Yearbook of Morphology: Vol 3 1990 [Reprint 2021 ed.] 9783112420744, 9783112420737

Citation preview

Yearbook of Morphology o xJ

Yearbook of Morphology Editors:

Geert Booij J a a p van Marie

Consulting Editors: Stephen Anderson (Baltimore) Mark Aronoff (Stony Brook, N.Y.) Laurie Bauer (Wellington) Rudie Botha (Stellenbosch) J o a n Bybee (Albuquerque, New Mexico) Wolfgang Dressier (Wien) Jack Hoeksema (Groningen) Rochelle Lieber ( D u r h a m , N . H . ) Peter Matthews (Cambridge, U.K.) Franz Rainer (Salzburg) T h o m a s Roeper (Amherst, M A ) Sergio Scalise (Bologna) Henk Schultink (Utrecht) Jindrich T o m a n (Michigan) Wolfgang Wurzel (Berlin)

Editorial address:

Editors, Yearbook of Morphology Vakgroep T a a l k u n d e , Vrije Universiteit P.O. Box 7161 1007 M C A m s t e r d a m , the Netherlands

Contributors are requested to submit manuscripts in three-fold with Diskette, , and to comply with the Instruction to A u t h o r s printed on the inside back cover. The deadline for submission of papers to Volume 4 is March 1, 1991. A u t h o r s receive 25 offprints of articles and 10 offprints of book reviews.

GeertBooijand Jaapvan Marle(eds)

Yearbook of Morphology

1990 FORIS PUBLICATIONS Dordrecht - Holland/Providence Rl - USA

Published by: Foris Publications Holland P.O. Box 509 3300 AM Dordrecht, The Netherlands Distributor for the U S. A. and Canada: Foris Publications USA, Inc. P.O. Box 5904 Providence RI 02903 U.S.A. CIP-DATA Yearbook Yearbook of morphology. - Dordrecht [etc.] . Foris ISSN 0922-3495 Appears annually Yearbook 3, 1990 / Geert Booij, Jaap van Marie (eds.) ISBN 90-6765-505-8 paper SISO 805.3 UDC 801 55(058) Subject heading: morphology (linguistics) , yearbooks.

ISSN 0922-3495 ISBN 90 6765 505 8 paper © 1990 Foris Publications - Dordrecht 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.

Table of Contents

Articles'. Haruo Kubozono Phonological Constraints on Blending in English as a Case for PhonologyMorphology Interface

1

Jerome Packard A Lexical Morphology Approach to Word Formation in Mandarin

21

Theme "When Words Happen To Be Phrases"

39

Guest Editors: Allard Jongman and Aditi Lahiri Allard Jongman and Aditi Lahiri Introduction

41

Geert Booij The Boundary between Morphology and Syntax: Separable Complex Verbs in Dutch 45 Rob Schreuder Lexical Processing of Verbs with Separable Particles

65

Pienie Zwitserlood Comments on the Paper by Schreuder

81

Lyn Frazier Identifying Structure under XQ

87

Joseph Bayer on the Paper by Frazier Comments

107

vi

Contents

Aditi Lahiri, Allard Jongman and Joan Sereno The Pronominal Clitic [dar] in Dutch: A Theoretical and Experimental Approach

115

Eric Reuland Head Movement and the Relation between Morphology and Syntax

129

Jaklin Kornfilt Comments on the Paper by Reuland

163

Manfred Bierwisch Verb Cluster Formation as a Morphological Process

173

Arnold Zwicky Words in Syntax and Moremes in Morphology, Simple and Composite

201

Wolfgang Klein Comments on the Papers by Bierwisch and Zwicky

217

Reply:Gavarrö Anna A Reply to Scalise on 'The Notion of "Head" in Morphology'

223

Book Reviews Antonietta Bisetto Review of Franz Rainer, I norm di qualitd nell' italiano contemporaneo Franz Rainer Review of Mervyn F.Lang, Spanish Word Formation. Productive Derivational Morphology Fritz Schweiger Review of Alfred Holl, Romarüsche Verbmorphologie und relationentheoretische mathematische Linguistik Book Notices received by Geert Booij Publications

229

234 238 241 243

Yearbook of Morphology 3 (1990), 1-20

Phonological constraints on blending in English as a case for phonology-morphology interface Haruo Kubozono 1. INTRODUCTION

The past few decades have seen a revival of interest in word formation in linguistic research. However, comparatively little is known about blending on the whole despite its productivity and its potential importance for word formation research in general. This is largely due to the tendency to disregard blending in morphological research because of its subtractive nature (cf. Uhlenbeck 1962). In this paper, however, this position is not taken for the following two reasons. First, blending is part of word formation (where new words are formed on the basis of existing words), which, in turn, is generally defined as a component of morphology. Second, blending exhibits various linguistic patterns, including one relating to the notion 'head', which are common to ordinary word formation processes like compounding. For these reasons, the position taken in this paper is that blending is a morphological process, although it may be distinct from ordinary morphological processes in many respects. With this understanding as a point of departure, this paper seeks to uncover the phonological constraints on blending in English, and to consider them in comparison with analogous constraints in Japanese. The term 'blending' used in this paper refers to a rather narrow range of word formations, as illustrated in (1), which involve "merging parts of words into one new word" (Marchand 1969: 451). (1)

a. b. c. d.

smoke / fog —> smog breakfast / lunch brunch lunch / supper —»lupper magazine / book —¥ mook

In principle, it is distinguished from other similar word formation processes, including those in (2). Specifically, it involves two source words in a paradigmatic relation, i.e., words that might substitute for one another, as opposed to words which occur side by side,1 and it is in this point that blending differs primarily from

Author's address: Dept. of British and American Studies, Nanzan University, 18 Yamazato-cho, Showa-ku, Nagoya, 466 Japan.

2

H. Kubozono

clipping and clipped compound (compound shortening), the two processes which tend to be confused with blending most often. Furthermore, I put 'syntactic blending' such as different than (< different from + other than) beyond the scope of this study. (2)

a. b. c. d.

compound formation: blackboard < black + board clipping: exam < examination clipped compound: breathalyzer < breath analyzer; sci-fi < science fiction acronym: PTA < parent-teacher association

The term 'blending* defined in this way covers two classes of processes, namely, 'spontaneous' and 'conscious' (or intentional) blending. I tentatively call the forms produced by these two processes blend errors and 'blends in word formation (WF)' respectively, although many of the now 'established' blend forms stem from spontaneous blend errors. While no clear borderline can be drawn between spontaneously-formed and consciously-formed blends, or between blend errors and blends in word formation, I nevertheless make this distinction on the ground that it cannot be justified to assume, a priori, one and the same strategy for these two types of blend processes. The English data analyzed and discussed in the present study come mainly from three sources. The source for blend errors is the collection by Fromkin, comprising 61 spontaneous blend errors reported in the Appendix to Fromkin (1973).2 As for consciously-formed blends, I chose two data sources: the collection by Wentworth (1934) - supposedly the biggest collection of this kind generally available now - and the data reported in Pound (1914). The former contains approximately 3600 blends (some of which are arguably blend errors), including 143 blend forms which are made from two monosyllabic source words. While data on English blends are abundantly reported in the literature, no previous study contains a collection of blends in Japanese, either spontaneously-formed or intentionally-formed. Accordingly, the Japanese data to be discussed in this paper come basically from the private collections by the writer himself and some others. This study has analyzed the 76 blend errors reported in the Appendix to Kubozono (1989) and 48 blends in word formation collected by the author.

2. LINGUISTIC STRUCTURE OF BLENDS

Before discussing the phonological rules underlying blending in English, let us briefly sketch the linguistic structure of blends in general. Previous studies have revealed that blending is subject to a variety of linguistic conditions, as are other types of word formation. These constraints fall into two types: those relating to the nature of the blended items, and those regarding the manner in which they are actually blended.

Phonological constraints on blending in English

3

The first type of constraint can further be classified into two kinds, syntactic and semantic. In syntactic terms, the two words to be blended belong to the same syntactic category, such that a noun interacts with another noun, a verb with another verb, etc. (cf. Terao 1984, Kubozono 1989). This feature can be easily understood when one recalls that blending involves two words in a paradigmatic relation: That is, words which are "in competition for the same slot in the utterance" (Crompton 1982: 138). The present study provides evidence that this rule holds with blends in English, both spontaneously-formed and consciously-formed ones. In fact, very few potential exceptions have been found to this rule: e.g. beflrst < before (used as a conjunction) + first (used as an adverb); wour < we + our.3 Table 1 gives the details of an analysis of Fromkin's blend error data and of the 143 blends reported in Wentworth's collection, both whose source words are monosyllabic: Instances which are ambiguous with respect to syntactic category (e.g. boast < best + most) are all classified as 'others'. The results in this table show that noun-noun blending is the most productive pattern in English, followed by verb-verb and adjective-adjective blendings. Interestingly, blends in Japanese show a basically identical tendency (cf. Kubozono 1989). Table 1. English blends classified according to syntactic category Category

No. of instances (%) Blend errors

Blends in WF

Noun-Noun Verb-Verb Adj-Adj Adverb-Adverb Others

24 12 15 4 6

(39%) (20%) (25%) (7%) (10%)

67 (47%) 57 (40%) 8 (6%) 2 (1%) 9 (6%)

Total

61 (100%)

143 (100%)

In semantic terms, blending appears to be constrained in the following two ways. First, it generally involves two words of a similar, if not identical, semantic content, although it sometimes involves merging two items of opposite semantic contents (cf. Terao 1984, Kubozono 1989). Second, there is a rule concerning the way in which the two source words are arranged: Blends "preserve the normal attributes of the compound such that the end-part is the thematic base to which the new initial part is related" (Quirk et al. 1985: 1583). In other words, the word whose non-initial component constitutes the non-initial component of blend forms serves as the 'head' of the whole expression, just as the righthand element constitutes the head of the whole expression in compound formation. The evidence that has been adduced for this 'righthand head rule (RHR)' (cf. Williams 1981) in blends includes motel (< motor(ist) + hotel), which is a kind of hotel rather than a kind of vehicle. Of these two semantic constraints, the first can be readily substantiated in the present study. In Fromkin's error data, for example, most of the instances involve

4

H. Kubozono

blending two semantically similar words (e.g. corright < correct + right), while some obviously involve a pair of antonyms: e.g. complify < complicate + simplify. By contrast, the RHR is difficult to confirm from the purely semantic viewpoint previously suggested. This is primarily because blending generally involves two words of an essentially identical status, and thereby yields what seems to be a complete mixture of the source words in semantic terms. This unfortunate situation can be remedied by phonological considerations, however, as will be discussed below. In addition to these constraints which apply at the 'input* of blending, there appear to be several factors which condition its output, that is, the ways in which the process occurs and/or the shapes of the resultant blend forms. These constraints fall into two types, namely, morphological and phonological.4 Leaving phonological constraints aside for the moment, the most productive formation rule underlying blending appears to be that in (3), whereby the initial part of one word is combined with the non-initial part of another (cf. Bauer 1983). (3)

AB / XY —> AY

In this formation rule, either B or X can be null, as illustrated in (4), especially when the two source words have some phonemic overlapping. (4)

a. b. c.

cor(rect) / right —> corright guest / star —> guestar cool / (warm)th —> coolth

The reality of this formation rule has been substantiated by the present study, where this rule alone accounts for a vast majority of blend forms in the two types of blending: 92% of the blend errors (i.e. 56 out of 61 instances), 96% of the noun blends in Pound's data (142 out of 148 instances), and 94% of the blends in Wentworth's collection which are made from two monosyllabic source words (134 out of 143 instances).5 Moreover, the same formation rule holds with the two types of blending in Japanese, as illustrated in (5) and (6) respectively (cf. Kubozono 1989), which is suggestive of the universality of this formation pattern. (5)

Blend errors a. peni(i) 'penny' / (pen)su 'pence* —» penisu b. ne(ko) 'cat' / (nja)nko 'kitty' —» nenko c. toma(re) 'stop!'/ (suto)ppu 'stop!' —»tomappu

(6)

Blends in Word Formation a. bini(iru) 'vinyl' / (nai)ron 'nylon' -4 biniron 'vinylon' b. o 'tail' / (si)ppo 'tail' -» oppo 'tail (colloquial)' c. mazu(i) 'damn it!' / (sima)tta 'oups!' -» mazutta 'damn it!' (colloquial)

Phonological constraints on blending in English

5

3. PHONOLOGICAL CONSTRAINTS

Having sketched the non-phonological constraints on blending, let us now consider the question of how this process is constrained phonologically. It is possible to assume, a priori, the following two phonological constraints on the forms that result from blending. One is a phonotactic constraint that prohibits forms which do not conform to the phonotactic structure of the language. Specifically, there is a set of co-occurrence restrictions between the peak (syllable-nuclear vowel) and the coda (the following consonant cluster) in English (cf. Fudge 1987). The other constraint rules out forms which happen to be identical in phonemic shape to either of the source words.6 These constraints seem effective in filtering out such blend forms as in (7) and (8) respectively. The effects of these constraints will be discussed in more depth later. (7)

a. b.

bi(t) / (few) ->*[bi] smo(ke) / (dri)nk —» *[smouT)k]

(8)

a. be(st) / (mo)st *best b. fi(st) / (hoi)st *fxst c. wh(at) / (wh)ich *which d. gl(ance) / (gl)impse —» *glimpse

A careful examination of the forms that actually result from blending in English reveals that the word formation process is subject to stronger phonological constraints. To be precise, the process is subject to two more phonological constraints; one concerning the possible switch point at which two source words are each split and consequently combined, and the other concerning the phonological length of blend forms. 3.1. Syllable structure constraint Analysis of the blend errors reported by Fromkin (1973) has led Crompton (1982) to propose what I call 'syllable structure constraint'; the idea that blending in speech errors is constrained by the constituent structure of the syllable. This constraint states that switches between the source words take place 'syllable-constituent by syllableconstituent',7 which has two separate effects. Given the syllable structure in (9), it prohibits switches within a syllable constituent, that is, switches which result in a split of syllable constitutents, e.g., an onset or coda split. In addition, this constraint implies that the blended items must switch in the same syllable position such that if one word is split in a given syllable position - between the onset and the peak, for example - the other word is split in the same position.

6

H. Kubozono

(9)

Syllable Onset

Peak

Coda

The reality of this constraint has been largely substantiated by the present study. In fact, very few instances in the present corpus constitute exceptions to the first part of Crompton's constraint. Typical exceptions are given in (10), the very first example being the only exception found in Fromkin's blend error data. Similarly, very few examples obviously contradict the second part of the constraint, although there are a couple of ambiguous instances.8 (10)

a. b. c.

sh(e)/(F)romkin —> shromkin (Onset split) s(cale)/climb -> sclimb, or sc(ale)/(c)limb —» sclimb (Onset split) slan(g)/(tal)k —» slank (Coda split)

Given the tripartite syllable model in (9), the syllable structure constraint suggested by Crompton implies that onset-peak boundaries and peak-coda boundaries are equally possible points within the syllable at which the blended items can switch. However, Crompton further suggests that the first type of boundaries is the more preferred of the two candidates, and thereby implies support for the more hierarchical syllable structure in (11), which is additionally supported by various kinds of independent evidence from English and other languages (cf. Pike 1967, Fudge 1969). (11)

Syllable Onset

Rhyme (Rime)

Peak

Coda

Stated conversely, this now established syllable constituent model leads to a stronger version of syllable structure constraint by implying cohesiveness between the peak and the coda as against between the onset and the peak, or preference for switches at the latter boundaries in blends over switches at the former boundaries. Crompton's observation has been previously expressed by Dressier (1976) as well,9 and has been statistically substantiated by Fudge (1987), who analyzed all types of speech errors reported by Fromkin (1973). While Fudge's analysis provides statistical support for Crompton's observation, it does not give a precise picture of what happens specifically to blend errors per se, nor does it allow for the effects of the other phonological constraints, illustrated in (7) and (8) above, in evaluating the data. In view of this, it is obvious that Fromkin's blend error data need to be reanalyzed more carefully in order to confirm the reality of the syllable structure constraint.

Phonological constraints on blending in English

7

Moreover, it is also necessary to see if the same constraint holds in consciouslyformed blends as well, not just in blend errors. 3.1.1. Blend errors Fromkin's (1973) blend error data include blends of polysyllabic source words as well as monosyllabic words. Accordingly, they yield three major switch patterns as illustrated in (12). (/./ represents syllable boundaries): (12)

Major switch patterns in blend errors a. switch at onset-peak boundary cl(ose) / (n)ear —» clear b. switch at peak-coda boundary swi(tched) / (cha)nged —> swinged c. switch at syllable boundary pop.(u.lar)/ (pub).lic pop.lic

While these three patterns combined account for 55 out of the 56 instances which follow the formation pattern in (3), the situation is complicated by the fact that many blends are ambiguous with respect to the switch point implied, the point at which the blended items are split and subsequently combined. This ambiguity arises because the blended items have part of their phonemic representations in common. If analysis is confined to the three switch points in (12), the present corpus includes 23 ambiguous instances, such as those in (13), as compared to 32 unambiguous instances, such as those in (12) above. Table 2 compares the productivity of the three switching patterns as regards the unambiguous instances, while Table 3 compares the three patterns as regards the ambiguous and unambiguous instances combined, with the former counted twofold or threefold, depending on how many interpretations each instance permits. (13) a. b. c.

stom.ach st(om.ach) sto(m.ach) stom.(ach)

/ / / /

tum.my (t)um.my (tu)m.my (tum).my

—» stummy -» stummy —> stummy —> stummy

Table 2. Switch point in blend errors in English (unambiguous instances only) Switch point

No. of instances (%)

(12a) (12b) (12c)

21 (66%) 7 (22%) 4 (13%)

Total

32 (100%)

8

H. Kubozono

Table 3. Switch point in blend errors in English (ambiguous and unambiguous instances combined) Switch point

No. of instances (%)

(12a) (12b) (12c)

43 (79%) 22 (40%) 18 (38%)

Total

55 (100%)

While the results in Tables 2 and 3 clearly show that onset-peak boundaries are a much more preferred switch point than peak-coda boundaries in blend errors, they include cases such as those in (14) and (15), which can be accounted for by the constraints illustrated in (7) and (8), respectively. (14)

b(it) / (f)ew -> [bju:]; bi(t) / (few) ->*[bi] (= 7a)

(15)

a. b.

b(est) / (m)ost boast be(st) / (mo)st *best (= 8a) wha(t) / (whi)ch -» watch wh(at) / (wh)ich *which (= 8c)

In these cases, a particular switch point seems to have been chosen because switching at the alternative point would have yielded forms which are either phonotactically impermissible in English (because of a strong phonotactic constraint between the peak and the coda mentioned earlier),10 or which are identical to one of the source words. These problematic cases need to be removed, to make a fair comparison of the competing patterns in (12a) and (12b). Table 4, accordingly, shows the results of such a reanalysis based on the unambiguous cases in Table 2. The switching pattern in (12a) remains predominant over that in (12b) (with a difference significant on the binomial test at a level below .03). It can reasonably be said, therefore, that blended items tend to switch at onset-peak boundaries in blend errors in English. Table 4. Switch point in blend errors in English (unambiguous instances only, reanalyzed) Switch point

No. of instances (%)

(12a) (12b)

13 (76%) 4 (24%)

Total

17 (100%)

3.1.2. Blends in word formation The productivity of the switch at onset-peak boundaries can be observed more clearly in blends in word formation. To see this point, the present study analyzed the 143 monosyllabic blends made from two monosyllabic source words reported in

Phonological constraints on blending in English

9

Wentworth's collection. Of the 143 blends, 19 instances are disyllabic blends like guestar (< guest + star), and nine others show formation patterns other than the basic pattern in (3) (e.g., flump smink smo(ke) / (dri)nk -> *[smomjk] (= 7b)

(19)

a. b.

f(ist) / (h)oist -> foist fi(st) / (hoi)st *fist (= 8b) gla(nce) / (gli)mpse —» glampse gl(ance) / (gl)impse —> *glimpse (= 8d)

Table 7. Switch point in blends in English word formation (unambiguous instances only, reanalyzed) Switch point

No. of instances (%)

(16a) (16b)

27 (82%) 6 (18%)

Total

33 (100%)

Out of the 33 instances which are unambiguous in every respect, 27 imply a switch at onset-peak boundaries as compared to only six instances which imply a switch at the alternative syllable point. This difference is significant on the binomial test at a level below .001. Again, the reanalyzed data remain suggestive of the superiority of the onset-peak boundary over the peak-coda boundary as an implied switch point. It can be concluded, therefore, that onset-peak boundaries are the most preferred switch point in English blends, both for blend errors and for blends in word formation. Seen in a theoretical context, this result supports the idea of the 'rhyme' as being an independent syllable unit and, hence, the syllable structure given in (11) above. (20) illustrates the most popular switch pattern in English in the light of this syllable structure. This observation, viewed from another standpoint, reveals that the vowel - which is the nucleus of the syllable - comes from the righthand source word rather than the lefthand counterpart in monosyllabic blends. This finding has a close bearing on the notion 'phonological head of a word', to be discussed below.

