Linguistics in the Netherlands 1989 9783110870060, 9783110133707

Citation preview

LINGUISTICS IN THE NETHERLANDS 1989

AVT PUBLICATIONS A V T Publications is a series s p o n s o r e d by t h e Algemene Vereniging voor Taalwetenschap (Linguistic Society of t h e Netherlands). In addition to t h e a n n u a l p u b l i c a t i o n of Linguistics in the Netherlands f u r t h e r publications, r e s u l t i n g f r o m other activities p r o m o t e d by t h e Society, may appear in t h i s series.

Still

available

Linguistics in the Netherlands 1977-1979 W i m Z o n n e v e l d a n d Fred W e e r m a n (eds) Dfl. 3 0 . — / U S $ 2 1 . 9 0 Linguistics in the Netherlands 1983 Hans B e n n i s a n d W.U.S. v a n Lessen Kloeke(eds) Dfl. 3 4 . — / U S $ 2 1 . 9 0 Linguistics in the Netherlands 1986 Frits B e u k e m a and A a f k e Hulk (eds) Dfl. 3 8 . — / U S $ 2 4 . 9 0 Linguistics

in the Netherlands

1988

Peter C o o p m a n s and A a f k e Hulk (eds) Dfl. 3 8 . — / U S $ 2 4 . 9 0 . —

Language Development Peter J o r d e n s a n d J o s i n e A. L a l l e m a n Dfl. 6 0 . — / U S $ 3 9 . 9 0

m&t&tp- rr

m

Hans Bennis Ans van Kemenade (eds)

1989 FORIS PUBLICATIONS Dordrecht - Holland/Providence RI - 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

CIP-DATA

ISBN 90 6765 420 5 © 1989 Algemene Vereniging voor Taalwetenschap seer.: J. Lalleman, Dutch Studies, Univ. of Leiden, P.O. Box 9515, 2300 RA Leiden. 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 List of contributors Kees-Jan Backhuys Adjectival compounds in Dutch

1

Hans Bennis & Teun Hoekstra PRO and the Binding Theory

11

Renée van Bezooijen Evaluation of an algorithm for the automatic assignment of sentence accents

21

Norbert Corver Left branch extractions and D P

31

Dicky Gilbers How to digitalize phonology

41

Anneke Groos Particle-verbs and adjunction Josée

51

Heemskerk

Morphological parsing and lexical morphology

61

Eric Hoekstra Binding, objects andof the structure of theinEnglish Jarich Hoekstra & Lâszlô Marâcz Some implications I-to-C-movement FrisianVP

71 81

Teun Hoekstra & Hans Bennis A representational theory of empty categories

91

René Kager & Hugo Quené A sentence accentuation algorithm for a Dutch text-tospeech system

101

vi

René Mulder & Pim Wehrmann Locational verbs as unaccusatives

111

I.H. Slis & M. van Heugten Voiced-voiceless distinction in Dutch fricatives

123

Norval Smith, Roberto Bolognesi, Frank van der Leeuw, Jean Rutten & Heleen de Wit Apropos of the Dutch vowel system 21 years on

133

Mieke Trommelen & Wim Zonneveld Stress, diphthongs, r in Dutch

143

Ludo Verhoeven Acquisition of clause linking in Turkish

153

Gertrud de Vries NPs and scope (Un)specified

163

W. Leo Wetzels Sonority, major class features, syllable structure, manner features: How much of the same? The case of Yakut

173

Wim van der Wurff Scope and movement in Bengali negation

185

Preface

This volume contains a selection of the papers presented at the twentieth annual meeting of the Linguistic Society, held in Leiden on January 21, 1989. 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 papers were presented. The nineteen 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. April 1989

Hans Bennis Ans van Kemenade

List of contributors

K-J. Backhuys University of Utrecht, Department of Dutch, Trans 10, 3512 JK Utrecht H. Bennis University of Leiden, Department of Dutch, P.O. Box 9515, 2300 RA Leiden R. van Bezooijen University of Amsterdam, Institute of Phonetic Sciences, Herengracht 338, 1016 CG Amsterdam R. Bolognesi University of Amsterdam, Department of Linguistics, Spuistraat 210, 1012 VT Amsterdam N. Corver University of Brabant, Department of Language and Literature, P.O. Box 90153, 5000 LE Tilburg D. Gilbers University of Groningen, Department of Dutch, Grote Kruisstraat 2-1, 9712 TS Groningen A. Groos University of Utrecht, Research Institute for Language and Speech, Trans 10, 3512 JK Utrecht J. Heemskerk University of Leiden, Department of General Linguistics/Phonetics Laboratory, P.O. Box 9515, 2300 RA Leiden M. van Heugten University of Nijmegen, Department of Language and Speech/Phonetics, P.O. Box 9103, 6500 HD Nijmegen E. Hoekstra University of Groningen, Department of Linguistics, Grote Rozenstraat 31, 9712 TG Groningen J. Hoekstra Frisian Academy, Coulonhüs, Doelestrjitte 8, 8911 DX Ljouwert (Leeuwarden) T. Hoekstra University of Leiden, Department of General Linguistics, P.O. Box 9515, 2300 RA Leiden R. Kager University of Utrecht, Research Institute for Language and Speech, Trans 10, 3512 JK Utrecht

X

F. van der Leeuw University of Amsterdam, Department of Linguistics, Spuistraat 210, 1012 VT Amsterdam L. Maräcz University of Groningen, Department of Linguistics, Grote Rozenstraat 31, 9712 TG Groningen R. Mulder University of Leiden, Department of General Linguistics, P.O. Box 9515, 2300 RA Leiden H. Quene University of Utrecht, Research Institute for Language and Speech, Trans 10, 3512 JK Utrecht J. Rutten University of Amsterdam, Department of Linguistics, Spuistraat 210, 1012 VT Amsterdam I. H. Slis University of Nijmegen, Department of Language and Speech/Phonetics, P.O. Box 9103, 6500 HD Nijmegen N. Smith University of Amsterdam, Department of Linguistics, Spuistraat 210, 1012 VT Amsterdam M. Trommelen University of Utrecht, Research Institute for Language and Speech, Trans 10, 3512 JK Utrecht L. Verhoeven University of Brabant, Department of Language and Literature, P.O. Box 90153, 5000 LE Tilburg G. de Vries University of Brabant, Department of Language and Literature, P.O. Box 90153, 5000 LE Tilburg P. Wehrmann University of Leiden, Department of Dutch, P.O. Box 9515, 2300 RA Leiden W.L. Wetzels University of Nijmegen, Department of French, Erasmusplein 1, 6525 GG Nijmegen H. de Wit University of Amsterdam, Department of Linguistics, Spuistraat 210, 1012 VT Amsterdam W. van der Wurff University of Amsterdam, Department of English, Spuistraat 210, 1012 VT Amsterdam W. Zonneveld University of Utrecht, Department of English, Trans 10,3512 JK Utrecht

Kees-Jan Backhuys

Adjectival compounds in Dutch 0. Introduction The central theme of this article is the morphology of adjectival compounds in Dutch. I start out in section 1. with a number of remarks on compounding in general. The topic of section 2. is a group of adjectives which are traditionally seen as relevant to the topic of A-compounding. I will discuss them by means of Trommelen & Zonneveld (1986), who give an entirely different view of the morphological status of these special formations. The research by Trommelen & Zonneveld leads to some implementations, the validity of which is investigated in the remainder of this article. The Trommelen & Zonneveld-view is placed opposite to two large groups of concatinated adjectives, the constructions with pseudo-suffixes (section 3.) and the large and very productive group of [w s]-labeled adjectives (section 4.) In 5. the question is answered "which is which?", and there I will demonstrate that all three groups in focus here, are in fact A-compounds, but that the group discussed by Trommelen & Zonneveld is not representative for A-compound formation in general. In section 6. the conclusions are given, and I will give a first indication how the insights gained can be implemented in an overall theory of Dutch generative morphology. 1. Nominal compounds: stress and linking morphemes One of the characteristic properties of nominal compounds is that they have a fixed stress pattern: a compound bears stress on its lefthand member: strHndbkl ("beach ball"), ijsbker ("polar bear"), driaidiur ("revolving door"), hikbljl (lit. 'hack axe'), grdotsprkak ("bragging"), duddom ("great-uncle"). Thus, compound stress in Dutch is [s w]. There are only very few, lexicalized, exceptions to this regularity; as a rule examples of these are confined to burgemSester ("mayor"), stadhuis ("city hall") and rookvlSes (litt. 'smoke beef'). Some of these exceptions could very well be 'lexicalised phrases': hdoglSraar ("professor"), flkuwekill ("bull shit"), pldttegrdnd ("street map"), dUbbelepUnt ("colon"), vllegende schdtel ("flying saucer"), stiakt het vuren ("cease-fire"), with a [w s]-pattern, as indicated. These formations form a limited group; although lexicalised phrases have scarcely been part of linguistical research, I dare to state aforehand that this process is not as productive as the very productive process of compound formation. The same [s »r]-pattern can be found in other lexical categories as well, since not only nominals can give rise to compound formation: vSldrijd("to cyclo-cross"), ijsbeer ("to walk about"/litt. 'to polarbear'), zSeziek^ ("seasick"), This leaves us with the following generalization for lSvensmoe("world-weary"). Dutch compound stress: (1)

the stresspattern of compounds is [s tr]

In this discussion I will elaborate on the question if (1), as far as A-compounds are concerned, can be maintained. A second property of compounds is the (internal) occurrence of so-called linking morphemes, -s- or -a-, handel-s-oorlog ("trade-war"), begroting-s-tekort ("bud-

2

get deficit"), bij-s-korf ("bee hive"), plag-a-hut ("turf hut"). Insightful analyses of this p h e n o m e n o n are hard to come by but at least the following appears to be true: (2)

linking morphemes go w i t h lefthand

members of compounds

This can be illustrated by a w e l l - k n o w n D u t c h memory aid: "How do y o u write dorpsstraat ("village street")? W i t h 1 or 2 s-es?" The problem is, that the sound by itself gives no definite answer, because of to the s of straat, combined w i t h the fact that D u t c h has no geminates. The trick is to change the w o r d in question into one w h i c h does enable you to hear the -s-, for instance dorp-s-plein, dorp-s-weg ("village square / road"). If y o u hear an s in these formations, y o u also write it in dorp-s-straat. It is obvious that this m e t h o d is based on the implicit assumption that it is the lefthand m e m b e r of the compound that imposes the intermediate phoneme. Furthermore, (2) is supported by the following observation. D u t c h has (at least) two lexical entries land: the first meaning "nation", for instance "all of H o l land", and the second meaning "country / meadow". There are also several lexical entries zin, one meaning "sentence", and others meaning "reason", "sense" or "feeling". N o w observe that the semantic differences coincide w i t h different linking morphemes: (3)

land-s-grens land-s-belang land-s-man

"national border" "national interest" "compatriot"

vs.

land-grens land-weg land-man

zin-s-bouw zin-s-ontleding zin-s-deel

"syntax" "parsing" "sentence

vs.

"sensual zin-genot zin-tuig "sense" zin-9-beeld "symbol"

constituent"

"country border" "country road" "peasant" pleasure"

I will not elaborate u p o n the question why, in a language such as Dutch, where a large part of the morphology goes by the Righthand Head Rule (cf. Trommelen & Zonneveld (1986)), lefthand members appear to be dominant w i t h respect to linking morphemes, nor will I specify the tendencies that rule the assignment of - s - or - a - . I will confine myself to (2).

2. A-compounds Consider the D u t c h derived adjectives in (4): (4)

aardedonker doodstil flinterdun ijskoud klinkklaar kurkdroog

'earth-dark' 'death-silent' 'wafer-thin' 'ice-cold' 'sound-clear' 1 cork-dry'

oliedom pij lsnel ragf ijn stokoud vlijmscherp witheet

1

oil-stupid' 'arrow-quick' 'cobweb-fine' 'stick-old' 'f1 earn-sharp' 'white-hot'

These adjectives have an 'intensifying' lefthand part. The lefthand m e m b e r s have only a vague semantic contribution; they only supply notions s u c h as "very", "terribly" and so forth. Therefore bloedheet {'blood-hot') means "very hot", where it is irrelevant that bloed is an N, and not generally 'hot'.) Usually the members of this type of adjectives are a 'fixed' pair, and therefore semantically unproblematic formations such as *sneeuwkoud ('snow-cold'), *bijlscherp ('axe-sharp') and *flinterfijn ('wafer-fine') are ungrammatical. A l t h o u g h not intirely unproductive, this type of adjective formation is morphologically (and semantically) limited. It is commonly assumed that the formations in (4) are compounds (Van den Toorn (1983 a,b), Schultink (1980)) and (thus) have compoundstress, [ s w ] : stdkdud,

3 rSgfijn-, see for instance Schultink (1980: 211) and the various lemmata in Van Dale (the most authorative dictionary for the Dutch language). For years this view of the facts has been the accepted one. A-compounds were considered, as far as stress was concerned, identical to N-compounds. Notice that these stolcoud-adjectives, as I will refer to them, also contain linking morphemes (5a); and that the righthand m e m b e r itself may be suffixed (5b). The relevance of this will become clear later. (5) a.

b.

aap-a-trots beer-a-sterk boord-a-vol dood-s-bang duim-3-dik

'ape-proud' 'bear-strong' 'brim-full' 'death-scared' 'inch-thick'

hemel-s-blauw hond-s-brutaal huiz-a-hoog raaf-a-zwart steek-o-blind

'heaven-blue' 'dog-brazen' 'house-high' 'raven-black' 'itch-blind'

brood-[nood-ig] dol-[geluk-ig]

'bread-necessary' ' mad -happy'

goud-[eer-lijk] kots-[mis-el ijk]

'gold-honest' 'puke-sick'

Trommelen & Zonneveld (1986), argue that in Dutch A-compounds are very rare. They support this claim by arguing that stokoud-a.dj ectives are not compounds at all, but 'lexicalized phrases'. The argument involves the stresspattern of this construction and its morphosyntactic properties with respect to further specification. With respect to the former, they observe that the true situation concerning the stresspattern of these adjectives is different from what is usually suggested: an [s v]-pattern is found in attributive position, and a [w s]-pattern in predicative position. This is precisely the pattern found w i t h more common intensifiers such as zeer or erg ("very"): (6) a.

[w s]

Hij vond haar "He found her

[bloed m 6 o i ] / [zeer m 6 o i ] 'blood-beautiful' / very beautiful"

Zij maakte het [beeld sch6on] / [erg sch6on] "She made it 'statue-pretty' / very clean" b. [s w ]

Het [blóed m o o i e ] meisje / Het [zéer m o o i e ] meisje "The 'blood-beautiful' girl" / "The very beautiful girl" Het [béeld schone] raam / Het [èrg schone] raam "The 'statue-pretty' window" / "The very clean window"

The [w s]-stresspattern in predicative position can also be illustrated by the ambiguity of the sentences in (7). The lefthand ones contain an obsystematic ject NP-adj ective-combination which has sentential final main stress by the Dutch Nuclear Stress Rule. The righthand ones contain stoicoud-adjectives. (7)

Hij vindt "He finds

[ons bloed] móoi / Hij vindt ons [bloed m ó o i ] [our blood] beautiful" / "He finds us ['blood-beautiful']"

Zij maakte [het beeld] schóon / Zij maakte het "She made [the statue] clean" / "She made it Hij maakt "He makes

[beeld schóon] ['statue-pretty']"

[haar kip] lékker / Hij maakt haar [kip [her chicken] tasty" / "He makes her [ 1

lékker] chicken-fit']"

Hij vond [jullie peper] duur / Hij vond jullie [peper d ù u r ] "He found [your pepper] expensive" / He found you ['pepper-expensive']" Hij hield [het ijs] kóud / Hij hield het [ijs k ó u d ] "He kept [the ice] cold" / "He kept it ['ice-cold']" With regard to the second argument of Trommelen & Zonneveld, notice that, since the lefthand part of the stokoud-compounds is intensifying, and means "very/terribly", such specifications cannot themselves precede these formations:

A (8)

Hij maakte het [erg lekker] "He made it very delicious"

Hij maakte het *zeer [erg lekker] "He made it *terribly very delicious"

Hij vond haar [kip lekker] "He found her ' chicken-fine"1

Hij vond haar *zeer [kip lekker] "He found het *very ' chicken-fine'"

Formations of the stokoud-type are thus, according to Trommelen en Zonneveld, not A-compounds at all, but lexicalized phrases, which receive [w s]-stress due to the Nuclear Stress Rule, whereas the Dutch variant of a 'Rhythm Rule' is responsible for the observed [s w]-pattern in attributive position. The name of adjectival compounds is reserved by Trommelen and Zonneveld for the consistently [s w] adjective type lSvensmoe, zSeziek mentioned earlier, w h i c h can also be intensified with erg (erg zeeziek, etc.)

