection are set predicates, just like N is. That is, in an NI the levels in the V/I-projection serve as the restriction of the nominal variable (cf. Reuland 1988, 332). In Reuland's theory, the structure of the NI (13a) will be as in (13b): (13)a

Het brieven beantwoorden van Jan [DP het [„. [„. [ 1P (PRO) [ V V I . [„pbrieven] [v/ibeantwoorden] ] ] (x) ] van Jan ]] This structure shows a strong resemblance to the structure in (10), but makes no use of a nominal inflectional category such as D. In the next section, I will try to embed Reuland*s analysis in a DP-like framework. b

6. An analysis of nominal infinitives In a theory that presupposes the existence of a category D of determiners and also assumes that nominal phrases contain a variable, the canonical structure of a nominal phrase (a DP) can be represented as in (14): (14)

[ DP [„• D [„, [„. N (x) ]]]]

107 In (14) not only N", but NP as well is to be considered a set expression. This is in accordance with Abney's assumption that an NP in the DP-analysis can be put on a par with an N' in the standard analysis. In a theory such as the DPanalysis we can say that an NP cannot occur on its own because it contains a variable that is not bound. The determiner that takes the NP as its complement acts once more as the binder of the variable. This suggests that the canonical structure of NIs with a determiner should be (IS): (15)

[ DP [D. D. [ HP [„. IP (x) ]]]]

In this structure, the nominal variable is the head of the NP that is the complement of D. IP is to be considered a set predicate and N' a set expression. Finite IPs are not considered set predicates. An IP with nonfinite inflection, however, contains a non-lexical subject, because nonfinite I is unable to assign case to a lexical subject (I leave open the question of whether in this case the empty subject is PRO or another empty category). The empty subject behaves as a variable that is not bound within IP. So, we can consider an IP containing an infinitive a set predicate. Thus looked upon, an IP with non-finite inflection is an open sentence (cf. also Van Zonneveld 1990, in which it is claimed that an NI is a projection of a 'bare S', i.e. a VP with a subject variable).4 There is a problem with an analysis along these lines which I will not attempt to solve here: we have to relate the variable that is the head of the NI and the empty subject of the IP. This problem arises as well in an analysis of prenominal participles such as (16): (16)a b

deze artikelen schrijvende taalkundige this articles writing linguist [DP[D• deze [NP[N. [ lp (ec) artikelen schrijvende] taalkundige (x) ] ] ] ]

I will assume that the nominal variable is coindexed with the empty subject of IP. One way to look upon this coindexation is the following. As Higginbotham (1985) points out, we can consider the variable or open place in a nominal phrase as a representation of a thematic role. Intuitively speaking, we can identify this role and the role assigned to the empty subject: both roles correspond to the individual(s) of which the property expressed by the IP is predicated. Formally, we can account for this identification if we assume that the role corresponding to the variable binds the role of the empty subject (in the sense of e.g. Williams 1989). So, the thematic role of the subject is discharged (in the sense of Higginbotham 1985) via the variable. Following Higginbotham, I assume that a thematic role can be discharged in two ways: either it can be bound (by a determiner), or it can be assigned to a lexical NP. Following a suggestion in Higginbotham (1983), I assume that the thematic role represented by the variable in a nominal phrase is discharged in both ways if the determiner is a genitive NP, as in John's cat. In this case, the genitive NP binds the role, which is at the same time assigned to that same NP. Furthermore, in a (Dutch) nominal phrase such as de kat van Jan ('the cat of John') the role is bound by the determiner and assigned to the NP in the PP. In the nominal domain the thematic role represented by the variable is not assigned obligatory, it is sufficient that it is bound by a determiner, as in the cat, where the role is not assigned, but only bound. The same holds for NIs. In (17a) and (17b) Rj, the thematic role of the empty subject (coindexed with the thematic role represented by the variable) is both bound and assigned, but in (17c) it is only bound (x is a variable over events):

108 (17)a b c

Jans fietsen stelen the x: steal (x) & Rx (Jan,x) & R 2 (bicycles,x) Het fietsen stelen van Jan the x: steal (x) & Rx (Jan,x) & R 2 (bicycles,x) Het stelen van fietsen the x: steal (x) & R 2 (bicycles,x)

An analysis along these lines is in contrast with the analysis that only NIs without a determiner contain a non-lexical (PRO) subject, as is assumed in Hoekstra & Wehrmann (1985), Hoekstra (1986) and Van Haaften et al. (1985). In Hoekstra & Wehrmann (1985) it is claimed that (18a) is grammatical because Che (implicit) argument associated with leuk is able to control the PRO subject of the NI, while (18b) is ungrammatical because schiin does not have an (implicit) argument that is able to control the PRO subject of the NI. In their analysis, the ungrammaticality of (18c) is unexpected, because the NI should not contain a PRO subject. If, however, we assume that (18c) contains a non-lexical subject as well as (18b), we can explain its ungrammaticality: (18)a b c

boeken lezen is leuk books reading is fun *boeken lezen is slechts schijn books reading is only appearance *het lezen van boeken is slechts schijn the reading of books is only appearance

This suggests that the thematic role of the non-lexical subject in an NI is assigned outside the NI° if it is not assigned within it. Again, we can account for this in terms of a coindexation of thematic roles, as proposed above. Contrary to the structure in (10), (15) contains a nominal projection. The presence of this nominal projection seems to be necessary, because otherwise the nominal properties cannot be accounted for. So, the structures of the NIs (lb) and (lc), repeated here as (19), become those in (20) :5 (19)a b (20)a

Het the Het the

stelen van fietsen (is een misdaad) stealing of bicycles (is a crime) fietsen stelen van Jan (bracht hem in grote problemen) bicycles stealing of John (took him into big trouble) DP

b

DP

N het

ec

N"

I" VP

DP

ec

I' VP

I

I

V

DP

V

stelen

fietsen

stelen

109 presence of the variable. If there is no determiner, there is no variable, because it cannot be bound. If there is a determiner, there is also a variable, because vacuous quantification is forbidden. Reuland assumes that in NIs such as (la), repeated as (21), the NI is not an NF (or DP), but an IP: (21)

Fietsen stelen (is spannend) bicycles stealing (is exciting)

Contrary to Reuland, I assume that NIs such as (21) do contain a determiner, viz. a (phonologically null) generic determiner, and that they also have a nominal variable as a head. So, the generic determiner acts as the binder of the variable. This assumption can be based on the striking resemblance between the behaviour of NIs and mass nouns, as is shown in (22) and (23):4 (22) a b c (23)a b c

Melk is gezond milk is healthy *De melk is gezond the milk is healthy De melk van Friese koeien is gezond the milk from Frisian cows is healthy Roken is ongezond smoking is unhealthy *Het roken is ongezond the smoking is unhealthy Het roken van sigaretten is ongezond the smoking of cigarettes is unhealthy

In Hoekstra & Wehrmann (1985) , it is suggested that the mass noun melk and the infinitive roken both refer to an unstructured mass and that we cannot refer to this mass by using an NP (or a DP) with a definite article, because this presupposes a structuring of the mass. Uhen, however, the mass is split up (cf. (22c) and (23c)), we can refer to one o°f these parts by using a DP with a definite article. So, the explanation of the pattern in (22) and (23) would in essence be semantic. It is, anyway, not just a problem for a theory of nominal infinitives. I claim, then, that the structure of the NI (21) is (24): (24)

[ DP [D. D [ MP [„, [ IP [vp [v,[DPfietsen][vstelen]]]] (x) ]]]]

As (24) shows, I assume that NIs without an overt determiner contain a variable just like NIs with an overt determiner do. This variable can be bound by the generic determiner. The thematic role of the empty subject cannot be discharged, because this type of NI cannot contain a PP with the external argument of the verb. An analysis along the lines proposed here enables us to generalize over the different types of NIs: every NI is to be considered a DP. Differences between NIs with and NIs without an overt determiner should not be explained in terms of structural properties of the constructions, but in terms of properties of the determiners involved. 7. Conclusion In this article I have shown that previous analyses of NIs, which all have their shortcomings, can be extended in such a way that the solution to certain problems comes closer. In particular, I have suggested an analysis in which the distinction between NIs with and NIs without an overt determiner is explained more adequately than in previous analyses.