Phonological constraints on blending in English

11

3.1.3. Japanese blends Having justified the notion 'rhyme', it is worth noting that the same notion does not always play a significant role in the blending process of other languages. As pointed out by Kubozono (1989), for example, blends in Japanese do not show a tendency to split between the onset and the peak, although they are otherwise very similar to English blends. To be more specific, while there is a certain switch point within the syllable which is most preferred in Japanese blends, this switch point is not onsetpeak boundaries, or syllable boundaries, but mora boundaries. This is illustrated in (21) and (22), where mora and syllable boundaries are marked by /-/ and /./, respectively; all syllable boundaries are also mora boundaries, by definition, although not vice versa. (21)

Blend errors a. pe.ni-(i) / (pe-n).su —> pe.ni.su (= 5a) b. ne.(ko) / (nja)-n.ko —> ne-n.ko (= 5b) c. to.ma.(re) / (su.to)-p.pu —»to.ma-p.pu (= 5c)

(22)

Blends in Word Formation a. bi.ni-(i.ru) / (na-i).ro-n —> bi.ni.ro-n (= 6a) b. o / (si)-p.po o-p.po (= 6b) c. ma.zu-(i) / (si.ma)-t.ta —» ma.zu-t.ta (= 6c)

In Japanese, not just the word formation process of blending, but many other phenomena - morphological, phonological, and phonetic - can be generalized by the notion 'mora' (see Kubozono 1989 for the details). I have interpreted these facts as evidence for the unit 'mora', a prosodic unit smaller than the syllable, whose relevance as the unit for measuring phonological distance in 'mora languages', like Japanese, has been substantiated by various lines of linguistic evidence (cf. McCawley 1978). Although blends in Japanese show a crucial difference from English counterparts in this respect, it is worth noting that blends in English and in Japanese alike prefer a certain switch point within the syllable over other logically possible points. I have interpreted the difference between the two languages as arising from a difference in syllable structure proper. In other words, I have taken the position that blending is equally constrained by syllable constituency in English and Japanese alike, but that the precise effect of this constraint varies from language

12 H. Kubozono to language because of interlanguage variations in syllable structure per se (cf. Kubozono 1989). In terms of syllable constituency, this interpretation implies that the rhyme is not a universal syllable unit, as often hinted in the literature (cf. Fudge 1987), thereby casting doubt upon the hierarchical structure model in (11). While this interpretation may not be the only possible explanation of mora phenomena in Japanese (cf. Poser 1990, Kubozono, in preparation), it is still worth emphasizing that Japanese and English show both a striking similarity and difference in the patterning of blending. 3.2. Length constraint 3.2.1. Blend errors In addition to the syllable structure constraint just discussed, blends in English are subject to yet another phonological constraint. A careful analysis of Fromkin's blend error data, which have both monosyllabic and polysyllabic words for source words, reveals that there is a high degree of correspondence between source words and resultant blend forms with respect to phonological length. To be more specific, the righthand source word and the resultant blend form consist, in most instances, of the same number of syllables. The effect of this length rule, tentatively formalized in (23), can be seen from Table 8. (23)

In AB / XY -> AY, XY and AY are equal in phonological length.

Table 8. Length relation in blend errors in English (measured in terms of the number of syllables) Pattern I. II. III. IV. V. Total

No. of instances (%) AB AB AB AB AB

= * = = *

XY XY AY XY XY

= = * -t *

AY AY XY AY AY

34 (61%) 14 (25%) 5 (9%) 2 (4%) 1 (2%) 56 (100%)

One remarkable fact about this table is the high productivity of the first pattern, whereby the blend form agrees in phonological length with both of the source words.11 Equally interesting is the productivity of the second pattern - pattern II and its contrast with the relatively low productivity of the third pattern, III. There are 19 cases where the blend form equals only one of the source words in phonological length, the majority of which belong to pattern II (cf. (12c)): The difference between these two patterns is significant on the binomial test at a level below .05. Because of this high productivity of the second pattern, the first two patterns, combined account for nearly 90% of the instances conforming to the formation rule in (3).

Phonological constraints on blending in English

13

Moreover, half of the remaining instances show switching of the source words with a common phonemic representation as their switch point (e.g. population + pollution —> populution), which suggests that the 'exceptions' can be handled by some other independently-motivated principles. Given this syllable-based generalization in English, one may wonder, quite reasonably, if the same generalization cannot be captured by phonological units other than the syllable, such as the phoneme. This possibility cannot be substantiated well, however. In Table 9, for example, an analysis is illustrated whereby the phoneme is assumed as the unit to measure phonological distance. Obviously, the phoneme does not allow us to capture any length agreement between the blend form and the blend sources, either the righthand or the lefthand word. Table 9. Length relation in blend errors in English (measured in terms of the number of phonemes) Pattern I. II. III. IV. V. Total

No. of instances (%) AB AB AB AB AB

= * = = *

XY XY AY XY XY

= = * * #

AY AY XY AY AY

10 19 10 3 14

(18%) (34%) (18%) (5%) (25%)

56 (100%)

All in all, it can be said that there is a strong tendency for the righthand source word to maintain its phonological length in the resultant blend form, as shown in (23), and that this tendency can be generalized if (and only if) the syllable is recognized as the unit by which to measure phonological distance in English. A more careful examination of the length rule in question reveals that there is an agreement in phonological length between A and X, that is, between the initial part of the lefthand source word which remains intact in the blend form, and the initial part of the righthand counterpart which is omitted. This allows us to assume the following picture as the general mechanism underlying the length rule in (23) and, hence, the blend error process in English. The fact that A and X in (23) agree in phonological length can conceptually be understood if one assumes that the speaker intended to produce the righthand source word, XY, but somehow confused it with the lefthand source word, AB, in the course of the utterance. To be more specific, the speaker actually started to produce the wrong word but, realizing that it was not the word he had intended, switched from that word to the word he had originally intended. What is crucial in this 'correction' process is that the speaker does not go back to the very beginning of the right word, but makes a switch from the wrong word to the right one at the given syllable-internal point where the error was detected. If, for example, the error is detected at the point where the onset of the initial syllable of the wrong word has been produced, the onset of the word is subsequently combined with the peak of the corresponding syllable of the right word.

14

H.

Kubozono

This is illustrated in (24a), which accounts for the pattern in (12a). Similarly, if the detection is made between the peak and the coda of the initial syllable in one word, the switch takes place in such a way that this onset-peak portion is combined with the coda of the corresponding syllable of the right word. This is illustrated in (24b), and can be seen to underlie the switch pattern in (12b). If this kind of switch takes place at the syllable boundary, then there must be an agreement between the number of the syllables taken from the wrong word and the number of syllables to be omitted from the right one. This is illustrated in (24c) and is exemplified in (12c). (24)

a. b. c.

Onset-(Peak-Coda)... (Onset)-Peak-Coda ... Onset-Peak-(Coda-Onset ... (Onset-Peak)-Coda-Onset ... Onset-Peak-Coda-(Onset-Peak-Coda)... (Onset-Peak-Coda)-Onset-Peak-Coda ...

This assumed mechanism for the generation of blend errors may look similar to Crompton's syllable structure constraint sketched above. They are crucially different, however, in that the former refers not only to the syllable-internal position, but also to the location of the syllable within words in which the switch takes place. In other words, the length rule proposed here states that if two words switch in blending, the switch must take place not only in the corresponding syllable-internal position but in the corresponding syllables as well. Seen in this way, the rule in question can be understood as a general rule regulating the phonological quantity, so to speak, of the portions involved in blending. The validity of this conceptual explanation can be substantiated by the fact that blend errors in Japanese exhibit a length regularity which is strikingly similar to the length rule underlying the English counterparts. My previous analysis of blend errors in Japanese (Kubozono 1989) shows a high level of correspondence in phonological length between the righthand source word and the blend forms (cf. (21)). Thus, the length rule accounts for more than 80 % of the blend errors in Japanese, as shown in Table 10, while exceptions to the rule are mostly blends of morphologically complex source words, such as compound nouns. Table 10. Length relation in blend errors in Japanese (measured in terms of the number of morae) Pattern I. II. III. IV. V. Total

No. of instances (%) AB AB AB AB AB

= * = = *

XY XY AY XY XY

= = * * *

AY AY XY AY AY

30 (39%) 35 (46%) 6 (8%) 0 (0%) 5 (7%) 76 (100%)

Phonological constraints on blending in English

15

In fact, the only substantial difference between blend errors in English and Japanese with respect to the length regularity is that the rule in Japanese is to be generalized not in terms of the syllable but in terms of the mora. In the instances given in (21), for example, the first part of the lefthand source word, i.e. A in (3), consists of the same number of morae as the corresponding part of the righthand source word, X, so that the resultant blend form, AY, comes to have the same number of morae as the righthand source word, XY. As shown by Kubozono (1989), this line of generalization cannot reasonably be captured if the syllable is assumed as the unit of phonological length in Japanese. Given these facts, it seems reasonable to assume that the length rule in (23) is a rather general rule in the morphological process of blending, although its precise content - i.e., the definition of the notion 'phonological length' - can vary from language to language, depending upon what counts as the unit for measuring phonological length in that particular language. 3.2.2. Blends in word formation The observation that the length rule in (23) constrains blends in English is further borne out by an analysis of blends in word formation. Pound (1924) contains 148 noun blends, of which 142 instances conform to the basic formation rule given in (3) above. As shown in Table 11, the length rule in (23) accounts for nearly 80% of these instances, with most of them attributed to pattern II. This is exemplified in (25) as well as by the examples in (1). Table 11. Length relation in blends in WF in English (measured in terms of the number of syllables) Pattern I. II. III. IV. V.

No. of instances (%) AB AB AB AB AB

= * = = *

= = * * *

AY AY XY AY AY

24 (17%) 87 (61%) 8 (6%) 6 (4%) 17 (12%) 142 (100%)

Total

(25)

XY XY AY XY XY

a. b. c.

ar.gu.(ment) / (sig.ni).fi.ca.tion —» ar.gu.fi.ca.tion in.d(e.pen.dent) / (dem).o.crat —> ind.o.crat need / (ne).ces.si.ty —» need.ces.si.ty

Moreover, of the 31 instances which do not conform to this rule, 18 examples can be attributed to the independent principle sketched earlier (e.g. animal + mule —» animule, bungalow + loafer —> bungaloafer), leaving only a handful of 'genuine' exceptions to the length rule. These results conclusively show that the length rule in (23) constrains blends in word formation as well as blend errors in English. As in the case of blend errors, the same rule accounts for most blend forms in word formation in Japanese as well as in English. This observation, exemplified in (22),

16

H.

Kubozono

is summarized in Table 12, where my own collection of 48 blends is analyzed. Here again, many of the exceptional instances can be attributed to some other independent principles, including those relating to morphological complexity and phonological overlapping, suggested above (see Kubozono, in preparation, for details). Table 12. Length relation in blends in WF in Japanese (measured in terms of the number cf morae) Pattern I. II. III. IV. V. Total

No. of instances (%) AB AB AB AB AB

= * = = *

XY XY AY XY XY

= = * * *

AY AY XY AY AY

11 (23%) 21 (44%) 3 (6%) 4 (8%) 9 (19%) 48 (100%)

4. ARGUMENTS FOR THE 'HEAD' IN BLENDS

The existence of the length rule in (23) has several implications to note. First, the fact that a significant generalization can be made based on the concept 'syllable' in English reinforces the relevance of this unit in the phonological descriptions of the language (cf. Kahn 1976), particularly as a unit for measuring phonological length. It must also be noted in this regard that what counts as the unit of phonological distance can vary from language to language, supporting the view previously expressed by McCawley (1978). The length rule in question also has a significant implication concerning the notion 'head' in word formation processes. The fact that the righthand source word, XY, determines the phonological length of the resultant blend forms both in English and Japanese can be taken as implying that this word is the more important element phonologically, or the 'head' of blends in a phonological sense. The notion 'head (of a word)' and, hence, the RHR, are primarily a morphological notion and rule, to be sure. However, if one considers the phonological patterns of blending in general, it seems reasonable to discuss the notion from a phonological viewpoint as well as from a morphological (or morphosyntactic) one, and thereby to introduce the notion 'phonological head of a word'. In the discussion of blend errors, specifically, I have implicitly correlated this concept of 'phonological head' with the morphological concept 'head of the word' by interpreting the righthand source word as the target word which the speaker intended.12 The same strategy seems to be employed in the intentional process of blending, where the same length agreement is observed to hold between the righthand source word and the resultant blend form. Thus, while the length rule in question is essentially a phonological rule - or, better yet, a phonological constraint on the word formation process - the existence of this kind of rule can be taken as evidence that the RHR holds in blends in English and Japanese.

Phonological constraints on blending in English

17

The conclusion that the RHR holds in a subtractive morphological process like blending just as in additive processes (cf. Martin 1988), is also supported by the evidence for the switch at onset-peak boundaries, discussed above, namely, that monosyllabic words are split in English blending into two parts, onset and rhyme (peak + coda). Since the vowel constitutes the head of the syllable, the fact that this dominant part comes from the righthand source word rather than the lefthand counterpart allows us to define the righthand word as the more important element phonologically. The same argument does certainly not apply to the blending in Japanese, where the most preferred switch point is mora boundaries. However, there is some evidence which suggests that the righthand source word plays the greater role in the blending of Japanese as well. This evidence relates to the fact that, both in blend errors and blends in word formation, the larger part of the resultant blend forms comes from the righthand source word (cf. Kubozono, in preparation). This can be seen from Table 13, which shows that the cases where Y - the portion taken from the righthand source word - is longer (in mora length) than A — the portion portion that comes from the lefthand source - outnumber the cases where A is longer than Y (i.e. pattern III vs. pattern I). (The fourth and fifth patterns represent cases in which more than one mora boundary can be interpreted as the switch point implied). This tendency can be taken as suggesting that the righthand source word tends to contribute more to the blend forms. Table 13. Comparison of A and Y in Japanese blends (measured in terms of the number qfmorae) No. of instances (%)

I. II. III. IV. V.

Pattern

Blend errors

Blends in WR

A A A A A

14 (18%) 24 (32%) 29 (38%) 0 (0%) 5 (7%)

3 (6%) 18 (38%) 16 (33%) 0 (0%) 2 (4%)

76 (100%)

48 (100%)

> = < >
mer; c(oi)ns / (ch)a(nge) -» canes. Lewis Carroll's famous blend, slithy (< slimy and lithe) should also be regarded as an atypical blend in this respect 6. In speech errors, blends of this type may not necessarily be ruled out as impossible because we would simply not be able to see these blends even if they occurred. 7. Apart from this syllable structure constraint, there is only one phonological constraint found in the literature. Bauer (1983) suggests that blending may not produce forms identical in phonemic shape to other existing words of a similar semantic content This constraint is claimed to rule out the form duck [d/vk] for the blend of dove and hawk because it "could lead to all sorts of misunderstandings" (Bauer 1983: 235). This constraint, however, cannot be well substantiated by the present study; the particular example adduced can be properly handled by the syllable structure constraint, to be discussed shortly. 8. In (10a), for example, the first word can be analyzed as implying a split between the onset and the peak or within the onset (assuming an empty element in the second half of the onset, i.e. [sh + 0]), while the second source word is split obviously within the onset 9. Providing evidence for an onset-rhyme division in blending. Dressier (1976) further makes several interesting claims which partially anticipate the observations made in this paper. 10. I agree with Fudge (1987) that absence of an analogous constraint between the onset and the peak constitutes evidence for the cohesiveness between the peak and the coda and, thus, for the syllable unit 'rhyme'. 11. A similar observation has been made by Terao (1984) about blend errors in Japanese, although it has been reformulated by Kubozono (1989). 12. This explains in a straightforward manner the suggestion made by Quirk et al. (1985) that the stress pattern of blend forms tends to follow that of the righthand source word.

20 H. Kubozono REFERENCES Adams, V. 1973. An Introduction to Modern English Word-formation. London: Longman. Bauer, L. 1983. English Word-formation. Cambridge: Cambridge University Press. Crompton, A. 1982. "Syllables and segments in speech production''. In A. Cutler (ed.), Slips of the Tongue and Language Production. The Hague: Mouton, 109-162. Dressler, W. 1976. 'Tendenzen in kontaminatorischen Fehlleistungen". Die Sprache 22, 1, 1-10. Dressler, W., L. Tonelli, and E. Magno Caldognetto. 1987. "Analisi contrastiva dei lapsus e delle paiafasie fonologiche rispetto alia sillaba," Parallela 3, 54-60. Tübingen: Narr. Fromkin, V. 1973. "The non-anomalous nature of anomalous utterances''. In V. Fromkin (ed.), Speech Errors as Linguistic Evidence. The Hague: Mouton, 213-242. Fudge, E. 1969. "Syllables". Journal of Linguistics 5, 253-286. Fromkin, V. 1987. "Branching structure within the syllable". Journal of Linguistics 23, 359-377. Jespersen, O. 1922. Language: Its Nature, Depelopment, and Origin. New York: H. Holt & Company. Kahn, D. 1976. Syllable-based Generalizations in English Phonololgy. Ph.D. dissertation, MIT. Kiparsky, P. 1982. "Lexical morphology and phonology". In I.-S Yang (ed.), Linguistics in the Morning Calm. Seoul: Hanshin, 3-91. Kubozono, H. 1988. "Constraints on phonological compound formation". English Linguistics 5,150-169. Kubozono, H. 1989. "The mora and syllable structure in Japanese: Evidence from speech errors". Language and Speech 32, 3, 249-278. Kubozono, H. In preparation. "Foot structure and syllable structure in Japanese". Marchand, H. 1969. The Categories and Types of Present-day English Word-formation. München: C.H. Beck'sche Verlagsbuchhandlung. Martin, J. 1988. "Subtractive morphology as dissociation". Proceedings of the Seventh West Coast Conference on Formal Linguistics 7, 229-240. McCawley, J. 1978. "What is a tone language?" In V. Fromkin (ed.), Tone: A Linguistic Survey. New York: Academic Press, 113-131. Mohanan, K.P. 1982. Lexical Phonology. PhD dissertation, MIT. Namiki, T. 1982. "The notion 'Head of a Word' and core and periphery: Interactions between affixation and subcategorization". Studies in English Linguistics 10, 21-41. Pike, K.L. 1955/67. Language in Relation to a Unified Theory of the Structure of Human Behavior. The Hague: Mouton. Poser, W. 1990. "Evidence for foot structure in Japanese". Language 66, 78-105. Pound, L. 1914. "Blends, their relation to English word-formation". Anglistische Forschungen XLII. Quirk, R., S. Greenbaum, G. Leech, and J. Svartvik. 1985. A Comprehensive Grammar of the English Language. London: Longman. Scalise, S. 1988. "The notion of 'head' in morphology". Yearbook of Morphology 1, 229-245. Shibatani, M. and T. Kageyama. 1988. "Word formation in a modular theory of grammar: Post-syntactic compounds in Japanese". Language 64, 3, 451-484. Soudek, L.I. 1978. "The relation of blending to English word-formation: Theory, structure, and typological attempts". In W. Dressier and E. Meid (eds.), Proceedings of the Twelfth International Congress of Linguists (Vienna, 1977). Innsbruck, 462-466. Terao, Y. 1984. "On blends", (in Japanese) Tsukuba Working Papers in Linguistics 3, 15-31. Uhlenbeck, E.M. 1962. "Limitations of morphological processes: Some preliminary remarks". Lingua 11, 426-432. Wentworth, H. 1934. Blend Word in English. Doctoral dissertation, Cornell University. Williams, E. 1981. "On the notions 'lexically related' and 'head of a word". Linguistic Inquiry 12, 245274.