3.

Pseudosuffixes

I will accept the above-mentioned properties of the stoJcoud-adjectives, that is Trommelen & Zonneveld's view as far as the stress-pattern and the ( i m p o s s i b i l i ties of futher specification of such constructions are concerned. My contribution here, however, will be threefold: I will argue first that the zeezieJc-type does not represent the A-compound type fitting the trio strindbal, vSldrijd, zeeziek; second that A-compound formation is by no means as unproductive as Trommelen & Zonneveld suggest; and third, that [s w ] is not the usual stress-pattern for A-compounds. This discussion goes as follows. The zeezieJc-type of adjective is especially frequent with the righthand members -ziek, -vol, -graag and -rijk, and less frequent with -arm and -vrij. These righthand m e m b e r s are called 'semisuffixes' ("halfsuffixen", for instance Van den Toorn (1983a,b)) or 'pseudosuffixes'. They are also mentioned in the A N S (p. 365), but unfortunately without the stressobservations that are so relevant to the discussion: (9)

-ziek "-sick"

z£e-zlek, spil-ziek, twist-zlek 'sea-sick', 'waste-sick', 'quarrel-sick' "sea-sick", "extravagant", "quarrelsome"

-vol "-ful"

smaak-vdl, wiarde-vdl, liefde-vdl 'taste-ful', 'worth-ful', 'love-ful' "tasteful", "useful", "loving"

-graag "-happy"

grljpgrdag, schietgr&ag, pr£atgr&ag 'grab-happy', 'shoot-happy, 'talk-happy' "itchy", "trigger-happy", "talkative"

2 r6em-rljk, invloed-rljk, kl6ur-rljk 'fame-rich', 'influence-rich', 'colour-rich' "famous", "influential", "colourful" "-poor", -vrij "-free" (see (10) for the rare [s wr]-examples)

-rijk "-rich" -arm

The semantics of these constructions is comparatively transparent. The fact that these pseudosuffixes can also occur in isolation suits us well, as it allows an instructive contrast with the stoJcoud-adjectives of the previous section. Recall that the former phenomenon is unproductive, the latter productive. We are looking for examples in w h i c h the lexical material is the same, but the two apparently identical adjectives belong to different classes. Thus, consider steenrijk ( - "stone" + "rich"). This can be a stoJc S66* ("-bound") (ANS: 367), -lustig ("-lusty") (ANS: 368), -vijindig ("-hostile") (ANS: 365-6), -kliar ("-ready") (ANS: 365). These adjectives exhibit the [w s]/[s v] rhythmically alternating stress-pattern of the stokoud-group, but they can be qualified with erg, zeer, and so on. Such a stress-behaviour can also easily be combined with those longer -end-cases ([veei] [bel6vend] , ('much-promising') [klemtoon] [aantrSkkend] , ("stress-attracting"), etc.) which frequently crop up in the Yiterature about stress attraction. These formations are also liable to the rhythmic stress-shift in attributive position which I illustrated in (6). Also belonging to the same group of adjectives are possibly the data of the very productive formation of the (pseudo-) past participle on the righthand side, which are given below. (13) ademben&nend bloeddoorl6pen computergestuurd doortrfipt eensgezind eigenger6id godgeklaagd hulpbeh6evend ingewikkeld kortger6kt minderbedeeld mismaakt noodgedw6ngen

"breathtaking" "bloodshot" "computer-controlled" "cunning" "united" "headstrong" "disgraceful" "needing" "complicated" 'short-skirted' 'less fortunate' "deformed" "from necessity"

omstr^den onderv6ed oudb&kken overs6kst prinsgezind ruimd£nkend stemger6chtigd uitgebr£id vold&an vold6ngen volstrikt welgemanierd zonoverg6ten

"controversial" "undernourished" "stale" "oversexed" "Orangist" "broad-minded" "entiteled to vote" "extensive" "satisfied" "accomplished" "complete" "well-mannered" "sunswept"

Formations like (12)-(13) are discussed in Van den Toorn (1983a,b) where the different types mentioned here are ordered in different ways, and often given with an inaccurate stresspattern, if stress is mentioned at all. The constructions in (13) coexist with less frequent [s w]-formations like tiitgekookt ("clever"), Hitgeput ("exhasted"), dpgewonden ("roused"), telSurgesteld ("disappointed"), which are conspicuous for the fact that they generally have a particle to the left, which, combined with their deviant stresspattern, suggests that they have an entirely different status. If the formations under (12)-(13) really have the stresspattern I claim they have, they should be liable to the test above in (7) by Trommelen & Zonneveld, which appears to be the case: (14) a. Zij vond [the milieu] vriendelijk / the [milieu vriendelijk] "She found [the surroundings] friendly / it [ecologically sound]" Wij noemden [jullie wereld] vr6emd / jullie [wereld vriemd] "We called [your world] strange / you ['world-strange']" Hij behandelde [haar soort] gelljk / haar [soort gelijk] "He treated [her species] equal / her ['species-equal' (-identical)]" Wij maken [haar kleur] 6cht / haar [kleur Icht] "We make [her colour] real / her ['colour-real']" Hij maakte [jullie vrouw] vij£ndig / jullie [vrouw vijindig] "He made [your wive] hostile / you ['women-hostile']"

7 (14) b. Hij vond [haar tijd] gebónden / haar [tijd gebónden] "He found [her time] bound / her ['time-bound']" Hij zocht [het doel] gericht / the [doel gericht] "'He searched [the target] aimed1 / it ['target-aimed']" The stress-pattern of the formations (12)-(13) can also be demonstrated by putting such formations opposite to equal compounds with category N, whereby they appear to have a contrastive stresspattern. Formations that are accidentally N ([s w]) as well as A ([w s]) can, with a little luck and imagination, indeed be found (special thanks to Wim Zonneveld): (15) De wâterpasgaf aan dat de vloer waterpisliep "The leveling rod indicated that the floor was level" Gezeten op een schâakmat^, zette hij hem schaakmât "Sitting on a 'chessmat', he 'put him checkmate' ( - checkmated him)" Hij zocht zich het âpezuur^, en keek apezûur "He searched 'himself the apesour' (-like hell), and looked 'ape-sour"' Zijn épetrots maakte hem apetróts "His 'ape-pride' made him 'ape-proud"1 This demonstrates convincingly that two 'stress-wise' different groups of concatinative adjectives can be distinguished, the zeezielc-type (constructed with pseudo-suffixes) and the vrowrvljandig-type discussed in this section, the latter appearing to be a major productive group of derived adjectives that should not be excluded from the discussion, as Trommelen & Zonneveld (1986) tends to do.

5. Two groups of A-compounds The outlines of the corpus of derived A-constructions now become clear. The most important question remains the precise status of the aforementioned constructions; those with a pseudosuffix in section 3, and the large group of derived adjectives from the previous section. I will show that there are four arguments to assume that the constructions of both cases have to be regarded as compounds, of which the [w s]-labeled adjectives represent the main group. The first argument concerns the above mentioned linking morphemes. We already saw under (5) that the appearance of such morphemes is not uncommon in the stokoud-cases, but they also appear in the constructions from the previous section, under (16a), and in the more isolated 'pseudosuffixed' cases of (11), see (16b), (of which I fail to grasp for the time being why in the former case 1 could not find any examples of -s- and whether this is relevant.) (16) a. arbeid-s-ongeschikt beweging-s-gerieht doop-s-gezind gezag-s-getróuw gezag-s-vólgend hervorming-s-gezind hemel-s-bréed hond-s-d61 inkomen-s-afhânke1ij k kapitaal-s-krâchtig leven-s-lâng uitkering-s-geréchtigd wezen-s-vréemd

'work-unfit' / "incapacitated" 'movement-directed' 'baptism-denominated' / "Baptist" "law-abiding" 'law-following' 'reform-minded' / "reformist" 'sky-broad' / "as the crow flies" 'dog-crazy' / "rabid" "income-related" 'capital-strong' / "whealthy" "life-long" "entiteled to a benefit" 'core-strange' / "alien"

8 (16) b. geslàcht-s-rijp kléur-a-blind léven-s-moe méns-a-schuw

"sexually mature" "colour-blind" 'life-tired' / "world-weary » 'people-shy' / "shy"H

Notice that these linking morphemes, as well as those in (5a) are 'lefthand bound'; compare [beer-9]stSrk vs. [ijzer\stSrk ('iron-strong') and [beer-a]stSrk vs. [beer-»]g£il ('bear-horny'), and those like (16): [leven-s]¿cht ('life-real') vs. [kleur]¿cht ('colour-real') and [leven-s]Scht, [leven-s]m6e ('lifeweary') vs [1 even-s]gev£arlijk ('life-dangerous'), and that this is in line with the behaviour of nominal compounds I discussed in section 1. In this respect — and that is not without significance — identical lefthand members in N-compounds and these A-constructions behave similarly: [plicht-s-getrouir]/ [plicht-s-besef] ('duty-faithful', 'duty-consciousness'). The appearance of these linking morphemes, and their behaviour according to those in N-compounds, can be seen as support for the status of real compounds for these formations in question. In the second place the righthand members of both types can also appear outside N-suffixes: beteken-nis^-vol , raad-sel^-voluitkeer-ing^-s-gerechtigd , gewoon-te^getrotnr ("mean-ing-tul", "enigmatic*/1riddle-ful', 'custom faithfull'), which is in contradiction with the properties of A-suffixes: *ken-nis^-lijk , *leid-ing^-ig^ ('knowledge-ly', 'drain-y'), but completely compatibel with the status of compounds for these formations. A third difference is the fact that real compounds allow the lefthand member to be a 1 super'-compound (with a recursively stached lefthand member), and this is also possible for both classes of derived adjectives: hoofdklemtoon-aantrSkkend, h6ofdklemtoon-arm, zuidwesterstorm-gevSelig, natuursteen-rijk ('main stress attracting/poor', 'south west storm sensible', 'natural stone rich'). The unability of the stoicoutf-adjectives to undergo such an enhancement (*natuursteen-rijk, *walnootijs-k6ud ('walnut-ice-cold') — one of arguments by Trommelen & Zonneveld) is just a restriction of the in other respects also restricted group of section 2, and not of the major groups in section 3. and 4. Finally we can not deny the productivity of these formations, especially those of section 4. As I mentioned before, vast productivity is a property which we would associate with compounds rather than with lexicalised phrases. In my opinion the only justified conclusion from all this is that the derivations with pseudosuffixes as well as the [w s]-adjectives are in fact A-compounds, and the stokoud-cases from section 2. form a limited subclass of the latter. Thus there are two possible ways in which A-compounds can be constructed: Type I: the righthand member is part of a closed group of which at least -vol, -rijk, -ziek, -graag are members, and the stresspattern is [s v]: zSeziek, invloedrijk, and they are semantically transparent. Type II: the righthand member is not from a fixed list, and can be almost everything, for instance: • a regular adjective; underived: gastvrij, vuurvSst, bladstil or derived: vrouwvijSndig, milieuvrlendelijk, lichtgevdelig (a subgroup of these are the morphologically constrained cases with intensifying lefthand members: stokdud, apetrdts, dolgelukkig, etc.), • or a non-adjective or even an unclear morpheme: openbiar, schaakmit, goedkSop, • or a (pseudo)participle: computergestuurd, stemgerSchtigd, and has the [w s]-stresspattern, and the formation is liable to rhythmic stresshift in attributive position.

9 The formations of the type stokSud, bladstil, etcetera have a righthand member that is not part of the closed group of possible pseudo-suffixes, and thus it would be predicted that they behave like the second group. This prediction is borne out. There is therefore, even without further theoretical motivation, on the basis of the empirical data only, and the subsequent assumption of a bifurcation within the corpus of concatenated adjectival constructions, an explanation for the behaviour of the stoJcoud-adjectives. It is also immediately explained why the intuitions of native speakers are the way they are with formations in -vol, -rijk etcetera. They are pseudosuffixes, but since they also serve the restrictions of a type II derived adjective, it is not always clear to the native speaker what the correct stresspatern is. So, ldodvrij is a type I-adjective, (with transparent semantic structure: "free of lead",) but also I sometimes observe that speakers, especially in predicative position, say loodvrij. If we accept the distinction between type I-derived adjectives and type II-derived adjectives (where the I-type is productive with a fixed righthand member, and the II-type is productive as a process), this explains oppositions like zHurvrij ( - type I, [s w]-stress, and meaning "free of acid") vs. gastvrij (» type II, [v s]-stress, rhithmic stressshift, and a non-transparant meaning) are allowed, and why the properties of the subsequent formations are as they are. Of course there are a lot more of these instructive contrastive examples, like for instance: broodndchter (- "[+lntens.] sober", type II) vs. brdoddronken (- not "l+intens.] drunk", but an isolated type I-case). Furthermore it is relevant that from now on, it is impossible to see, by means of stress alone, which cases must be phrases, and which are type II-compounds. Therefore we must conclude that bijdehand ("quick-witted"), onverrichterzake ("empty-handed") and spelenderwijs ("without effort") are suspected to be phrases, but we can not prove this, if stress structure is accepted as the only decisive criterium.

6. Conclusion What we see is that over the years the view of A-compounding has changed. Traditionally the stoJcoud-adjectives were the sole representatives of this group, with initial stress, as was generally assumed. Trommelen and Zonneveld take stress as criterion for the question whether a construction is a compound, and they observe that the stokoud-a.dj ectives are not compatible with compound stress. This, combined with the morphological limitations of these formations, leads them to the conclusion that these are not compounds. Since they implicitly assume that the formations in question are the sole representatives of the class of (aledged) A-compounds, they more or less automatically reach the conclusion that adjectival compounds are not alowed in Dutch. My approach of the topic is entirely different. My view is that, although Trommelen and Zonneveld's empirical observations are correct, they unfortunately reach the wrong conclusion, because their research was aimed at this morphologically restricted subclass which has always been wrongly assumed to be representative of the phenomenon of adjectival compounds in Dutch. On closer inspection the group discussed by Trommelen & Zonneveld is part of a much larger group of A-constructions with a [w s]-pattern; and what is more, it is a group which as a whole is very productive. This insight, combined with the great morphological similarity between these A-constructions and N-compounds (linking morphemes, sequential differences, 'super'-compounds to the lefthand side) leads to the inevitable conclusion that these constructions are indeed compounds and that therefore the starting point of stress as a distinguishing factor is invalid.

10 This forces me to the conclusion that in Dutch (17)

N-compounds are [s w] and A-compounds are [y s].

And what about the nice generalization under (1) that all compounds in Dutch serve the [s v]-pattern? In the face of the overwhelming empirical evidence to the contrary, this idea has to be abandoned. But I think we can make another generalization in return: The corpus of compounds falls into two parts, as far as stresspattern is concerned: N and V are [s w] (N: ijsbeer, strindbal; V: Ijsbeer, grdsduin ("gloss")), A and P are [w s] (A: goedkSop, volstrSkt; P: onder&an ('under-on1), tussendSor ("in between")). In the near future (Backhuys (forthcoming)) I wil elaborate on this idea. There I will also demonstrate how this idea yields insight into two other intriguing problems of Dutch morphology: stress attraction and the behaviour of -loos ("-less") and -achtig ("-ish"), and into how this, including the insights aquired here, can be explained within the framework of Lexical Morphology.

NOTES * This article is one of the results of research founded by the Stichting Taalwetenschap, which is supported by the Dutch Organization of Scientific Research NWO, project number 164-300-010. I am grateful to Wim Zonneveld for help and support, to Thérèse Bill for her inspiring remarks ('zonoverkapt', 'geldovergoten') and to both of them for being there when they were most needed. I also thank Ans van Kemenade for her help. 1. I thought it best to give formations of this type literal glosses. To distinguish them from the other glosses they are put in Italics and between single quotation marks: aardedonker - "pitch-dark" / 'earth-dark'. Other literal glosses, not of the stokoud-type, are also between single quotation marks, but are not italicized: brôoddronken - "unruly" / 'bread-drunk' 2. With -rljk I know the stressshifted formation omvingrijk ("voluminous"), for which I have no immediate explanation. It is striking however that this formation has the compound 6mvang, and compounds are usually not followed by pseudosuffixes . REFERENCES Backhuys, K-J, ..., dissertation, Dutch departement RU Utrecht, (forthcoming) Geerts, G., (e.a.), (eds.) Algemene Nederlandse Spraakkunst (ANS), Groningen. (1984) Schultink, H., 'Boundaries, wordclasses, and the accentuation of derived words (1980) in Dutch', in: Zonneveld, W., F. van Coetsem & O.W. Robinson (eds.), Studies in Dutch Phonology, 1980, p. 205-222. van den Toorn, M.C. , 'Halfsuffixen', in: De nieinre taalglds 76/b, p. 547-552. (1983a) van den Toorn, M.C., 'Enkele types van samengestelde adjectieven', in: Leuvense (1983b) bijdragen 72 (1983), p. 257-271. Trommelen, M. & W. Zonneveld, 'Dutch Morphology: evidence for the Righthand (1986) Head Rule,' in: Linguistic Inquiry 1711, p. 191-200.