110 Notes: *

I thank Peter Coopmans, Eric Reuland, Ron van Zonneveld and Jan-Wouter Zwart for comments and discussion. This research was supported by the Netherlands Organization for Scientific Research (NWO), project number 300163-036, which is gratefully acknowledged.

1. Usually, NIs are considered the counterparts of English gerunds. In this paper I will not pay particular attention to gerunds. I believe, however, that my analysis can (with some adaptations) be applied to gerunds as well. 2. Van Haaften et al. (1985) consider NIs such as (4a) ungrammatical. They explain this by assuming that when a head is able to assign more than one thematic role, it has to assign them in the same direction. A construction such as het bloemen geven aan mijn moeder (' the flowers giving to my mother') is derived via PP-over-V. In my opinion, (4a) is acceptable. 3. We could, alternatively, consider the infinitive as the lexical item that gets inserted at D-structure. Here, -en is no Inflectional affix. This enables us to maintain the variant of the DP-analysis in which a VP is complement of D. I will not pursue this possibility. 4. Chierchia (1987) claims that even certain (embedded) sentences with finite inflection are open sentences expressing a property. For example, in John thinks that he is a genius the embedded sentence is an open sentence. 5. As (19) shows, I assume that the direct object is base generated as sister of the verb and afterwards adjoined to N' . This movement is incompatible with the theory of movement outlined in Chomsky (1986), according to which adjunction to an X' is not allowed. There is, however, evidence that such an adjunction has to be allowed. In this paper, I will not go into the details of this. 6. The generalization suggested by (22) and (23) seems to be undermined by the fact that NIs with a generic determiner cannot contain an adjective, whereas mass nouns can, cf. koude melk is pezond ('cold milk is healthy'). NIs, however, share characteristics with neuter nouns (cf. Abraham 1989). Neuter mass nouns can be preceded by an adjective, but this adjective never bears the adjectival affix -e when the determiner is generic, cf. *groene pras (' green grass' ) . So, we could even argue that NIs with a generic determiner, such as hard lopen ('fast walking') contain an adjective. We also have to explain why in the case of NIs with a generic determiner the direct object cannot follow the head, as in *roken van siearetten ('smoking of cigarettes'). Here, again, the generalization seems to be undermined, because melk van Frlese koeien ('milk from Frisian cows') is acceptable. In this case, however, the head is not followed by one of its arguments, but by a modifier. NIs with a generic determiner can contain a modifier as well, cf. roken in de tuln (is oneezond') ('smoking in the garden is unhealthy1).

References : ABNEY, S.P. 1987 The English Noun Phrase in its sentential aspect, dissertation, MIT

Ill ABRAHAM, W. 1989 Verbal substantives in German, in C. Bhatt, E. Löbel and C. Schmidt (eds) Syntactic phrase structure phenomena in Noun Phrases and Sentences. Amsterdam/Philadelphia, 79-98 CHIERCHIA, G. 1982 Nominalization and Montague grammar: a semantics without types for natural language, Linguistics and Philosophy 5, 303-354 1984 Topics in the syntax and semantics of infinitives and gerunds, dissertation, Univ. of Massachusetts 1987 Attitudes and anaphora de se, ms., Cornell University CHOMSKY, N. 1981 Lectures on Government and Binding. Foris, Dordrecht 1986 Barriers. LI Monograph 13, MIT Press 1988

Some notes on economy of representation and derivation, ms., MIT

DIK, S.C. 1985 Nederlandse nominalisaties in een funktionele grammatika, Forum der Letteren 26, 81-107 FUKUI, N. 1986 A theory of category projection and its applications, diss., MIT HAAFTEN, T. VAN, S. VAN DE KERKE, M. MIDDELKOOP & P. MUYSKEN 1985 Nominalisaties in het Nederlands, GLOT 8, 67-104 HIGGINBOTHAM, J. 1983 Logical form, binding and nominals, Linguistic Inquiry 14, 395-420 1985 On semantics, Linguistic Inoulrv 15, 547-593 HOEKSTRA, T. 1986 Deverbalization and inheritance, Linguistics 24, 549-584 HOEKSTRA, T., & P. WEHRMANN 1985 De nominale infinitief, GLOT 8, 257-274 JACKENDOFF, R. 1977

X-bar-svntax: a study of phrase structure. LI Monograph 2, MIT Press

POLLOCK, J.-Y. 1989 Verb movement, Universal Grammar, and the structure of IP, Linguistic Inquiry 20, 365-424 REU1AMD, E. 1986 A feature system for the set of categorial heads, in P. Muysken and H. van Riemsdijk (eds) Features and projections. Foris, Dordrecht, 41-88 1988 Relating morphological and syntactic structure, in M. Everaert et al. (eds) Morphology and modularity. Foris, Dordrecht, 303-337 WILLIAMS, E. 1989

The anaphoric nature of thematic roles, Linguistic Inquiry 20, 425-456

ZONNEVELD, R. VAN 1990 Syntactic nominalization, to appear in W. Abraham and E. Reuland (eds) Proceedings of the 5th conference on comparative Germanic syntax. Groningen ZWART, J.-W., & E. HOEKSTRA 1989 Functionele projecties in nominalisaties, ms., Univ. of Groningen

Anneke Neijt

Prosodic Structures and Phonetic Findings - The Case of Equally Stressed Adjectives 0. Introduction* Van Heuven (1986, 1987) presents data on pitch, duration and intensity in the pronunciation of different classes of adjectives in Dutch. He draws attention to one special class of adjectives, compound forms such as beeldschoon, which are claimed to display a pattern with two main stresses (1-1). Such a pattern is in conflict with the proposal in generative phonology to represent prominence as a set of binary oppositions (both in metrical and in grid theory such oppositions are currently used to express the culminative nature of stress). Van Heuven therefore remarks that his findings are a "challenge to generative phonologists" (1986, 85) and ends his article with the question "Would they be prepared to revise their rule mechanisms so as to allow the generation of two primary stresses within a single domain; and if so, how can this be done?" This paper takes issue with this question. A detailed account will be presented of the relation between phonetic data and phonological structures. The answer to the question posed by Van Heuven will be straightforward: the question is ill-conceived, since the experiment shows that not one, but at least two different representations of prosodic prominence need to be distinguished. A non-culminative autosegmental structure with high and low tones accounts for pitch movements in focused constituents, and a culminative metrical structure accounts for the prominence expressed by duration. Both phonetically and phonologically, therefore, prosodic prominence is a nonprimitive notion. 1. The experiment The experiment discussed in Van Heuven (1986) includes 10 bisyllabic adjectives of five different classes: (1) a. strongly rising (0-1): b. c. d. e.

gemengd (x3m£t]t, 'mixed') gepast (xSpdst, 'fit') lightly rising (2-1): bizar (bizar, 'bizarre') concreet (kOJJkret, 'concrete') equal (1-1): lichtgrije (HxtYrCis, 'light grey') beeldschoon (beltsxon, 'very pretty') lightly falling (1-2): komisch (korais, 'comical') logisch (loyis, 'logical') strongly falling (1-0): pittig (pltSx, 'spicy') mager (may3r, 'meagre')