Yearbook of Morphology 3 (1990), 21-37

A Lexical Morphology Approach to Word Formation in Mandarin1 Jerome L. Packard Words in Mandarin Chinese may be formed by combining morphemes from virtually every form class category. However, these complex words vary in their ability to undergo certain grammatical operations. The nature of these grammatical restrictions has gone largely unanalyzed in previous works on Mandarin morphology. Previous treatments have been mostly descriptive lists of complex words which have in common either internal structure or derived form class. For example, in his seminal treatment of Mandarin morphology, Lu (1964) concentrates mostly on the structure of nominal compounds.2 Chao (1968) provides an exhaustive description of word formation in Mandarin, especially of the various types of compounds, which he classes primarily by internal structure. Li & Thompson (1981) also contain a complete listing of morphological types, listing compounds once again by internal structure. The cited works do not discuss certain morphological process constraints that occur in Mandarin, and also do not deal in depth with productivity and distributional differences that obtain between processes of affixation (including inflection) and compounding. In this paper I will argue that certain morphological process constraints in Mandarin result from properties of the components which make up the word, and that a judicious arrangement of word formation rules in the lexicon accounts for the distributional facts. The way I propose to account for these facts is to order the various types of complex words with respect to the affixation processes. I will achieve this ordering by relying on the theory of lexical morphology. In lexical morphology, word formation rules are ordered in the lexicon according to level, and the ordering of these levels accounts for many generalizations in phonology and morphology. The insight that the lexicon may be divided into ordered blocks of wordformation rules is originally due to Siegel (1974). It was further developed by Allen (1978), and more recently by Kiparsky (1982), Scalise (1984) and Mohanan (1986). Interestingly, some of these investigators have used the theory mainly to account for phonological phenomena, by ordering phonological rules with respect to the word-formation rules. In this paper, I wish to account solely for word-formation phenomena in Mandarin, and will leave the discussion of phonology for some future time.3

Author's address: Dept. of Oriental Studies, University of Pennsylvania, Philadelphia, PA 19104, U.S.A.

22 JJL. Packard Table 1 gives a brief description of how English is analyzed according to lexical morphology. The English lexicon is usually divided into four levels (e.g., Halle & Mohanan 1985). Class 1 derivational affixes (those that attach to stems and often affect the sound of a word) are at level I. Class 2 derivational affixes (those that attach to words and usually do not affect the sound of a word) are at level II. Compounding occurs at level III and regular inflection occurs at level IV. Table 1: Brief Outline of Lexical Morphology (English)

Level I

Class 1 deriv. affixes: (e.g., a-, in-, -ate, active -»activ-afe, —» activ-jty tender -»tender-i'ze,

Level II

Class 2 deriv. affixes: (e.g., de-, un-, re-, activate -»deactivate activity —> activity-/ess2 tenderize -»tenderiz-er 2

-ous, -ity, -ize, -ion, -ic, -al...)

non-, -less, -ness, -ship, -ful, -er...)

Level III Compound words: (e.g., lighthouse, blackbird, fire engine, meat tenderizer...)

Level IV Regular inflection: (e.g., -ed, type —> cat —> meat tenderizer —>

-s,...) typed cats meat tenderizers

As a general rule in lexical morphology, those processes that are 'closest to the syntax' and that are most productive are located at 'shallower' levels (i.e., Ill & IV). This seems to be borne out, in that class 2 affixes are much more productive than class 1 affixes. Also, the process of inflection is arguably closely related to syntax, and is therefore at level IV. Now let us consider how this theory (or a modification thereof) might apply to Mandarin Chinese. Before we look at the structure of the Mandarin lexicon that I am proposing, let us first go over some assumptions. The first is that all Mandarin words have heads, and that the head of a word is canonically determined by position, according to the form class of the word, rather than by the relation between categories as in syntax (c.f. Selkirk 1982, DiSciullo & Williams 1987). The rule for Mandarin is listed in (1). (1)

Canonical Head Rule (Mandarin): In a word-internal configuration: for all [[x] [y]]v, x is the head of V for all [[x] [y]] sv , x is the head of SV (stative verb) for all [[x] [y]]N, y is the head of N

23

Word formation in Mandarin

The rules in (1) simply state that in Mandarin, all verbs, including stative verbs (SV), are canonically left-headed, and all nouns are canonically right-headed. The notion of a head which is defined canonically according to position may seem strange at first glance, but there is good reason to believe that it is correct for word formation. Briefly, the primary evidence from Mandarin is that when speakers decompose words through reanalysis, they treat the canonically defined head as though it were a member of the form class of the entire word, even when that head is transparently a member of a different form class. For a full justification of this position with evidence from Mandarin, see Packard (1989). In (2) we have the principle that affixation (which subsumes inflection) occurs on the heads of words. According to this principle, the head of a word may only accept affixation if the head is 'visible', i.e., if its internal brackets have not been erased as a result of passage from one lexical level to the next. In addition, affixation at level IV is different from that at other levels, in that there is a default application just in case the head is not visible. (2)

Affixation/Inflection Principle All affixation (including inflection) applies to the head of a word, but only if the head is 'visible' (i.e., if its brackets have not been erased due to interlevel movement). At level IV, default application is to the word, just in case the head is not visible.

A final assumption before considering the structure of the Mandarin lexicon is the rule which assigns exocentric compounds to lexical level, given in (3). This rule defines exocentrism, and states that exocentrics are assigned to the deepest level of the lexicon. According to the present analysis, a word is considered exocentric if its form class does not match the form class of the morpheme in canonically-defined head position. For example, the verb zuoyou (NB: this is not the far more commonly used zuoyou which means 'approximately') in (4) does not have a verb as the left member of the compound, and so is exocentric. Also, the stative verb zuiying in (5) does not have a stative verb in the left-hand position, and so it is also considered exocentric. Notice that zuoyou would be considered exocentric according to the traditional definition, while zuiying would not. Since zuiying is a stative verb and its right-hand constituent is a stative verb, according to traditional analyses its head would be considered to be on the right, and it would be considered endocentric. (3)

Exocentric Level Assignment: All compounds which are transparently exocentric (i.e., [-V x] v [-SV x] s v [x -N] n ) are formed at the 'deepest' (i.e., most opaque) level of the lexicon (viz., level I).

(4)

zuoyou [[ M M v

left-right

'to influence'

24 J.L. Packard (5)

zuiying [[ INC Isvlsv

mouth-hard

'stubborn'

For Mandarin, the structure of the lexicon I am proposing is as follows. At level I (the 'deepest' level) are the words which have been formed using the least productive of the word formation processes in Mandarin. For example, words that have only a vaguely discernable or no internal structure, such as the conjunctions suiran 'although', keshi 'however', and the adverbials yiqian 'before' and yihou 'after'. Also at this level are all exocentrics, and also words that belong to the class of non-productive, non-infixing resultatives (Chao 1968: 435). At level II are compounding and affixation4 processes that are more highly productive, such as regular (i.e., non-exocentric) nominal compounding, and regular resultative verb compounding. Also at this level are the resultative verb interfixation process, and nominal affixes such as -zi, -r and -tou. At level III are all regular and stative verb compounding processes, and also the question-formation reduplication and interfixation process. Finally, level IV contains only inflectional affixation processes. Let us consider in detail how the word formation processes operate at each level. At level I, consider first compounds that are exocentric. The verbal compounds zuoyou, gechang, and zuiying seen in (6), (8) and (10) (and exemplified in (7), (9) and (11)), are bisyllabic regular or stative verbs which are considered exocentric by our definition, because they are verbs whose left-hand members are not verbal. (6)

zuoyou

left-right

'to influence'

[[ ] N [ M v (7)

Neige zi bu zuoyou juzi de yisi that word not influence sentence MOD meaning 'That word doesn't affect the sentence's meaning'

(8)

gechang [[ M Ivlv

(9)

Gechang women weida zuguo sing our great fatherland 'Sing the praises of our great fatherland'

(10)

zuiying

song-sing

mouth-hard

[[ INC Isvlsv (11)

Neige ren hen zuiying that person very stubborn 'That person is very stubborn'

'to sing'

'stubborn'

Word formation in Mandarin

25

Ordinarily, regular verb compounds may form questions by means of a word-internal inflection process (Huang 1988,1989). This inflectional process duplicates the head, and inserts the negative marker bu between the original and duplicated syllables (V1-bu-V1-V2, or 'A-not-A'), as examples (12)-(13), (15)-(16) and (18)-(19) illustrate. Examples (14), (17), and (20) crucially demonstrate that the affixation process may apply only to the head of the word, as predicted by the affixation rule in (2) above. (12)

Neige zi bu gaibian juzi de yisi that word not revise-change sentence MOD meaning "That word doesn't change the sentence's meaning'

(13)

Neige zi gai-ftu-gaibian juzi de yisi? that word revise-not-revise-change sentence MOD intent 'Does that word change the sentence's meaning?'

(14)

*Neige zi gaibian-ftu-bian juzi de yisi?

(15)

Yongzan women weida zuguo sing-praise our great fatherland 'Sing the praises of our great fatherland'

(16)

Yong-ftw-yongzan women weida zuguo sing-not-sing-praise our great fatherland '(Do we) sing the praises of our great fatherland?'

(17)

*Yongzan-6«-zan women weida zuguo?

(18)

Neige ren hen gaoxing that person very high-elated "That person is very happy'

(19)

Neige ren gao-ftu-gaoxing? that person high-not-high-elated 'Is that person happy?'

(20)

*Neige ren gaoxing-bu-xing?

When we try to apply this question operation to the (canonically-defined) heads of the exocentric compounds, the result is ungrammatical as seen in (21)-(23). (21)

*Neige zi zuo-bu-zuoyou juzi de yisi?

26 J.L. Packard (22)

*Ge-òu-gechang women weida zuguo?

(23)

*Neige ren zui-ftw-zuiying?

This question-forming inflectional process cannot apply to exocentric compounds because these compounds are formed at level I, while the question-forming inflectional process (as we shall see) is located at level III. The inflectional process therefore does not have access to the heads of the exocentric compounds, and so cannot apply to them. Continuing our discussion of level I, note that there is a special type of verb compound in Mandarin, called the resultative verb compound, which takes a potential infix {-de- 'can' or -bu- 'cannot') indicating the ability or inability of the verbal action to take place. Examples (24a)-(26a) show examples of the potential infix applying to resultative compounds. Note once again that the process may apply only to the head of the word, as seen by the unacceptability of (24b)-(26b). (24)

a.

gai wan correct-finish 'finish correcting'

—» gai-feu-wan correct-not-finish 'unable to finish correcting' —» *gaiwan-bu

jiangchu speak-emit 'speak out'

—> jiang-dc-chu speak-can-emit 'able to speak out' —» *jiangchu-de

b. (25)

a. b.

(26)

a.

bancheng do-finish 'finish'

b.

ban-feu-cheng do-not-finish 'unable to finish' *bancheng-bu

However, there is a restricted class of resultative verb compounds (as observed by Chao 1968: 435) that cannot take the potential infix, as illustrated in (27)-(31) (nor can they take the V r b u - V r V 2 question inflection operation, as the expansion to the right of the second arrow in examples (27)-(31) demonstrates). (27)

gaishan correct-proper 'improve'

—> *gai-de-shan —» *gai-6u-gaishan

Word formation (28)

in Mandarin

27

shuoming

—> *shuo-d«-ming —> *shuo-¿>«-shuoming

speak-clear 'explain' (29)

—» *gai-de-liang *gai-Z>u-gailiang

gailiang correct-good 'reform'

(30)

—» *jiang-£w-he —» *jiang-fcn-jianghe

jianghe speak-harmonious 'reconcile'

(31)

bantuo

—» *ban-/>M-tuo —> *ban-b«-bantuo

do-proper 'do properly' The potential infixation process cannot apply to these restricted resultatives because they are formed at level I, and their heads therefore are not visible to the infixation process, which is located at level II. They also cannot take the V 1 -bu-V 1 -V 2 question inflection operation because that operation is located at level III. Moving on to level II, at this level are located all (endocentric) nominal compounding processes, the regular resultative verb compounding process, nominal affixation, and the -bu-l-de- resultative infixation process. First, consider the nominal affixes -zi -r and -tou. The function of these affixes is to form nouns when combined with a (usually nominal) root morpheme. The change in meaning of the resulting noun is subject to considerable variation, but often has a meaning of diminution or 'smallness' (Chao 1968 p 220 ff.). Some examples are: ping 'vase' —> pingzi 'jar, bottle'; shui 'water' —» shuir 'watery part'; jie 'to come together' —» jietou 'joint'. Note in (32)-(34) that that the affixes attach to the end of the word, but may not 'infix', or attach to the leftmost member of the word. (32)

a.

b.

zhuozi shuzhuo(zi') zhuoqiu zhuoz/qiu gangz/ shuigang(zi)

book-table table-ball

'table' 'desk' 'table-tennis'

water-jar

'jar, crock' 'water-jar'

28 J.L. Packard gangpen •gangzj'pen c. pingzi jiuping(zi) pinggai •pingzi'gai (33)

(34)

ge(r) shan'ge(r) gequ geci *gerqu *gerci b. hua(r) tuhua(r) huabao *huarbao c. ban(r) heiban(r) banbi *banrbi

jar-basin

'basin'

wine-bottle bottle-cover

'bottle' 'wine-bottle' 'bottlecap'

mountain-song song-song song-word

song 'folk song' 'song' 'lyrics'

chart-picture picture-paper

'picture' 'drawing' 'pictorial'

black-board board-wall

'board' 'blackboard' 'partition'

soft-bone bone-frame bone-marrow

'bone' 'coward' 'skeleton' 'bone-marrow'

ten-word-street street-road

'street' 'intersection' 'street, neighborhood'

back-rear rear-station

'rear' 'behind-the-back' 'the rear station'

a.

a.

gatou ruangutou guge gusui •guioMge •guioMsui b. ji ttou shizijieto« jiedao •jietowdao c. houtoM beihou(tou) houzhan •houfoMzhan

The explanation for these facts is that the nominal affixation process operates on the head of the word (as stipulated in (2)). Since the head is on the right for nouns (as stipulated in (1)), the affixation process therefore will not apply internally (i.e., on the left) even though it has access to the internal structure of the word by virtue of being located at the same level.

Word formation in Mandarin

29

Next, (35)-(48) (as in (12)-(13), (15)-(16) and (18)-(19) above) demonstrate that while regular compound verbs may take V1-bu-V1-V2 question inflection ((35)-(40)), this is not possible for resultative verb compounds ((41)-(48); also true for restricted resultatives, see (27)-(31) above). This is true whether the resultative form has undergone the potentializing (-bu-l-de-) infix process or not, as seen in (45)-(48). (35)

Women jintian yiding taolun zheige wenti we today surely discuss-speak this-Cl problem 'We'll surely discuss this problem today'

(36)

Women jintian tao-òu-taolun zheige wenti? we today discuss-not-discuss-speak this-Cl problem 'Are we discussing this problem today?'

(37)

Ta xialibai yao kaoshi he next-week will test-try 'He'll take a test next week'

(38)

Ta xialibai kao-ft«-kaoshi? he next-week test-not-test-try 'Will he take a test next week?'

(39)

Neige wenti hen qingchu that-Cl problem very clear-neat 'That problem is clear'

(40)

Neige wenti qing-ftw-qingchu? that-Cl problem very clear-neat 'Is that problem clear?'

(41)

Women jintian yiding tanwan zheige wenti we today surely talk-finish this-Cl problem 'We'll surely finish talking about this problem today'

(42) (43)

(44)

*Women jintian tan-few-tanwan zheige wenti? Ta mingnian yiding kaoshang daxue he next-year surely test-ascend college 'He'll surely test into college next year' *Ta mingnian kao-bw-kaoshang daxue?

30 JL. Packard (45)

(46) (47)

(48)

Women jintian yiding taiufewan zheige wenti we today surely talk-can-finish this problem 'We can surely finish talking about this problem today' * Women jintian tan-ftu-tandewan zheige wenti? Ta mingnian yiding kacxfeshang daxue he next-year surely test-can-ascend college 'He can surely test into college next year' *Ta mingnian kao-bu-kaodeshang daxue?

The explanation for the inability of the resultative verbs to take the V1-bu-V1-V2 question inflection is that the resultatives are formed at level II, while the question inflection process is located one step higher, at level III. The inflection operation cannot therefore apply to the head (i.e., the left-hand member) of resultative verb compounds because the internal structure of these compounds is invisible to that operation. Moving on to level III, at this level are located all 'regular' (i.e., non-resultative, non-exocentric) verbal and stative verbal compounding processes, tightly-bound verbobject compounding processes, and the V1-bu-V1-V2 question inflection operation. As is expected, since regular verb compounds are formed at this level, they may be converted into questions by the V1-bu-V1-V2 question inflection operation as seen, e.g., in (12)-(13), (15)-(16), (18)-(19) and (35)-(40) above. This is also true of 'tightly-bound' verb-object (V-O) verb compounds such as those in (49) below, as exemplified in (50)-(55). (49)

chuban guanxin zhuyi

emit-edition concern-heart attend-meaning

(50)

Ta meinian dou chuban wenzhang he every-year all emit-edition article 'He publishes articles every year'

(51)

Ta meinian chu-Zw-chuban wenzhang? he every-year emit-not-emit-edition article 'Does he publish articles every year?'

(52)

Ta hen guanxin nide qingkuang he very concern-heart your situation 'He is concerned with your situation'

'publish' 'be concerned with' 'notice'

Word formation in Mandarin (53)

Ta guan-feM-guanxin nide qingkuang? he concern-not-concern-heart your situation 'Does he care about your situation?'

(54)

Neige xiaohair hen zhuyi chezi that-M child very attend-meaning car 'That child is very careful about traffic'

(55)

Neige xiaohair zhu-fen-zhuyi chezi? that-M child attend-not-attend-meaning car 'Is that child careful about traffic?'

31

However, while these 'regular' verb compounds and tightly-bound V-O verbal compounds may undergo the V1-bu-V1-V2 question inflection operation, they may not undergo the potentializing (-bu-l-de-) infixation process, which operates on resultative verb compounds such as tanwan and kaoshang in (41) and (43) to give us tandewan and kaodeshang respectively, as seen in (45) and (47). (56)-(67) demonstrate that the potentializing operation results in ungrammatical forms when applied to regular verb compounds such as taolun or kaoshi ((56)-(59)), stative verb compounds such as qingchu or gaoxing ((60)-(63)), or tightly-bound V-O forms such as chuban or guanxin ((64)-(67)). (56)

(57) (58)

(59) (60)

Women jintian yiding taolun zheige wenti we today surely discuss-speak this problem 'We'll surely discuss this problem today' *Women jintian yiding tao-de-lun zheige wenti Ta xialibai yiding kaoshi he next-week surely test-try 'He'll surely take a test next week' *Ta xialibai yiding kao-de-shi Dao shihou yiding hen qingchu arrive time surely very clear-neat 'When the time comes, (it) will surely be clear'

(61)

*Dao shihou yiding qing-de-chu

(62)

Ta mingtian hui gaoxing he tomorrow can high-elated 'He will be happy tomorrow'

32 JL. Packard (63) (64)

(65) (66)

(67)

*Ta mingtian gao-de-xing Wo jinnian yiding chuban neipian wenzhang I this-year surely emit-edition that-M article 'I will surely publish that article this year' *Wo jinnian yiding chu-de-ban neipian wenzhang Ta yiding guanxin nide qingkuang he surely concern-heart your situation 'He is surely concerned with your situation' *Ta yiding guan-de-xin nide qingkuang

The preceding facts reveal an interesting generalization about word formation in Mandarin. That is, there is a relationship of complementarity between resultative and regular (including stative and verb-object) verbal compounds with respect to their ability to accept V1-bu-V1-V2 inflection versus -bu-/-de- potential infixation. Regular verbal compounds accept only Vj-bu-Vj-V2 inflection, while resultative verb compounds accept only potential infixation. The explanation for this complementary distribution of morphological processes is simply that these two operations both inflect the heads of verbs (i.e., the left-hand or 'internal' member), but they operate at different, mutually exclusive, lexical levels. Moving on to level IV, this is where all regular inflectional processes occur. These 'regular' inflectional processes are those that are totally productive, and may apply to any verb (excluding those with which they are semantically incompatible). This class of inflectional processes includes the aspect marker -le, the marker of past experience -guo, the potentializing endings -deliao and -buliao, the marker of continuing aspect -zhe, and so on. Ordinarily, verbal affixation applies to the head of a word, which would be the left-hand member in the case of verbal compounds. But since the inflection processes are located at level IV, they cannot 'see' that the heads of these compounds are located word-internally on the left, and any attempt to apply the inflection wordinternally is therefore ungrammatical, as seen in (68)-(87). Unlike affixation at other levels, affixation at level IV (inflection) applies to the word by default just in case the head is not visible, as stipulated in (2). Examples are given using 'regular' ((68)(74)) and 'regular resultative' ((75)-(81)) verb compounds, and tightly-bound V-0 verbal compounds ((82)-(87)) below. (68)

jiejue xuexi shoushi

resolve-decide leam-review collect-pick

'solve' 'study' 'clean up'

Word formation in Mandarin (69)

(70) (71)

Women zuotian jiejue-/e zheige wenti we yesterday resolve-decide-ASP this problem 'We solved this problem yesterday' *Women zuotian jie-te-jue zheige wenti Zheixie hanzi wo dou xuexi-guo this-Cl:pl character I all learn-review-ASP 'I've studied these characters before'.

(72)

*Zheixie hanzi wo dou xue-guo-xi

(73)

Wuzi shoushi-/e ma? room collect-pick-ASP Q 'Has the room been cleaned up?'