Hans Bennis & Teun Hoekstra

PRO and the Binding Theory 0. Introduction In the standard GB-approach (cf. Chomsky 1981) binding theory (henceforth BT) offers an explanation of the distribution of PRO. The interpretation of PRO is left to control theory. In this article we shall argue that PRO is best analyzed as an anaphor with respect to its binding properties, i.e. that BT identifies PRO by binding it to an antecedent in a local domain. In section 3 we shall present an alternative account of the distribution of PRO. We shall argue that the distribution of PRO can be made to follow from a representational theory of empty categories. This theory is discussed in more detail in Hoekstra & Bennis (this volume). 1. The distribution and interpretation of PRO; the PRO-theorem The standard GB-account of the distribution of PRO runs as follows: PRO has properties of both anaphors and pronominals; it can thus be characterized as a pronominal anaphor; this characterization rests on an apparent contradiction since anaphors are bound within their governing category, whereas pronominals are free within the same domain; this contradiction is resolved if we assume that PRO is ungoverned; if PRO is ungoverned, it has no governing category and the requirement that PRO be both bound and free in its governing category is vacuously satisfied; from this account it follows that the distribution of PRO is limited to ungoverned positions, such as the subject position of most infinitival clauses. This account of the distribution of PRO, generally known as the PRO-theorem, is appealing since the distribution of PRO is made to follow from the independently motivated BT. However, the major problem is the paradox that the interpretation of PRO cannot be made to follow from BT. Although the binding properties of PRO constituted the starting point of the PRO-theorem, these properties cannot be accounted for in BT, given that PRO is considered as a pronominal anaphor. It has often been observed that PRO has several properties which are characteristic of anaphors (cf. Koster 1984, Lebeaux 1985 Manzini 1983, Williams 1980). Relevant properties of the antecedent are: c-command, obligatoriness and locality. These properties are illustrated in (1). (1)

a. John, tries [PRO. to go home] b. [The father of John-) ]. tries [PRO^... to go home] c. John^ tries to force ''BiHj [PRO^^j-' to go home]

If PRO were to be treated as an anaphor these properties would follow directly from BT. If the interpretation of PRO is removed from BT as a consequence of its status of pronominal anaphor, we have to assume that another module is able to account for the properties in (1). In general it is held that control theory accounts for the interpretation of PRO. We are now faced with an undesirable redundancy in the theory. Control theory has to copy a substantial part of BT in order to account for the interpretation of PRO. From this we may conclude that the PRO-theorem is able to provide an explanation of the distribution of PRO, but is faced with the problem that no conditions on the interpretation of PRO may be derived from BT.

12

If we drop the characterization of PRO as a pronominal anaphor, we may account for the interpretation of PRO in terms of BT. On the other hand, we then have to provide an alternative account of the distribution of PRO. We will adopt this approach here for two reasons. First, applying BT to the interpretation of PRO has interesting consequences and removes the redundancy between BT and control theory. Secondly, we do not need to make reference to the notion of government in the account of the distribution of PRO. 2. The interpretation of PRO; PRO as an anaphor Several linguists have argued that PRO is or can be an anaphor. Williams (1980) and Koster (1984) argue for a distinction between obligatory control and non-obligatory control. They take PRO in the obligatory control configuration to be an anaphor. Manzinl (1983) and Lebeaux (1985) take PRO to be an anaphor in all the constructions they discuss. They extend the theory in such a way that BT allows a correct interpretation. In this section we will argue that it is possible to consider PRO to be an anaphor in all instances and that we do not have to revise the theory in order to be able to let BT determine the domain of interpretation of PRO. Since for independent reasons government does not play a ^significant role in our theory, we will define principle A of BT as in (2). (2)

Principle A of the Binding Theory a. An anaphor must be bound in its binding category b. A category A is a binding category for B iff A is the minimal maximal projection which contains B and an opacity factor different from B ^ c. Opacity factors are subject and [+finite]

The definition in (2) is in no significant way different from existing proposals concerning the interpretation of anaphors. It is our claim that this definition is sufficient for the interpretation of PRO. We will restrict our attention in this section to PRO as the subject of infinitival clauses. The definition in (2) allows a straightforward explanation for the facts in (1). The binding category for PRO is the matrix clause, since the matrix clause is the first maximal projection with an opacity factor. The obligatoriness of the antecedent follows from the fact that an anaphor must be bound, the c-command requirement follows from the fact that an anaphor must be bound and the locality requirement follows from the fact that an anaphor must be bound in its binding category. It is important to observe that the anaphoric status of PRO does not determine the interpretation of PRO entirely. BT determines a domain within which PRO should have an antecedent. The actual choice of a particular antecedent within that domain does not belong to BT. This choice is dependent on different kinds of lexical factors, such as the choice of the matrix verb. Unfortunately, there is not much insight into the way in which this selection of antecedent takes place. As is usual, we will refer to this set of properties as the theory of control, although would be more appropriate to refer to this as the non-theory of control. In a sentence such as (3a) BT determines that both John and Mary are possible antecedents, but Bill and Sue are not. The distinction between binding and control is confirmed by the fact that another choice of the matrix verb allows us to change the antecedent from Mary in (3a) to John in (3b). Another structure of the infinitival clause results in a similar change of controller, as can be seen from the difference between (3a) and (3c). However, it is not possible to arrive at a construction in which either Bill or Sue are interpreted as the antecedent of PRO by changing the matrix verb or the complement structure.

13

(3) a. Bill told Sue that John has asked Mary. [PRO. to go home] b. Bill told Sue that John, has promised 1 Mary [PRO. to go home] c. Bill told Sue that John i has asked Mary [PROj to 1 be allowed to go home] This indicates that within the domain determined by BT, control further restricts the choice of antecedent. This interpretation of control allows us to extend the domain of control to other cases of binding. In many respects the situation in (4) is similar to the situation in (3). (4) De leraren vertelden de ouders dat de jongens de meisjes naar elkaars oordelen vroegen The teachers told the parents that the boys asked the girls for each others judgements In (4) BT dictates that neither the teachers nor the parents can be the antecedent of each other. Within the subordinate clause both the boys and the girls are potential antecedents, even if the girls is to be taken as the actual antecedent in (4). We may consider this to be a case of object control. We will now consider a number of objections to the interpretation of PRO in the way indicated above. a) It has been suggested that there exists a difference between the interpretation of PRO and the interpretation of lexical anaphors: whereas the choice of an antecedent for a lexical anaphor is free within the binding category, the choice of an antecedent for PRO is uniquely determined. There are two ways to argue against this difference: one may either show that the interpretation of PRO is not as uniquely determined as claimed or that the selection of an antecedent for lexical anaphors is not as free as is claimed. Both lines of argument are in fact applicable. First, the antecedent of PRO is either an object, a subject or both, depending on either the matrix verb or the infinitival complement. This is shown in (5). (5) a. Jan. belooft Marie. PRO. naar huis te gaan John promises Mary J to house to go 'John promises Marie to go home' b. Jan. dwingt Marie. PRO. naar huis te gaan JohA forces Mary ^ •'to house to go c. Jan. stelt Marie, voor naar huis te gaan JohA proposes Maiy ^ to house to go d. Jan. belooft Marie. PRO. naar huis te mogen gaan John promises Mary-' to house to may go The sentences in (5) do not argue for a unique determination of the antecedent of PRO. Similarly, we saw in (4) that with lexical anaphors the selection of the antecedent is not completely free either. We may conclude from this that there is no real difference between lexical anaphors and PRO in the determination of their antecedent. b) It is evident that there are constructions in which there is no lexical antecedent for PRO. This so-called arbitrary PRO appears to be a very strong argument against a theory in which PRO is always an anaphor. Some examples are given in (6). (6) a. Het is leuk om PRO een boek te lezen It is nice for a book to read 'It is nice to read a book' b. Er wordt geprobeerd om PRO de deur open te maken There is tried for the door open to make 'Someone tries to open the door'

14

We will argue that in these cases PRO is bound by an implicit argument in the matrix clause. This implies that we^ extend the class of potential antecedents to implicit, non-lexical arguments. In order to motivate this position, there are again two strategies. First, we will show that there are indications that it is indeed a non-lexical argument that is the antecedent of PRO. Next we will show that lexical anaphors can be bound by implicit arguments as well. In the sentences in (6) the implicit argument of the adjective/passive participle can be added in a PP, as in (7). (7) a. Het is leuk voor mij om PRO een boek te lezen It is nice for me for a book to read b. Er wordt door mij geprobeerd om PRO de deur open te maken There is by me tried for the door open to make It is clear that in these sentences PRO is obligatorily bound by the NP in the added PP, mij. If the sentences in (6) really involve arbitrary PRO, it is completely unexpected that the addition of an optional PP turns the sentence into a configuration of obligatory control. Moreover, the interpretation is not always influenced by the addition of a PP, as is shown in the sentence in (8).

(8) Er wordt voor mij geprobeerd om PRO de deur open te maken There is for me tried for the door open to make Although a PP is added to (6b), PRO is not interpreted as bound by mij. The difference between (6) and (7) and between (8) and (7) is that the Implicit argument of the adjective/participle is lexically absent in (6) and (8), but lexically present in a PP in (7). The interpretation of PRO in (6)-(8) can be accounted for quite easily if we allow implicit arguments to bind anaphors. This interpretation remains a mystery in a theory in which arbitrary PRO is allowed. An additional argument for the idea that the implicit argument is actually the antecedent of PRO in (6) can be derived from the ungrammatically of the sentences in (9) (cf. Hoekstra & Wehrmann 1985). (9) a.*Het is waarschijnlijk om PRO morgen naar huis te gaan It is probable for tomorrow to house to go (Het is waarschijnlijk dat men morgen naar huis gaat) b.*Er blijkt om PRO morgen naar huis te gaan There appears for tomorrow to house to go (Er blijkt dat men morgen naar huis gaat) The difference between (6a) and (9a) resides in the type of adjective. Although both adjectives are allowed in a construction with a finite subordinate clause, adjectives of the type leuk (nice), moeilljk (difficult), interessant (interesting), mogelijk (possible) combine with infinitival clauses, whereas adjectives of the type waarschijnlijk (probable), zeker (certain) do not. At the same time the adjectives of tne type waarschijnlijk do not combine with a voor-PP, as in (7a). This strongly suggests that these adjectives do not have an implicit argument. Under our approach the ungrammaticality of (9a) is due to the fact that no antecedent, explicit or implicit, is available as an antecedent for PRO. The same explanation can be given for the ungrammaticality of (9b). Impersonal passives such as (6b), have an implicit argument which may bind a PRO subject in an infinitival complement. Raising verbs, such as blijken (appear), do not have an implicit argument. The impossibility of an infinitival CP complement in (9b) then follows from the violation of principle A of the BT. If our approach is correct, we expect that implicit arguments can be the antecedent for lexical anaphors as well. The sentences in (10) illustrate that this situation does indeed occur.

15

(10) a. Er wordt gerekend op elkaars medewerking There is counted on each others cooperation 'People are counting on each others cooperation' b. Er wordt in kroegen alleen maar over zichzelf gesproken There is in pubs only about himself spoken 'People speak in the pub only about themselves' Another relevant observation is that this theory directly accounts for the fact that so-called arbitrary PRO seems to have features, such as [+human]. In sentences such as (11) the subject of the infinitival clause is interpreted as human. Since the verb gebeuren (happen) requires a non-human subject, (lib) is ungrammatical. (11) a. Het is niet leuk om te vallen It is not nice for to fall (= it is not nice for someone to fall) b.*Het is mogelijk om tegelijkertijd te gebeuren It is possible for simultaneously to happen In our view this property follows from the fact that the implicit argument of an adjective, such as leuk or mogelijk, is human for the reason that only human beings have the capacity to evaluate. In a theory in which (11a) contains an arbitrary PRO the human interpretation of PRO is unexpected, since both humans and non-human objects have the capacity to fall. A final argument in favour of the analysis presented here concerns the fact that there is no expletive PRO. We do not find sentences such as (12a), although the finite counterpart in (12b) is perfectly acceptable. (12) a.*Het is leuk om PRO gedanst te worden It is nice for danced to be b. Het is leuk dat er gedanst wordt It is nice that there danced is 'It is nice that people are dancing' Under our analysis the impossibility of (12a) follows directly from the fact that PRO is an anaphor. It is clear that there are no expletive implicit arguments. Therefore PRO is Interpreted as (+human]. This results in a rather strange interpretation in which people can be danced. With these observations we have turned an apparent problem for the anaphoric status of PRO into an advantage for our theory. Given that binding by an implicit argument is motivated, the anaphoric status of PRO allows an explanation of a number of otherwise puzzling phenomena. c) The third potential argument against the anaphoric status of PRO is the lack of strict c-command in (13). (13) a. Jan vroeg [aan Piet^] om PRO^ naar huis te gaan John asked to Pete for to house to go b. Het is leuk [voor Jan.] om PRO. naar huis te gaan It is nice for John for to house to go The antecedent in (13) is embedded in a PP. In this way the strict c-command requirement on binding is violated. The facts in (13) are problematic only if it can be shown that there is a distinction between binding of lexical anaphors and binding of PRO. However, it is well known that the strict ccommand requirement should be relaxed for the binding of lexical anaphors as well. This is illustrated in (14). (14) Ik sprak [met Piet.] over zichzelf. I spoke with Peter about himself

16

d) Another potential argument against the anaphoric status of PRO is the possibility of non-local binding. In a sentence such as (15), the antecedent appears to be outside the binding category of PRO. (15) Jan, zegt dat het gezond is om PRO. te zwemmen John says that it healthy is for 1 to swim If the binding relation is indeed as is indicated in (15), this would be a serious problem for our theory. However, it can be shown that there is no direct relation between Jan and PRO in (15). The fact that they can be interpreted as coreferential is due to the fact that Jan can be interpreted as the implicit argument of mogelijk. If we add a non-coreferential implicit argument in a PP or if we replace the adjective mogelijk by waarschiinlijk, which has no implicit argument, the coreference between Jan and PRO becomes impossible. This is illustrated in (16). (16) a.*Jan. John b.*Jan. John

zegt says zegt says

dat het that it dat het that it

voor mij. gezond is om PRO. for me ^ healthy is for waarschijnljk is om PRO. te probable is for to

te zwemmen to swim zwemmen swim

If PRO could take a long distance controller, as suggested by (15), the ungrammaticality of (16b) would come as a total surprise. The facts in (16) therefore indicate that there is no non-local binding of PRO. e) A final potential argument against our analysis is the possibility of split-antecedents. Several matrix verbs (for instance voorstellen (suggest) and aandringen (urge)) allow PRO to be interpreted as the subject and the object of the matrix clause together. This is shown in (17). (17)a. Jan. stelt Piet. voor om PRO. . met elkaar naar de film te gaan John suggest Pexer for with each other to the movies to go b. Ruslandj dringt er bij Amerika. op aan om PRO. . elkaar te vertrouwen Russia urges America for each other to trust In the literature the property of split-antecedent is known as a typical pronominal property. Here too, there are ways to approach this argument. Either it is shown that PRO cannot have a split antecedent, or that lexical anaphors may have split antecedents as well. The first strategy is highly implausible since PRO in (17) must be plural as the antecedent of elkaar. Since the matrix clause does not contain a potential plural antecedent we have to assume that PRO has a split antecedent. ^Hence, it must be shown that split antecedents may occur with lexical anaphors . The example in (18) shows that the anaphor elkaar can indeed have a split antecedent. (18) Rusland. drong bij Amerika. aan op elkaars. . medewerking bij het toezicht op ontwapening ^ Russia urged America for each others cooperation with the supervision of the disarmament We have thus shown that all arguments against an anaphoric status of PRO and in favour of a pronominal status of PRO can be disputed. Furthermore, we have seen that a theory in which PRO is always anaphoric provides us with a number of interesting explanations concerning the interpretation of PRO. Moreover, such a theory allows us to eliminate the redundancy between BT and control theory. However, we have lost the explanation for the initial motivation to consider PRO to be a pronominal anaphor: the distribution of PRO. In the next section we will discuss this subject.