114 These adjectives were embedded at the position of the dots in ' + focus contexts' (2) and focus contexts' (3). The variable places of stress are represented by differently shaded blocks, a representation also used in the discussion to follow. (2) + Focus context a. Stress neither left nor right • : Hil je ... een keer zeggen ('Would you ... once more say') b. Stress left only : Hil je heel ... een keer zeggen ('Would you quite ... once more say') c. Stress right only Q : Wil je ... ding een keer zeggen ('Would you ... thing once more say') d. Stress both left and right Wil je heel ... ding een keer zeggen ('Would you quite ... thing once more say') (3) - Focus context a. Stress neither left nor right • : Wil je ... een beetje HARder zeggen ('Would you ... a little LOUder say') b. Stress left only |] : Wil je heel ... een beetje HARder zeggen c. Stress right only [] : Wil je ... ding een beetje HARder zeggen d. Stress both left and right Wil je heel ... ding een beetje HARder zeggen The complete set of test sentences, read twice in random order by one male and one female speaker, was recorded. For each of the syllables in the adjectives acoustic measurements were performed on the basis of which duration and pitch excursion differences between the two syllables were established. The measurements were converted to relative difference measures. The duration difference was calculated by dividing the duration of the longer vowel in the word by that of the shorter, and subtracting 1. The result (%) was given a negative sign if the first vowel was shorter than the second. The pitch excursion difference (i.e. pitch span) was calculated by subtracting the smaller excursion from the larger one in the word, again with a negative sign if the first syllable contained the smaller value. (See Van Heuven 19B6, 1987 for further details.) The experiment showed only minor differences between the two rising classes and the two falling classes. Henceforth we will ignore this distinction, in order not to complicate the discussion. (Different metrical structures at the level of feet or below could be given, but the distinction does not seem to play a role in the rules to be discussed here.) Figures (4) and (5) summarize the results of the experiment in so far as pitch span and duration are concerned. The original figures included information on intensity differences as well, since intensity will play no role in the discussion here, this part of the figures has been eliminated. Rising patterns are in the left half of the pictures; falling ones are in the right half. In the remainder of this paper, both duration and pitch will be called rising and falling. Duration is called rising if the second syllable of the adjective is longer than the first; pitch is called rising if the second syllable of the adjective is higher than the first.

115 (4) Difference in pitch span between syllables (semitones) a. in + focus contexts (2) rising

V V

equal

7

T A A

^

falling -5 b.in

- focus contexts (3) V

rising

t7

^A

equal falling

.

.

-5

•

.

0

Difference in vowel duration between syllables,% a. in + f o c u s contexts (2) rising

T

V7V

equal falling

o • -100

0

Oo 100

b. in - focus contexts (3) rising

V

TVV

equal falling

• -100

Legend to (4) and (5) V A

rising adjectives equal adjectives falling adjectives

0

Ooo 100

• stress neither left nor right [] stress left only [J stress right only | stress both left and right

116 Observe that the findings in (4) and (5) must be interpreted relatively: the point of equal pitch span and duration need not be represented by zero, but may vary for each group of adjectives, since the figures are not corrected for segmental differences between the groups of adjectives. In other words, inherent pitch and duration differences between the first and second syllable of the adjectives may have influenced the results. This fact complicates the discussion considerably, but fortunately will do so only in the case of duration. As will be shown in 2.1., there is evidence that zero in (4) indicates the point of equal pitch span division. No such conclusion is possible for the duration figures (5). Strictly speaking, for the data in (5) only the difference in distribution within one group of adjectives may be subjected to further discussion. Nevertheless, the fact that rising adjectives are more to the left and falling ones are more to the right will be taken to indicate that the point of equal duration will be somewhere in between, though not necessarily at the same point for both classes of adjectives. With these preliminaries in mind, the corresponding phonological rules will be discussed in the next section.

2. Discussion of the results Publications on prosodies in generative phonology are exclusively concerned with metrical structure (Nespor and Vogel 1986, Baart 1987, Halle and Vergnaud 1987, Kager 1989, Visch 1989), or tonal structure (Pierrehumbert 1980, Gussenhoven 1984). Implicit in such work is the hypothesis that there exists a notion of 'prosodic prominence', and that one representation, metrical or autosegmental, will be able to account for prosodic prominence. Two separate lines of research have thus been developed. Selkirk (1984) forms an exception to the rule. Within her metrical account, she draws attention to the autosegmental analysis of pitch movements by Pierrehumbert (1980), and presents some phonological rules for the relation between metrical structure and pitch accents. Selkirk, however, claims that pitch movements affect metrical structure (cf. the Pitch Accent Prominence Rule, 1984: 276). The data presented in the figures (4) and (5) do not support this. If pitch span and duration were different phonetic realizations of the same phonological representation, they would display a parallel behaviour in all contexts. No such parallel behaviour is observed, as illustrated for instance by the rising adjectives, which display a mirror image distribution of white • and black • contexts in (4a) and (5a). Moreover, pitch movements disappear in a context where duration differences still hold (figure (4b) versus (5b)), which is strong evidence for the independence of pitch and duration. In the next two paragraphs a detailed account of the distribution in (4) and (5) will be presented on the basis of a set of rules which are neither new nor surprisingly exotic. For the sake of clarity, a description of these rules will be proposed, but no doubt future research will show that these descriptions need refinements. 2.1. Pitch accent Pitch accents mark constituents with focus, i.e. those parts of an utterance which contain new information. The experiment shows that no difference between adjective classes can be observed in the - focus context (4b). The fact that all adjective classes in all contexts centre around zero suggests

117 that no pitch movement of Importance occurs here, and that zero in the pitch span figures (4) is indeed the point of 'equal pitch'. Pattern (4b) is perhaps spectacular, but the reason for this pattern is less exciting: 'nothing happens' in the adjectives, since a tone movement occurs much later, in the following phrase (sen beetje HARder, 'a little LOUder'). Perhaps in such an area declination (or downdrift) occurs, which may explain why most adjectives are to the right of zero in (4). The distribution in (4b) is uniform, and thus receives a uniform explanation. The distribution of pitch contours in (4a), the + focus contexts, is much more complex, but will also be explained by a uniform treatment* Notice that in the b- and d-contexts, the experiment may have been ambiguous between a reading in which only heel is in focus and a reading in which the complete phrase heel Adj. (ding) is in focus. In what follows, I assume the latter. The constituents within the focused constituent in which pitch accent will occur is partly determined by syntax, and partly by phonology. Assume that the pitch accents involved in the experiment are high tones, and that the following rule accounts for the place of these toneB: (6) Realization of focus The constituent in focus receives a non-floating high tone (H*). This tone is associated with the metrical head of the syntactic head of this constituent. The place of such high tones is thus dependent on both metrical structure and syntactic structure, since both metrical and syntactic headedness are involved. Examples of the effect of (6) will be presented below. In the contexts (2a) and (2b), adjective phrases (AP) are in focus; in (2c) and (2d) noun phrases (NP) are in focus: cf. the structures in (7). (7) a.

rpi

AP

NP

rpi

AP

rKPn

d.

r " ~ i

"

AP heel

ding

A heel

ding

First consider the NPs: in (7c-d) the syntactic head is the noun ding, 'thing', which receives a high tone (H*) by (6). The adjectives preceding this noun will be realized with another tone, represented in (8) by the variable X. (8) Tonal contours of NPs in focus X (heel) (heel) (heel) (heel) (heel) (heel)

H* biconlichtbeeldkolo-

zar creet grija schoon misch gisch

ding ding ding ding ding ding

Whether X is L or H, the syllables associated consecutively will be lower by declination, predicting a slightly falling pattern for all classes of adjectives. On the basis of representation (8) one expects all adjectives within NPs to

118 be in the vicinity of point zero, which is by and large what the experiment produced, cf. the figures with black right halves [J and • in (4a). That some are not as close to zero as in (4b) may be explained by the fact that the adjectives in the + focus contexts are near to the boundary of tone X, which may be an area of independent tonal movements. Next consider the tonal contours associated with the APs in focus. Focus will again be realized on the syntactic heads, which are the adjectives to be filled in on the dots in (7a) and (7b). Since these adjectives are polysyllabic, metrical structure determines with which syllable H* will be associated. The metrical structures of the adjectives are given under (9) (M abbreviates for 'mot', word; 2 is foot; s and v are strong and weak respectively): (9) a. Rising adj.