(74)

*Wuzi shou-/c-shi ma?

(75)

kanjian zuowan zhuadao

(76)

Ta zuotian kanjian-/e Lao Li he yesterday look-perceive-ASP Lao Li 'Yesterday he saw Lao Li'

(77) (78)

(79) (80)

(81) (82)

look-perceive do-finish grab-arrive

'see' 'finish doing' 'capture'

*Ta zuotian kan-/e-jian Lao Li Wo shangxingqi yijing zuowan-/e naxie gongke I last-week already do-finish-ASP that-Cl homework 'I already finished that homework last week' *Wo shangxingqi yijing zuo-Ze-wan naxie gongke Didi zhuadao-/e laoshu brother grab-arrive-ASP mouse 'Little brother caught a mouse' *Didi zhua-/e-dao laoshu Zheiben shu mei chuban-guo this-M book not emit-edition-ASP 'This book has not been published before'

34

J.L.Packard

(83) (84)

(85) (86)

(87)

*Zheiben shu mei chu-gMo-ban Ta conglai mei guanxin-gwo wode shi he ever not concem-heart-ASP my affair 'He has never been concerned with my affairs before' *Ta conglai mei guan-guo-xin wo de shi Wo zhuyi-/e ta shenmo shihou zou de I attend-meaning-ASP he what time leave MOD 'I noticed what time he left' *Wo zhu-/e-yi ta shenmo shihou zou de

A summary of the Mandarin lexicon is given in table 2 below. Table 2: Mandarin Lexicon Notes: a. all complex nouns right-headed b. all complex verbs (incl. stative (=SV)) left-headed c. levels crucially ordered d. no ordering within levels e. structure formed at one level is invisible at subsequent levels level I (compounding only; no affixation) All exocentric compounding processes [-V x] v [-SV x] s v [x -N] n (e.g., zuoyou, zuiying) Restrictive Resultative Compounds (allow no potential) = [V V] v (where 2nd V is bound, and is a result) (e.g., gaishan, shuoming) level II (compounding and affixation) Compounding: [x N] n (i.e, [SV N] [N N] [V N]) = all endocentric nominal compounding Regular Resultatives (allow potential) = [V V] v (where 2nd V is free, and is a result) Affixation: -zi, -tou, -r nominal suffixation (operates on head) -bu-/-de- potential infix for [V V] v (operates on head) level III (compounding and affixation) Compounding: [V V] v (where 2nd V is not a result) = all regular verbal compounds (e.g., xihuan) [SV x] s v = all endocentric SV compounds (e.g., gaoxing) Highly bound [V 0 ] v (inflect outer rather than inner verb) (e.g., chuban, zhuyi, guanxin, zhidao) Affixation: Q-word formation: V1-bu-V1-V2 (operates on head)

Word formation in Mandarin

35

level IV (inflectional affixation only; no compounding)

Inflection: (-zhe, -guo, -le, -buliao, -men, etc.) I have presented some data, and accounted for it using a theoretical framework. If this were as far is it were to go, the framework could simply be considered nothing more than a fancy descriptive or notational device which was constructed only to account for distribution patterns in the data. In other words, it could be argued that this method of assigning processes to different levels has no independent motivation, other than attempting to account for distributional and co-occurrence restrictions seen in the data. In fact, there is an independent criterion for assigning the various morphological processes to different lexical levels. That criterion, as alluded to earlier, is productivity. In general, the morphological processes become more productive as we move from level I to level IV. For affixation processes, productivity simply means the propensity of a given affixation process to apply to words. For the compounding (i.e., non-affixation) processes, we may consider two kinds of productivity: the first is the word formation productivity of the compounding process itself, and the second is the propensity of a particular compounding process to accept affixation. Let us briefly go through the productivity of processes at each level. Exocentric compounds are at level I because they are productive in neither sense of the word. That is, exocentric compounding neither accepts affixation, nor is it a highly productive process of word formation. Class 1 ('restricted') resultatives are assigned to level I because they do not constitute a productive process (that is, they are rare), and also they accept no internal affixation. Nominal compounding is assigned to level II because while it is fairly common in forming words, it is nonetheless only semi-productive in accepting suffixation. The number of noun compounds which takes nominal suffixes is limited. Class 2 resultative compounding is assigned to level II because it is a fairly but not totally productive word formation process. Also, the meanings of resultative compounds are often not fully predictable, i.e., they are often idiosyncratic (e.g., chiguang, eatbright, 'to completely eat up'). The -de-l-bu- potential infix is at level II because although it productively applies to regular resultatives, it is clearly less productive than V r b u - V r V 2 question formation, which is at level III. Both regular verbal and stative verbal compounding are at level III because they are very productive processes of word formation in Mandarin. In addition, both regular verb and stative verb compounds take V r b u - V r V 2 question inflection without exception. Vj-bu-Vj-V2 question formation is at level III because it is obviously highly productive, applying without exception to three separate classes of verbal compounds: stative verb compounds, regular verb compounds, and verb-object compounds. Regular inflection is located at level IV because it is totally productive. It is possible for any verb to accept verbal inflection, subject only to semantic restrictions.

36 JL. Packard In addition, the placement of inflection at level IV is consistent with the theory of lexical morphology, since most previous investigators have placed inflection processes at the most 'shallow' level, that is, the level closest to the syntax. It stands to reason that if lexicons have similar structures across languages, then they should share certain cross-linguistic properties. Some of the more obvious ones we have seen are the proximity of inflection to syntax, and the reduction in productivity as we move deeper into the lexicon. Other cross-linguistic properties that we might expect to find would be similarities in psycholinguistic processing, first-language acquisition and aphasie loss of elements at a given lexical level (see, e.g., Packard, in preparation). In conclusion, I have presented some distributional facts regarding word formation in Mandarin, and have proposed a system which accounts for these facts. I hope that this analysis demonstrates that, contrary to what many scholars have contended, Mandarin Chinese does have morphology, and the morphology that it has is very interesting indeed.

NOTES 1. This paper is an extensively revised version of a talk presented at the first Northeast Conference on Chinese Linguistics (NECCL), at Columbus, Ohio, on May 6, 1989.1 especially thank Claudia Ross for helpful comments. I also thank Tom Ernst, Shengli Feng, Jack Hoeksema, Frank Hsueh, Shizhe Huang, Rich Janda, Paul Lloyd, Bob Sanders, Chilin Shih, Ziqiang Shr, Jim Tai, Shou-hsin Teng, Xinping Zhou, Nancy Eng and Chioko Takahashi for helpful comments. Thanks also to two anonymous reviewers for Yearbook of Morphology, and to its editors, for many astute observations and helpful comments. 2. The term 'compound' is traditionally used in Chinese to denote bisyllabic words which do not have an affix as one of the two syllables. This is because virtually every non-affix syllable in Mandarin is a clearly isolable morpheme of the language. I will continue to use the term here, and offer criteria for the definition of 'affix' in Mandarin in note 4. 3. For an example of a lexical treatment of Mandarin phonology, see Chan (1984). 4. As noted above, compounds are bisyllables which do not contain affixes. Affixes are defined in Mandarin as i) bound forms that are ii) at least somewhat productive, with a iii) fairly constant meaning, and iv) occupy a consistent position in the formation of words.

REFERENCES Allen, M. 1978. Morphological Investigations. PhD Diss., University of Connecticut. Chan, M. 1984. "Word formation in Mandarin: a preliminary sketch". Paper presented to the West Coast Conference on Linguistics (WESCOL). Vancouver, Canada. Chao, Y-R. 1968. A Grammar of Spoken Chinese. Berkeley: University of California Press. DiSciullo, A.-M. and E. Williams. 1987. On the Definition of Word. Cambridge: MIT Press. Halle, M. and K.P. Mohanan. 198S. "The segmental phonology of English". Linguistic Inquiry 16,57-116. Huang, C-T.J. 1989. "Modularity and explanation: the case of Chinese A-not-A questions". In M. Chan, and T. Ernst (eds.), Proceedings of the Third Ohio State University Coherence on Chinese Linguistics. Bloomington, Indiana: Indiana University Linguistics Club.

Word formation in Mandarin

37

Huang, C-TJ. 1988. "Chinese A-not-A questions: a modular approach", (in Chinese). Zhongguo Yuwen 5, 247-264. Kiparsky, P. 1982. "Lexical morphology and phonology". In The Linguistic Society of Korea (ed.), Linguistics in the Morning Calm. Seoul: Hanshin. Li, C. and S. Thompson. 1981. Mandarin Chinese. Berkeley: University of California Press. Lu, Z-W. 1964. Chinese Morphology (in Chinese). Beijing: Scientific Publishing Co. Mohanan, K.P. 1986. The Theory of Lexical Phonology. DordrechtReide! Packard, J. In preparation. The Linguistic Analysis of Aphasie Chinese Speech. Packard, J. 1989. "On the notion 'head of a word' in Mandarin". Ms., U. of Pennsylvania. Scalise, S. 1984. Generative Morphology. Dordrecht: Foris Publications. Selkirk, E. 1982. The Syntax of Words. Cambridge: MIT Press. Siegel, D. 1974. Topics in English Morphology. PhD Diss., MIT.

Yearbook of Morphology 3 (1990), 41-43

Introduction Allard Jongman and Aditi Lahiri

The collection of papers in the thematic part of this volume addresses the issue of the interaction of linguistic components in grammar and processing. Versions of these papers were presented in a workshop entitled "When words happen to be phrases" (18-20 December, 1989) as part of the Interface Project at the Max Planck Institute for Psycholinguistics. Discussants' reports, following some papers, have been included. Since some of the papers presented at the workshop were on areas other than those appropriate for this journal, we were unable to include them here.* The workshop focused on the structural intricacies and incongruities of the morphology/syntax relation and their processing consequences. The differences and similarities between word-like things and phrase-like things are not as clear-cut as one could hope for. For instance, there could be two competing views on the properties of complex words. Complex words are essentially different from phrases. They are formed according to rules and principles which have virtually nothing in common with those that guide operations in the phrasal domain. Or it may be the case that complex words follow approximately the same rules and principles as phrasal syntax. This includes drawing comparisons between morphologically and syntactically complex words, verb clusters, and phrasal units. The papers in this volume certainly do not all propagate a single view, and indeed, they succeed in bringing out the controversial nature of the problem by examining crosslinguistic data. Booij discusses the interpretation of separable complex verbs (SCV, 'op-bellen') in Dutch which show word-like properties while the initial particles show prefix-like properties. He draws a distinction between SCVs and the comparable inseparable complex verbs (ICV) with similar initial constituents and claims that SCVs are not words like ICVs. Neither can SCVs be treated as syntactic phrases since they do not follow normal phrasal properties. Booij claims that SCVs are best thought of as phrases that are created in the lexicon by a special kind of phrasal construct, namely [[Prt][V]]v,, where the V* node comes between V and the projection of V. Thus, they are not constructed by morphological rules, nor are they phrasal structures at the sentence level. Schreuder also discusses the properties of SCVs, but from a processing perspective and argues that as far as the recognition lexicon is concerned, the particle and the verb of the SCV are not accessed as single units like the ICVs. The approaches are quite compatible in that under both theories the separable complex verbs are created

42 A. Jongman and A. Lahiri in the lexicon and are not syntactic phrases. Schreuder views the SCVs as having an integration node in the recognition lexicon where the prefix and verb stem are put together, enabling the entire SCV to be recognized. In Booij's model, the integration takes place under the V* node which is a special phrasal construct also created in the lexicon. Frazier supports the view that the major syntactic category of a word must be available before phrasal analysis begins. Thus, the parsing of compounds is a matter of the syntactic parser, not the morphological parser. This accounts for the fact that in strings like the warehouse fires, no syntactic structure is assigned until the following word disambiguates the X° category of fires. Each potential X° is minimally attached to the syntactic phrase marker. Thus, for an ANN sequence such as crispy almond tarts, the prefered parsing is one in which tarts becomes the head of the right-branching structure [crispy] [almond tarts]. However, these preferences are overridden by lexicalized compounds. These are preferred and become available before the parser processes strings as syntactic compounds. Therefore, fast food delivery would have the natural reading [fast food] [delivery] since fast food is lexicalized. In this view, then, compounds other than lexicalized compounds are parsed according to the principles of phrase structure parsing. Lahiri, Jongman, and Sereno discuss verb plus pronominal cliticization which optionally triggers word-boundary phenomena. When attaching to verbal hosts, Dutch pronominal clitics can behave in two different ways. The two syntactic words (host and clitic) can combine to form different prosodic units, with different surface phonological consequences. In a psycholinguistic experiment, the parsing and recognition of the verb-plus-clitic constructions were shown to be determined by the relation of the surface string to the underlying phonological representation of the verb rather than the prosodic characterization of the string. Reuland discusses the role of word structure in the morphology/syntax interdace. Under current assumptions about the relation between word formation and syntax, the word is at best a derivative notion, without any independent status in the grammar. Reuland, however, argues for a level of word structure as the interface between morphological structure and S-structure. He proposes a morphology-based approach, in which complex verbal categories are formed in the lexicon (i.e., presyntactically) and projected according to principles of X'-syntax. In Dutch, an SOV language, the domains of V and I can overlap. According to current assumptions which are mostly built on analyses of English and other VO-languages, V and I project independently. In Dutch, however, no element which is not part of the verb - for instance a complement - can intervene between the main verb and the auxiliary. To allow the overlap, Reuland argues that V and I are merged in the morphology and are projected simultaneously in the syntax. Bierwisch brings up the question of whether coherent verbal constructions have both word-like and phrase-like properties, being intermediate between morphology and syntax. Bierwisch argues that verb clusters in German such as gehen lassen wollen are complex words, created by morphological processes in the lexicon, similar

Introduction

43

to those of a stem and a suffix within a complex word. He draws attention to a number of syntactic properties which these verb clusters share with simple verbs. Zwicky draws a distinction between morphological words ('moremes') and syntactic words ('Ws'), either of which could be simple or composite. Composite phrasal expressions like a slept all day look, or verb clusters similar to those discussed by Bierwisch, can behave word-like syntactically but not necessarily morphologically. Moreover, Zwicky draws attention to the fact that the morphologysyntax interface can instantiate a single moreme by a sequence of one or more Ws, as well as a single W by a composite of two or more moremes. He also discusses ways in which word-like units could be detected. He distinguishes three types of morphosyntactic "intervention constraints" which clarify possible and impossible interruptions of different types of words. While this set of papers does not present a unitary account of morphology/syntax phenomena, it certainly reflects the fact that the study of interfacing grammatical components is an exciting and relatively new area of research. The inherent complexity of studying the relations between two or more modules of the grammar makes this area a prime candidate for an interdisciplinary approach that must include both theoretical and experimental investigation. We hope that this volume will provide an incentive for further research in this difficult domain. * We would like to thank the editors of the journal, Geert Booij and Jaap van Marie, the reviewers, and Josef Bayer, Edith Sjoerdsma and Gertie de Groen for making this volume possible.

Yearbook of Morphology 3 (1990), 45-63

The boundary between morphology and syntax: separable complex verbs in Dutch* Geert Booij

1. INTRODUCTION: THE PRINCIPLE OF LEXICAL INTEGRITY

Probably the most fundamental principle of the lexicalist theory of word formation is the Principle of Lexical Integrity that is formulated as follows by Lapointe (UK)*): 1 (1)

Principle of Lexical Integrity No syntactic rule can refer to elements of morphological structure.

This formulation of the Principle of Lexical Integrity does not exclude all types of interaction between morphology and syntax. In particular, it does not exclude the possibility that phrases form parts of words. This is in accordance with ample empirical evidence that the so-called 'No Phrase Constraint' (a term from Botha 1984) is incorrect, as is shown in e.g. Dressier (1988) and Hoeksema (1988). Not only can lexicalized phrases form inputs for both compounding and derivation (as in the Dutch words God-is-dood-theologie 'God is dead-theology* and ban-debommer 'ban the bomb-er'), but certain types of phrases can productively be used as the specifier part of compounds in Dutch (e.g. NP's with adjectival specifiers as in [[[0wde]A[7nannen]N]N.[/it«j]N]N 'old men's home'). Similar evidence from Afrikaans is given in Botha (1984: 28).2 Note that the Principle of Lexical Integrity does, however, predict that wordinternal phrases are not accessible to the rules of syntax. In this paper I will deal with one of the claims implied by the Principle of Lexical Integrity, namely, that material cannot be moved out or into words. This hypothesis has received a great deal of discussion in the recent literature, notably by Baker (1988a, b) and Di Sciullo & Williams (1987). Baker allows for syntactic rules of incorporation, which move X°-categories (both lexical morphemes and bound morphemes) and adjoin them to other words, thereby creating complex words in surface structure, whereas Di Sciullo & Williams want to exclude such analyses. I will focus here on the other side of the coin, namely the claim that sublexical elements cannot be moved by syntactic rules out of the complex words to which they Author's address: Vakgroep Taalkunde, Vrije Universiteit, P.O. Box 7161, 1007 MC Amsterdam, The Netherlands.

46 G. Booij belong in deep structure. I will show that the morphological and syntactic behaviour of an important class of verbs in Dutch, the so-called separable complex verbs (e.g. opbellen 'to phone') does not conflict with the prohibition on syntactic movement of sub-lexical elements out of words. I will argue that the separable complex verbs are phrasal constructs, and that the two parts, therefore, can be separated. Their word-like properties will be shown to follow from independently motivated principles. Moreover, I will show that these verbs cannot be analyzed as the result of a syntactic rule of incorporation.

2. PHRASAL PROPERTIES AND WORD PROPERTIES OF SEPARABLE COMPLEX VERBS

Dutch, like the related languages of German and Afrikaans, has a large class of separable complex verbs (SCV's) illustrated here by opbellen 'to phone' (in Dutch orthography SCV's are written as one word when the two constituents are adjacent). SCV's behave as phrases in that the two parts can be separated by rules such as Verb Second and Verb Raising, as illustrated in (2) and (3) respectively. Verb Second is the rule that moves the tensed verb into second position in main clauses, and presupposes that the underlying word order of Dutch is SOV (cf. Koster 1975). Verb Raising is the rule that raises the verb of an embedded clause to the right of the verb of the dominating clause where it forms a unit with that verb (Evers 1975). (2)

DS John me op belde SS John beldej me op t; John rang me up 'John phoned me'

(3)

DS dat John [PRO me op bellen] wil SS dat John [PRO me op t j wil belief that John [PRO me up t j wants ring; 'that John wants to phone me'

Note that instead of bellen it is also possible to raise opbellen: (4)

dat John [PRO me t^ wil opbellen; that John [PRO me t j want up ring 'that John wants to phone me'

However, this does not imply that opbellen is one word. The facts are also explained if we assume that SCV's are minimal projections of V because verb projections can also be raised, as pointed out by Haegeman & van Riemsdijk (1986: 419). For instance, in West Flemish both V and VP (i.e. V + direct object) can be raised. For standard Dutch we therefore assume that both V and V* ( = the minimal projection

Separable complex verbs in Dutch

47

of V, itself dominated by the VP-node and the dominating node for SCV's) can be raised. In other words, I propose to assign the structure [P V]v* to SCV's, where P stands for Particle. The separability of SCV's can also be observed in the way they form their past participles. In Dutch, past participles are formed by prefixing ge- and simultaneously suffixing t/d to the verbal stem. However, in SCV's ge- occurs in between the particle and the verb: op-ge-beld. This follows directly from a phrasal analysis of such constructions, whereas it forms a problem in an analysis that assigns word status to SCV's. It was on the basis of such an observation that De Rooy-Bronkhorst (1980) proposed to assign phrasal status to SCV's. This separability can also be observed with respect to the occurrence of the infinitival particle te that appears before verbal infinitives: (5)

John belooft te komen 'John promises to come' John belooft op te bellen 'John promises to ring'

That is, in infinitival forms of SCV's, te appears in between the two parts: op te bellen. Again, this argues in favor of a phrasal analysis of SCV's. In what is probably the first generative analysis of verb-particle combinations, Koster (1975: 171) proposes to interpret them as compound verbs. For instance, opbellen is assigned the following structure:

To account for the fact that only the base-V is moved by V-second, leaving the particle behind, Koster assumes that V-second only applies to the tensed part of compound verbs. This means that we have to allow for syntactic rules that move sublexical elements out of words. Moreover, this solution presupposes that the feature [tense], for some reason, does not percolate from the head to the dominating V-node, contrary to standard assumptions about the percolation of morpho-syntactic features (cf. Lieber 1989). This problem is also discussed by Groos (1989) who analyzes SCV's as cases of 'deep structure adjunction' (in opposition to Baker's (1988a) analysis of complex verbs as cases of surface structure adjunction). The hypothesis of deep structure adjunction means that verbs are (obligatorily or optionally) subcategorized for one or more particles that are adjoined to the verb, resulting in the following syntactic structure for SCV's:

48 G. Booij Vo

(7)

Vo

Part

Verb Raising can now apply to either the lower or the higher V. Verb Second applies only to the lower V because the feature [tense] only occurs on the lower V. This is based on the assumption that "feature percolation is limited to projections of heads, excluding thereby percolation to adjunction categories" (Groos 1989: 51). In my opinion, Groos's analysis has a number of problematic aspects. Firstly, it remains unclear how a structural distinction is to be made between cases of deep structure adjunction and those of morphologically derived compound verbs, such as slaapwandelen 'to sleep-walk', which have a parallel structure: Vo

(8) y° I slaap

yo I wandelen

In other words, how can the percolation conventions see that wandelen is the head of slaapwandelen, yet that be lien is not the head of opbellenl Secondly, if V-raising can apply to either the lower or the higher V o , why is it incorrect to only raise wandelen?: (9)

DS dat hij [PRO slaapwandelen] ging that he sleep-walk went that he went sleep-walking SS *dat hij [PRO slaap t¡] ging wandelen¡

Finally, the deep structure adjunction hypothesis does not account for the effects of the addition of particles on the argument structure of the resulting SCV (see Section 3). The problem is that argument structure may not be changed in syntax, since this would violate the Projection Principle. This issue will be discussed in greater detail below. Let us now discuss the observations that have led people to assign word status to SCV's, a position also reflected in the orthographical convention of writing Dutch SCV's as one word. Before going into the details, I present a representative list of words that occur as first constituents of SCV's, and of comparable inseparable complex verbs (ICV's) with similar first constituents, as well; they are presented in their citation form, the infinitive, with the infinitival suffix -en:

Separable complex verbs in Dutch (10)

SVC: ian-blijven 'stay on' ichter-blijven 'stay behind' if-branden 'burn down' bij-betalen 'pay extra' d6or-boren 'go on drilling' in-ademen 'inhale' nl-kijken 'check' n6er-gooien 'throw down' 6m-blazen 'blow down' 6nder-gaan 'go down' 6p-komen 'come up' 6ver-komen 'come over' r6nd-brieven 'tell around' t6gen-gaan 'combat' uit-ademen 'breathe out' uit6en-Iopen 'diverge' v6or-komen 'occur' v6ort-duren 'continue' w6er-keren 'return'

49 ICV: aan-bfdden 'worship' achter-Mlen 'recover' door-b6ren 'perforate'

om-bldzen 'blow around' onder-gian 'undergo' over-k6men 'happen to'

voor-k6men 'prevent' weer-gilmen 'echo'

Most of the words that occur as first constituents can also function as prepositions/ postpositions. However, mis, neer, uiteen, voort, and weer do not occur as adpositions, but only as adverbs. I will use the usual term 'particle' to refer to this class of separable adpositions (P) and adverbs (Adv). There is also a small set of SCV's, the first constituent of which is either an adjective (e.g. goed-keuren 'to approve', vol-houden 'to go on'), a noun (e.g. ademhalen 'to breathe', stof-zuigen 'to vacuum clean', feest-vieren 'to have a party'), or a morpheme that does not exist as an independent word (e.g. gade-slaan 'to watch'). Two of these adjectives, vol 'full' and mis 'wrong', also occur as the first constituent of ICV's: (11)

SCV mis-grijpen 'miss one's hold' v61-houden 'maintain'

ICV mis-stian 'not suit' vol-br6ngen 'accomplish'

In the case of Noun-Verb combinations their unity can be observed in that they take niet as their negative element (= the negative element for verbs), whereas a syntactically independent NP takes geen as its negative element: (12)

a.

Hij kon niet / geen adem halen 'He could not breathe'

50 G. Booij b.

Hij kon *niet / geen adem krijgen 'He could not get breath'

The fact that in (12a) both niet and geen can occur follows from the ambiguity of adem halen. It is an established SCV (and thus selects niet), but can also be interpreted as a VP with the generic noun adem as its direct object (with concomitant selection of geen). Also compare the following sentences: (13)

a. b.

Hij kan niet stofzuigen 'lit. He cannot dust-suck' 'He cannot vacuum-clean' (= He cannot work with a vacuum-cleaner) Hij kan geen stof zuigen 'He cannot suck dust'

In (13a) we find the established SCV stofzuigen with the idiosyncratic meaning 'to vacuum-clean', whereas in (13b) we are forced to assign a literal interpretation to stof zuigen. Given the remarkable similarity between the SCV's and the ICV's illustrated in (10) it becomes clear that it will be hard to account for their difference in separability if both types are assigned word status. For instance, why is it possible to separate the two parts of overkomen 'to come over', whereas this is impossible in overkomen 'to happen to'? This difference follows straightforwardly if SCV's and ICV's have the following structures: SCV: ICV:

t[X][V]]v, [[X][V]]v

(X = P, Adv, A, N)

The existence of such minimal pairs is therefore the main empirical stumbling block for the proposal by Le Roux (1988) to consider SCV's as verbal compounds and to interpret the rule of V-second as a case of Head Movement in the sense of Baker (1988a). Why then are people inclined to assign word status to SCV's? One reason is that they frequently function as inputs for deverbal word formation processes (14a) and compounding (14b): (14)

a.

b.

aanbied-en 'to offer' aankom-en 'to arrive' aantrekk-en 'to attract' aanton-en 'to demonstrate' opberg-en 'to store' doorkijk-en 'to see through'

aanbied-er 'offerer' aanbied-ing 'offer' aankom-st 'arrival' aantrekk-elijk 'attractive' aantoon-baar 'demonstrable' opbergdoos 'store box' doorkijkbloes 'see through blouse'

Separable complex verbs in Dutch

51

However, this does not contradict a syntactic analysis of SCV's. As we saw in Section 1, not all word formation is exclusively word-based: the No Phrase Constraint is empirically incorrect, and phrases may occur as parts of complex words. On the other hand, this does not mean that all word formation rules accept syntactic inputs. For instance, prefixes do not attach to syntactic phrases, and, more generally, phrases normally do not occur as the heads of words in Dutch.3 This is very relevant to the problem under discussion. For instance, the deverbal prefix vercan be attached to ICV's, but not to SCV's: (15)

a. b.

with IVC's:

veronderstdllen 'presuppose' veroverh6ersen 'overrule' vervolmdken 'make perfect' with SVC's: *ver6vermaken, *verv61maken

This simply follows from the prohibition on phrases as heads or as inputs for prefixation, if we assign SVC's the V* status.4 A second reason for assigning word status to SCV's is the fact that their semantic interpretation is usually partially or completely unpredictable. This is similar to what we find for English verb-particle combinations such as to look for, to look after etc. Some examples are given in (16). (16)

vangen 'to catch' slaan 'to hit' wenden 'to turn' heffen 'to raise'

aanvangen 'to begin' aanslaan 'to start barking' aanwenden 'to use' aanheffen 'to start singing'

Semantic idiosyncrasy, however, is not a proper criterion to distinguish words and phrases, since there are many idiosyncratic expressions for which there is no doubt as to their status of syntactic constructs (cf. Di Sciullo & Williams 1987). The only consequence is that many SCV's have to be listed in the lexicon. This is also necessary for formal reasons: some SCV's contain parts that do not occur as independent words, 'cranberry morphs' in the terminology of Aronoff (1976): (17)

a. first part non-existent. gade-slaan 'to watch' b. second part non-existent: na-bootsen 'to imitate' om-kukelen 'to fall down' aan-tijgen 'to accuse' op-kalefateren 'to restore'

52 G. Booij A third observation that has led linguists to consider SCV's as words is the fact that the particles, like (real) prefixes, seem to have the power to change the syntactic category of the input words. In Dutch, prefixes such as be- and ver- derive verbs from verbs, nouns, and adjectives, and hence have category-changing power, which is characteristic of derivational morphology. The separable particles also seem to have this power, witness the data in (18): (18)

input A: dik 'thick' zwak 'weak' snel 'fast' diep 'deep'

SCV: indikken 'to thicken' afzwakken 'to weaken' toesnellen 'to rush forward' uitdiepen 'to deepen'

input N: burger 'citizen' aap 'monkey' brief 'letter' huwelijk 'marriage'

SCV: inburgeren 'to settle' naapen 'to imitate' rondbrieven 'to tell around' uithuwelijken 'to marry off

Suppose that these facts would induce us to consider the formation of SCV's as part of morphology. It would then be impossible to consider them as compounds, because in Dutch compounds it is the right constituent that is the head and that determines the syntactic category of the resulting complex word, whereas in the examples listed in (18), the right constituent does not determine the syntactic category. Hence, words like in-, af-, na-, toe- and rond- would have to be considered as derivational prefixes, because derivational prefixes can change the syntactic category of their base words. We then end up with a situation in which prefixes, although they are bound morphemes, can nevertheless be separated from the base word to which they are attached. It will be clear that this is a major drawback for a prefixal interpretation of the first constituents of SCV's, and therefore, another analysis is called for. My analysis is as follows: in Dutch, the conversion of adjectives and nouns into verbs is productive. Hence, a verb like zwakken as used in afzwakken is a possible word of Dutch. This possible word is only used in combination with the particle af. As has been argued in Booij (1977) and Allen (1978), for example, possible though not existing words are often used in creating complex expressions. Therefore, what I assume here is that these possible verbs, created from adjectives and nouns by conversion, are available for the creation of particle-verb combinations. Independent evidence for this analysis is that conversion to verbs in Dutch normally only applies to simplex adjectives and nouns (Van Marie 1985). This makes the correct prediction that, although both verbal prefixes (19a) and particles (19b) can occur with complex verbs, the apparently category-changing particles are only found to co-occur with mono-morphemic words.5

Separable complex verbs in Dutch (19)

a. b.

[on[geluk]N]N [[ootmoed]Nig]A [on[schuldig]A]A [ver[huren]v]v [[analys]Neren]v

'accident' 'humble' 'innocent' 'let' 'analyze'

53 verongelukken verootmoedigen verontschuldigen6 onderverhuren dooranalyseren

'to die in an accident' 'to humble' 'to excuse' 'to sublet' 'completely analyze'

A final argument for giving a morphological interpretation to the creation of SCV's is that the addition of particles quite often affects the syntactic valency of the verb, just like (real) prefixes. This phenomenon is discussed in detail in the following section. There it will be argued that changes in syntactic valency follow from changes in the lexical-conceptual structure of linguistic expressions. Hence, since the addition of particles affects the lexical conceptual structure of the particle-verb combination, the syntactic valency of an SCV may be different from that of the verb that it contains. As will be argued below, these observations do not force us to consider the formation of SCV's as cases of prefixation, but they do show that the formation of SCV's cannot be a matter of syntax. In sum, the word-like properties of SCV's and the prefix-like properties of particles as discussed so far, are not in conflict with a phrasal analysis of SCV's, an analysis that also accounts for the differences in both syntactic and morphological behavior between SCV's and ICV's.

3. SYNTACTIC VALENCY CHANGE AND SYNTACTIC ACCOUNTS OF SCV'S

Dutch verbal prefixes may have effects on the syntactic valency of the input words. For instance, the prefix be- systematically creates obligatorily transitive verbs from both intransitive and transitive input verbs (cf. Booij & Van Haaften 1988). A parallel phenomenon occurs with certain particle-verb combinations. (20)

bellen (opt. tr.) 'phone' / iemand opbellen 'phone somebody' lopen (intr.) 'walk' / de straten aflopen 'tramp the streets' rijden (intr.) 'ride' / de auto inrijden 'run in the car' wonen (intr.) 'live' / een vergadering bijwonen 'attend a meeting' juichen (intr.) 'cheer' / iemand toejuichen 'cheer somebody' zitten (intr.) 'sit' / een straf uitzitten 'serve one's time'

Transitivization is not the only effect on syntactic valency of the addition of a particle to a verb. In other cases, the effect is 'ergativization', i.e. the creation of a (superficially intransitive) verb that selects zijn as its auxiliary. According to Hoekstra (1984), such ergative verbs do not have an external argument. The internal argument is moved to the subject position in surface structure. Hence we have to distinguish between intransitive verbs (external argument only), transitive verbs (internal and external argument), and ergative verbs (internal argument only). The

54 G. Booij ergativization effect is illustrated in (21) and (22) - the difference between the two classes to be discussed below: (21)

glijden 'glide' / uitglijden 'slide' vliegen 'fly' / uitvliegen 'fly away' marcheren 'march' / afmarcheren 'march away' lopen 'walk' / weglopen 'walk away' zwaaien 'wave' / afzwaaien 'leave military service'

(22)

bijten 'bite' / uitbijten 'erode away' doezelen 'doze' / uitdoezelen 'doze away' studeren 'study' / afstuderen 'finish one's studies' koelen 'cool' / afkoelen 'cool down'

Although syntactic valency changes do occur as a result of derivational processes, it is not necessarily the case that valency change always indicates the process at hand to be a derivational one. Since argument structure is a projection of Lexical Conceptual Structure (Zubizarreta 1987, Booij & van Haaften 1988, Booij 1989, Carrier and Randall, ms.). valency changes can be seen as the effects of the semantic changes brought about by the combination of particles and verbs. However, before defending this position in more detail, I will discuss another, purely syntactic approach to Particle-Verb combinations, because this will serve to elucidate the range of facts that have to be accounted for. Following a suggestion in Kayne (1984), Hoekstra et al. (1987) propose a Small Clause analysis for SCV's. Their main argument is that the SC-analysis explains the transitivization effect illustrated in (20): (23)

a. b.

dat that dat that

ik I ik I

[Peter oplgc bel Peter on phone = 'that I phone Peter' [de auto in]^ rijd the car in run = 'that I run in the car'

That is, sentences with SCV's are analyzed as containing a resultative small clause of which the particle is the predicate, and thus they are claimed to be similar to resultative small clauses such as (24)

dat ik [de deur groen] sc verf that I the door green paint = 'that I paint the door green'

In other words, what look like the internal arguments of opbellen and inrijden, respectively, are the subjects of small clauses. That is, the transitivization effect is analyzed as the effect of the necessary presence of an overt subject in the resultative small clause.

Separable complex verbs in Dutch

55

For ergative SCV's such as wegsterven 'die out', Hoekstra et al. (1987: 68) propose the following analysis, illustrated by (25): (25)

dat het geluidj [tj weg]^ sterft that the sound away dies = 'that the sound fades away'

That is, since stervert is an ergative verb with no underlying subject, the subject of the small clause is moved to the subject position. This analysis covers two kinds of effect on syntactic valency: (i) transitivization and (ii) no change for SCV's that contain ergative verbs. However, we also find cases of ergativization, as shown in (21)-(22). The verbs in (21) are either intransitive or ergative, i.e. they occur with both hebben and zijn. They are ergative when combined with a directional adverbial. Compare (26a) with (26b): (26)

a. b.

Ik heb gevlogen 'I have flown' Ik ben naar Amsterdam gevlogen 'I have flown to Amsterdam'

Hoekstra et al. (1987) analyze naar Amsterdam in (26b) as the predicate of an SC. Hence, the ergative variant of movement verbs such as vliegen can be used in combination with a directional SC: (27)

dat ikj [tj naar Amsterdam] gevlogen ben

If this analysis is correct, it might also be used to explain the ergativity of the Particle-Verb combinations in (21), since the particles may be claimed to be directional predicates. However, the ergativity effect also occurs in the SCV's in (22), where the base verbs do not express some kind of movement. That is, the ergativity effect in this case cannot be said to follow from already available syntactic generalizations. Rather, what seems to be going on here is that the SCV expresses a change of state, unlike the corresponding single verbs, and changes of state are usually expressed as ergative verbs in Dutch (e.g. breken 'to break, erg.', sterven 'to die'). The crucial role of semantic (i.e. lexical-conceptual) structure in the determination of the syntactic valency of SCV's is also supported by the two uses of aflopen and afslaan (from ¡open 'to walk' and slaan 'to hit'): (28)

De wekker loopt af (erg.) "The alarm clock goes off Hij liep de hele tentoonstelling af (tr.) 'He did the whole exposition'

56 G. Booij (29)

De motor slaat af (erg.) 'The engine stops' Hij sloeg de uitnodiging af (tr.) 'He turned down the invitation'

Although af may have a directional interpretation (as in afmarcheren 'to march away') the ergativity or transitivity of a Prt-V combination completely depends on its meaning. Since both aflopen and afslaan have (at least) two semantic interpretations, they have two syntactic valencies. A comparable minimal pair is afnemen (erg.) 'to decrease' versus afnemen (tr.) 'to take away'. More generally, particles with a directional interpretation do not always enforce ergativity, as is shown, for example, by the particle uit 'outward'. The following Prt-V combinations are all transitive: (30)

uitbazuinen 'to trumpet' uitbesteden 'to farm out' uitstomen 'to dry-clean' uitvegen 'to wipe out'

Another empirical problem for the SC analysis of Prt-V combinations is that, in many cases, they keep the intransitivity of their verbal constituent. In such cases the SC analysis is impossible, as illustrated by the SCV doorwerken 'to go on working'. Note that werken is an intransitive verb, and that the intransitive doorwerken selects hebben as its auxiliary. (31)

dat ik [? door] sc gewerkt heb that I through worked have = 'that I went on working'

As Hoekstra et al. (1987: 68) point out, the subject of the small clause cannot be PRO since it is a governed position. Hence, it would have to be interpreted as an adjunct small clause, comparable to (32): (32)

dat Jan [PRO bedroefd] zijn brood at that John sad his bread ate = 'that John ate his bread sad'

However, particles cannot be interpreted as the heads of adjunct SC's. Firstly, particles only occur before verbs whereas adjunct SC's can also occur before the subject or the direct object: (33)

Jan at zijn brood op/bedroefd Jan at bedroefd/*op zijn brood Bedroefd/*op at Jan zijn brood

Separable complex verbs in Dutch

57

Secondly, the particle does not predicate anything of the antecedent of PRO. This problem for the SC analysis does not bear an incidental character: a number of particles preserve the syntactic valency (in these examples the intransitivity) of the verbs, e.g., door, omhoog, op, voor, voorop, rond: (34)

door-zeuren 'to go on nagging' omhoog-schreeuwen 'to shout upwards* op-kijken 'to look up' voor-proeven 'to taste beforehand' voorop-fietsen 'to cycle in front' rond-fietsen 'to cycle around'

This problem might seem to be solved if one were not to consider these particles as the heads of small clauses, but as adverbs that have been incorporated into the verb by X°-movement instead. However, as I will argue below, an incorporation analysis is also problematic. To sum up what we have seen so far in this section: an empirical obstacle for an SC analysis of Part-V combinations is that, in many cases, the predicted transitivization effect does not occur, and that the ergativization effect does not follow from available generalizations concerning the change of non-ergative verbs into ergative ones. This suggests a crucial role for the semantic properties of the Prt-V combinations in accounting for their syntactic valency. Another problem for the SC analysis is the following: As shown by Verb (Projection) Raising, Prt-V combinations behave as a unit. Therefore, as Groos (1989) points out, the SC analysis requires that, before Raising applies, the particle (i.e. the predicate of the small clause) is adjoined to the V of the dominating clause, presumably by Head Movement (Baker 1988a: 54), which adjoins a word of the category X° to its proper governor Y°. Groos shows that this consequence of the SC analysis creates new problems: While particles are free to incorporate, i.e., to move along with the verb under V[erb] R[aising], intransitive prepositions are not. Consider the following sentences containing intransitive prepositions: (a) (b)

Jan heeft altijd boven willen wonen 'Jan has always up(stairs) wanted to-live' *Jan heeft altijd willen boven wonen 'Jan has always wanted up(stairs) to-live'

In the ungrammatical sentence above, [(b)], the intransitive preposition has been incorporated into V and moved along with the verb under VR. (Groos 1989: 53)

Consequently, the application of Head Movement has to be restricted by lexical information: which elements can be incorporated (and what the meaning of the resulting SCV is) has to be lexically encoded on the verb. Although this is not

58 G. Booij impossible, it is an unattractive solution given the productivity of the class of SCV's (see below). Moreover, it still does not explain why the rule of Verb Second does not apply to the whole Prt-V combination, but to the verbal part only. Finally, note that the SC analysis does not cover those SCV's that do not form Prt-V combinations or A-V combinations, i.e. SCV's in which the first constituent is a noun, such as paardrijden 'to ride horseback'. An SC analysis would require a structure of the following kind, in which paard is the predicate of the SC: (35)

dat Jan [? paard] sc reed 'that Jan horse rode'

Clearly, such a structure is impossible due to the absence of a proper subject for the SC. In Van Riemsdijk (1978) we find another syntactic approach to SCV's. Van Riemsdijk proposes to consider particles as intransitive pre/postpositions in underlying structure. They are then adjoined to the following verb by a rule of incorporation. This proposal, of course, reminds us of the more general proposal of Baker (1988a) to interpret incorporation as X°-movement to the head. As Groos (1989) correctly points out, such a rule of Incorporation would have to be governed lexically, since incorporation normally is not optional. Compare (36), in which the intransitive preposition is proposed with (37), in which the particle must be incorporated and cannot be preposed: (36)

a. b.