17

3. The distribution of PRO Having established that PRO is an ordinary anaphor, we can no longer derive its distributiongfrom BT, i.e. the requirement that PRO be ungoverned is no longer a theorem. Ve propose an alternative, which is based on properties of chain formation (cf. Hoekstra & Bennis, this volume). Starting point is that there is not only a complementary distribution between PRO and lexical NP, but a complementarity of the interpretation of e as either trace or PRO as well. We maintain that there is only one empty category (ec). If this ec can be connected to an antecedent through a chain, i.e. if the ec is not the head of a chain, the interpretation of the ec depends on the interpretation of the head of the chain. This is the ec which is normally referred to as t or e. If no chain can be formed, i.e. if the ec itself is the head of the chain, the interpretation of the ec is dependent on the theory which characteristically relates two independent chains, viz. BT. We will now turn to a first approximation of the algorithm which determines whether a chain can be formed or not. Suppose we have the following formulation of chain formation: (19) Chain Formation (1st preliminary version) a. link an empty category to a c-commanding NP if possible b. CP breaks a chain Let us see how this principle explains the ungrammatically of the examples in (20) and (21). The ungrammaticality of (20b-f) and (21) follows from Chain Formation. In all these cases a chain can and hence must be formed between the ec and an antecedent (John in (20b-f) and i_t in (21)) since no CP intervenes between them. The ungrammatically of the sentences in (20b-f) then follows from the ©-criterion because one chain receives two 6-roles. The sentences in (21) are grammatical only if it^ is an argument, related to the ec in the position of PRO through a chain. In that case there is a chain with only one 6-role. (20) a.*PR0 walks b.*John sees PRO c.*John counts on PRO d.*John considers PRO intelligent e.*John believes PRO to be intelligent f.*John hears PRO sing a song (21) a.*It was kissed PRO b.*It seems PRO to be intelligent c.*It was believed PRO to be intelligent The ungrammaticality of (20a) follows for a different reason. Chain^ formation is not applicable in this case. There is a local relation with AGR , but this relation is insufficient to identify the ec in a non-pro-drop language. However, it is an advantage of our theory that the explanation of (20a) is different from the explanation of the ungrammaticality of (20b-f), since it allows us to explain gthe fact that in pro-drop languages an ec is possible in subject position only . Before we proceed to the refinement of Chain Formation, we will illustrate in which way the theory developed thusfar accounts for the properties of a simple control configuration as in (23). (23) John promises Mary [ e to go home] The ec in (23) is the subject of an infinitival CP. It follows that Chain Formation cannot be applied in such a way that the ec is connected to an

18

antecedent. Consequently the ec is the head of its chain, hence an anaphor. The binding category of the ec is the matrix clause. Within this clause e must find an antecedent. There are two potential candidates (John and Mary). The theory of control determines that in this case John is the antecedent of e. It is easy to show that the formulation of Chain Formation given in (19) is not correct. The statement that a CP breaks a chain implies that CP is an absolute barrier to chain formation. A simple case of long wh-movement shows that this cannot be true. In order to allow the ec in (24) to be connected to its antecedent who, we will change (19) into (25). (24) Who. do you think that John saw e. (25) Chain Formation (2nd preliminary Version) a. link an empty category to a c-commanding NP if possible b. CP breaks an A-chain It is not only the case that the type of chain is relevant to chain formation, the status of the category from which extraction takes place is relevant as veil. In general adjuncts are islands to extraction. Moreover, adjuncts allow PRO-subjects. The sentences in (26) show the impossibility of extraction from adjuncts. The sentences in (27) show the possibility of a PRO-subject in adjunct clauses and adjunct small clauses. (26) a.*Who. did John go home [after he had met e.] b.*What. did Mary jump on the table [afraid of e.] (27) a. John 1 went home [after e meeting Mary] b. Mary jumped on the table [e afraid of mice] To account for these facts we will add to (25) the claim that adjuncts, including CP adjuncts, break a chain. Finally, we will add NP, or rather DP, to the set of categories which break a chain. DP may contain a PRO-subject and no movement from DP is allowed. We will not discuss this point here. Ve change (25) into (28). (28) Chain Formation (3rd preliminary version) a. link an empty category to a c-commanding NP if possible b. CP breaks an A-chain c. Adjunct XP and DP break a chain It is clear that the condition in (28) does not have much explanatory value itself since it just lists the different conditions under which chainformation can(not) take place. In Hoekstra & Bennis (this volume) we show that conditions b. and c. in (28) are independently motivated. Interestingly, the so-called strong binding requirement on empty operators follows straightforwardly from our approach. Such empty operators are found in [SPEC,CP] of an adjunct clause. Under our approach such an empty operator is not yet a new type of ec, but can be taken as a specific instance of e, the properties of which follow again from chain formation and BT. According to (28c), an ec in this position cannot enter into a chain. Hence, it is the head of a chain, which can only be identified through binding. Let us limit ourselves here to two specific instances of empty operator constructions, viz. infinitival relat lves, as in (29), and tough movement constructions of the type in (30). For arguments that the phrase following easy in (30) is an adjunct, cf. Bennis & Wehrmann (1987). Given that these ec's are identified through binding, the interpretation of the sentences in (31) follows. (29) a. a man [ e^ [ e^ to fix the sink ]] b. a book [ e. [ e. to throw away e. j] (30) John is easy [ e~! [ e. to please e. j]

19

(31) a. Jan^ lijkt mij. [e^ vervelend (voor N P ^ H e ^ . ^ . om [e^ e^ te zien]]] 1 1 John seems (to)-Ve boring for NP ^ for to see te b. Ik^ vind [ Jan. vervelend (voor NI\ ) t®*-/- o m [®|, zien]]] 1 1 1 I consider John- boring for NP - for ^ to see Given (28), the ec in object position of the infinitival clause in (31) cannot be the head of a chain. The infinitival clause being an adjunct, the head of the chain of this ec should be found within the infinitival clause. We thus have two ec's in one clause which are both the head of their chain. These categories occupy the position of [SPEC,CP] and [SPEC,IP]. Both ec's are anaphors. They have to find an antecedent within their binding category. In (31) it is the matrix small clause which is the first XP containing an opacity factor (subject). Given that (31a) is a raising construction we correctly predict the raised subject Jan to be the antecedent of the empty anaphor in [SPEC,CP]. In (31b) the SC-subject (Jan) occupies its base position. Consequently this subject is the antecedent for the ec in [SPEC,CP] of the infinitival adjunct. The ec in [SPEC,IP] will have the implicit argument or the NP in the voor-phrase as its antecedent. Our approach thus provides an adequate account of the locality of strong binding in these empty operator cases. In the theory outlined above the class of ec's is partitioned into a class of ec's which are the head of a chain, and those which are not. While the latter are chain-identified, the former are subject to BT for their identification. This analysis eliminates the redundancy between binding and control in the standard theory, and reduces the inventory of ec's. The partitioning factor is the principle of Chain Formation. In Hoekstra & Bennis (this volume), this principle is extensively motivated. Notes 1. Note that we use the notion Binding Category instead of Governing Category in our formulation of BT (cf. Chomsky 1981:220). In our approach the notion of government plays no independent role in the determination of domains relevant for BT. 2. This probably does not exhaust the class of possible opacity factors. As this paper is not concerned with the actual way in which binding is established, we shall not go into this issue. 3. Although it is non-structural which appear to suggested for Kayne (1981).

true that the actual choice of antecedent seems to involve factors, it is by no means excluded that the properties be relevant are ultimately structurally represented, as is the alternation between subject and object control in e.g.

4. It is unclear to us at this point what the representation of implicit arguments should be taken to be. The minimal position, from the point of view of the Projection Principle, would be that implicit arguments are structurally instantiated, i.e. projected from the argument structure into the syntax. This raises questions concerning their licensing, on which we have nothing of interest to say. Alternatively, they are not projected, which would have obvious consequences for the characterization of BT, which go beyond the approach advocated here, certainly if we are right in claiming that lexical anaphors may also be bound by implicit arguments. 5. The point about split antecedents is that such an interpretation should be a matter of coreference rather than binding, if binding is construed as a coindexation relation between two NPs. If our claim is correct that not only alleged PRO, but also lexical anaphors can take a split antecedent, this conception on binding will have to be adapted.

20 6. In the literature (cf.Manzini 1983) it is suggested that instead of government, the theory of Case might provide an explanation for the distribution of PRO. Aside from the fact that an account of the distribution of PRO in terms of Case is ad-hoc, it would not exclude PRO in the sentences in (21). 7. Note that we have to assume that the presence of AGR-coindexation prohibits an identification of e by means of BT. We would otherwise predict that e can occur as the subject of finite embedded clauses even in non-pro-drop languages. This suggests that AGR-coindexation is just a specific case of chain formation, i.e. between AGR and its specifier. 8. Apart from pro in subject position, some languages also allow null objects. This raises the question, within our approach, how these are licit. Note that our approach predicts that these null objects should enter into a chain with an A-bar antecedent, as they are themselves in a theta-position. For hull objects of the Chinese-Portugese type, this appears to be essentially correct. For arbitrary null objects, as found in Italian and other languages, our approach leads us to expect that these are licensed by a possibly empty clitic. References BENNIS, H. & P.VEHRMANN 1987 Adverbial Arguments, in F.Beukema & P.Coopmans (eds) Linguistics in the Netherlands 1987, Foris Publications, Dordrecht CHOMSKY, N. 1981

Lectures on government and binding, Foris Publications, Dordrecht

HOEKSTRA, T. & H.BENNIS 1989

A Representational Theory of Empty Categories, this volume

HOEKSTRA, T. & P.VEHRMANN 1985

De Nominale Infinitief, in GLOT 8, 257-275

KAYNE, R. 1981 Unambiguous Paths, in J.Koster & R.May (eds) Levels of Representation KOSTER,Foris J. Publications, Dordrecht 1984

On Binding and Control, in Linguistic Inquiry 15, 417-461

LEBEAUX, D. 1985

Locality and Anaphoric Binding, in The Linguistic Review 4, 343-364

MANZINI, R. 1983

On Control and Control Theory, in Linguistic Inquiry 14, 421-446

WILLIAMS, E. 1980 Predication, in Linguistic Inquiry 11, 203-238

Renée van Bezooijen

Evaluation of an algorithm for the automatic assignment of sentence accents 0. Summary In this contribution an algorithm for the automatic assignment of sentence accents in written Dutch text, developed by Kager and Quene, was evaluated experimentally. Results showed that the output of the algorithm is considered significantly more adequate than accents randomly distributed over content words but significantly less adequate than the accents produced by a trained news broadcaster and those realized by the majority of the subjects themselves. Insight was sought into the basis of the adequacy judgments and suggestions are made for the improvement of the algorithm. 1. Introduction The central aim of the national Dutch SPIN-program "Analysis and Synthesis of SPeech" (ASSP) is to improve the quality of text-to-speech conversion for Dutch. In one of the ASSP research projects, carried out by Kager and Quene, an algorithm has been developed for the automatic assignment of sentence accents in written text. In this contribution an experiment is described in which the adequacy of the algorithm ("PROS" from now on) was evaluated, both on paper and by ear. A more detailed description of the experiment can be found in Van Bezooijen (1989). Four components can be distinguished in PROS (for more details, see Kager and Quene, this volume). In the first component, it is assessed which words in the input text occur in the function word lexicon, which contains about 400 words. The words not found in the function word lexicon are labeled as content words. In the second component, a limited number of word labels are attached. Labeling information pertaining to function words is copied from the function word lexicon; labels pertaining to content words are derived from formal word characteristics and contextual analysis. In the third component, prosodic boundaries are determined on the basis of word labels (for example, a boundary is placed before a subordinate conjunction) and on the basis of punctuation, such as comma's. The fourth component contains a set of about 20 accent assignment rules that determine in which circumstances words should receive an accent. In these rules, use is made of the three types of information mentioned above: the distinction between function and content word, word labels, and prosodic boundaries. An example of an accent assignment rule is: verbs preceded by a pronoun and followed by a prosodic boundary are accented. The PROS accent assignment rules only use information explicitly available w i t h i n a sentence. For example, the word(s) following zulke ('such') is (are) taken to contain given information, and is (axe) therefore not accented. However, implicit semantic/pragmatic relationships b e t w e e n sentences are not taken into account. In practice, virtually no function words and the majority of the content words (about 85%) receive an accent in PROS. In an average text the total percentage of words to which an accent is assigned lies between 45% and 50%. The aim of the experiment described in this paper was to assess the adequacy of the PROS output. The material used as a carrier for the accents consisted of single sentences and sentences embedded in texts. The evaluation was carried out in three ways.

22

In the first place, the PROS accents were formally compared to two optimal accent versions, one derived from the accents assigned to the sentences by a group of twelve subjects (SUBJECTS version) and the other derived from the reading of a professional news broadcaster (PB-version). This comparison provided "objective" information on the differences between PROS accents and two supposedly highly adequate accent representations. In the second place, a paper-and-pencil evaluation was carried out in which the same twelve subjects rated two accent versions of the 32 sentences - accented words capitalized - on an adequacy scale. One version consisted of the PROS accents, the other contained accents which had been randomly distributed over the content words, the total number of PROS and RANDOM accents being equal. This task served to separate the evaluation of the linguistic adequacy of the PROS accents, in terms of place and number, from the adequacy of their acoustical realization. In the third place, an auditory evaluation was conducted in which all four accent versions of the 32 sentences were presented: PROS, RANDOM, SUBJECTS, and PB. The speech utterances were generated by means of diphone synthesis. One aim of this task was to compare the adequacy ratings for accents presented on paper with those for accents presented via the auditory channel. This could be done for the two versions, PROS and RANDOM, appearing in both evaluations. The second and most important aim of the auditory evaluation was, of course, to assess the relative adequacy of PROS when applied in a context for which it was developed, namely synthetic speech. The structure of this paper is as follows. In Section 2 information is provided on methodological aspects of the experiment: sentence material (2.1), experimental design (2.2), and each of the three experimental tasks, i.e. subjects' accent assignments (2.3), paper-andpencil evaluation (2.4), and auditory evaluation (2.5). In Section 3 results are presented and discussed, 3.1 being devoted to a formal comparison of PROS with the two optimal accent representations, 3.2 to the paper-and-pencil evaluation, and 3.3 to the auditory evaluation. Finally, Section 4 contains some concluding remarks. 2. Method 2.1 Sentence material The sentence material included eight single sentences and 24 sentences embedded in texts. The total number of words was 588. The eight single sentences consisted of short messages that one may hear in daily life via radio, telephone, or intercom, messages related, for example, to train arrivals and traffic conditions. These messages can and are likely to be pronounced without a preceding or following context. The mean number of words per sentence was 18.5, with a standard deviation of 6.9 words. The embedded sentences were part of eight short texts, which were taken from Dutch newspapers. Subject matter varied from weather forecast to illegal fishing. The number of sentences per text fluctuated between two and five. The mean number of words per sentence was 18.4, with a standard deviation of 5.4 words. 2.2 Experimental design The experiment consisted of three tasks: (1) Subjects' accent assignments: the subjects were asked to indicate which words should receive ail accent in the 32 sentences presented to them in written form. From the individual accent assignments one representative accent structure for each sentence was derived (SUBJECTS accent structure). (2) Paper-and-pencil evaluation: the sentences were presented and evaluated in written form in two different versions, one with the accents yielded by the PROS algorithm and one with accents distributed randomly over the content words, the total number

23

of PROS and RANDOM accents being identical. The RANDOM version served as a "fair" minimum reference. (3) Auditory evaluation: the sentences (diphone synthesized) were presented and evaluated auditorily in four different versions, namely with the RANDOM accents (minimum reference), with the PROS accents, with the SUBJECTS accents obtained in (1) (first maximum reference), and with the accents realized by the trained news broadcaster P B (second maximum reference). In all three tasks the sentence material described in Section 2.1 was used. The texts were always presented in their integral form, with the constituting sentences in the correct order. The single sentences were presented individually. Texts and single sentences were presented blockwise. Within the blocks there were two orders of presentation, varying per task. The subjects were randomly divided into two equal-sized subgroups, each of which received one of the two orders. All tasks were administered to the same twelve subjects, all trained listeners, in the order given above. The time interval between (1) and (2) was a couple of days, that between (2) and (3) about two months. In the next section, methodological aspects of the three parts of the experiment will be presented in more detail. 2.3 Subjects' accent assignments In this task the subjects were asked to indicate for each sentence the most adequate accent structure by underlining those words that they thought should be accented if the sentences were read aloud. Since it would be difficult to explain the method of the auditory evaluation in Section 2.5 without information on the SUBJECTS accent structure - one of the four versions presented auditorily - it is more appropriate to present the data related to the consistency of the subjects' accent assignments and the procedure used to derive the S U B J E C T S accent structure in this method section rather than in the results section. The mean number of accent assignments per sentence was 6.3 with a standard deviation of 1.1. The total number of accents assigned per subject, summed over all 32 sentences, ranged from 162 to 238. Apparently, there are widely differing views on how many accents a text should contain, and where. Differences among subjects seemed to result, among other things, from differences in the interpretation of what constitues given/new information and rhythmic preferences. To obtain one representative accent structure for each of the 32 sentences it was decided to consider those words as accented which were designated as such by at least seven out of the twelve subjects. This appeared to be a good cut-off point because the number of SUBJECTS accents for each sentence obtained in this way is virtually identical with the mean number of accent assignments p)er sentence, averaged over subjects. The total number of S U B J E C T S accents is 198. 2.4 Paper-and-pencil evaluation In this written p>art of the experiment the 32 sentences were presented to the subjects in two different versions (accented words capitalized): with the accents yielded by the PROS algorithm and with "randomly" assigned accents. Of course, the p>ossibilities for constructing a minimum reference are virtually endless. To obtain a qualitatively and quantitatively "fair" minimum reference, there were two restrictions on the "random" accent assignments: first, just like (in principle) in the PROS algorithm, only content words (as defined by PROS) could be accented, and, second, the total number of accents was identical with that in the PROS sentences, namely 274. The subjects were asked to rate the accent structures of each of the 64 sentences on a 10-point scale, going from 1 = very inadequate to 10 = very adequate. A 10-px>int scale was opted for because of the frequent use of 10-pjoint evaluation in the Dutch school system, with well-established definitions of the scale terms.