M

ri

Compound adj.

M

c. Falling adj.

r i

bi zar con creet

M I

r*~\

licht grijs beeld schoon

o ,8

o |W

ko lo

misch gisch

The structures of the rising and falling adjectives in (9) are noncontroversial, but structure (9b) for compound adjectives has been subject to some debate. The alternative is to consider compound adjectives as phrasal constituents, also with a stronger rightmost constituent (see Trommelen and Zonneveld (1986), and the dismissal of this proposal by Backhuys (1989)). The choice is of no importance to the argument developed here, since in either case the righthand member of the compound adjective will be associated with H*: in (9b) because the righthand member is the metrical head — in a phrasal structure because the righthand member is the syntactic head. The tonal representation of compound adjectives is thus similar to the tonal representation of rising adjectives, cf. (10a), and falling adjectives are represented as in (10b): (10) Tonal contours of APs in focus a. Rising and compound adjectives X H* X

biconlichtbeeld-

zar creet grijs schoon

b. Falling H*

H*

heel heel heel heel

biconlichtbeeld-

zar creet grijs schoon

heel heel

ko- misch lo- gisch

adjectives

X

ko- misch lo- gisch

Given that X is lower than H*, these structures correctly predict the rising and falling patterns of (4a) (cf. the distribution of these adjectives in the and [] contexts). The fact that compound adjectives in these contexts • pattern with rising adjectives is predicted by their similar tonal representation.

119 2.2. Duration Assume that metrical structure represents durational properties of an utterance: both the position of pauses and the relative duration must then be predicted by metrical structure. The experiment included measurements of duration differences between the first and second syllable in different contexts. Consider first the test results for the rising and falling adjectives: the strong feet in the rising adjectives and the strong syllables of the falling adjectives are relatively longer. These results confirm the intuitive idea that strong and weak in the metrical tree correspond with more or less length. Not unexpectedly, falling and rising adjectives display a similar pattern in the + and - focus contexts. Consider next the patterns in some more detail. The rising and falling adjectives show a remarkable opposition: rising adjectives are more rising in the black • contexts (2d) and (3d); falling adjectives are more falling in the white • contexts (2a) and (3a). Why is this so? An explanation can be achieved by taking phonological phrasing into account. Observe that the test sentences contain a piece of metalanguage, (heel) ... (ding), which presumably forms a separate phonological phrase. The utterances tested thus consist of three phonological phrases (), as indicated in (11). (11) Wil je

3y[MAN(y) & b. Vx[HUMAN(x) ~> 3y[MAN(y) & c. Vx[HUMAN(x) --> 3y[MAN(y) &

likes him TELL(LKE(x)(MARY)) (y)(x)]] TELL(LKE(v)(MARY)) (y)(x)]] TELL(LIKE(he)(MARY))(y)(x)]]

d. 3y[MAN(y) & Vx[HUMAN(x) - > TELL(LIKE(x)(MARY)) (y)(x)]] e. 3y[MAN(y) & Vx[HUMAN(x) - > TELL(LIKE(v)(MARY)) (y)(x)]] f. 3y[MAN(y) & Vx[HUMAN(x) --> TELL(LIKE(he)(MARY))(y)(x)]] Each group is subdivided with regard to the interpretation of the pronoun 'him', which can be interpreted in three different ways, respectively related to 'everyone', to 'a man' or to some other discourse referent not mentioned in the sentence. (In the latter case the pronoun is represented by a special variable 'he'-) 3 The interpretation of (25) strongly depends on the interpretation of (24). A first incomplete description of the semantic content of (25) will be twofold, depending on the order of quantifiers. (25) everyone told a boy that Suzy does. A. Vx,[HUMAN(x,) - > 3y,[BOY(y,) & TELL(does(SUZY))(y,)(x,)]] B. 3y,[BOY(y,) & Vx,[HUMAN(x,) ~> TELL(does(SUZY))(y1)(x1)]] Now it turns out that the interpretation of (25) must be constructed on the basis of its parallelism with (24). First of all, as in the foregoing examples, the VP anaphor induces structural parallelism: the quantifier structure in the interpretations of the two sentences must be the same. The first interpretation for (25), (25)A., is only possible if (24) has been assigned an interpretation from the first group; interpretation (25)B. is possible only if (24) has received an interpretation from the second group. Note that parallel quantifier structures cannot be obtained by means of some notion of copying: both clauses have their own quantifier structure which will have to be 'checked' for the required parallelism. Using FMG, quantifier distribution is attached to the main

127 verb so the semantics of the main verbs of the clauses involved will have to match to obtain parallel quantifier distribution. This example shows that notions of copying or deletion of verb )hrases are too coarse to account for VP anaphora because clausal contexts need to be checked or certain properties. Secondly it is important to notice that the interpretation of the pronoun 'him' in (24) forces a parallel interpretation on (25). For instance if 'him' refers to some discourse entity 'John' (interpretations c. or f.), the pronoun in the resolution of the VP anaphor must refer to the same entity 'John'. Moreover, if in (24) the pronoun 'him' is interpreted as being bound by 'a man' (interpretations b. or e.), the interpretation of (25) will have to be parallel, binding the pronoun in the interpretation of the VP anaphor to 'a boy'. In order to account for this indexical parallelism, the binding structures (particularly which quantifier binds the pronoun) of the clauses involved have to be parallel. It is not possible to treat the binding structure of a clause as a property of the verb, i.e. these parallelisms cannot be 'attached to the verb' in an FMG way. We conclude that semantic constraints on ellipsis cannot be limited to VP identity. Complicated, embedded, VP anaphora in an ambiguous quantifier structure induce clausal parallelism concerning quantifier- and binding-structures. In order to account for this parallelism, we propose to match the syntactic/semantic structures of the clauses involved.

i

3 Matching of Clausal Structures Having established the relevance of clausal parallelism for the resolution of VP anaphora, we now propose a computational mechanism to account for these observations: a matching mechanism that works on the syntactic/semantic structures of the clauses. The general idea of this matching mechanism is to provide a precise means to calculate what two (structured) clauses have in common. If we would only think about sentence-level syntax and semantics, a matching of clausal structures might seem a rather drastic extension of the required machinery. But of course we must realize that VP anaphor resolution is in fact a discourse-level process, and that the mechanism for accomplishing it should be embedded in a mechanism which is capable of assigning a structure and an interpretation to a discourse. From the discourse perspective, the situation looks rather different, and more attractive: it turns out that one of the structural relations which have been postulated to play an important role at the discourse level is exactly the relation of syntactic/semantic parallelism, and that the process that is necessary for VP anaphor resolution is parasitic upon the matching process that is needed for establishing discourse structure and discourse coherence. Syntactic/semantic parallelism, i.e. clauses sharing syntactic/semantic structure, is a well-known notion in the discourse literature. It has been invoked to account for pronoun resolution (see for instance Cowan (1980) and Kameyama (1986)), for topic/focus articulation, and for coherence (Polanyi (1985)). In Scha (1988) it is assumed that parallelism plays a significant role in the structuring of discourse. Despite the importance of this notion, no one, as far as we know, has come up with a precise formalization of this concept. (A first attempt can be found in Scha (1988), where semantic arallelism is embodied in tne grammar rules. However, this is not worked out in much detail.)