(37)

Ik wil met voor wonen 'I do not want to live in front' V66r wil ik met wonen

a.

Ik wil niet voor sorteren 'I do not want to get in lane' b. *V66r wil ik niet sorteren

The problem is not so much that Incorporation would be lexically governed, in the sense that the lexicon would define which incorporated structures are acceptable (this assumption is also made in Baker 1988a), but that the obligatory application of this rule would not follow from other principles. The words that are subject to incorporation are lexical morphemes, and hence their obligatory incorporation does not follow from their being bound morphemes. An incorporation analysis would also cause problems for SCV's with nouns as first constituents, because it predicts the wrong adjectival form geen instead of the adverbial form niet (see (13)): (38)

DS: dat ik geen stof zuig SS: *dat ik geen ^ [stof zuig]

Separable complex verbs in Dutch (39)

59

dat ik niet [stof zuig] 'that I not vacuum clean'

This shows that stof and zuig should already form a unit at deep structure, in order to receive the right form of the negative element. As we saw above, SCV's are frequently inputs for compounding and derivation. This is also a problem for the incorporation analysis as proposed by Van Riemsdijk (and presupposed in the SC analysis), because the SCV's in such analyses only form a unit at the level of surface structure, whereas there is no evidence that the derivational and compounding rules involved can apply as part of the syntactic derivation. On the contrary: since derivational rules, such as deverbal -er-affixation, affect the argument structure of the input verbs, the Projection Principle requires such derivational processes to take place pre-syntactically. Finally, note that Van Riemsdijk's incorporation analysis does not account for the effects of particle incorporation on the syntactic valency of the verbs. In sum, an analysis is called for that analyzes SCV's as phrasal constructs without creating them by means of syntactic movement rules.

4. A LEXICAL ACCOUNT OF SCV'S

It is a fairly generally accepted assumption that lexicalized syntactic phrases may be listed in the lexicon. However, it is incorrect to use this assumption to account for SCV's. Clearly, not all SCV's are lexicalizations. An account in terms of lexicalization is probably all right for SCV's with nouns and adjectives as their first constituents, such as [goed]A [keuren]v 'to approve' and [adem]N [halen]v 'to breathe', because these SCV's may have arisen through reanalysis of syntactic structure, e.g. dat ik [dat goed] keur 'lit. that I that right judge' —> dat ik dat [goed keur] 'that I approve of that'. Note that the class of SCV's with nominal or adjectival first constituents is not productive. Some particle-verb combinations may also have arisen through lexicalization. On the other hand, the class of SCV's with particles is very productive, and can be extended unintentionally, with transparent meanings. For instance, the particle door can be used with all kinds of action verbs, with the predictable meaning 'to go on V-ing': (40)

doorschrijven 'to go on writing' doorkopen 'to go on buying' doortikken 'to go on typing' doordrinken 'to go on drinking'

60 G. Booij Similarly, the particle mee has the meaning 'to join in V-ing': (41)

meedenken 'to join in thinking' meeschrijven 'to join in writing' meefietsen 'to join in cycling' meeschelden 'to join in cursing'

This suggests that the stock of expressions in the Dutch lexicon can be extended in two ways: by morphological rules that create complex words dominated by an X°category, and by rules that create a specific kind of phrasal construct, namely [[Prt][V]]v*, where V* stands for the node in between V and the projection of V that dominates V and its possible objects. Whereas phrasal structure at the sentence level, in particular the VP constituent, may be seen basically as a projection of the lexical properties of the words, this is not the case for the minimal phrase [Prt V]. For instance, there is nothing in the argument structure of the verb bellen 'to phone' that requires the particle op to be there. This, of course, supports the view that SCV's are accounted for by rules in the lexicon, because lexical rules are optional. Instead of assuming a rule for each particle (in the spirit of Aronoff 1976), one could also provide specific lexical entries for particles, entries that specify both their particular meaning contribution to SCV's, and the fact that they combine with V's (e.g. by [— V]v»), similar to what Lieber (1980) proposed for affixes. Note that the meaning of productive particles such as door is much more specific than their adverbial/ adpositional counterparts. For instance, the particle door either means 'to go on V-ing' or 'to V completely '. As far as the changes in syntactic valency brought about by adding particles to verbs, the following should be remarked. As already pointed out above, the argument structure of a verb should be seen as a projection of its lexical-conceptual structure. Whether the particle affects syntactic valency depends on its meaning. For instance, the particles door and mee (cf. 40-41) affect the lexical-conceptual structure of the verb given their meanings 'go on with V-ing' and 'to join in V-ing', respectively, where V stands for an intransitive verb; but this change in meaning has no effect on argument structure, because the number of participants involved is not changed. Therefore, the same argument structure will be projected, and hence the syntactic valency does not change. Both door and mee require that the verb with which they combine be an intransitive one, as witnessed by the ungrammaticality of sentences such as (42): (42)

a. *dat we de appels dooreten that we the apples through eat = 'that we go on eating apples' b. *dat we de appels meeeten that we the apples with eat = 'that we join in eating apples'

Separable complex verbs in Dutch

61

Therefore, SCV's with door and mee are intransitive. There are also cases, however, in which the syntactic valency does change. Above, we saw the example afstuderen 'to finish one's studies' that denotes a change of state, whereas studeren 'to study' denotes an activity. In Dutch, changes of state are expressed as ergative verbs, and activities as (optionally) transitive verbs. This explains the change in syntactic valency from studeren to afstuderen. The same contrast is found in nemen 'to take' (obi. tr.) versus afnemen 'to decrease' and toenemen 'to increase': it is the idiosyncratic interpretation of the SCV's with nemen that predicts the syntactic ergativity of such verbs. The transitivization effect of certain particles, as discussed in Section 3, follows from the fact that by adding a particle such as op in opbellen 'to phone' or of in aflopen 'to tramp' it is expressed that the action mentioned by the verbs is directed towards an object in such a way that the object is somehow affected by that action. Therefore, the well known generalization that verbs with affected objects are expressed as transitive verbs (cf. Booij & van Haaften 1988) can also be used to explain the transitivization effect of particles such as op and af on some of the verbs with which they occur.

5. CONCLUSIONS

In this paper I have shown that SCV's should not be considered as words, but rather as phrases that are created in the lexicon. Hence, we are not forced to assume that rules such as Verb Raising and Verb Second are allowed to move sub-lexical elements. This analysis, therefore, is in agreement with a specific subclaim of the Lexical Integrity Hypothesis, namely, the claim that sublexical elements cannot be moved out of words. The analysis defended here also implies the rejection of a syntactic incorporation analysis of SCV's, but this, of course, does not allow us to conclude that incorporation analyzes for other phenomena should similarly be rejected. That is, we can only conclude here that it is possible to give an account of SCV's that is in agreement with both the claim that sublexical elements cannot be moved out of words, and the claim that (sub)lexical elements cannot be moved into words. Further investigation is required to determine how the notion 'minimal projection of V', a level in between V and the level that includes objects, can be incorporated in the theory of projection levels in a non-ad hoc manner. For instance, one might ask whether there are languages other than just Dutch and some other Germanic languages, that require such a level of projection, either for V or for another lexical category.7

62

G. Booij

NOTES * Previous versions of this paper were read at a meeting of the Jonge Onderzoekers of the University of Utrecht, at a workshop of the Max Planck Institut fiir Psycholinguistik in Nijmegen, and at the 4th International Morphology Meeting, Veszprdm, Hungary. I would like to thank the audiences at those meetings, the anonymous referees, and the guest editors for their useful comments. 1. Compare the following definition of the Lexicalist Hypothesis in Anderson (1989: 1): "The syntax neither manipulates nor has access to the internal form of words". 2. A consequence of the fact that word formation allows for syntactic inputs is that the Principle of Lexical Integrity does not simply follow from an organization of the grammar in which morphology and syntax are assigned to completely separated components, as was suggested by Di Sciullo & Williams (1987: 46). 3. Exceptions are those phrases which function as names, such as vliegende schotel 'flying saucer' from which the compound nep-vliegende schotel 'fake flying saucer' can be coined (example suggested to me by one of the referees). 4. This restriction holds only for unstressed prefixes. The prefix her- 're-' that bears main stress can combine with SCV's, e.g. her-aan-besteden 'to put out to bid again', her-in-delen 'to rearrange' etc. 5. The only exception in the data at hand (cf. 18) is uithuwelijken since the noun huwelijk is derived from the verb huw(en) 'to marry'. Note, however, that huwelijk has probably lost its morphological and semantic transparency, because -elijk is an unproductive suffix, and the verb huwen is rather obsolete, in contrast to the word huwelijk itself. 6. In the prefix sequence ver-on a t is inserted after on-. 7. One of the referees pointed out to me that a similar notion, the notion 'verboid' is used in Coopmans & Everaeit (1988) for the unit formed by the verb laten 'to let' plus a following verb. The verb laten is similar to particles in SCV's in that it can affect the argument structure of the following verb, but nevertheless exhibits the syntactic behaviour of an independent syntactic unit.

REFERENCES Allen, M.R. 1978. Morphological Investigations. Diss., Univ. of Connecticut. Anderson, S.R. 1989. "Lexicalism and the distribution of reflexives". Yearbook of Morphology 2, 1-19. Aronoff, M. 1976. Word Formation in Generative Grammar. Cambridge Mass.: MIT Press. Baker, M. 1988a. Incorporation. Chicago: Chicago University Press. Baker, M. 1988b. "Review of Di Sciullo and Williams 1987". Yearbook of Morphology 1, 259-283. Booij, G.E. 1977. Dutch Morphology. A Study of Word Formation in Generative Grammar. Dordrecht: Foris Publications. Booij, G.E. 1989. "Morphology, semantics and argument structure". (VU Working Papers in Linguistics #33). Booij, G.E. and T. van Haaften. 1988. "On the external syntax of derived words: evidence from Dutch". Yearbook of Morphology 1, 29-44. Botha, R.P. 1984. Morphological Mechanisms. Lexicalist Analysis of Synthetic Compounding. Oxford etc.: Pergamon Press. Carrier, J. and J. Randall. Ms. From Conceptual Structure to Syntax. Coopmans, P. and M. Everaeit. 1988. "The simplex structure of complex idioms: the morphological status of laten". In M. Everaert et al. (eds.), Morphology and Modularity. Dordrecht: Foris Publications, 75-104. De Rooy-Bronkhorst, A. 1980. "Past participle ge-deletion and the role of stress in Dutch complex verbs". In W. Zonneveld et al. (eds.), Studies in Dutch Phonology. The Hague: Martinus Nijhoff, 159-204.

Separable complex verbs in Dutch

63

Di Sciullo, A.-M. and E. Williams. 1987. On the Definition of Word. Cambridge Mass.: MIT Press. Dressler, W. 1988. "Preferences vs. strict universals in morphology: word-based rales". In M. Hammond and M. Noonan (eds.), Theoretical Morphology. Orlando etc.: Academic Press, 143-154. Evers, A. 1975. The Transformational Cycle in Dutch and German. Ph.D. Diss., Univ. of Utrecht Groos, A. 1989. "Particle-verbs and adjunction". In H. Bennis and A. van Kemenade (eds.), Linguistics in the Netherlands 1989. Dordrecht: Foris Publications, 51-60. Haegeman, L. and H. van Riemsdijk. 1986. "Verb projection raising, scope, and the typology of rules affecting verbs". Linguistic Inquiry 17,417-466. Hoeksema, J. 1988. "Head-types in morpho-syntax". Yearbook of Morphology 1, 123-38. Hoekstra, T. 1984. Transitivity. Dordrecht: Foris Publications. Hoekstra, T. et al. 1987. "Complexe verba". Glot 10, 61-78. Kayne, R. 1984. Principles of Particle Constructions. Ms. Koster, J. 1975. "Dutch as an SOV language". In A. Kraak (ed.), Linguistics in the Netherlands 19721973. Assen: Van Gorcum, 165-177. Lapointe, S. 1980. A Theory of Grammatical Agreement. Doct diss, Univ. of Amherst. Le Roux, C. 1988. On the Interface of Morphology and Syntax. Stellenbosch Papers in Linguistics # 18. Lieber, R. 1980. On the Organization of the Lexicon. Doct. diss., MIT [Published by Indiana Univ. Linguistics Club, 1981]. Lieber, R. 1989. "On percolation". Yearbook of Morphology 2, 95-138. Marie, J. van. 1985. On the Paradigmatic Dimension of Morphological Creativity. Dordrecht: Foris Publications. Riemsdijk, H. van. 1978. A Case Study in Syntactic Markedness. Dordrecht: Foris Publications. Zubizarreta, M.-L. 1987. Levels of Representation in the Lexicon and in the Syntax. Dordrecht Foris Publications.

Yearbook of Morphology 3 (1990), 65-79

Lexical processing of verbs with separable particles* Robert Schreuder

1. INTRODUCTION

The central topic of this paper is the processing of verbs with separable particles. Booij (this volume) gives an overview of the linguistic complexity of these verbs. Some linguistic problems of verbs with separable particles are also discussed by Bierwisch (1987). In the present paper, I will show that these verbs also pose special problems for language processing. I will begin by briefly describing the possible role of morphological structure in lexical access. I will then describe the general method of the experiments to be discussed here, and the results of previous research suggesting that verbs with separable particles have a special 'status' within the mental lexicon. I will present an explanation for these results in terms of connections between lexical structures. The results of two further experiments in which some implications of this proposal are explored will then be discussed. In conclusion, I will suggest that the problem of processing these verbs cannot be solved by the lexicon alone.

2. THE ROLE OF MORPHOLOGICAL STRUCTURE DURING LEXICAL ACCESS

The basic question in the psycholinguistic study of the role of morphological structure in lexical access is the following: does every morphologically complex word have its own access representation? In psycholinguistic models of word recognition, modality-specific access representations are often assumed to mediate between the acoustic, or visual, input and semantic, syntactic, and phonological information about the word stored in the mental lexicon (cf. Frauenfelder & Tyler 1987). Is the input representation of a morphologically complex word mapped onto a single access representation, or several access representations? Consider a word form like redefines. Does this get mapped onto a single access representation, , two access representations, and or even three access representations, , , and ? Do inflectionally-complex forms have

Author's address: FTT/IWTS, University of Nijmegen, Wundtlaan 1, 6525 XD Nijmegen, The Netherlands.

66 Robert Schreuder Researchers generally agree that the lexical representations of morphologically related words must be either shared or connected. The results of, for example, Andrews (1986), Fowler, Napps, and Feldman (1985), or Stanners, Neisser, Hernon, and Hall (1979) can all be interpreted "in terms of cross-activation within the lexicon between words sharing morphological attributes" (Andrews 1986). There is little agreement, however, on how morphological information should be represented in the mental lexicon. One way of representing this information is that discussed above, with decomposed access representations and parsing procedures that allow the components of morphologically complex words to be mapped onto the relevant access representations. Another way of representing morphological information might be to list each morphologically complex form in the lexicon and assume morphologically related forms to be connected in some way. Experiments may show morphological information to be represented in the mental lexicon, but the exact form of the representation still has to be decided. A review of the experimental results relevant to the issue of morphologically decomposed access representations in the mental lexicon falls beyond the scope of the present paper (see Henderson 1985, for an excellent overview). To summarize briefly: Some researchers show evidence for decomposed access representations for derivationally-complex words (e.g., Taft & Forster 1975, Taft 1981, Lima 1987); some researchers show evidence for decomposed access representations for inflectionally-complex words (e.g., Caramazza, Laudanna, & Romani, 1989; Jarvella, Job, Sandstrom, & Schreuder, 1987); other researchers show no such evidence (e.g., Andrews 1986, Monsell 1985); and still others argue against decomposition because of methodological problems (Henderson 1985, Smith 1988). The issue of decomposed access representations seems to be undecided at this moment. Morphology is clearly represented in the mental lexicon in some way, but whether this representation consists of direct decomposed access structures, or becomes available only after lexical access as a result of multiple links between morphologically related words is as yet an unsolved question (Andrews 1986, Henderson 1985). If we assume each access representation to have its own lexical representation in terms of meaning and syntactic categorization, then a fully decomposed mental lexicon would imply an advantage in terms of amount of storage, as compared to a lexicon with a full listing of access representations. Decomposed access representations, with their own store of syntactic and semantic information, also imply compositional rules. Compositional rules are necessary to compute the final meaning and the syntactic category of the complex form. However, complex forms with idiosyncratic meaning aspects cannot be stored decompositionally in such a model, a problem to which we will return. At this point, it should be stressed that the answers to the questions raised might depend upon the language under consideration (Cutler 1985). That is, the organization of morphological representations in the mental lexicon for people speaking Finnish or Turkish may differ from that of people speaking English or Dutch. In fact,

Lexical processing of verbs with separable particles

67

there are empirical results suggesting differences in the morphological organization of the mental lexicons for speakers of different languages (Jarvella et al. 1987). The present paper is a follow-up on Schreuder, Grendel, Poulisse, Roelofs, and van de Voort (1990). Using an experimental technique known to be particularly sensitive to the role of morphological structure during the access phase of word recognition, these authors found an interesting difference in the processing of verbs with bound prefixes compared with verbs with separable particles. These results form the starting point of the present paper and will be discussed in section 4. However, before discussing the results of Schreuder et al. (1990), this experimental technique has to be discussed since the interpretation of their results demands some insight in the way the experiments were set up, and also because the experiments reported here use the same technique. This particular technique avoids many of the pitfalls associated with other techniques (Smith 1988).

3. EXPERIMENTAL TECHNIQUE

In the experiments the following technique was used. Subjects are asked to fixate on a point in the center of a computer screen, which is replaced by a word after 750 milliseconds (ms). The subject is asked to read this word out loud as soon as he or she sees it. When the word is presented, a clock is started, which is then stopped when a voice-activated relay is triggered by the subject starting to pronounce the word. In this way a reaction time (RT) to the word is obtained. This task is typically called a naming task. In certain conditions, a number of letters in the word are 'previewed' for 60 ms. The rest of the word becomes visible after 60 ms, and the timer is started when the full word is visible. The part of the word being previewed is typically referred to as the prime. In non-priming trials of the experiment, the word is presented all at once and the timer therefore also starts at once. That is, timing starts in both cases when the full word is visible. The RT in the priming condition is subtracted from the RT in the non-priming condition, giving a measure of the facilitatory or priming effect of the prime. This method, called the short Stimulus Onset Asynchrony (SOA) priming task, has shown promising results in studying the role of morphological structure during the early stages of lexical processing (Jarvella et al. 1987, Schreuder et al. 1990). But why would such a technique be particularly sensitive to morphological structure? As an example I will use the complex form betonen ('to show' or 'to manifest'), which consists of the verb stem tonen, and the prefix be. In the priming condition we present either the stem or the prefix for 60 ms and then add the respective prefix or stem. We compare the RT obtained in the priming condition with the RT obtained when the word is presented all at once. If independent access representations exist for and , then presentation of either segment should speed the naming of betonen relative to simultaneous presentation. That is,

68 Robert Schreuder a morphologically complex word is presumably decomposed into its constituents, and by presenting stem and prefix asynchronously in time we will facilitate the mapping of these constituents onto their respective access representations and . This benefit should show up as a shorter naming time in the case of a priming condition relative to when a word is shown all at once. How can we be sure, however, that this facilitation is the result of the morphological structure of the access entries in the mental lexicon and not simply a result of the previewing of a few letters? There is ample evidence, for instance, for orthographic priming (see Forster 1989). In order to rule out this possibility, a monomorphemic control word was sought for each of the word forms in these experiments - for betonen in this case the word begonia (same as English 'begonia'). In this way we compare the priming effect of the morpheme be on the complex verb form betonen with the priming effect of the pseudo morpheme be on the monomorphemic word begonia. Similarly, we compare priming by tonen on betonen with the priming of gonia (a pseudo word) on begonia. If significantly more priming is obtained when the prime is a morpheme in the word to be named than in the case where the prime is not a morpheme in the word to be named, we can conclude that the priming is, in actual fact, not just the result of the orthographic overlap between prime and target. We will call this genuine morphological priming and argue that this is consistent with a model assuming decomposed access entries for such word forms.