24

2.5 Auditory evaluation The advantage of presenting the sentence material on paper is that one obtains information on the adequacy of accentuation without the interference of the possibly incorrect acoustical realization of the accents in terms of FO-movements or the inferior quality of the non-natural speech. On the other hand, the task is rather abstract and far removed from daily life, in which accents are exclusively auditory phenomena. It was therefore decided to evaluate the algorithm auditorily, in a context in which it will eventually be applied, namely synthetic speech. Because of its relatively good quality, diphone speech (speaker HZ) was opted for rather than allophone speech. The syllables to be accented were indicated by hand. The stimulus utterances were generated with a neutral intonation contour in which the last two accents before a prosodic boundary are linked to form a "flat hat", that is a rise on the first accented syllable and a fall on the second accented syllable, joined by a stretch of high declination. The preceding accents are produced with a rise on the accented syllable followed by a slow fall until the next accented syllable ("sawtooth" pattern) (Van Rijnsoever, 1988). Four different auditory accent versions were made: RANDOM, PROS, SUBJECTS, and PB. The accents for the last version were obtained by having the trained newsreader PB read the material aloud and have the recordings transcribed with respect to accent by four phoneticians experienced in this type of task. Those words were considered to be accented which were indicated as such by at least three of the four transcribers. There were 261 accents which met this criterion. Just like in the paper-and-pencil evaluation the subjects were asked to rate the accent structure of each sentence on a 10-point adequacy scale. During the auditory task, the subjects had the written material (of course without accent indications) in front of them. The time interval between sentences was 5 seconds. There were four practice sentences to accustom the subjects to the task and to give them an impression of the range in accent structures in the test material. The task took about 50 minutes. 3. Results 3.1 Formal comparison of PROS with SUBJECTS and PB One way to assess the adequacy of PROS is to compare the accents it contains with the accents in SUBJECTS and PB. It appears that the number of accents in PROS (N=274) is very close to PB (N=261), whereas the number of accents in SUBJECTS is much smaller (N=198). A similaritiy index was calculated for each pair of versions by dividing the total number of words with an identical accent status (+ or - accent) by the total number of words in the stimulus material. Because of the high number of unaccented words in all versions, the differences between the similarity coefficients are small. The highest similarity is found between the two maximum references PB and SUBJECTS (88.6%). The similarity between PROS and PB is fairly high as well (88.3%); it is 5% higher than the correspondence between PROS and SUBJECTS (83.3%). Of course, a comparison can go much further than counting. It is possible to draw up an inventory of all differences between accent versions, and try to categorize them. In this way detailed information can be obtained on the frequency of occurrence of particular types of differences. Not all differences could be unambiguously assigned to a particular category, but 61 differences of the total of 98 differences (61.2%) between PROS and SUBJECTS, and 36 of the total of 69 differences (52.2%) between PROS and PB were covered by the nine categories distinguished below. 1. Measures. PROS very consistently accentuates names of measures (kilo, graad, dag, etc.) whereas SUBJECTS and PB do not. This regularity would be relatively easy to

25 include in PROS, by listing the relevant words in a lexicon. However, the generality of the phenomenon would first have to be checked by looking at a much larger corpus of spoken material. In the present material, the inclusion of a "measure" rule in PROS would reduce the number of differences with SUBJECTS and PB with seven. 2. G i v e n i n f o r m a t i o n . PROS very consistently accentuates words which can be considered to contain given information, whereas SUBJECTS consistently does not and the performance of PB is variable. Obviously, the relationship between accentuation and the concept of given/new information is complex and difficult to capture in a formal rule, even with an exhaustive syntactic analysis. However, a small part of the problem may be solved pragmatically by deaccenting literal repetitions of content words in consecutive sentences. Inclusion of such a rule in PROS would have led to two differences less with SUBJECTS and one difference less with PB. 3. S e n t e n c e final participles. There are nine sentence final participles in the material, none of which is accented in PROS. In five cases the absence of accent is shared with SUBJECTS and PB, in one case with SUBJECTS but not PB, and in three cases with neither SUBJECTS or PB. Of the latter three, two could be explained taking recourse to semantics; the third case has to do with the structure of the sentence in terms of predicates, arguments, and conditions. With the present state of affairs, only the last case could be captured in a rule. 4. S m a l l w o r d s . There are three "small words" in the material, namely nu, meer, and zo, which are always accented in PROS but never in SUBJECTS nor (except for one case) PB. Of course, these words c a n be accented in Dutch, but apparently not in all contexts or functions. Perhaps exceptions could be formulated by making use of certain context cues, such as the presence of mogelijk to the right of zo However, the gain of such a specification is relatively small. 5. D o u b l e n a m e s . In all seven cases of "double names", such as Dominicaanse Republiek or Bondsrepubliek Duitsland, PROS assigns an accent both to the first and second element, whereas the first element is never accented in SUBJECTS and about half of the times in PB. The variability in PB suggests that the accents in question may be optional, but if one should opt for agreement with SUBJECTS, it would not be difficult to formulate a rule, profiting from the fact that the two name elements usually begin with a capital. 6. A f t e r n e g a t i o n . There are three cases where a word preceded by a negation receives an accent in PROS but not in SUBJECTS. However, the deaccentuation seems to be a matter of preference since PB does assign an accent in two of the three cases. It would probably take quite a bit of research to detect possible regularities in this phenomenon. 7. A f t e r k n o w n - q u a l i f i e r s . There is a rule in PROS which makes use of so-called "known-qualifiers", such as deze and dergelijke, to deaccentuate words supposedly containing given information. In one case the rule works out well; in two cases, where both SUBJECTS and PB place an accent, it does not. Perhaps the PROS rule is too general, but it might be difficult to limit its application. 8. W r o n g w o r d label. There are eleven cases where a difference between PROS and SUBJECTS and/or PB seems to be the result of a word labeling error rather than defective accent assignment. All pertain to noun < - > verb confusions. With a correct syntactic analysis (some of) these differences could perhaps have been avoided. 9. A d j e c t i v e s . The material contains eleven adjectives which are accented in PROS but not in SUBJECTS. However, the performance of PB suggests that the accents in question are optional, so that there seems to be little sense in formulating a rule with a view to deaccuenting adjectives. Moreover, an unrestricted adjective deaccentuation rule would be so general that it would first have to be assessed in which cases (de)accentuation is obligatory and in which cases it is optional. The data given above show that there are cases of PROS accentuation which differ both from SUBJECTS and PB and that there are cases which differ from only one of the two. Assuming that the accentuation of the two maximum references is correct or at least acceptable - some caution may be in place for a few idiosyncratic PB-accents - it

26

could be hypothesized that the former are accentuation "errors", whereas the latter are not. Thus an attempt to improve the PROS algorithm should concentrate on the cases where PROS differs both from SUBJECTS and PB. But even these cases should not be taken at face value. After all, it is difficult to estimate the representativeness of the certainly limited sentence material used in the present evaluation. This means it is somewhat risky to draw definite conclusions as to the (in)adequacy of particular PROS rules solely on the basis of the results reported here. What would help are statistics on accentuation from a large corpus of spoken Dutch. Impressionistic data based on a limited number of observations may lead to wrong generalizations. That some of the differences between PROS on the one hand and SUBJECTS and/or PB on the other are indeed relevant in terms of adequacy, i.e. do not (all) fall under the heading of free variation, appears from the results of the two evaluation tasks described below. 3.2 Paper-and-pencil evaluation As mentioned in Section 2.4, in the paper-and-pencil evaluation the sentence material was presented (randomly mixed) in two versions: PROS and RANDOM. The subjects were asked to rate each of the 64 sentences on a 10-point adequacy scale. The reliability of the ratings, adjusted for subject bias, was high: an Ra of .90 for PROS and of .87 for RANDOM, which means that the subjects agreed very well on the relative adequacy of the different sentence structures (for an explanation of the reliability coefficient, see Asendorpf and Wallbott, 1979). Table 1. Paper-and-pencil evaluation. Mean adequacy scores, averaged over subjects, and standard deviation (in parentheses) for two accent versions.

Text sentences Single sentences Overall

RANDOM

PROS

4.8 (1.4) 4.7 (1.8) 4.8 (1.7)

6.1 (1.8) 6.8 (1.6) 6.3 (1.7)

As appears from Table 1, the PROS accents were judged to be more adequate than the RANDOM accents. An analysis of variance proved the difference to be significant at the 1% level. In terms of the meaning of the scale terms, the average adequacy of the PROS accents can be termed "sufficient" and that of the RANDOM accents "insufficient". However, for both accent versions adequacy varies considerably as a function of sentence. The variation in the scores for the RANDOM sentences is self-evidently due to chance, and therefore not worth looking into. However, the variation in the scores for the PROS sentences must result from specific properties of the PROS algorithm which lead to incorrect accentuation decisions. One way to gain insight into the relationship between the adequacy scores and properties of the PROS accents is to compare PROS with SUBJECTS by assessing the number and type of differences in accentuation and trying to estimate their effect on the subjects' adequacy scores. Three predictors were constructed. In the First place, on the basis of the difference in the total number of accents between PROS and SUBJECTS it can be hypothesized that PROS often assigns too many accents in a sentence and that this has a negative effect upon the subjects' adequacy judgments. This hypothesis was tested by correlating the adequacy scores for each sentence, averaged over subjects, with predictor "n", defined as the number of accents in PROS minus the number of accents in SUBJECTS. If the hypothesis is correct, the correlation coefficient should be negative: the larger the difference in the number of accents

27

between PROS and SUBJECTS for a particular sentence, the lower the adequacy score. The correlation coefficient was -.34 (p tot een HOOGTE van ongeveer VEERTIG CENTIMETER. BOVENDIEN werd m e t NAME o p TERSCHELLING VEEL HINDER ondervonden van USVORMING. Hierdoor konden d e MEDEWERKERS van de VOGELWACHT VEEL PLAATSEN

waar zich OLIESLACHTOFFERS bevonden NIET bereiken. ONDANKS de BARRE WEERSOM-

STANDIGHEDEN slaagde men er in DERTIG EIDEREENDEN te vangen.

Notice that (13) induces a new type of error in (14), viz. verbs incorrecdy left unaccentuated (such as bereiken and slaagde). 4. PROS 4.1 Introduction We will now discuss another series of improvements on simple CW-FW algorithms, which have been implemented in PROS. In addition to accentuation which is based on category membership and individual words, PROS contains three types of rules to improve accentuation. The set-up of PROS is as below:

105

(15)

PROS Accentuate CWs except for (i) verbs and (ii) words specified in the lexicon as [-accent]. Adjust accentuation by means of three types of rules: a. Rhythmic deaccentuation. b. Deaccentuation of words in contexts signalling given information. c. Default accentuation of VERBs in specific contexts. In addition to such accentuation rules, PROS contains rules for word class assignment and prosodic phrase construction. These modules will not be discussed here, but we will assume their output to be accessible to the accentuation module. See for more specific information on labelling and phrase formation Kager and Quen6 (1987). In the following sections we will discuss (15a-c). 4.2 Rhythmic deaccentuation If a word occurs between two accentuated words (of specific categories), then this word is deaccentuated. All three words must belong to the same prosodic phrase. The common general context of rhythmic deaccentuation rules is in (16), which also shows two instantiations of this context. (16)

Rhythmic-DEaccentuation (general context) [+ACC] [-ACC] / t+ACC] [+ACC] Q-TERM CW CW "three thousand years" drie duizend gulden half miljoen "half million guilders" tempe raturen zeer lage "very low temperatures" CW ADV PART3 buren droevig kijkende "sad looking neighbours" zeer verbaasde toeschouwers "very amazed spectators" spoorlijnen lopende "parallel running tracks" evenwijdig Rhythmic deaccentuation is restricted to contexts where it is certain that it should apply, in order to avoid incorrectly deaccentuating the middle term in e.g. an ADV-ADJ-NOUN sequence such as ongelooflijk simpele middelen ("incredibly simple means"), or typisch tijdelijke arbeid ("typically temporary employment"). 4.3 Deaccentuation in contexts signalling given information The second type of rule in (15) is deaccentuation of words in contexts signalling given information. Qualifiers such as dit, deze ("this"), dergelijke, zo'n, zulke ("such") imply that the referent of the qualified term has been introduced earlier in the text. Therefore, these words signal given information, or [-focus]. Two separate classes of deaccentuating qualifiers are distinguished, for reasons related to default verb accentuation. (17)

Deaccentuation of words in contexts signalling given information. [+ACC] -» [-ACC] / KNOWN-QUALIFIER a. {dit, deze, dergelijk(e), zo'n, zulke,...} b. {ander(e), volgend(e), vorig(e), (de)zelfde} "different" "next" "previous" "same" Obviously, the basis of this type of rule is probabilistic. In a minority of relevant cases, the qualified term is [+focus], but the examples of (18) certainly represent the majority: (18) a. Het GEVOLG van zulke temperaturen is VRESHLUK The CONSEQUENCE of such temperatures is AWFUL

106

b.

Ze heeft de VOLGENDE trein genomen She has the NEXT train caught ("She has caught the next train")

4.4 Default accentuation of verbs in specific contexts According to rule (15c), verbs are accentuated in specific contexts. As will be clear from the discussion of thematic structure and accentuation above, a verb (being a predicate) will be accentuated if it is in a focus-domain on its own. This will happen in two cases. First, if a [+focus] adverbial term intervenes between an argument and the predicate (cf. 7c). Second, if no [+focus] argument is present to form a domain with the predicate, for instance in case the argument is either [focus] or simply absent. Let us now discuss these cases one by one. 4.4.1 Adverbial constituents In order to mimick the accentuation effects of blocked focus-domain assignment after adverbials, both thematic and focus information is needed. Again, we take advantage of heuristics. Adverbial terms can consist either of single adverbs or of adverbial phrases, which are often PP's. In the former case, an adverb will trigger accentuation of a following veib: (19) Default accentuation of VERBs following AD Verbs INT VERB -> [+ACC] / ADV Note that there is no direct or indirect check on the ADV being [+focus]. The reason for this is probabilistic again, as we have found that even particle-like ADVs such as al are usually accompanied by stressed verbs. The rule refers to a following INT (intonational phrase boundary), which is inserted in the prosodic phrase structure module. It serves to guarantee that the verb to be accentuated is syntactically clause-final and not clause-medial, where it would typically be followed by an argument. Dutch has a basic SOV word order, but the finite verb is moved to second position in the matrix clause. See the following example (cf. 3a): (20)

De BUREN hebben

HEFTTGGEPROTESTEERD

ADV VERB INT PROS can detect adverbial phrases consisting of PP's through their initial preposition. That is, prepositions such as ondanks, wegens, sinds introduce constituents which take the place of ADV in (19). To check the constituency of a sequence of woids starting with these prepositions, PROS contains heuristics for detecting (simple) constituents, which will not be discussed here. An example of a sentence containing an adverbial PP is below: (21) De BUREN hebben ONDANKS onze HULP GEPROTESTEERD [ ADV ] VERB INT The simple constituent ondanks onze hulp is detected as an adverbial phrase, triggering default accentuation of the verb by rule (19). 4.4.1 Deaccentuated constituents We will now turn to the interesting effect of default accentuation of verbs accompanied by [-focus] arguments. As (17) gives us a probabilistic indication of [-focus] terms, we can employ these cues to arrive at a more adequate accentuation of CW verbs. In the rule below, KQ(17a) is the KnownQualifier, and CW a deaccentuated term following it: (22) Default accentuation of VERBs following Known-Qualifiers VERB -> [+ACC] / KQ(17a) CW INT

107

Notice that only one type of qualifier (17a) is referred to in the rule. This is because we have found that the other type of qualifier (17b) frequently takes the integrative accent in an argument-predicate structure. Another type of default accentuation is required for verbs whose argument is pronominal, hence can be assumed to be [-focus]. See (23) below: (23)

Default accentuation of VERBs following PRONouns VERB -> [+ACC] /PRON INT Examples of verbal default accentuation by (22), (23) are below: (24) a.