?he formalization that we propose now is based on the notion of 'generalization'. Generalization

is a computational notion that is the dual of 'unification' (Cf. Robinson (1965)). Given two structures S! and S2, a generalization of those structures is a third structure S of which both S! and S2 are instances. In other words, structure S is at least as general as both S, and S2. We want to employ the notion of generalization to compute a formula which indicates what two structures have in common. We shall do this by means of the standard notion of 'most specific eneralization' (msg). ormally, generalization is based on 'antisubstitution' (Cf. Knight (1989)). An antisubstitution y is a mapping of terms into variables. Two terms a and b are generalizable if there exists an antisubstitution y such that y(a)=y(b). In that case 7(a) is called a generalization of a and b. A generalizer y of terms a and b is called the most specific generalizer of a and b if for any other generalizer T|, there is an antisubstitution £ such that £fy(a))=ri(a). In the latter case, 7(a) is the most specific generalization of a and b. The maximalization of common information of structures that we pursue in our concept of generalization differs somewhat from the standard concept of 'most specific generalization'.

128 Therefore we introduce the notion of 'most specific common ground' (mscg). The mscg is the generalization over structures that extracts the shared structure of clauses as specifically as possible. This especially effects underspecified elements such as anaphora. In the semantic representations, anaphoric terms that are discourse referent searchers (such as pronouns and VP anaphors) are represented by special context variables. A context variable is semantically a constant, but dependent on the context for its interpretation. Though they are constants in the semantic representations, they are considered to be variables from a computational point of view; in the matching process context variables may be instantiated, making them coreferent with their antecedents.4 A context variable is a placeholder for future instantiation. Another common way to look at a context variable is to regard it as the set of all its possible instances. From that point of view, a clause that contains an underspecified element, such as a VP anaphor, is considered a manifold ambiguous clause: Each possible instance of the variable gives an interpretation. This entails that when anaphors are involved, the search for common information between a pair of clauses results in a set of msg's. The most specific common ground of two structures S, and Sj selects that msg that incorporates the maximum of common information. In particular, the mscg of an anaphoric clause, being a manifold ambiguous clause, and its antecedent clause is the most specific of their msg's. Formally tne set of msg's of a constant c and a context variable v consists of the msg's of all pairs from (c)xV (if V is the set of all possible instances of v). The most specific common ground is defined as the most specific element in this set, i.e. the constant c. We illustrate the application of the notion of most specific common ground with some examples. Consider the syntactic/semantic structures of clauses (26) and (27), S, and S, respectively. (The simplest (basic) translation for the verb 'like' is used because no quantifier ambiguity is at stake.) (26) John likes Mary (and) (27) Fred likes Mary. S, (LIKE(MARY))(JOHN)

JOHN

(LIKE(MARY))(FRED)

Xx. (LIKE(MARY)Xx)

XyXx.(LIKE(y))(x)

FRED

MARY

Xx.(LIKE(MARY))(x)

XyXx.(LIKE(y))(x)

MARY

What the two subject terms ('John' and 'Fred') share is that they are both quantifying terms. So the subjects generalize to a variable for a quantifying term Cvar,').5 The rest of the structure S, is the same as S2 (alphabetic variance of semantic variables considered). The mscg of these substructures is therefore that very substructure: (LIKE(MARY))(var,)

mscg(S„Sj)

var.

Xx.(LIKE(MARY))(x)

X.y\x.(LIKE(y))(x)

MARY

In example (26)/(27) no anaphors (and therefore no context variables) are involved. Compare, however, the following simple example, which contains a VP anaphor.

129 (26) John likes Mary (and) (31) Fred does (too). S, (LIKE(MARY))(JOHN)

S3

JOHN

FRED

/

\

Xx.(LIKE(MARY))(x)

/

XyXx.(LIKE(y))(x)

\

does(FRED)

/ \

does

MARY

The result of the matching process on these two structures is given below. As described above, the context variable 'does' functions computationally as a variable. The most specific common ground of the substructures of the VP's in (26) CXx.(LIKE(MARY))(x)') and (31) (the context variable 'does') is the substructure labelled 'Xx.(LIKE(MARY))(x)'. mscg(S,,S3)

XyÄjc.(LIKE(y))(x)

MARY

The most specific common ground found in (26)/ (31) is thus the same as the one found in (26)/(27).' More important than the resulting structure is the fact that the VP anaphor is resolved because it is instantiated during the matching process. The computational notion of matching implements the clausal parallelism requirements on the resolution of VP anaphora. We have argued in section 2 that clausal parallelism comprises two aspects: structural parallelism - the quantifier structure of the anaphoric clause parallels the quantifier structure of the antecedent clause - and indexical parallelism - variables can only be 'copied' if a structurally parallel binding quantifier is present (copying of free variables in VP anaphor resolution gives wrong results). These parallelism constraints on the relation between the structured clauses can now be formulated in terms of conditions on the required matching. As we have seen the matching process takes syntactically composed semantic structures of the relevant clauses as input. The fact that only the syntactically composed semantic structures are considered, accounts for the observation (in section 1) that VP anaphora cannot refer to non-constituents. In case of an anaphoric clause, the structure of the anaphoric clause must be subsumed by the structure of the antecedent clause, i.e. there must be a match. For this match the following conditions hold: Syntactic Agreement - The auxiliaries must match.7 Parallelism - The quantifier structure of the anaphoric clause parallels the quantifier structure of antecedent clause. - The binding structures are parallel. (Semantic variables can only be "copied" if a structurally parallel binding quantifier is present.) The parallelism constraint concerning quantifier structure is implemented as a condition on the verbs that take care of quantifier distribution: The (FMG) raising and lowering operations that are applied to the basic translation of the main verb in order to attacn quantifier distribution must be identical. The constraint on the binding structures is a structural condition on the matching. Using the matching mechanism and considering the relevant conditions, VP anaphors are resolved in the required parallel way. We illustrate this with some examples. As we have shown in section 2, the quantifier distribution in the ambiguous sentence (10) (we take (10a) as an example) is incorporated in the semantics of the verb 'love' by raising its translation. This

130 also determines the quantifier structure in the interpretation of the following clause (11). Consider the syntactic/semantic structures (10a) and (11'). (10) Every man loves a woman. (10a) Vx[MAN(x) - > 3y[WOMAN(y) & LOVE(y)(x)]] XQVx[MAN(x)->Q(x)]

XTj.T2(Xx.3y[WOMAN(v) & LOVE(y)(x)])

/

X.T1XT2.T2(Xy.T,(Xx.LOVE(y)(x)))

(11) Every boy does (too). (11') does(XPVx,[BOY(x1)->P(x1)])

/

XPVx,[BOY(x1)->P(x1)]

\

XQ.3v[WOMAN(y) & Q(y)]