4. RESULTS OF THE SCHREUDER ET AL. (1990) STUDY

The particular word forms investigated in this study were inflectionally suffixed verb forms (present tense and past tense), past participle verb forms, verbs with inseparable prefixes, and verbs with separable particles. The results for verbs with separable particles and verbs with bound prefixes were found to differ from each other1 and are shown in Table 1. Table 1. Summary of the results of Schreuder et al. (1990f Control

Experimental NO-PRIM

Ml

M2

NO-PRIM

CI

C2

Inseparable prefix (infinitive)

(betonen) 538

(be) 515 [23]

(tonen) 512 [26]

(begonia) 534

(be) 508 [26]

(gonia) 512 [22]

Separable particle (infinitive)

(opzien) 505

(op) 479 [26]

(zien) 467 [38]

(optie) 510

(op) 494 [16]

(tie) 499 [11]

' Mean naming latencies (ms) for the experimental and control words without a prime (NO-PRIM) or primed by either their first morpheme (Ml) or the second morpheme (M2), or primed by a nonmorphemic string in the control conditions (CI and C2, respectively). Numbers in square brackets indicate the amount of priming obtained, in ms.

Lexical processing of verbs with separable particles

69

The statistics showed a significant genuine morphological priming effect only in the case of verbs with separable particles. That is, the 23 ms benefit of priming verbs like betonen with their prefix be does not significantly differ from the priming effect of 26 ms obtained by priming the monomorphemic control word begonia with be. Similarly, the 26 ms benefit of showing the verb stem tonen does not differ significantly from the 22 ms obtained when priming begonia with the control string gonia. For the verbs with separable particles, the same type of comparison showed significant differences; that is, the 26 ms priming obtained with a particle was found to be significantly larger than the 16 ms priming obtained with the control, and the 38 ms priming obtained with a verb stem is significantly larger than the 11 ms obtained with the control. The conclusion from these results is that the mental lexicon has a full listing of verbs with inseparable prefixes and decomposed entries for verbs with separable particles, one for the particle and one for the verb. However, this conclusion tells us little that we didn't already know. The particles in these verbs are free morphemes, just as their verb stems. In other words, the verbs and the particles must have their own access representations. More interesting, however, is the question whether a separable particle verb like opzien (to look up, to fear) is mapped onto the access entries for op and zien in the mental lexicon. Or must we postulate a separate entry for the complex form opzien in the mental lexicon? However, as said before, wherever we have separate entries, the semantics of their combination should be compositional.

5. COMPOSITIONALITY

As others have argued before (e.g., Baayen 1986, de Vries 1985), verbs with separable particles often have idiosyncratic meaning aspects. Consider for instance the simple verb vallen (to fall) as part of separable verbs such as aanvallen (to attack), afvallen (to lose weight) and opvallen (to attract attention). Complex verbs such as these certainly are not compositional. This poses an interesting problem for the human language processing system, as I will argue. The particle and the verb in these combinations can be separated within a sentence (see Booij, this volume). The particle can come first - Dat AAN te willen VALLEN lijkt mij gevaarlijk (wanting to attack that seems dangerous to me) - or second - zij VIELEN ondanks alle tegenslagen opnieuw AAN (They attacked again despite all the setbacks). The problem for the language processing system is the following. At time t, the verb vallen, for example, is accessed and its meaning will be retrieved. The system cannot know at that moment, however, whether a particle will occur later in the sentence. At moment t+x, a particle is accessed. Its meaning will be accessed and perhaps the parser will tell the lexicon that the verb and particle have to be combined, a point to which I will return. In the case of non-compositional verbs this simply isn't enough. In this case, it seems, lexical access must somehow be repeated, to retrieve

70 Robert Schreuder the non-predictable meaning of the combination. These complex verbs then, constitute an exception to the general process of word recognition in that their lexical access must be distributed over time. Notice that the amount of intervening linguistic material is, in principle, of unbounded length. How does the lexicon know with which earlier occurring word form to combine the present form? How does the lexicon detect that the two forms should be combined? Finally, how does the lexicon find the meaning representation for the combined word forms?



h

K

LEX. REPRESENTATIONS bevallen

Figure 1. Access representations for inflected forms of the verb afvallen, the particle af, the verb vallen, and the verb bevallen. Node / integrates activation over time. When both af and one of the inflection forms of vallen are encountered, node I will start to activate the lexical representations belonging to afvallen and inhibit the lexical representations belonging to vallen and af. In this simplified picture, links to morphosyntactic information (e.g., related to tenses), have been omitted.

Lexical processing of verbs with separable particles

71

6. CONNECTIONS IN THE LEXICON

In our previous work (Schreuder et al. 1990) we proposed the following: access representations of simple verbs and those of free-occurring particles, like prepositions and adverbs, are connected with the access representations of the complex verbs that contain their combination. When such a simple verb or a particle is accessed, activation will spread to the access representations of the complex verbs to which they are connected. When, separated in time, both are accessed, their combined influence on the access representation should be enough to trigger it, and so make available the idiosyncratic information that is unique for THIS particular combination of verb and particle.

These connections, in other words, explain why morphological priming only occurred in the experiment for verbs with separable particles. Figure 1 shows a sligthly modified version of this account. A weak point in the model of Schreuder et al. (1990) was the assumption that connections exist between access representations. The particle and verb can occur widely separated in space and time, which requires that the access representations for these word forms be activated for a relatively long period of time - an assumption that runs counter to all activation models of word recognition, where rather rapid decay is always assumed to be the case. The starting assumption for the model presented in Figure 1 is that the access representations for neither derivationally complex word forms, nor inflectionally complex word forms are decomposed. The access representations of derivationally complex verb forms and compound verb forms are directly linked to their lexical representations (containing the relevant semantic, syntactic, and phonological information). Inflected forms of simple verbs and particles are, however, also connected to an 'integration' node whenever they can combine together into a complex verb with a separable particle. When an inflected simple verb and a particle with which it can combine are both encountered within a sentence, the integration node is triggered and, in turn, the lexical representation of the complex verb. That is, the integration node needs two inputs before it starts to activate other representations. When the integration node is triggered, the lexical representations of the simple verb and of the particle are inhibited. Integration nodes, moreover, are assumed to decay much slower than access representations. Notice that this model implies that each complex verb with a separable particle needs its own integration node and, therefore, we call this model the Morphological Integration Model. A further assumption is that presenting the particle and verb stem with a very short SOA, as in the experiments of Schreuder et al. (1990), results in activation of access representations of the particle, the verb stem, but also in activation of the access representation of the complex verb form. The lexical representation of the complex verb form is thus activated twice - both by the integration node and by the access representation of the complex verb form. This assumption is necessary to be

72 Robert Schreuder able to predict the genuine morphological priming effects obtained for these verbs in contrast to verbs with bound prefixes. This assumption of the model still needs independent testing, since the model is post hoc. A testable prediction of the Morphological Integration (MI) model is the following. When the verb stem of a separable verb has been accessed, and the particle not yet encountered, activation of the lexical representation of the complex verb containing this verb stem should not occur. After the access of vallen, the lexical representations of aanvallen or afvallen should not be activated. Similarly, when aan is encountered first, the lexical representations of complex verbs containing the form aan should not be activated. This prediction has yet to be tested. A further prediction of the MI model is that simple verbs such as vallen (to fall) should not prime morphologically complex verbs with an inseparable prefix like bevallen (to please). Though bevallen is non-compositional, it needs no connections because it always occurs directly together with its bound prefix. That is, vallen never needs to be accessed and only later associated with be to come up with bevallen (or vice versa, be followed by vallen). The MI model bears some resemblance to Bybee (1988), who assumed connections between entries that are semantically related and entries that are phonologically related. Where semantic and phonological connections coincide in her model, morphological identity can be established. Notice that in our proposal links are also often established between entries that have little or no semantics in common (e.g., between vallen en afvallen).

7. BOUND PREFIXES VS. SEPARABLE PARTICLES

The above account leads to the following empirical prediction. When we measure the priming effect of vallen on afvallen, we predict that it will be larger than the priming effect of vallen on bevallen. We prime with exactly the same string, a simple verb, in both cases, but only in the case of a separable particle verb, however, are connections assumed to exist and activation of these connections therefore expected to produce priming. Recently, an experiment was carried out in which such words were directly compared. In this work we compared directly the priming effect of a simple verb on the naming of a complex verb with a separable particle with the priming effect of the same simple verb on a complex verb with an inseparable prefix. That is, we compared the priming effect on aanvallen (to attack) with the priming on bevallen (to please) of the simple verb vallen (to fall). Similarly, we compared the effect of aan on aanvallen with that of be on bevallen. The general procedure of this experiment was quite similar to that employed in the previous experiments, and 48 of such pairs were used. The details of this experiment and the following can be found in Schreuder (forthcoming), and the results of the present experiment can be found in Table 2.

Lexical processing of verbs with separable particles

73

Table 2. Bound prefixes vs. ¡inseparable particlesf BOUND PREFIX (bevallen)

SEPARABLE PARTICLE (aanvallen) 500

NO-PRIMING

538

VERB PRIMING

519

[19]

472

[28]

PREFIX OR PARTICLE PRIMING

516

[22]

463

[37]

' Mean naming latencies (in ms) for the three conditions (NO-PRIMING, VERB PRIMING, PREFIX or PARTICLE PRIMING, for the two different types of verbs. The amount of priming obtained is shown in brackets.

The amount of facilitation caused by verb-stem priming was significantly larger for complex verbs with a separable particle than for complex verbs with an inseparable prefix (37 ms vs. 22 ms). Also, the priming effect for a separable particle was significantly larger than the priming effect for an inseparable prefix (28 ms vs. 19 ms). It should be noted, however, that in the latter case the primes were different (aan vs. be) and that the latter results must, therefore, be interpreted with caution. These results suggest - in a more rigorous way than the preceding findings, and in keeping within the MI model - that connections between access representations and the integration node (representing the combination of these representations) are present for complex verbs with separable particles and absent for verbs with inseparable prefixes.

8. THE ROLE OF COMPOSITIONALITY

It might be necessary to qualify the conclusion of the last paragraph. There are verbs with separable particles that seem to be perfectly compositional. Verbs with separable particles such as doorsnijden (to cut-through), meenemen (to take along), and omdraaien (to turn around) are clearly compositional. One could argue that connections may not be necessary for such compositional verbs - the meaning can clearly be computed once the respective parts have been encountered. A second experiment was undertaken, therefore, to determine whether separate access processes might be used for compositional versus non-compositional verb complexes. This experiment involved the same verb stem combined with different particles, the result being compositional in one case and non-compositional in the other case. Examples are: leiden (to lead) in wegleiden (to lead away) vs. opleiden (to educate); brengen (to bring) in meebrengen (to bring along) vs. ombrengen (to kill); keren (to turn) in omkeren (to turn around) vs. uitkeren (to pay out). If it is true that only semantically non-compositional verbs require links between their complex forms and respective morphemes, then a non-compositional verb should show a larger priming effect when primed with a verb stem, than a compositional verb primed with the same stem. Sixty subjects participated in an experiment employing

74 Robert Schreuder exactly the same procedure as the previous experiment. The results are presented in Table 3. Table 3. The Role of Compositionaiirf COMPOSITIONAL VERBS NON-COMPOSITIONAL VERBS (omkeren) (uitkeren) NO-PRIMING

507

VERB PRIMING

488

500 [19]

483

[17]

* Mean naming latencies (in ms) for compositional and non-compositional verbs in no-priming and verbpriming conditions. The amount of priming obtained is shown in brackets.

The data analyses showed no significant differences between the priming effects found in the two cases (17 ms vs. 19 ms). The same verb stem equally facilitated the non-compositional combination and the compositional one. This result suggests that connections exist within the lexicon for all separable verbs, regardless of whether they are compositional or not. Why would this be the case? One reason might be that most separable verbs are non-compositional. Another reason might be that many verbs have both a compositional and a non-compositional reading. The existence of a single noncompositional reading, however, necessitates postulating connections for that particular verb in the lexicon (e.g. meegeven has a compositional reading (to give) and a non-compositional one (to yield). To summarize: the results of the experiments presented here, together with the results of Schreuder et al. (1990), suggest that verbs with separable particles (be they compositional or not) have special connections in the mental lexicon. That is, the access representations of particles, and the access representations of the inflected forms of the verb stems, are indirectly connected to the lexical representation of their combination. In the model presented in Figure 1, these connections are mediated by a so-called integration node. In the following, I will argue that the MI model is a step towards understanding how the lexical representation for complex verbs is accessed, but that there are still many problems to be solved. In the preceding, for example, I assumed that as soon as the second part of a complex verb was accessed - be it the verb or the particle - the meaning of the complex verb would also be retrieved. As will be seen, however, the connections hypothesized in the MI model appear to be necessary, but not sufficient for the retrieval of the meaning of a complex verb. It should be noted that the following arguments are rather speculative and that the results of the experiments presented in the previous section do not bear directly on the issues to be discussed.

Lexical processing of verbs with separable particles

75

9. PROBLEMS FOR THE MENTAL LEXICON

The mental lexicon is that device which, when given an access code, provides information about the meaning and syntactic properties of that code. Consider the sentence fragment in (1): (1)

Ikzouop... 'I might on*

The mental lexicon has no way of knowing whether op is a preposition at this point, or whether it is a particle of a complex verb. Fragment (1) could, for instance, be continued as either (2) or (3): (2)

Ik zou op de vloer kunnen vallen 'I might fall on the floor'

(3)

Ik zou op de vloer op kunnen vallen 'I might attract attention on the floor'

In both (2) and (3), the mental lexicon is confronted with not only op but also vallen. According to the MI model presented in Figure 1, both sentences (2) and (3) will activate the lexical representation of the complex verb 'opvallen'. For sentence (2), however, this leads to the wrong interpretation. At the moment that op is accessed, integration nodes connected to op are activated because it might be necessary to combine op with a verb stem at some point later in the sentence. In the case of (1), it is unclear whether op will be a preposition (2) or a particle (3). It is clearly the syntactic structure of the sentence that determines whether op is a particle or a preposition. In sentences (2) and (3), for example, the parser can only indicate whether op is to be considered a preposition (2) or a particle (3) after confronting vallen. Sometimes this decision can only be made quite late in a sentence or after the start of the next sentence. That is, the correct lexical representation may depend on more than just the verb plus particle/preposition. Consider the following fragment: (4)

Zij kwamen om ... 'They died ...' or 'They came to ...'

This fragment could be continued as (5) or (6): (5)

Zij kwamen om de verwarming te repareren 'They came to repair the heating system'

76 Robert Schreuder (6)

Zij kwamen om in deze verschrikkelijke kou 'They died in this terrible cold'

In these cases, both kwamen and om have already been accessed, but the syntactic information that determines whether a particular case is an instance of a separable verb or not is only available much later. This is even more clear in (7) below, where the change of syntactic structure actually changes the status of verb plus article combination found in (6): (7)

Zij kwamen om in deze verschrikkelijke kou de verwarming te repareren "They came in this terrible cold to repair the heating system*

It appears, thus, that the task of the lexicon is to keep the idiosyncratic meaning of the combination of the particle plus verb accessible while the syntactic processor determines the exact status of the various constituents and whether the idiosyncratic meaning is needed or not. It seems unreasonable to attribute this syntactic disambiguation to the mental lexicon, in part because the information resulting from the disambiguation becomes available after the individual word forms have already been accessed by the mental lexicon. Such an assumption, moreover, would entail that the mental lexicon itself parse the incoming sentence, which is not a very sound idea - to say the least. It might also be argued that the semantic context of the sentence helps in many cases to determine the right interpretation. Such an assumption, however, also leads to problems. Consider examples (8), (9), and (10). (8)

Ik breng de man om geen tijd te verliezen naar huis 'In order not to lose time I take the man home'

(9)

Ik breng de man om met mijn geweer2 'I kill the man with my gun'

(10)

Ik breng de man om mijn geweer te laten zien ... 'I take the man in order to show my gun ...'

The meaning of 'gun' is clearly related to the meaning of ombrengen (to kill). When confronting the combination in the context of guns, therefore, it might be tempting to take om as a particle - as in (9). As can be seen in (10), however, this is not always the case. In this example, the syntactic information clearly indicates that breng and om do not have a verb-particle relation. These examples suggest that the special problems for the language processing system raised by verbs with separable particles can only be resolved through the combined efforts of both the lexicon and the syntactic processing system. The details of this interaction are a matter for further research, not only because they shed light

Lexical processing of verbs with separable particles

77

on the processing of particle verbs (e.g., how syntactic processing can be interfaced with the MI model of lexical access), but also because they may shed light on the processing of idiomatic expressions in general. Non-compositional verbs with separable particles in languages such as Dutch, German, or Afrikaans are often argued to be similar to verbal idioms (e.g., Baayen 1986). Most of the psycholinguistic research in this area, however, has dealt with the manner in which the meaning of the idiom gets stored and whether this meaning is retrieved directly, or only after literal interpretation is attempted (e.g., Gibbs & Gonzales 1985). The problem of the access of such expressions, however, is similar to that of separable particle verbs, as is illustrated in (11), (12), and (13). (11)

Hij viel ondanks alle voorzorgen toch door de mand (idiom) 'Despite all precautions he made a poor show'

(12)

Hij viel ondanks alle voorzorgen toch door het gat (compositional) 'Despite all precautions he fell through the hole'

(13)

Hij viel ondanks alle voorzorgen toch niet af (non-compositional) 'Despite all precautions he did not lose weight'

An idiomatic expression such as that in (11) also poses a problem for lexical access similar to that for verbs with separable particles. Example (11) shows that the parts of the idiom can be separated by a number of intervening linguistic elements. That is, access to the mental lexicon must presumably take place at two different points in time and similar mechanisms can therefore be assumed to be at work in these cases, as well as in the case of particle verbs.

10. CONCLUSION

To summarize, the results of Schreuder et al. (1990) together with the results presented here indicate that the access representations of the constituents of verbs with separable particles are connected in some way to the lexical representation of these complex verbs. The MI model, as introduced in the present paper, also suggests that so-called integration nodes might be responsible for the triggering of the lexical representation of the complex verb. The elements of a verb plus particle combination may be confronted at different points in time, making the often idiosyncratic noncombinatorial meaning of such a structure difficult to access. This problem, however, can be dealt with by assuming the MI model of lexical access with specific integration nodes for specific verb-particle combinations. A different problem is to determine whether a particular sequence of forms should be taken as an instance of a verb plus article or not. It seems that this decision is beyond the capabilities of the mental lexicon because the necessary information often becomes available only after

78 Robert Schreuder the constituent morphological parts have been accessed. The decision, moreover, often entails syntactic disambiguation at a relatively global level. The MI model, in other words, is in need of further testing and should be supplemented with an interface to a syntactic parsing mechanism in the future.

NOTES * Special thanks are due to Lee Ann Weeks for her assistance in transforming earlier versions of this paper into the one presented here. Any remaining errors or inconsistencies are, of course, my own. 1. In other parts of the Schreuder et al. research (1990), various forms of inflectionally complex verb forms were examined. The results showed no differences in the priming of inflectionally complex verbs versus control verbs. Schreuder et al. (1990) conclude on the basis of these results that inflectionally complex verbs have their own access representations. 2. In contrast with the English translation of example 9, the Dutch sentence is unambiguous. The phrase met mijn geweer (with my gun) can only be interpreted as the instrument of ombrengen (to kill), (cf. Rayner, Carlson, & Frazier, 1983).

REFERENCES Andrews, S. 1986. "Morphological influences on lexical access: Lexical or nonlexical effects?". Journal of Memory and Language 25, 726-740. Baayen, H. 1986. "Syntactic word formation: A study on the interaction of lexicon and syntax". Free University of Amsterdam Working Papers in Linguistics no.22. Bierwisch, M. 1987. "A structural paradox in lexical knowledge". In E. van der Meer and J. Hoffmann (eds.), Knowledge Aided Information Processing. Amsterdam: Elsevier Science Publishers, 141-172. Bybee, J.L. 1988. "Morphology as lexical organization". In M. Hammond and M. Noonan (eds.), Theoretical Morphology: Approaches in Modern Linguistics. London: Academic Press, 119-141. Caramazza, A., A. Laudanna, and C. Romani. 1988. "Lexical access and inflectional morphology". Cognition 28, 297-332. Cutler, A. 1985. "Cross-language psycholinguistics". Linguistics 23, 659-667. Forster, K. 1989. "Basic issues in lexical processes". In W. Marslen-Wilson (ed.), Lexical Representation and Process. Cambridge MA.: The MIT Press, 75-107. Fowler, C.A., S.E. Napps, and L. Feldman. 1985. "Relations among regular and irregular morphologically related words in the lexicon as revealed by repetition priming". Memory & Cognition 13, 241-255. Frauenfelder, U.H. and L.K. Tyler. 1987. "The process of spoken word recognition". Cognition 25,1-20. Gibbs, R.W. and G.P. Gonzales. 1985. "Syntactic frozenness in processing and remembering idioms". Cognition 20, 243-259. Henderson, L. 1985, 'Towards a psychology of morphemes". In A.W. Ellis (ed.), Progress in the Psychology of Language, Vol.1. London: Lawrence Erlbaum Associates, 15-72. Jarvella, RJ., R. Job, G. Sandström, and R. Schreuder. 1987. "Morphological constraints on word recognition". In A. Allport, D.G. MacKay, W. Prinz, and E. Scheerer (eds.), Language Perception and Production: Relationships between Listening, Speaking, Reading, and Writing. London: Academic Press, 245-265. Lima, S.D. 1987. "Morphological analysis in sentence reading". Journal of Memory and Language 26, 84-99.