Ze hebben b.

dergelijke

temperaturen OVERLEEFD

KQ CW ("They have survived such temperatures")

De ZEEHONDEN hebben

het

INT

OVERLEEFD

VERB INT) ("The seals have survived it") Finally, we will discuss an extension of verbs, as occurring in the above rules, to complex verbs, of which the stem can occur separately from the particle. Stem and particle are consistently together only in participles (25a). But in infinitives, the infinitive maricer te can be interposed (25b), whereas in finite forms in matrix clauses, the verbal stem in verb-second position can be far away from the particle in clause-final position (25c): (25)

PRON

VERB

De buren hebben zich gisteren hevig opgewonden ("Yesterday the neighbours were highly agitated") b. De buren begonnen zich gisteren hevig op te winden ("Yesterday the neighbours started to get highly agitated") c. De buren wonden zich gisteren hevig op ("Yesterday the neighbours were highly agitated") An interesting property of separable verbs is that the particle part is accentuated, even if the verbal stem is far away in verb-second position in matrix clauses. (26)

a.

a.

De BUREN hebben zich gisteren hevig OPGEWONDEN

c.

De BUREN wonden zich gisteren HEVIG OP

b.

De BUREN begonnen zich gisteren HEVIG OP te winden

The conditions for accentuation of these verbs are identical to those holding for non-separable verbs, and incorrect non-accentuation of a verbal particle is accordingly a serious error. However, particles occurring in separable verbs are often identical to prepositions {op being an example of such an ambiguous form). This presents difficulties in detecting verbal particles, especially when these are 'stranded' in clause-final position. PROS contains rules which fairly accurately locate clause boundaries, so that clause-final prepositions, as potential particles, can be accentuated. The rules of verbal accentuation will generally correctly accent such clause-final forms. We will now give an impression of the performance of PROS. When the text used earlier is fed into PROS, the output in (27) results*: (27)

De ZOEKACTIE op TERSCHELLING naar de met OLIE besmeurde VOGELS is dit weekeinde ERNSTIG BEMOEILDKT door de ZWARE SNEEUWVAL ZATERDAGNACHT en ZONDAG. D e SNEEUW stapelde zich OP tot een HOOGTE van ongeveer VEERTIG CENTIMETER. BOVENDIEN werd met NAME op TERSCHELLING VEEL HINDER ondervonden van IJSVORMING.

Hierdoor konden de MEDEWERKERS van de VOGELWACHT VEEL PLAATSEN waar zich OLIESLACHTOFFERS bevonden NIET BEREIKEN. ONDANKS de BARRE WEERSOMSTANDIGHEDEN slaagde men er in DERTIG EIDEREENDEN te vangen.

Clearly, the number of errors has decreased as compared to the outputs of the simple accentuation algorithms discussed in section 3. Note especially the correct accentuation of verbs (besmeurde,

108

BEMOEIUJKT, stapelde, ondervonden, bevonden, BERElKEN, vangen), and the verbal particle op. But certainly the output of PROS cannot be considered ideal, as will be clear from (27). Deaccentuation of 'known qualifiers' over-applies in dit weekeinde, where the anaphor dit does not introduce a known or given referent. As we have remarked above, however, this use of anaphoric qualifier represents a minority of the actual cases attested in written text. Conversely, the known, or [-focus] status of De sneeuw is not signalled by an explicit 'known qualifier'5, and therefore SNEEUW is incorrectly accentuated. The noun plaatsen is preferably unaccentuated, probably since it is semantically fairly empty. However, accentuating it does not produce a severe error. Finally, PROS leaves the verb slaagde unaccentuated, as it does not detect the prosodic boundary after er in. Had this boundary been detected, then slaagde would have been the only CW in a prosodic domain, and it would have been accentuated (by a rule that has not been discussed here). 5. Conclusions PROS represents an attempt to arrive at fairly acceptable sentence accentuation of raw text. In this respect, it performs considerably better than algorithms using only lexical information, viz. FWCW distinctions and the accentuation properties of categories and individual words. In addition to lexical information, PROS takes into account accentuation properties of sequences of words. This is achieved by means of probabilistic rules; the principal ideas behind these rules, however, are wellgrounded in theoretical descriptions of sentence accentuation and prosodic phrase structure.

Notes 1 The ordering between the first two clauses of (6a) reflects the preference for analysing A-P-A into A-(PA) over (AP)-A. A sequence of Subject-VERB-Object will be organised into two focus domains: the Subject, and the predicate phrase. 2 The gloss reads: "The search on [the island of] Terschelling for the oil-polluted birds has been seriously hampered dining the weekend by the heavy snowfall on Saturday night and Sunday. The snow piled up to a level of some 40 centimeters. Also, on Terschelling in particular, much hindrance was experienced from ice-formation. This kept bird watch co-operators from reaching the oil victim's sites. In spite of the severe weather conditions, one succeeded in catching 30 eiderducks." 3 The PARTiciples of (16b) are accentuated (in spite of their being VERBs) by rule (19). 4 We have not indicated phrase boundaries. 5 The definitite article De is too weak an indicator of [-focus] status. References ALLEN, J., M.S. HUNNICUTT, and D. KLATT 1987 From Text to Speech: The MITalk system, Cambridge, UK: Cambridge University Press BAART, J.L.G. 1987

Focus, Syntax, and Accent Placement, dissertation Rijksuniversiteit Leiden

BAART, J.L.G. and J. HEEMSKERK 1988 The problem of ambiguity in morphological analysis for a Dutch text-to-speech system, Proceedings (7thand FASE symposium), Edinburgh, 959-965 BERENDSEN, E., S.SPEECH'88 LANGEWEG, H. VAN LEEUWEN 1986 Computational phonology: Merged, not mixed, Proceedings COLJNG'86, Bonn, 612-615

109

BEZOODEN, R. van 1989 Evaluation of an algorithm for the automatic assignment of sentense accents, in H. Bennis and A. van Kemenade (ed) Linguistics in the Netherlands 1989, Dordrecht: Foris (this volume) DAELEMANS, W. 1987 Studies in language technology: An object-oriented computer model of morpho-phonological aspects of Dutch, Ph.D. thesis Leuven University GUSSENHOVEN, C. 1983 On the grammar and semantics of sentence accents, Dordrecht: Foris 't HART, J. and R. COLLIER 1975 Integrating different levels of intonation analysis, Journal of Phonetics 3:235-255 KAGER, R. and H. QUEN6 1987 Deriving prosodic sentence structure without exhaustive syntactic analysis, in J. Laver and MJ. Jack (eds) Proceedings European Conference on Speech Technology, Edinburgh, 1:243-246 KAGER, R. and E. VISCH 1987 Metrical constituency and rhythmic adjustment, Phonology 5:21-71 KERKHOFF, J., J. WESTER, and L. BOVES 1984 A compiler for implementing the linguistic phase of a text-to-speech conversion system, in H. Bennis and W.U.S. van Lessen Kloeke (eds) Linguistics in the Netherlands 1984, Dordrecht: Foris, 111-117 KRUYT, J.G. 1985 Accents from speakers to listeners. An experimental study of the production and perception of accent patterns in Dutch, dissertation Rijkuniversiteit Leiden. LAMMENS, J. 1987 A lexicon-based grapheme-to-phoneme conversion system, in J. Laver and M.A. Jack (eds) Proceedings European Conference on Speech Technology, Edinburgh, 1:281-284 NOOTEBOOM, S.G., and J.G. KRUYT 1987 Accents, focus distribution, and the perceived distribution of given and new information: An experiment, Journal of the Acoustic Society of America 82(5):1512-1524 O'SHAUGHNESSY, D. 1976 Modelling fundamental frequency and its relationship to syntax, semantics, and phonetics, Ph.D. thesis, M.I.T., Cambridge MA TERKEN, J.M.B. 1985 Use and function of intonation: Some experiments, dissertation Rijksuniversiteit Leiden VAN WIJK, C„ and G. KEMPEN 1985 From sentence structure to intonation contour, in B.S. Muller (eds) Sprachsynthese: Zur Synthese von natürlich gesprochener Sprache aus Texten und Konzepten. Hildesheim: Georg Olms, 157-182

Rene Mulder & Pim Wehrmann

Locational verbs as unaccusatives 1. Introduction* Dutch has a class of verbs that allow the AP vol (met) NP in their complements, forming sentences such as 'De kamer staat vol mensen', 'The room stands full (of) people'. In this paper we will argue that this construction has a representation which is comparable to that of sentences containing verbs of the well-known smear/swarm-class. From this analysis it follows that Dutch locational verbs like zitten, liggen, staan and hangen ('sit', 'lie', 'stand' and 'hang'), all of which allow vol in their complements, are unaccusatives. We will present further support for this conclusion. 2. The vo/-alternation Consider the sentences in (1) and (2). (1) a Hij giet het water in het bad He pours the water in(to) the bath b Hij giet het bad vol water He pours the bath full (of) water (2) a Het water stroomt in het bad The water runs in(to) the bath b Het bad stroomt vol water The bath runs full (of) water In each of these sentences the same elements are related, the one (het bad) being the location of the other (het water). What distinguishes the a- from the b-sentences is that location and entity-to-be-located occupy different syntactic positions. The locative element can either be expressed in a PP, as in the a-sentences, or in a position marked with accusative or nominative Case, as in (lb) and (2b), respectively. The element to be located appears either with accusative or nominative Case, as in (la,2a), or in the complement of the adjective vol, as in (lb,2b). What we observe is that a specific structural position, which is marked accusative in (1) and nominative in (2), can contain the locative element (henceforth: LOC) as well as the material to be located (henceforth: MAT, using the terminology of Hoekstra & Roberts 1988). Exactly the same situation can be observed with sentences containing verbs of the class of which smear, load, swarm and crawl are some of the typical examples (for a discussion, see e.g. Anderson 1971 and Fillmore 1968). Two Dutch examples are given in (3) and (4): (3) a

Hij smeert boter op zijn brood He spreads butter on his bread b Hij besmeert zijn brood met boter He (aff)-spreads his bread with butter (4) a De bijen zwermen in de tuin The bees are swarming in the garden b De tuin zwermt van de bijen The garden is swarming with (lit. 'of the') bees Once again, each of these sentences expresses a relationship between an entity (de boter, de bijen) and its location (zijn brood, de tuin), both of which can occupy the same structural position, marked with accusative or nominative Case, as in (3) and (4), respectively.

112

There are further resemblances between the sentences in (1,2) and those in (3,4). First observe that when LOC is expressed as an accusative or nominative NP, as in (3b,4b), MAT appears in the complement of one of the prepositions met Svith' or van 'of. The same is possible in sentences of the type in (lb,2b). For instance, the bare NP in the complement of vol can be replaced by a mei-PP in (lb,2b), while in other cases, vol also allows a van-PP in its complement: (5) Hij giet het bad vol met water He pours the bath full of (lit. Svith') water (6) De kamer staat vol met/?van rook The room stands full of smoke "The room is filled with smoke' A further similarity between (1) and (2) on the one hand, and (3) and (4) on the other hand, concerns the notion of 'completeness'. As discussed in Anderson (1971) and other work cited there, the a- and b-sentences of (3) and (4) differ slightly in meaning. While the b-sentences only have a completive (or: 'holistic') reading, in the sense that they claim that the entire location is covered with some material, the a-sentences allow a reading in which only some part of the location has material in or on it. Exactly the same contrast shows up between the a- and b-sentences of (1,2), an observation also made in Van Oosten (1986). Here too, when the LOC element occupies the position marked with structural Case, i.e. in the b-sentences, the action described in the predicate is interpreted as being 'completive' in the sense that the entire location is affected by it, not just a part, as is possible in the a-sentences. The fact that the sentences in (3,4) and those in (1,2) are alike in a number of respects suggests that we may assign them comparable representations. Recently, specific analyses have been proposed by Hoekstra, Lansu & Westerduin (1987) for sentences as in (3), and Hoekstra & Roberts (1988) for sentences as in (4). What these analyses have in common is that they both can be interpreted as taking the relation between LOC and MAT as a structural one, expressible in terms of a Small Clause (SC)-complement to the verb. Either LOC or MAT can be the subject of this SC, deriving the two variants of the pairs in (3,4), as shown in (7): (7) a V [NP mat LOC] s c b V[NP l 0 C XMAT] S C The main difference between (3) en (4) is that the verb assigns an external 6-role in the former sentence, but not in the latter. Under Burzio's generalization, the verb in (3) assigns accusative Case to the subject of the SC, which can be either MAT or LOC. In (4), the verb has a 9'-subject and consequently fails to assign Case to the SC-subject. This NP, either MAT or LOC, moves to the matrix subject position to receive nominative Case. We propose that this analysis be extended to the cases m (1,2), taking the a-sentences to be instances of (7a), and the b-sentences to be cases of (7b). Note that in (7b) it is not the MAT-entity that constitutes the head of the SC, but rather the element X, which means "complete". This element may be be-, as argued by Hoekstra et al (1987) for (3b), or vol, as in (lb,2b). The PP expressing the MAT-entity cannot be the head of the SC, since it may occur postverbally: (8) a dat hij het bad vol (met water) giet (met water) (cf. 5) b dat hij zijn brood (met boter) besmeert (met boter) (cf.3b) That vol must be the head is shown by the fact that it can incorporate into a higher verb, which is not possible for adjuncts. This is shown in (9), where the matrix-verb, like any modal, allows only verbal elements on its right. Also, its presence is obligatory, contrary to the MATPP. (9) a b (10)

dat hij de tank vol wil gooien/wil volgooien met benzine that he the tank full wants (to) throw/wants (to) full-throw with gas dat hij de kast snel wil schilderen/*wil snel schilderen that he the closet fast wants (to) paint/wants (to) fast-paint dat hij de tank *(vol) gooit (met benzine)

1 13

The MAT-PP might be analysed as an adjunct, coindexed with the head of the SC by some interpretative mechanism (cf. Hoekstra et al 1987). Alternatively, it could be construed as a complement to the adjective. Since generally, extraposition is possible for PP-complements, the wordorder facts in (8) would then be captured. A property of the constructions in (3) and (4) that has not been mentioned yet is the fact that sometimes a third sentence can be added to these pairs. So in addition to (4a) and (4b), we find the following possiblity: (4) c Het zwermt van de bijen in de tuin It is swarming with (lit. 'of the') bees in the garden Hoekstra & Roberts (1988) discuss this construction extensively. They argue that (4c) can be related to (4b) if it is assumed that the locative PP in de tuin indirectly functions as the SC subject, through coindexation with the semantically empty element het 'it'. Het subsequently moves to the sentential subject position to receive nominative Case. Here they follow Bennis & Wehrmann's (1987) proposal that coindexation between so-called 'dummy'-/zei and a locative adjunct PP can account for the fact that this PP can be understood to bear the same thematic role as an argument-NP in a corresponding structure, as in (11). On this assumption, (4c) can be schematically represented as in (7c): (11) a b