\does

When the syntactic/semantic structures of (10) and (11) undergo matching, the result incorporates the relevant interpretation of the verb phrase. In the matching process, the 'does' vanalsle is instantiated by its antecedent VP. Thus the relevant readings of (11) are obtained: parallel quantifier distribution is simply incorporated because it is attached to the antecedent VP (by means of the raised translation of the verb). The condition on binding structures is irrelevant in this example. The theory makes essential use of the compositionality of the FMG semantics in that the complete semantics of the anaphoric clause depends on the compositional construction of both clauses. Consider again example (24)/(25): (24) Everyone told a man that Mary likes him a. Vx[HUMAN(x) - > 3y[MAN(y) & TELL(LKE(x)(MARY)) (y)(x)]] b. Vx[HUMAN(x) ~> 3y[MAN(y) & TELL(LDCE(y)(MARY)) (y)(x)]] c. Vx[HUMAN(x) ~> 3y[MAN(y) & TELL(LKE(he)(MARY))(y)(x)]] d. 3y[MAN(y) & Vx[HUMAN(x) - > TELL(LDCE(x)(MARY)) (y)(x)]] e. 3y[MAN(y) & Vx[HUMAN(x) - > TELL(LKE(y)(MARY)) (y)(x)]] f. 3y[MAN(y) & Vx[HUMAN(x) - > TELL(LKE(he)(MARY))(y)(x)J] (25) everyone told a boy that Suzy does. A. Vx,[HUMAN(x,) - > 3 yi [BOY( yi ) & TELL(does(SUZY))(y,)(x1)]] B. 3y,[BOY(y,) & Vx,[HUMAN(x,) - > TELL(does(SUZY))(y,)(x,)]] The VP anaphor can be resolved only if the syntactic constraints are fulfilled (which is clearly the case) and, if, moreover, quantifier and binding structure are parallel. As before, quantifier structure is attached to the verb. Contrary to the foregoing example, the VP anaphor is situated in an embedded structure. The requirement of parallel quantifier structure functions as a condition on the matching of the main verbs of the clauses. The condition on binding structures in this case results in tne following. All Noun Phrases are treated as quantifiers and are incorporated as such (as being a quantifier) in the mscg. The interpretation of the pronoun in (24) (represented in the syntactic/semantic structures as alternatives x, y and he) is 'passed on' to the anaphoric clause. The 'logical variable' interpretations (represented by x and y) are only passed on to the interpretation of the VP anaphor if a structurally parallel binding quantifier is present, whereas the context variable (he) is only included in the VP anaphor interpretation if it is resolved. Thus non-parallel interpretations are 'filtered' on basis of the formulated conditions.

131 4 Coda The matching mechanism which manipulates clausal syntactic/semantic structures is not an ad hoc machinery for the resolution of VP anaphora. It is a crucial mechanism in the construction of discourse structure and discourse semantics. In the integration of a sentence in its context, matching provides a formal means for tracing coherence and building discourse structure and discourse semantics. In fact, the resolution of (both pronominal and) VP anaphora can be seen as a side-effect of the process of building discourse structure using matching of clausal structures. In the current paper we have argued for a matching approach to account for the interplay of clausal parallelism and VP anaphora. We refer to Priist & Scha (1990) for the integration of this treatment of VP anaphora in a general framework for discourse. Notes 1. Although the role of the word 'too' is important, we shall not deal with it in this paper. One of the effects of the word 'too' is that it 'balances' parallelism of subject and verb phrase: The semantics of the subject of the anaphoric clause can not subsume the semantics of the subject in the antecedent clause whereas the semantics of the verb phrase of the second clause must subsume the semantics of the first clause verb phrase. Compare for instance the following two examples (in contrast with (6), example (7) implies that John is no man): (6) John loves a woman. Every man does. (7) John loves a woman. Every man does too. 2. Partee (1981, page 464-468) also provides evidence for the fact that ellipsis cannot be characterized in terms of identity of predication. 3. In principle, a pronoun receives a twofold translation: it is translated into a special context variable (he.he^he,...) and into a standard (logical) variable (x,y,x„yi,...). The latter, of course, only when it is bound by a quantifier. This twofold translation reflects the dichotomy of 'bound variable' versus 'referential' pronouns. The overhead due to this twofold translation of pronouns can be reduced substantially on the basis of syntactic constraints on the distribution of bound variable versus referential pronouns (Cf. Reinhart (1983)). 4. Context variables are also called 'context-dependent constants' (Cf. Janssen (1980)). We use 'does', 'he', 'he/ etc. in order not to clutter the semantic structures with types. 5. Of course the subject terms of (26) and (27) also share the information that they both refer to a male person. The standard notion of generalization is based on a rather poor hierarchy of terms as it always introduces variables when some information conflicts, thereby loosing further common information. For instance the terms 'John' and 'Fred' generalize to a variable of the appropriate type, thereby loosing the information that both terms refer to a male person. We shall assume a more sophisticated notion of generalization that incorporates this kind of lexical information. Such a notion would also provide a means to formalize the notion of coherence in the sense of (Polanyi (1985)). The latter notion covers, for example, the fact that sentences (28) and (29) both describe 'physical attributes' of John (Polanyi (1985), page 316). (28) John is a blond. (29) He weighs about 215. (30) He's got a very nice disposition. The contribution of (30), in this respect, results in the establishment that (28) through (30) all describe 'generally known attributes' of John. Such a notion of coherence can be moulded into a concrete form by computing the common ground of clauses on the basis of a lexical thesaurus. 6. In a flat notation: the result of matching [s LKE(MARY)XJOHN)[OT JOHN] [ w Xx.(LIKE(MARY))(x) [v XyXx.(LIKE(y))(x)] and [ s does(FRED) U FRED] U doesl

MARY]]]

132 is: U var,

U var,][y, Xx.(LEKE(MARY))(x) [ v \y/U.(LIKE(y))(x)] [OT MARY]]]

7. This constraint also accounts for facts such as in (31)/(32). (31) John must drive 100 miles an hour. (32) Sam must too. Due to ambiguity of the auxiliary, (31) has two readings (the 'permission' reading and the 'possibility' reading, Cf. Lightfoot (1982)). The sequence (31)/(32), however, has only two readings (and not four) because parallelism forces the same interpretation of the auxiliary. (N.B. Whether modals are considered part of the VP or not (Cf. Lightfoot (1982)) does not affect our account of ellipsis because the parallelism constraints are formulated for adjacent clauses.) References COWAN, J. 1980 The Significance of Parallel Function in the Assignment of Intrasentential Anaphora in: Papers from the Parasession on Pronouns and Anaphora, CLS 1980, page 110 - 120 JANSSEN, T.M.V. 1980 On Problems Concerning the Quantification Rules in Montague Grammar in: Time. Tense and Quantification: Proceedings of the Stuttgart Conference on Logic of Tense and Quantification. Rohrer (ed). page 111 - 134. Max Niemeyer Verlag. Tübingen HENDRIKS, H. 1987 Flexibele Montague Grammatika in GLOT 10 KAMEYAMA, M. 1986 A property sharing constraint in centering in: Proceedings of the ACL, page 200 - 207 KLEIN, E. 1987 VP Ellipsis in DR Theory in Studies in Discourse Representation Theory and the Theory of Generalized Quantifiers. Groenendiik, de Jongh & Stokhof (eds), Groningen-Amsterdam Sudies in Semantics (GRASS), Foris, Dordrecht KNIGHT, K. 1989 Unification: A Multidisciplinary Survey in ACM Computing Surveys. Vol.21, No.l LIGHFOOT, D.W. 1982 The Language Lottery: Toward a Biology of Grammar, MIT Press, London PARTEE, B. & BACH, E. 1981 Quantification, pronouns and VP anaphora in: Formal methods in the study of language, Groenendijk, Stokhof en Janssen (eds), Mathematisch Centrum, Amsterdam POLANYI, L. 1985 A Theory of Discourse Structure and Discourse Coherence in: Papers from the general session of the Chicago Linguistic Society. CLS 21, page 306 - 322 PRÜST, H. & SCHA, R. 1990 A Discourse Perspective on VP Anaphora, in Proceedings of the 7th Amsterdam Colloquium. University of Amsterdam: Institute for Language, Logic and Information REINHART, T. 1983 Coreference and Bound Anaphora in: Linguistics and Philosophy 6, page 47-88. ROBINSON, J.A. 1965 A Machine-Oriented Logic Based on the Resolution Principle, in: Tournal of the ACM 12(1)

SAG, I. 1977 Deletion and Logical Form, Ph.D. Thesis University of Pennsylvania SCHA, R. & POLANYI, L. 1988 An Augmented Contextfree Grammar for Discourse, in Proceedings of the 12th International Conference on Computational Linguistics (COLING), page 22-27 WILLIAMS, E. 1977 Discourse and Logical Form in: Linguistic Inquiry Volume 8 Number 1. page 101-139

Willebrord

Sluyters

The Derivation of Geminate Blockage Effects 0. Introduction It has been observed repeatedly that geminates behave as single units with respect to the application of phonological rules. For example, geminates are seen not to undergo epenthesis rules, nor may rules apply to only one part of a geminate. This phenomenon is usually referred to as "geminate integrity". Various authors have argued that integrity is a natural consequence of the autosegmental representation of geminates and of constraints regarding autosegmental association (for an overview, cf. Schein and Steriade 1986; for criticism and revision of earlier approaches, cf. Sluyters 1990a). Another peculiar aspect of geminates is that they often block the application of a phonological rule, without an obvious reason. Whereas the concept of integrity forces a rule to apply simultaneously to both parts of a geminate, it does not predict whether a rule will or will not apply to a geminate. The phenomenon of an otherwise unmotivated non-application of a rule is known as "geminate inalterability" or "geminate blockage" (I will use the latter term). Hayes' Linking Constraint (1986, henceforth LC) and Schein and Steriade's Uniform Applicability Condition (1986, henceforth UAC) are principles proposed to account for this anomalous behavior. In section 1 of this paper data on geminate blockage will be presented. The discussion in section 2 will make clear that both principles mentioned above are empirically adequate, because they correctly predict blockage and nonblockage in these cases. Nevertheless, I will argue that for various reasons they should be dispensed with. In section 3 a proposal will be worked out to derive geminate blockage from a feature percolation mechanism and the Obligatory Contour Principle, both of which must be part of UG to account for other phenomena. 1. Data. A particularly interesting case, which illustrates blockage by partial geminates and non-blockage by full geminates at the same time, is the first stage of the set of changes known as the 'Old High German Consonant Shift', which took place in about the 6"1 century a.C. (data from Penzl 1969:65-67, Braune and Eggers 1975:86-88,156 and Davidsen-Nielsen 1976-'78); asterisks indicate reconstructed forms). I will limit the discussion to the development of /p/, /t/ and /k/ (that of /b/, /d/, /g/ is irrelevant to the point to be made).

(1)

a.

Germanic op an slapan etan latan makon tekan

*[op"pl,an] *[ slaphphan] »[etVan] *[lathtl,an] *[makVon] *[tsekhkhan]

Old High German [offan] [slaffan] [ezzan] [lazzan] [ma/xon] [tsefyxan]

'open' 'to sleep' 'to eat' 'to let' 'to make' 'sign'

134 (table (1) continued) b. skip hwat ik c. helpan thorp herta holt werk d. plegan tiohan korn e. skeppian settian wekkian f. spilon stark scat naht fehtan luft craft

*[skiph] *[*vath] *[ikh] *[help"an] • t^orp"] *[hert"a] *[holth] *[werkh] *[p"legan] *[thioxan] *[khorn] *[skepVjan] *[sethtl,jan] *[wekVjan] *[ spilon] *[ stark"] *[skath] *[na/t] *[fejftan] *[luft] *[kraft]

[skif] [was] [IX] [helpfan] [t"orpf] [hertsa] [holts] [werk*] [pflegan] [tsioxan] [k/orn] [skeppfan] [settsan] [wekkjfan] [spilon] [stark*] [skats] [najt] [fehtan] [luft] [kraft]

'ship' 'what' 'I' 'to help' 'village' 'heart' 'wood' 'work' 'to take care 'to pull* 'corn' 'to create' 'to sit' to wake' 'to play' 'strong' 'coin' 'night' 'to fight' 'air' 'force'

A few comments are in order. Germanic /p/, /t/ and /k/ show up in Old High German as fricatives (la,b), affricates (lc,d,e) or remain unchanged (If). All commentators agree (cf. Penzl 1969:66 and Braune and Eggers 1975:86-87, fn.l) that aspiration constituted the opening step in the development to fricatives and affricates. In view of the exceptional behavior of the clusters in (If)1, this is quite plausible. It is often observed that aspiration rules do not apply to clusters that have identical laryngeal features. For example, in the First Germanic Consonant Shift (or Grimm's Law, about 5th century b.C.) /p/, /t/ and /k/ became aspirates as well and here the clusters /sp/, /st/, /sk/, /pt/ and /kt/ did not participate in the development (Penzl 1969:44-46 and DavidsenNielsen 1976-'78:46). Also, the Ancient Greek /s/^/h/ rule is blocked in case of a cluster with homorganic laryngeal features (Uetzels 1986:314-315). Since the aspiration of /p/, /t/ and /k/ in Old High German (and the posterior development of fricatives and affricates) is an otherwise exceptionless process, the behavior of the consonant clusters can only be due to the homorganicity of the laryngeal features. In other words, it must be a case of geminate blockage, caused by a partially shared matrix. At the same time, we observe (cf. (le)) that geminate /p/, /t/ and /k/ do undergo aspiration. I will consider the aspiration stage only, not the subsequent developments. Since all /p/, /t/, /k/ (except for those which are part of a cluster) underwent this process, geminates included, there may not be any prosodic condition on the target segment and since no adjacency requirement is necessary, we need not refer to the skeleton either (the relevance of these points will become clear in section 2). I assume a feature geometry along the lines of Clements 1989, with but one difference: [conti] is, non-crucially as far as the present paper is concerned, assumed to dominate the place features, as proposed in Wetzels 1990. (2)

.___root laryng^..^-^-"'"^ [-voice] [+spfead]

.. supralaryng ! oral [-cinti]

135 L e t us subsequently consider a case of b l o c k a g e involving full geminates. In T u s c a n Italian, the postlexical rule of 'Gorgia Toscana' affects the voiceless , [0] a n d [h] respecplosives /p/, / t / a n d /k/, w h i c h become continuants tively. G o r g i a T o s c a n a applies w o r d i n t e r n a l l y as w e l l as across wordboundaries. The process shows a c e r t a i n degree of geographical a n d social v a r i a b i l i t y w h i c h n e e d n o t concern us here. The data in (3) b e l o w have b e e n taken f r o m Tekavcic 1972 a n d Nespor a n d V o g e l 1986, occasionally supplemented w i t h m y o w n observations .