Dit café is gezellig Het is gezellig in dit café It is cosy in this café (7) c [hetj X MAT] SC LOC¡ On the assumption that the analysis proposed for (3) and (4) can be extended to cases like (1) and (2) with the adjective vol, we expect predicates of the latter type to appear in (7c) as well. This prediction is borne out, as is shown by the following example: (12) a De mensen zijn in de kamer b De kamer is vol met mensen c Het is vol met mensen in de kamer It is full of (lit. Svith') people in the room Turning now to the external syntax of the vo/-phrases under discussion, we correctly predict that they will occur in exactly those contexts where SCs generally appear. Restricting our attention to the complement of verbs, we find two classes of verbs selecting a SC with vol-. First, we find unergatives like leggen 'lay', zetten 'put', doen 'do', gieten 'pour', or gooien 'throw' (other examples are listed in Paardekooper 1986:265). These verbs assign accusative Case to the subject of the SC, which can be either MAT or LOC or, in some cases, het 'it' coindexed with an adjunct-PP expressing LOC. Examples of this construction are the sentences in (1), involving the unergative verb gieten 'pour', and those in (15) below. The second class of verbs selecting a SC with vol includes a number of unaccusative verbs, such as zijn 'be', raken 'become', stromen 'stream, run', komen 'come'. Some examples are given in (2) and (12). That these verbs are unaccusatives is evident from the fact that they select zijn 'be' as their perfect auxiliary, as shown in (13) below: (13) Het bad is vol water gestroomd The bath is full water streamed Interestingly, the verb hangen 'hang' appears in two different paradigms: (14) a De slingers hangen in de zaal The streamers hang in the hall b De zaal hangt vol met slingers c Het hangt vol met slingers in de zaal (15) a Hij hangt de slingers in de zaal He hangs the streamers in the hall b Hij hangt de zaal vol met slingers c Hij hangt het vol met slingers in de zaal

114

In (15), hangen is unergative. The more interesting case is (14), where in line with the analysis so far, we would have to say that hangen is unaccusative. The problem with this claim, however, is that hangen fails to show the properties that are characteristic of unaccusative verbs in Dutch, for example selection of zijn ('be') as the perfect auxiliary. With other locational verbs, like zitten, liggen and staan, the same situation obtains. Their occurrence in paradigms like in (12) and (14) suggests that they are unaccusative verbs. On the other hand, these verbs form their perfect tense with hebben, not zijn, and fail to exhibit other characteristics of unaccusativity as well. Consider the pair in (16): (16) a b

De mensen zitten in de zaal De zaal zit vol met mensen The hall sits full of (lit. Svith') people In (16b) the subject and the verb are clearly not thematically related, which can only indicate that the subject must be a derived subject. Now if the subject is not a deep structure subject in (16b), and if we take the proposed structural relationship between sentence pairs such as (16a) and (16b) seriously, then the subject cannot be a deep structure subject in (16a) either. I.e., zitten is an unaccusative verb with a SC-complement even in a case like (16a), for which an unergative analysis would have been conceivable as well. Of course, this argument rests on the assumption that (16b), for which an unaccusative analysis is the only one conceivable, and (16a) are really related. Evidence for this comes from two sources. First, as we have already seen, sentences like (16a) and (16b) are very similar to sentence pairs involving verbs of the smear/crawl-class, which are undoubtedly related. Second, sentence pairs like (16a) and (16b) must be related because they share the same selectional requirements. For example, in many cases in which English simply uses be to express the location of something, Dutch uses one of the locational verbs zitten, liggen or staan. The question as to when to use these verbs is a difficult one, the answer depending on the nature of what we have so far been calling LOC, MAT or the combination of them. For discussions of the factors involved, see Van den Toorn (1975) and Van Oosten (1986). For our purposes, a few examples will do. When MAT is a round object, liggen is selected, as in (17). But when LOC is an enclosed space, zitten is selected (cf. Van Oosten, p. 146), as in (18). When MAT and LOC have something to do with printing or writing, staan is selected, as in (19). (17) a

De knikkers liggen/*zitten/*staan op de tafel The marbles lie/sit/stand on the table b De tafel ligt/*zit/*staat vol knikkers (18) a De knikkers *liggen/zitten/*staan in het zakje The marbles lie/sit/stand in the bag b Het zakje *ligt/zit/*staat vol knikkers (19) a Het nieuws *ligt/*zit/staat in de krant The news lies/sits/stands in the newspaper b De krant *ligt/*zit/staat vol nieuws What concerns us here is that the a- and b-sentences always choose the same option. From this we conclude that they are structurally related. To account for the facts in (17)-(19) we can assume that in these cases selectional requirements are imposed on the relation between the verb and the SC as a whole, irrespective of the SC-internal order of MAT and LOC. This concludes our discussion of the voZ-construction. It is clear that the analysis presented so far can only be maintained if we adopt the assumption that locational verbs are unaccusative. We will present further evidence for this claim in the next section. 3. Locational verbs as unaccusatives As noted earlier, the assumption that lifxen, zitten, hangen and staan are unaccusatives is problematic in view of Hoekstra's (1984) claim that there is a one-to-one correspondence between unaccusativity and selection of zijn as the perfect auxiliary. Therefore, it is necessary to supply independent evidence for the view that locational verbs are unaccusative in Dutch.

1 15

3.1. Small Clauses The first argument concerns the causative counterparts of the locational verbs, shown in (20). The interpretation of the b-examples is 'make/do V', where V corresponds to the locational verbs in the a-examples. Uncontroversially, we analyse the complement of the causative verbs in (20b) as SCs. The relevant structure is NP V [NP [P NP]]\ (20) a

Het kind zit in de stoel The child sits in the chair Het boek ligt op tafel The book lies on table De jas hangt in de kast The coat hangs in the closet De lamp staat in de hoek The lamp stands in the corner

That we are dealing with SC-complements here not be omitted (21) or extraposed (22): (22) a (21) a * Ik zet het kind I set the child b b * Ik leg het boek I lay the book c * Ik hang de jas c I hang the coat

Ik zet het kind in de stoel I put the child in the chair Ik leg het boek op tafel I lay the book on table Ik hang de jas in de kast I hang the coat in the closet Ik zet de lamp in de hoek I put the lamp in the corner is shown by the fact that the PPs in (20b) may * dat ik het kind zet in de stoel that I the child set in the chair * dat ik het boek leg op tafel that I the book lay on the table * dat ik de jas hang in de kast that I the coat hang in the closet

Given the correspondence between the locational verbs and their causative counterparts, it follows that the relations between the NPs het kind, het boek, de jas on the one hand, and the PPs in de stoel, op tafel, in de kast on the other hand in (20a) are identical to those in (20b). If these subject-predicate relations are to be represented as SCs in (20b), the same must be the case in (20a), in accordance with Baker's (1985) UTAH. This means that locational verbs are unaccusative. As a consequence, we predict that the locational PPs in (20a) may not be omitted or extraposed either. Taken at face value, this appears to be incorrect, since the examples in (23,24) are not ungrammatical, although they do sound a little odd: (24) a Het kind zit dat het kind zit in de stoel The child sits that the child sits in the chair b ?Het boek ligt b ? dat het boek li^t op tafel The book lies that the book lies on table c ?De jas hangt c ? dat de jas hangt in de kast The coat hangs that the coat hangs in the closet d ?De lamp staat d ? dat de lamp staat in de hoek The lamp stands that the lamp stands in the corner This complication is easily resolved, however, by assuming that locational verbs may select either a SC or an NP-argument. If the argument is a NP, it is necessarily interpreted as performing the action denoted by the verb. In some cases, i.e. abstract NPs, this is impossible. Therefore, such cases can only be interpreted as involving a SC-complement. Hence, PPextraposition will then be completely impossible, as predicted, which is shown in (25). Also, the PP may not be omitted, viz. the contrast between (24) and (26): (23) a

(25) a

dat het nieuws *staat in de krant/in de krant staat that the news stands in the newspaper dat de humor *ligt op straat/op straat ligt that the humour lies on (the) street (26) a * Het nieuws staat b * De humor ligt We claim that the examples in (20a) are structurally ambiguous as to whether the verb is combined with a SC or with an NP. In section 3.4. we will return to this ambiguity.

1 16

The SC-argument gains force when we look at idioms. In (27a) we have an idiomatic expression in the complement of leggen and zetten. Crucially, the idiomatic reading is preserved when the causatives in (27a) are replaced by the corresponding locational verbs, as in (27b), modulo the nominative Case on the SC subject: (27) a Ik leg hem aan de ketting b Hij ligt aan de ketting I lay him on the chain He lies on the chain 'I'm giving him less room' Ik zet hem voor het blok Hij zit voor het blok I set him before the log He sits before the log 'I'm forcing him to decide' Ik zet hem voor paal Hij staat voor paal I set him for pole He stands for pole 'I'm making a fool out of him' A similar case can be made for intransitive adjectives like vast: (28) a Ik leg de boot vast b De boot ligt vast I lay the boat tight The boat lies tight 'I secure the boat' "The boat is secure' From this case, it is even more evident that locational verbs involve raising, since the S-subject has to be the subject of the secondary predicate at D-structure, in order to allow for the idiomatic reading. 3.2. The German /asjen-passive A further argument draws on German data; it is based on the assumption that the nature of locational verbs is the same in German as in Dutch. It is well known that passivization of causative and perception verbs is subject to severe restrictions across languages. In Dutch, for instance, it always leads to a bad result: (30) a Ik laat (de loodgieter) mijn gootsteen repareren I let (the plumber) my sink repair b Ik laat Marie vertrekken I let Mary leave c * De loodgieter wordt mijn gootsteen laten repareren The plumber gets [lit. becomes] my sink let repair d * Mijn gootsteen wordt (de loodgieter) laten repareren My sink gets (the plumber) let repair e * Marie wordt laten vertrekken Mary gets let leave In German, passivization of lassen 'let' is permitted, but only if the embedded verb is unaccusative, according to Webelhuth (1982:m 20): (31) a *Er wurde tanzen gelassen He was dance let b Die Werkzeuge wurden fallen gelassen The tools were fall let It need not concern us here how this contrast should be accounted for. What is relevant here is the fact that locational verbs pattern with unaccusatives: (32) Das Buch wurde liegen gelassen The book was lie let This supports our claim that locational verbs are unaccusative.

117

3.3. Other languages Locational verbs behave like unaccusatives in other languages as well. To see the relevance of this, one should consider how children learn to classify verbs as unergative or unaccusative. In languages like Italian, where there is an abundance of positive evidence, this may be a trivial question, but for a languages such as English it certainly is not. Here the relevant data are not easily accessible to the language learner. Diagnostics such as auxiliary selection, necliticization, the possibility of impersonal passives etc. are not available in English. However, the distinction does play a role even in English, e.g. in the existential f/iere-construction: (33) a There arrived/*danced three men This indicates that the distinction unergative-unaccusative can only be learnable if it can be derived from semantic primitives. This implies that as long as meaning is preserved, verbs cannot be unaccusative m some languages and unergative in others. Their status has to be the same across languages (cf. Grimshaw 1987, Levin & Rappaport 1988). Evidence from southern German, Italian, English and Finnish suggests that locational verbs behave as unaccusatives in these languages as well. The facts from southern German dialects (i.e. the dialects that are spoken in Austria, Switzerland and Bavaria, cf. Haider 1985), concern auxiliary selection. In those cases, the equivalent of be is selected as the perfect auxiliary, thus indicating that locational verbs are unaccusative. (34) Das Buch/Kind ist auf dem Tisch/Boden gelegen/gesessen The book/child is on the table/floor lain/seated We would expect similar facts to obtain in Italian. However, this can not be tested, since, for a reason we do not understand, locational verbs do not allow perfectivization at all. However, «e-cliticization gives the correct result: (35) a

Molta biancheria pendeva dal filo Much laundry was-hanging from-the line Ne pendeva molte del filo b Of-it was-hanging much from-the line Secondly, locational verbs behave like unaccusatives in English as well, i.e. they may appear in the existential Z/iere-construction: (36) a There are lying some books on the table b There were sitting three men on the bench Regarding Finnish, we take Belletti's (1988) claim that only indefinite internal arguments receive partitive Case as a proper diagnostic. From Fromm (1982) we give the following example: (37) oksilla istuu lintuj a on the twigs sit birds PART

Similar arguments can be drawn from the so-called 'genitive of negation' in Russian (cf. Babby 1980) and from locative inversion in Chichewa (cf. Bresnan & Kanerva 1989). Thus, we find evidence from several languages that supports our claim that locational verbs are unaccusative. 3.4. Dutch It is worth noting that even in Dutch, selection of zijn 'be' does not always lead to a bad result, as the following examples indicate: (38) a Het feestvarken is ?(in de versierde stoel) gezeten The feast-pig is in the decorated chair seated "The guest of honour was seated in the decorated chair'

118

b

Het huis is *(op een prachtige berghelling) gelegen The house is on a magnificent slope lain "The house is situated on a magnificent slope'

As indicated, the presence of a locative PP is required when zijn is selected. This behavior is on a par with observations concerning verbs of motion in Hoekstra (1984). Consider the contrasts in (39) and (40): (39) a b (40) a b

Jan Jan Jan Jan Jan Jan Jan Jan

heeft (naar Groningen) gewandeld has to Groningen walked is *(naar Groningen) gewandeld is to Groningen walked heeft (in de boom) geklommen has in the tree climbed is *(in de boom) geklommen is in the tree climbed

According to Hoekstra, the verbs as used in the a-examples are unergative. In the b-examples, however, they must be unaccusative, due to the presence of the SC-complement. Only participles of unaccusatives may be used attributively in prenominal position, as shown in (41). The examples in (42) and (43) show that the same holds for the sentences in (39,40), as well as for those in (38): * De gelachen man (41) a De man lachte The laughed man The man laughed De gevallen man b De man viel The fallen man The man fell (42) a De *(naar Groningen) gewandelde man The to Groningen walked man b Het *(in de boom) geklommen kind The in the tree climbed child (43) a Het *(in de versierde stoel) gezeten kind The in the decorated chair seated child b Het *(op een berghelling) gelegen huis The on a slope lain house It appears that locational verbs pattern like motional verbs in several respects. Hoekstra notes that the aspectual properties of motional verbs will be different depending upon their syntactic use. Consider example (44): (44) a Jan heeft (in de sloot) gesprongen Jan has in the ditch jumped b Jan is *(in de sloot) gesprongen Jan is in the ditch jumped While (44a) means that Jan performed an act of jumping, while being in the ditch, (44b) can only mean that Jan jumped from the bank into the ditch. So, in its unergative use a motional verb denotes an activity, performed by the subject. On the other hand, when the motional verb is an unaccusative selecting a SC-complement, the main statement made by the sentence is about the subject ending up at a certain location, the action performed by the subject being only of secondary importance. A comparable distinction can be observed with Dutch locational verbs. In the aspectual classification of Dowty (1979), locational verbs are classified as ambiguous between statives and activities. Presumably, they are unergative when used as an activity verb. This would mean that locational verbs can be used in two ways. They can either be unaccusatives selecting a SC, in which case they make a statement about the location of something. Alternatively, they can be unergatives, expressing the activity of a thematic subject, in which case a locative expression need not be present.