/sapere/ /la#pelle/ /prato/ /la#torre/ /amiko/ /la#kaza/ /lampa/ /kolpo/ /raspo/ /ltnto/ /molto/ /posto/ /stai^ko/ /parko/ /taska/ /awtostop/ /kapsule/ /dzfnit/ /ritmo/ // ik/ /tekniko/ /tsoppo/ /latte/ /bokka/

[sapere] [lamelle] [prado] [laSorre] [amiho] [laha: za] *[lampa] *[kolpo] *[raspo] *[UnSo] *[mol6o] *[pos6o] *[sta9ho] *[parho] *[tasha j *[awtostop] *[kapsule] *[dzini0] *[ri6mo] *[iih] *[t thnlko] 2 *[tsoyjoo] *[la6Qe] *[bohha]

' to k n o w ' 'the skin' 'greenfield' 'the tower' 'friend' 'the house' 'lamp' 'beat' 'frog' 'slow' 'much' 'place' 'tired' 'park' 'pocket' 'hitch-hike' ' capsule' 'zenit' 'rhythm' 'chique' 'technician' 'criple' 'milk' 'mouth'

The examples in (3a) show the regular a p p l i c a t i o n of G o r g i a Toscana, whereas those in (3b) indicate that the focus of the rule m u s t b e immediately p r e c e d e c b y a vowel. The process is formally stated in (4) below (cf. also Nespor ana Vogel 1986; I assume that / k / first b e c o m e s * / x / and t h e n [h] b y a later rule w h i c h delinks the v e l a r place node).

W

X ! [+vocoid]

^ ! root [+consj [ - v o i c e [ s u p r a l ] ! [oral] [ -conti] ->[+conti] condition: X, Xs adjacent, Xj in syllable onset

Y e t another case involving geminate blockage is Gallo-Romance lenition, w h i c h has b e e n extensively discussed in Jacobs a n d Wetzels 1988. For reasons of space, 1 cannot present all data o n this complex set of interconnected changes, b u t will h a v e to limit the discussion to one a s p e c t of lenition, the spirantization o f v o i c e d stops (the following data have b e e n taken from J a c o b s and W e t z e l s 1988; asterisks indicate reconstructed forms).

136 (5) a.

b.

c. d. e.

f.

Classical Latin habere nudum negare barbam tardum largum librum nigrum Rhodanum membrum mordere ungulam abbam addubbare

Gallo-Romance * I a/3ere ] *[nuSu] *[n^are] * [barba] *[tarduj *[largu] *[lißru] *[ne£ru] *[roonuj *[mtmbru] *[m ardrè] *[ui}gla] *[abba] *[addobberC]

' to have' 'naked' 'to deny' 'beard' 'late' 'large' 'book' 'black' 'Rhone' (top.) 'part, member' 'to bite' 'nail' 'abbot' 'to knight someone

J a c o b s a n d Wetzels note that the v o i c e d stops /b/, / d / a n d / g / are s p i r a n t i z e d if they are immediately p r e c e d e d b y a v o w e l (5a,c,d), b u t n o t if they are prec e d e d b y a consonant (5b,e), nor if the v o i c e d stop is a geminate (5f). There is a difference w i t h the T u s c a n case d i s c u s s e d above: whereas G o r g i a T o s c a n a requires the stipulation 'in syllable onset', the Gallo-Romance S p i r a n t i z a t i o n rule s h o u l d stipulate o n l y the adjacency of the target segment to a p r e c e d i n g vowel, because the syllable final stops i n (5d) participate in the process. I formalize the following rule (cf. Jacobs a n d Wetzels 1988:113). (6)

X, J [+vocoid]

X, 1 root [+conSj l+voISe"]^"—.. [supral] ( [oral] [ - conti ] +conti ] condition: X, Xj adjacent

2. The d e r i v a t i o n of blockage effects b y the LC and the UAC Let us consider the w a y in w h i c h the LC a n d the UAC derive the blockage effects o b s e r v e d in the case of O l d H i g h G e r m a n Aspiration, G o r g i a Toscana a n d GalloRomance Spirantization.

(7)

(8)

Linking Constraint (LC) A s s o c i a t i o n lines in structural haustively.

descriptions

must be

interpreted

ex-

U n i f o r m A p p l i c a b i l i t y Condition (UAC) G i v e n a node n, a set S consisting o f all nodes linked to n o n some tier T, a n d a rule R that alters the content of n: a c o n d i t i o n in the structural d e s c r i p t i o n of R o n any member of S is a c o n d i t i o n o n every member of S.

N o n - a p p l i c a t i o n of O l d H i g h G e r m a n A s p i r a t i o n to /sp/ etc. follows straightforw a r d l y from b o t h principles, because these clusters have shared laryngeal features .

137 (9)

/f.s.x/

/p.t.V root laryng

[-voice] The LC dictates exhaustive interpretation of association lines. The association line between the laryngeal node and the root is crucially mentioned in the statement of aspiration (cf. (2)), the same laryngeal node is also linked to the root of /f,s,)(/. Therefore, aspiration will not affect the clusters. The UAC dictates that both parts of the geminate structure satisfy simultaneouslv the conditions imposed by a rule. The laryngeal node is linked to a segment which satisfies the description of aspiration (/p/, /t/, /k/), but also to one which does not satisfy it (/f,s,jf/, since the rule imposes [-conti] on its target). Again, aspiration will not apply to the clusters. At the same time, full geminates are free to undergo the rule, because they are represented with one feature matrix. The only double association line in a full geminate is that of the root to the skeleton, but, as we observed, the skeleton is irrelevant to the rule. Correct results are also obtained in the Tuscan case. The association line between the skeleton and the root is crucially mentioned in the rule. Therefore, the LC predicts blockage. The UAC predicts blockage, because the second skeletal slot of a geminate does not satisfy the condition "adjacent to a preceding vowel". Obviously, geminate blockage in the case of Gallo-Romance Spirantization follows from the same adjacency condition. From the empirical point of view, then, the LC and the UAC are adequate principles (for the present, I will not examine in detail the differences between the LC and the UAC; for discussion cf. Sluyters 1990a). 3. An OCP-based derivation of blockage efects One might criticize Hayes (1986) and Schein and Steriade (1986) for adding to the theory an otherwise unmotivated principle which, moreover, is highly reminiscent of another principle of UG without formally reflecting this relation. Let us consider once more the representation of a geminate structure, which is as in (10a). (10)

X (a)

Why should a rule whose focus consists of the feature complex [X,-y] or [Xj-y) be blocked by the presence of [X,-y] or [X,-y] respectively? The informal answer to this question seems to be that [X,-y] and [Xj-y] are not autonomous in (10a), whereas they are so in (10b). Since the Obligatory Contour Principle is responsible for the unmarked status of (10a) as compared to (10b), one might argue that it is the OCP which has sacrificed the independence of the complexes [X,-y] and [Xj-y]. A mechanism which is formulated to account for geminate blockage should take this role of the OCP into account. The following is a proposal which aims at doing exactly that. At the end of this section, I will briefly illustrate its empirical consequences. Let me first establish which autosegmental tiers are relevant for the proper application of a given phonological rule. I propose the following convention.

138 (11)

Rule Interpretation Convention A phonological rule P w h i c h refers to the tiers [a,b,c,...] refers to tier [x] w h i c h immediately dominates [a,b,c,...) (where 'immediate' means that [x] does not dominate a tier [y] w h i c h itself dominates [a,b,c,...]). 3

Next, it m u s t be e s t a b l i s h e d w h i c h features or feature complexes are accessible at a given tier. I n i n d e p e n d e n t work (cf. Zubizarreta 1979), it has b e e n suggested that features m a y be percolated, for example, to a c c o u n t for metricallv conditioned h a r m o n y rules. I propose a general p e r c o l a t i o n convention. (12)

Percolation C o n v e n t i o n The content o f any tier is percolated to all tiers dominating it.

The effect of p e r c o l a t i o n is illustrated b y the partial r e p r e s e n t a t i o n in (13). (13)

.

supralaryngeal( [oral([contil [«*labial] etc.)])]} oral( [contil [-.labial] etc.)]) [contil [=