1 19

The process of middle formation provides a way of testing this. Note that middle formation is only possible with unergatives. We predict that middle formation of motional and locational verbs is possible only if the attention is focused on the action denoted by the verb, not on the relation between the location and the (unexpressed) agent. This is correct, as shown by the fact that the locative PPs in (45) below can easily be replaced by non-locatives: (45) a Het springt prettig op dit bed/met deze schoenen It jumps nicely on this bed/with this shoes b Het zit makkelijk in deze stoel/met je benen over elkaar It sits comfortably in this chair/with your legs crossed Assuming the parallel between locational and motional verbs to be correct, we should now ask what determines their auxiliary selection. When used as unergatives, both motional and locational verbs select hebben 'have'. When used as unaccusatives, motional verbs select zijn, whereas locational verbs select either hebben or zijn. We feel than within the class of unaccusatives zijn is selected when the predicate has perfective aspect (cf. 38), whereas hebben is selected when the predicate has imperfect aspect. Motional unaccusatives, with their subjects always ending up at a certain location, are necessarily perfective. Hence, they invariably select zijn. For locational unaccusatives, selecting either hebben or zijn, we correctly predict aspectual differences: (46) a

Het kind heeft (een half uur lang) in de versierde stoel gezeten The child has (a half hour long) in the decorated chair seated Het kind is (*een half uur lang) in de versierde stoel gezeten b The child is (a half hour long) in the decorated chair seated A perfective reading is incompatible with an adjunct of duration, such as een half uur lang for half an hour). Therefore, such a durative adjunct is impossible in (46b), but correct in 46a). The fact that the locational verbs pattern with verbs of motion provides extra evidence for the claim that they are, or at least can be, unaccusative. 4. Conclusion In this paper, we have argued that the representation of Dutch constructions containing the adjective vol involves a SC. We have argued that unergative verbs that select vol are comparable to verbs of the smear-class, while unaccusative verbs with vol are like swarm. One of the consequences of our proposal is that the locational verbs zitten, liggen, staan, and hangen are, or at least can be, unaccusative, which is problematic in view of the fact that they fail standard diagnostics such as auxiliary selection. We have supplied independent evidence, showing that the claim that locational verbs can be unaccusative can nevertheless be maintained. Notes * We would like to thank Hans Bennis, Frits Beukema and Teun Hoekstra for valuable comments. We also thank the members of the O.O.O. for stimulating discussion. The research for this article was supported by the Dutch Organization for Scientific Research (NWO), project numbers 300-163-033 and 300-163-034, which is gratefully acknowledged. 1. Notice first that LOC/MAT alternations can be found with predicates other than the adjective vol as well. There are many idiomatic expressions with the same meaning 'be completely filled/covered with', e.g. zwart zien van 'look black of, bol staan van 'bulge with' etc., entering the same alternation. Interestingly, also negative variants are possible, involving e.g. the adjective vrij 'free', combined with unergative houden 'keep', or unaccusative blijven 'stay' as in (i):

120

(i)

a

De auto's blijven uit de binnenstad The cars stay out off the inner city b De binnenstad blijft vrij van auto's The inner city stays free of cars c Het blijft vrij van auto's in de binnenstad It stays free of cars in the inner city Given this observation, we can supply the following considerations to the effect that the MAT-PP is a complement rather than an adjunct. I. Selectional restrictions are imposed on the preposition by the head of the SC, affixal or not: (ii) a vrij *met/van/*aan besmeren met/*van free with/of/to BE-smear with/of arm *met/*van/aan ontdoen *met/van poor with/of/to ONT-do with/of 'get rid of II. Some adjectives from the paradigm of vol do not allow a PP at all, which can be explained by assuming that they are intransitive, provided that the PP is an argument, hence a complement: (iii) Hij haalt de kast leeg (*van kleren) He gets the closet empty of clothes III. SCs like those under consideration here may also be headed by the preposition onder 'under', as shown in (iv): (iv) a Zijn gezicht zit onder de puisten His face sits under the pimples b Er zitten puisten op zijn gezicht There sit pimples on his face Interestingly, onder differs from vol in selecting only an NP and not allowing a PP to express MAT:

(v)

Hij zit onder (*van/*met/*aan) de puisten He sits under of/with/to the pimples In Dutch PPs may not be governed by prepositions in general (cf. Bennis & Hoekstra 1984). Hence, the fact that a PP is excluded in (v) follows immediately if the PP/NP is a complement. Note in passing that this argument also shows that vol must be an adjective in this sort of construction, although it can also occur in postnominal positions, which is normally not allowed for adjectives. 2. Under the SC-analysis, we correctly predict that the vo/-alternation can be found in the absolute mef-construction (i) and in adjunct position (ii): (i) a Met zoveel rotzooi op het strand is zwemmen niet leuk meer With so much junk on the beach is swimming not nice any more b Met het strand vol rotzooi is zwemmen niet leuk meer With the beach full junk is swimming not nice any more (ii) a Zij laten de rotzooi [PRO op het strand] achter They leave the junk on the beach behind b Zij laten het strand [PRO vol rotzooi] achter 3. We thank Anna Cardinaletti for helping us with these data. 4. This fact was brought to our attention by Frans van der Putten. References ANDERSON, S.R. 1971 On the role of deep structure in semantic interpretation, In Foundations of language 7, 387-396 BABBY, L. 1980 Existential sentences and negation in Russian, Ann Arbor, Karoma

121

BAKER, M. 1985 Incorporation, Diss. MIT BELLETTI, A. 1988 The Case of Unaccusatives, In Linguistic Inquiry 19, 1-34 BENNIS, H. & T. HOEKSTRA 1984 Gaps and parasitic gaps. In The Linguistic Review 4,1, 29-87 BENNIS, H. & P. WEHRMANN 1987 Adverbial arguments, In F. Beukema & P. Coopmans (eds), Linguistics in the Netherlands 1987, Dordrecht, Foris BRESNAN, J. & H. KANERVA 1989 Locative inversion in Chichewa, In Linguistic Inquiry 20 DOWTY, D.R. 1979 Word meaning and Montague Grammar, Dordrecht, Reidel FILLMORE, C.J. 1968 The Case for Case, In E. Bach, R.T. Harms (eds), Universals in Linguistic Theory, New York [etc], 1-88 GRIMSHAW, J. 1987 Unaccusatives-An Overview, In J. McDonough & B. Plunkett (eds), Proceedings of NELS 17 1986, vol 2, 244-260 FROMM, H. 1982 Finnische Grammatik, Heidelberg HAIDER, H. 1985 Über 'sein' und nicht 'sein', In W. Abraham (ed), Erklärende Syntax des Deutsche, Tübingen, Narr, 223-254 HOEKSTRA, T. 1984 Transitivity, Dordrecht, Foris HOEKSTRA, T., M.LANSU, & M.WESTERDUIN 1987 Complexe verba, In GLOT 10, 61-77 HOEKSTRA, T. & I. ROBERTS 1988 Middles in Dutch and English, paper presented at GLOW, Budapest LEVIN, B. & M. RAPPAPORT 1988 An approach to unaccusative mismatches, paper presented at NELS 19 OOSTEN, J. van 1986 'Sitting', 'Standing' and 'Lying' in Dutch, In J. van Oosten & J.P Snapper (eds), Dutch Linguistics at Berkeley, Berkeley, 137-160 PAARDEKOOPER, P.C. 1986 Beknopte ABN-syntaxis, 7th ed., Eindhoven TOORN, M.C. van der 1975 Over de semantische kenmerken van staan, liggen en zitten, In De Nieuwe Taalgids 68, 458-464 WEBELHUTH, G. 1985 German is configurational, In The Linguistic Review 4, 3

I.H. Slis & M. van Heugten

Voiced-voiceless distinction in Dutch fricatives 1.

Introduction

Recently, several aspects of the distinction between voiced and voiceless obstruents in Dutch have been the subject of extensive experiments, at least with regard to plosives (e.g. Slis 1985) and two-obstruent sequences (e.g. Van den Berg 1988 and Slis 1985). Intervocalic fricatives, however, have been neglected except for: (1) measurements of the onset and offset of the amplitude envelope of the vowels adjoining French and Dutch voiced and voiceless fricatives (Debrock 1977), (2) a pilot experiment by Van den Broecke and Van Heuven (1976) on voice activity during and duration of Dutch velar fricatives of 4 speakers, and (3) a more elaborate experiment on the perception by 63 listeners of the velar fricatives produced in the above-mentioned pilot (Van Heuven and Van den Broecke 1977) . Debrock (1977) concentrated mainly on the fortis aspect of the voiceless obstruents and the lenis aspect of the voiced ones. His measurements on the speech of 10 Dutch speaking Belgians showed that the rise time of the amplitude of vowels following the voiced fricatives /v/, /z/ and /3/ is longer (reflecting the lenis aspect) than of vowels following the voiceless fricatives /f/, /s/ and /¡/ (fortis). Similar results were obtained with 10 French speaking Belgians. No differences were observed between the rise times following Dutch /x/ (in orthography represented by ) and /y/ (in orthography represented by ). Debrock concluded that no fortis-lenis opposition is made in Dutch velar fricatives. With respect to aspects that are usually assiociated with the voiced-voiceless distinction, he observed that the duration of initial voiceless fricatives was longer than that of voiced ones in both French (59 out of 60 cases) and Dutch (58 out of 60 cases). A comparison of the two classes of fricatives with respect to frication noise intensity did not show consistent differences between voiced and voiceless fricatives. With respect to the occurrence of voicing it was shown that periodicity was absent in Dutch initial /z/, /v/ and /3/ in 6 out of 30 cases; in /y/ no voicing was present in 8 out of 10 cases. Debrock concludes from these data that "as regards to fricatives in Dutch, force of articulation could be a more constant characteristic than sonority" (Debrock, 1977:77). Several phoneticians and phonologists assume that the opposition between voiced and voiceless fricatives is gradually disappearing in Dutch. Especially with regard to the opposition between voiced and voiceless velar fricatives it is generally assumed that the difference between /x/ and /y/ is not made by speakers from the western regions of The Netherlands. The data presented by Debrock (1977) are in accordance with this view. For this reason Van den Broecke and Van Heuven (1976) concentrated on velar fricatives. They performed

124

measurements on the speech of 4 male subjects; six words with intervocalic

and that were part of carrier phrases were repeated 3 times. Voice activity was determined by means of an electrolaryngograph. No vocal cord vibration was observed in the productions. Small differences in the temporal structure of the words were observed. Vowels before voiceless fricatives were about 10 ms shorter than vowels before voiced ones (not significant). The difference in duration between the noise portions of /x/ and /y/ had to an inverse relationship with preceding vowel duration; /x/ proved be about 15 ms longer than /y/. The latter difference was statistically significant. In a subsequent experiment Van Heuven and Van den Broecke (1977) isolated nonsense words containing the embedded velar fricatives ( vs. ) from the utterances that were recorded in the experiment described above. In a first experiment the words were presented in random order; 63 subjects were asked to identify the embedded fricative. The results showed that identification of /x/ was 3% above chance and that of /y/ 2% above chance. A multiple regression analysis showed that "no (complex of) duration parameter(s) seems to govern their (..the subjects..) decision behaviour". After this first experiment a second was performed with 33 of the listeners that participated in the identification experiment. In this experiment the same stimulus words were presented in pairs differing with respect to the feature 'voice' of the fricative. Subjects were requested to indicate which member of a pair was voiced and which was voiceless. Average performance in this discrimination task was about random. The results of these three experiments support the view generally accepted among phoneticians that the voiced-voiceless opposition in Dutch fricatives cannot be maintained in all cases; especially with the velar fricatives this opposition seems to be (nearly) lost. From the phonological and phonetic handbooks it is obvious that more experimental evidence on the subject is needed. Van den Berg (1958), with for instance, states that velar fricatives are pronounced both and without voice activity. He doubts, however, whether a phonological opposition can be maintained, and suggests that the differently articulated velar fricatives can be regarded as combinatory variants. Booij (1981) states explicitly that he refrains from a choice regarding the need for a phonological distinction between [+voice] and [-voice] velar fricatives. Trommelen and Zonneveld (1982) use different phonemes /y/ and /x/ in their phonological system. As a working hypothesis we assume that the speakers intend to realize ¿he voicing indicated by the orthography. As a consequence, the phoneme notation used in this paper is based on the standard spelling of the words. From experimental research, it is known that not only the perceptible presence of voice activity leads to voiced judgements. Other perceptual cues may have the same effect. Duration of the consonant and of the preceding vowel, intensity of the frication noise, the fundamental frequency pattern in the adjoining vowels, and extension in the spectral domain and in duration of formant transitions in the adjoining vowels have been found to influence perceived voicing in intervocalic stop consonants (e.g. Slis and Cohen 1969, Slis 1985 or Lisker 1978) and two-obstruent sequences (Van den Berg 1988). Especially consonant duration proved to be a powerful cue.

125 In the experiment reported here, the voice activity during intervocalic fricatives and their durations were measured. The speakers were selected from two regions in the country. One group originates from the provinces of Noord Holland, Zuid Holland and Utrecht (to be called 'the west'), where it is assumed that the voiced-voiceless distinction is disappearing; the other group originates from the southern part of the province of Limburg (to be called 'the south-east'), where the voiced-voiceless distiction is still assumed to be operational (e.g. Cohen et al. 1971). Voicing of fricatives in standard Dutch (including regional variants) is the main subject of our measurements. Since it is known that men and women differ with respect to the voicing of their consonants in two-obstruent sequences (Slis 1985) we expect to find this effect with intervocalic fricatives as well; therefore, the measurements were done on speech of men and women. The effect of the factor sex on voicing of intervocalic fricatives forms the second goal of the experiment. In addition, it is known that stress has an effect on assimilation of voice (Slis 1986). It is, therefore, expected that presence or absence of voice activity during intervocalic fricatives is also influenced by stress; consequently, the measurements were done on fricatives before and after stressed syllables and between two unstressed syllables. The second parameter measured, viz. duration, also depends on other variables than phonological distinction between [+voice] and [-voice] alone, viz. stress and position in a sentence. The duration of stressed consonants or of speech sounds in the initial part of sentences is found to be longer than that of unstressed consonants or speech sounds in the final part of sentences. Therefore, the measurements were performed on fricatives in initial as well as in final parts of sentences under all three stress conditions. The central question we seek to answer is whether phonetic differences can be found between intervocalic [+voice] and [-voice] fricatives in Dutch which support an underlying phnological distinction. 2. 2.1.

Method Subjects

Forty subjects participated; 20 from the south-east and 20 from the west. Each of the two groups consisted of 10 men and 10 women. Most of the subjects were students; the remainder were employees of the Nijmegen University. All subjects used standard Dutch in their daily conversation, although all subjects from the south-east stated that they, as well as both their parents, occasionally spoke the south-east dialect. Five of the female and 4 of the male subjects from the west testified that they did not speak a dialect at all; nor did their parents. 2.2.

Speech material

The subjects were requested to read 25 meaningful sentences aloud in which all 6 fricatives, viz. /f/, /s/, /x/ (spelled ), /v/, /z/ and /y/ (spelled ), occurred in intervocalic position. The fricatives were preceded by a long vowel in three stress conditions,

126

viz. before and after a stressed syllable (in two-syllable words), and between two unstressed syllables (in three-syllable words). Since no suitable words with initial in a stressed syllable could be found, this condition is not present in the material. Fricatives preceded by a short vowel were only investigated after a stressed syllable, since no words with [+voice] fricatives before stress or between two unstressed syllables exist. All these conditions occurred twice in the material; once in words in the initial part of the sentences and once in words in the final part of the sentences. Speech was recorded in a sound treated booth on a two-channel professional tape recorder (Studer) with a tape speed of 19 cm/s. One channel was used for the audio signal, the other for the output of an electrolaryngograph. Both channels were displayed simultaneously with a UV-oscillograph (SE-oscillograph 6008). Duration and voicing of the fricatives were determined on the oscillographic display using the same operational definitions as in our experiments on assimilation of voice (e.g. Slis 1985 and 1986). A sudden decrease of the amplitude of the preceding vowel indicated the beginning of the fricative; this moment generally coincided with the beginning of the noise signal. A sudden increase of the amplitude of the vowel following the fricative indicated the end of the fricative; this moment generally coincided with the end of the frication noise. From the laryngographic signal the time intervals during which the vocal cords made no contact could be determined. In case the vocal cords ceased to make contact within the initial 50 ms of the fricative and did not resume vibration during the rest of the fricative, the consonant was regarded to be produced without voicing. If after the initial 50 ms of the fricative the vocal cords still made contact, the consonant was regarded to have been produced with voicing. It was observed that in almost all fricatives produced without voicing, voicing ended even within 10 ms after the vowel offset; the 50 ms criterion is therefore ample enough. 2.3.

Data analysis

As far as frequencies of voicing are concerned, a loglinear analysis was carried out on the multidimensional contingency table, with 0.5 added to all cells; the test was only run for [+voice] fricatives. Two separate analyses were carried out: group I consisted of fricatives after long vowels only, thus comparing the effect of the three different stress conditions, and group II consisted of fricatives after stressed syllables, thus comparing the effect of the preceding vowel length. Fricative duration was tested by means of an analysis of variance for both underlying [+voice] and [-voice] fricatives. The durations were analysed for each stress/vowel condition separately, viz. (1) after long vowels and after stress, viz. pasen/bazen, (2) before stress, viz. grafiek/gravin, (3) between two unstressed syllables, viz. basiliek/glazenier, and (4) after short vowels and after stress, viz. Bussum/puzzel. Group (2) was split up into two subgroups since words spelled with after long vowels and before stress did not occur; in group (2a) words with /f,s,v,z/, and in group (2b) words spelled with were analysed separately. In order to investigate the influence of the preceding vowel on fricative duration the data of group 1 and 4 were pooled. In order to investigate the influence of stress the durations of /f/, /s/, /v/ and /z/ for the conditions 1, 2a and 3 were pooled.

127

3. 3.1.

Results Frequency of voicing

Table I Observed frequencies of voiced realizations of 80 [+voice] fricatives per cell, broken down by place of articulation, sex and dialect consonant: west: women men south-east: women men

/y/

/v/

/z/

1 27

2 26

4

30 22

17 21

17 32

5

Table II Main effects and interactions found by applying a loglinear analysis to the observed cases of voice activity present in [+voice] fricatives; chi-squares are given for (I) three stress conditions after long vowels (left) and (II) for long and short preceding vowels after stress separately (right)

Main effects consonant (CONS): position in sentence (POS): sex of speaker (SEX): region of origin (REG): stress condition (STR): preceding vowel (VOW): significant interactions (CONS)x(REG): (POS)x(REG): (SEX)x(REG): (SEX)x(STR):

analysis I 3 stress conditions (after long vowels) chi-2 df P 6. 27 00 14 15 31 30 5 61

14 5 17 6

29 03 23 21

2 1 1 1 2

2 1 1 2