Lexical processing of verbs with separable particles

79

Monsell, S. 1985. "Repetition and the lexicon". In A.W. Ellis (ed.), Progress in the Psychology of Language, Vol.2. London: Lawîence Erlbaum Associates, 147-195. Rayner, K„ M. Carlson, and L. Frazier. 1983. "The interaction of syntax and semantics during sentence processing: Eye movements in semantically biased sentences". Journal of Verbal Learning and Verbal Behavior 22, 358-374. Schreuder, R., M. Grendel, N. Poulisse, A. Roelofs, and M. van de Vooit. 1990. "Lexical processing, morphological complexity and reading". In D.A. Balota, G.B. Flores d'Arcais, and K.Rayner (eds.), Comprehension Processes in Reading. Hillsdale, NJ: Lawrence Erlbaum Associates, 125-141. Smith, P.T. 1988. "How to conduct experiments with morphologically complex words". Linguistics 26, 699-714. Stanners, R.F., J J . Neisser, WP. Hemon, and R. Hall. 1979. "Memory representation for morphologically related words". Journal of Verbal Learning and Verbal Behavior 18, 399-412. Taft, M. 1981. "Prefix stripping revisited". Journal of Verbal Learning and Verbal Behavior 20,289-297. Taft, M., and K. Forster. 1975. "Lexical storage and retrieval of prefixed words". Journal of Verbal Learning and Verbal Behavior 14, 638-647. Vries, J.W. de 1975. Lexicale Morfologie van het Werkwoord in Modern Nederlands. Leiden: Universitaire Pers.

Yearbook of Morphology 3 (1990), 81-85

Comments on the paper by Schreuder Pienie Zwitserlood

1. INTRODUCTION

The question of how morphologically complex words are processed and represented has been the subject of quite some research in psycholinguistics during the past two decades. Models representing both ends of a continuum have been put forward; some advocate that words, when heard or read, are obligatorily decomposed into their (potential) morphemes, even if this segmentation produces the wrong analysis (Taft 1985, Taft & Forster 1975). This will be the case with pseudo-prefixed words: 'lime' is certainly not the root of the word 'sublime'. The other extreme view states that polymorphemic words are processed in essentially the same way as monomorphemic words (Tyler, Marslen-Wilson, Rentoul & Hanney 1988). Lexical access - the mapping of the auditory or visual input onto form (or access) representations in the mental lexicon - does not involve decomposition, and this processing view is complemented by a model of the lexicon containing only undecomposed surface forms (Butterworth 1983). As is common in science, experimental findings seldom unequivocally support extreme positions, and for the issue of morphology, numerous models of a more hybrid nature have been proposed (Jackendoff 1975, Manelis & Tharp 1977, Lukatela, Gligorijevic, Kostic, & Turvey 1980). In his contribution to this volume, Schreuder proposes his own view on the role of morphology and the lexicon. His Morphological Integration model has separate access representations for each surface form (valt, vallen, viel, etc.), and no decomposition of morphologically complex words. These access representations are connected to what Schreuder calls 'lexical representations', containing semantic and syntactic information. There is no morphological information coded in any of the representations in Schreuder's model. The access representation for unspeakingly, for example, does not know that it is made up of un, speak, ing, and ly, nor is its lexical representation morphologically structured (although it is at this level of representation that the information about form class must be specified). For the recognition of morphologically complex words, Schreuder argues, morphological coding at the level of access representations is not necessary, and his data show no evidence for decomposition during lexical access. An exception are the particle verbs in languages

Author's address: Max-Planck-Institute for Psycholinguistics, Wundtlaan 1, 6525 XD Nijmegen, The Netherlands.

82 Pienie Zwitserlood such as Dutch, German, Frisian, and Afrikaans, since verb and particle can become separated in surface structure. In the Dutch sentence Zij komen op afgrijselijke wijze om (They die horribly), the particle verb is omkomen, and the particle om is separated from the verb form komen. Schreuder's data show that these verbs behave differently. The recognition problem that the separability of these complex verbs poses is solved in his model not at the level of form or lexical representations, but rather through an additional 'integration node'. This node is connected to the access representations of the particle (e.g., aan), and of the simple verb (e.g., all surface forms of vallen). Moreover, the complex verb (e.g., aanvallen, aanviel, etc.) has its own set of access representations. The integration nodes only exist for separable complex verbs; bevallen, a prefixed verb, does not have an integration node. By implication, whereas the processing of a separable particle verb aanvallen can be affected by properties of the access node vallen, this is not the case for the unseparable prefix verb bevallen. Thus, in Schreuder's model, word-internal structure is indirectly represented, and for separable verbs only, through the existence of an integration node which collects evidence from subsidiary access representations. In what follows, I will confront Schreuder's proposal with several problems, the first concerning empirical evidence that seems incompatible with his model.

2. MORE DATA FROM DUTCH

In a number of experiments reported in Schriefers, Zwitserlood, and Roelofs (in press), we investigated the processing (in the auditory modality) and the representation of simple verbs, prefixed verbs, and particle verbs such as staan (stand), bestaan (exist), and toestaan (allow). A very persistent finding, obtained with different sets of material and two experimental tasks, is that complex verbs prefixed and particle verbs alike - are always recognized earlier in a gating task and responded to faster in a phoneme-monitoring task than their corresponding simple verbs. This finding is not in accordance with a decomposition view of lexical processing, which predicts no difference between the three verb types (for detailed arguments, see Schriefers et al.). Moreover, the results are not compatible with models advocating the undecomposed mapping of the sensory input onto full-form representations in the lexicon. One such model (Marslen-Wilson 1987) predicts that words - monomorphemic and polymorphemic words alike - will be recognised at their 'uniqueness point', the point in the sensory signal at which they diverge from other form representations in the mental lexicon. For a simple verb/particle verb pair such as staan and opstaan, this is at the /n/; the information about the /n/ separates the simple verb from other representations such as staal (steel), stapelen (stack), or staken (strike); the complex verb separates at the same /n/ from competing form representations such as opstapelen. So, these models predict no difference between complex and simple verbs in the amount of sensory input needed for identification

Comments on the paper by Schreuder

83

(gating task), or in the time needed to detect a pre-specified phonemic segment (phoneme-monitoring task). Our data, however, show a consistent processing advantage of complex words over corresponding simple verbs with the same uniqueness point. This effect obtained for particle verbs and prefixed verbs alike. As potentially confounding factors such as the frequency of occurrence of the full word forms in the language, the density of the lexical environment, and word length were controlled for, the only possible explanation for the processing advantage of morphologically complex words is in terms of a sensitivity of the complex word to the frequency of its composite simple verb. The frequency of occurrence in the language of a particular word is specified in the mental lexicon, and its effects are taken to have an early locus during word processing (Bradley & Forster 1987, Zwitserlood 1989). The fact that the frequency of a simple verb form determines the processing of verbs composed of that simple verb and a particle or a prefix leads us to believe morphological information has to be represented somewhere in the lexicon. How can Schreuder's model handle these results? The model has no problems in dealing with the particle verbs. Since the 'integration node' is connected not only to the access representation of the particle, but also to that of the simple verb, effects of lexical properties of this simple verb could, in principle, percolate to the integration node. One possibility would be that high-frequency simple verbs send more activation to the integration node they connect with than low-frequency items. Schreuder's model, however, cannot deal with the same type of effect for prefixed verbs. These have access representations directly connected to their lexical representations; effects of properties of the simple verb cannot be accounted for. Is it possible that the effects we obtained in our experiments reflect a different level of representation than the results reported by Schreuder in this volume? Schreuder's previewing task, as well as our phoneme-monitoring task, are supposed to tap into early phases of lexical processing. Schreuder's data showed a sensitivity to the complexity of particle verbs only. No effects were obtained for prefixed verbs. In our study, however, effects were obtained for both. A crucial difference between the two studies is the modality of stimulus presentation: visual stimuli in Schreuder's experiments, auditory stimuli in our experiments. Since the nature of stimulation is so different, it is obvious that different access representations are needed for spoken and written words. Could it be that morphological information is coded in auditory form representations, or in connections between them, but not in visual access representations? Although there is no independent evidence or motivation for this, this possibility cannot be ruled out.

3. SCHREUDER'S MODEL AND SYNTACTIC AND SEMANTIC ANALYSIS

Along more speculative lines, I see another set of problems for Schreuder's model of the lexicon, concerning the relationship between lexical information, syntactic

84 Pienie Zwitserlood parsing, and the construction of a semantic representation of the utterance. As Schreuder states, the mental lexicon alone cannot solve the problem of structural ambiguities at the surface level of the sentence. To decide which of two verbs is encountered (particle verb or simple verb), or whether an element like om is a particle or a preposition, the parser needs to be involved. Let us consider a slight variant of one of the examples given by Schreuder: (1)

Ze kwamen in deze verschrikkelijke kou tijdens de hevigste sneeuwstorm van de maand Januari om. 'They died in this extreme cold during the heaviest snowstorm of the month of January'

(2)

Ze kwamen in deze verschrikkelijke kou tijdens de hevigste sneeuwstorm van de maand Januari om de verwarming te repareren. 'They came in this extreme cold during the heaviest snowstorm of the month of January to repair the heating system*

At different points during the analysis of this sentence, Schreuder's version of the lexicon makes available the semantic information and the argument structure of two verbs: the simple verb and the particle verb. In Schreuder's model, the occurrence of kwamen, at an early point in the sentence, triggers the semantic and syntactic aspects of the simple verb. The parser can do one of two things: start computing an analysis on the basis of the argument structure of the simple verb, or wait. Only much later, the co-occurence of the particle om and the verb kwamen triggers the integration node to release its activation towards the lexical representation omkomen. It might seem that a parser with the potential of parallel computing now has two alternative possibilities: simple verb construction or particle verb construction. But this is not true in Schreuder's model: as soon as the integration node for the particle verb is triggered, the lexical node of the simple verb is blocked. Therefore, the only possible parse is in terms of the argument structure of the particle verb. If the sentence ends after om (sentence (1)), everything is fine. When, however, sentence (2) is presented, the occurrence of the NP de verwarming (heating system) renders this analysis of kwamen ... om in terms of the ergative particle verb invalid. The parser then has to backtrack, and for the appropriate re-analysis, it needs the lexical information about the structural properties of the simple verb which have been inhibited. One can easily see that a parser without a waiting strategy, committing itself to an early simple-verb analysis, has to backtrack twice. The problem with Schreuder's model is that it does not provide the parser with the relevant information. Moreover, it prohibits parallel computing: there is no information early in the sentence that a particle verb construction might be on its way, and as soon as a potential particle has triggered the integration node, the possibility of the simple verb construction is lost.

Comments on the paper by Schreuder

85

A related point concerns the availability of semantic information contained in the lexicon. In normal sentence processing, the semantic information contained in lexical representations is rapidly integrated into higher-level representations specifying the meaning of the utterance. With activation and inhibition of the lexical representations containing the crucial semantic specifications, this information will not be available at the right time for meaning integration. So, having integration nodes might seem an elegant solution for particle verbs, but it does not solve the parsing problem discussed above, nor does it help much to ensure the availability of semantic information for higher-level integration. Might it not be a better solution to have morphological information coded in the lexical representations, and to have access representations of simple verbs connected to lexical representations of particle verbs, without inhibition? Such a model could handle the Schriefers et al. data, and could also provide the parser with the option of more than one potential parse. Moreover, when the semantic information for both simple and particle verbs is simultaneously available, semantic integration can take place as soon as structural ambiguities have been resolved. Finally, it is intuitively obvious that the language user must have morphological information represented somewhere, otherwise he or she would not be able to produce and understand syllepses such as: Hier zet men koffie, over en af. (koffie zetten = make coffee; overzetten = ferry across; afzetten = cheat)

REFERENCES Bradley, D.C., and K.I. Forster. 1987. "A reader's view of listening". Cognition 25, 103-134. Butterworth, B. 1983. "Lexical representation". In B. Butterworth (ed.). Language production Vol. 2. London: Academic Press. Jackendoff, R. 1975. "Morphological and semantic regularities in the lexicon". Language 51, 639-671. Lukatela, G., B. Gligorijevic, A. Kostic and M.T. Turvey. 1980. "Representation of inflected nouns in the internal lexicon". Memory and Cognition 8 (5), 415-423. Manelis, L. and D.A. Tharp. 1977. "The processing of affixed words". Memory and Cognition 5 (6), 690-695. Marslen-Wilson, W.D. 1987. "Functional parallelism in spoken word-recognition". Cognition 25,71-102. Schriefers, H., P. Zwitserlood and A. Roelofs. In press. "The identification of morphologically complex spoken words: Continuous processing or decomposition?". Journal of Memory and Language. Taft, M. 1985. "The decoding of words in lexical access: a review of the morphographic approach". In D. Besner, T.G. Waller, and G £ . MacKinnon (eds.), Reading Research: Advances in Theory and Practice, Vol. 5. London: Academic Press. Taft, M. and K.I. Forster. 1975. "Lexical storage and retrieval of prefixed words". Journal of Verbal Learning and Verbal Behavior 15, 607-620. Tyler, L.K., W. Marslen-Wilson, J. Rentoul and P. Hanney. 1988. "Continuous and discontinuous access in spoken word-recognition". Journal of Memory and Language 27, 368-381. Zwitserlood, P. 1989. "The locus of the effects of sentential-semantic context in spoken-word processing". Cognition 32, 25-64.

Yearbook of Morphology 3 (1990), 87-105

Identifying structure under Xo* Lyn Frazier

0. INTRODUCTION

Investigations of syntactic processing typically abstract away from problems of lexical ambiguity, idealizing the parsing problem as the task of assigning structure to a lexically unambiguous or disambiguated string. Studies of word recognition generally focus on lexical access, or look-up of information in the lexicon, with concern at most for the morphological structure of already-encountered words. Consequently, little is known about where the lexicon and the syntax meet, or how novel lexical structures are identified. In principle, identification of the morphological and syntactic structure of novel inputs might fall exclusively under the domain of the syntactic parser. In effect, lexical analysis would be restricted to lexical look-up. In this case, the syntactic parser would require access to all productive morphological rules of the langauge. Alternatively, a lexical analyzer distinct from the syntactic processor might be responsible for the recovery of some or all word internal morphological structure. This lexical analyzer would presumably have access to all morphological regularities relevant for the identification of novel lexical forms, and it might be equipped with its own specialized processing principles, distinct from those of the syntactic processor. If distinct lexical and syntactic analyzers exist, knowing the division of labor between them should help to reveal the nature of the knowledge sources and operations available within each. Assume that X°s (nouns, verbs and the like) are the atoms of the phrasal syntax. The question to be addresssed here is how much, if any, structure under Xo the syntactic parser identifies? This question will be approached within the framework of a particular theory of syntactic parsing developed by my colleagues and myself (see Frazier 1987, for a review). The parser uses the grammar of a language to systematically attach new items into a phrase marker. At each step it postulates only those nodes required to preserve grammatical wellformedness (Minimal Attachment). For example, a postverbal noun phrase is first analyzed as a simple direct object of the verb because this analysis requires the postulation of fewer new nodes than does the alternative Author's address: Dept of Linguistics, University of Massachusetts at Amherst, Amherst, MA 01003, U.S.A.

88 L. Frazier where the noun phrase is the subject of a sentential complement to the verb. Given two analyses of an input which are equal in terms of the new nodes which must be postulated, new items are preferentially incorporated into the unit currently being processed (Late Closure), making revisions only when the first analysis is inappropriate (Revisions as Last Resort). In this model, much processing is accomplished by what I shall dub incidental decision making. Incidental decisions follow simply from independently required processing decisions or their interaction with grammatical requirements. Thus they are not made at explicit choice points, nor are they governed by a distinct set of principles. For example, clause boundary identification in this parsing model is not governed by any clause boundary strategy per se, but simply follows from minimally attaching each new item into the current phrase marker. As a result of certain attachments, it will turn out that under particular circumstances no subsequent item could be attached into an already postulated clause. This will effectively constitute clause boundary identification though the parser will not have made any special attempt during the parse to figure out where the clause boundary should be placed. Below I will suggest that syntactic identification of the head of a word proceeds in a like manner - it follows from independent attachment decisions, but is not otherwise governed by special principles. In other words, in the syntax, identifying the head of a word is a consequence or by-product of attachment, not vice versa. General principles of attachment govern the process, not syntactic processing principles specially designed to locate the head of a word. It will be argued below that whether walnut or pizza is the head of the first word in the string walnut pizza is a consequence of attachment principles, not head identification principles. Finally, the parsing model will operate with a theory of phrase structure (X'theory) which assumes only the head and maximal projection of a phrase are required by the general theory of phrases. Hence, when only the head is present, projections intervening between the head and maximal projection need not be postulated. Simple phrases may thus be represented as in (la) below rather than (lb). Of course, further articulation of the phrase may be required under some circumstances, e.g., when the node branches or because some grammatical principle would be violated if the projection were omitted (see Frazier, in press). (1)

a. N-max I I I N I girls

V-max I I I

b.

N-max I N' I

V

N

sleep

girls

I

I

V-max I V' I V I sleep

Identifying structure under X°

89

Turning to the relation between lexical and syntactic processing, very little evidence is available at present. The few relevant experimental studies that I am aware of will be discussed below. The results of these studies, together with intuitive evidence about processing ambiguous structures, will be used to argue for the principles in (2) below. However, given the paucity of evidence available, the generalizations proposed here must be considered tentative until further evidence is obtained. Principle 2a claims syntactic analysis of an item doesn't proceed until its major category features have been determined. This has been argued for independently (see Frazier & Rayner 1987, and below). Principle 2b has already been discussed. The structural configuration resulting from general attachment principles determines the syntactic processor's hypotheses about what is the head of a word. Principle 2b' (Sister Attachment) is simply one specific consequence of Minimal Attachment. It follows in a system incorporating Minimal Attachment, providing the rule of compound formation, or some other rule forming X°-s, is accessible to the syntactic processor, as specified in (2c). (2)

a.

Primacy of syntactic category resolution The syntactic processor operates on X°-s disambiguated with respect to major category features (Frazier and Rayner, 1987).

b.

Incidental head identification In the syntax, identification of the head of a word results from independently required attachment decisions. Percolation of features of heads cannot precede head identification in cases of ambiguity.

b'. Sister attachment A constituent may be added under X° if it results from a minimal attachment licensed by a grammatical rule (or is forced to salvage a structure which violates a basic grammatical principle, and the resulting unit is 'lexically viable', see below). c.

Syntactic compound formation The syntactic parser has access to the rule of compound word formation: X -» Y - Z, where X, Y and Z are words (potential X°-s in the syntax) and Z is the head of X.

In arguing for these principles, it will be necessary to rely on intuitions until online experimental evidence is obtained. It is important, therefore, to consider at the outset precisely what intuitive evidence is relevant. The focus here is on structures generated by productive rules, not identification of prestored structure in forms already present in the mental lexicon. Thus, initially it will be necessary to set aside lexicalized compounds. Also, to begin it will be crucial to set aside intuitions about ambiguous structures which contrast in their semantic or pragmatic plausibility, since

90 L. Frazier intuitions about such forms may only reflect late stages of structure selection, or structure revision based on plausibility considerations. The role of these factors will be addressed separately below in Section 3.1 Section 1 examines the conditions permitting syntactic (phrasal) parsing of an input to begin. Section 2 takes up questions of how an input to the syntax is attached into a larger structure. The competition between an analysis based on pre-existing lexical structure vs. phrasal analysis of an input is delayed until Section 3. Nominal compounds form the motivation for the principles in (2). Section 4 examines the implications of these principles for so-called 'reanalysis' structures involving complex verbs. Verbal compounds (cutthroat) are not addressed here, because for the most part they are not structures that can confuse the parser and thus, in the first instance at least, they yield no readily interpretable evidence about the processor's operation, or the location of the boundary between lexical analysis and syntactic analysis. Though clearly and directly relevant in principle, deverbal compounds (truck driver) and hyperstatic compounds (in-flight meal) lie beyond the scope of the present investigation.

1. WHEN DOES PHRASAL ANALYSIS OF AN INPUT BEGIN?

Several studies suggest that the parser proceeds with its syntactic analysis even when the strict subcategory (Chomsky 1965) or the minor syntactic category of the input has not yet been disambiguated. Thus, garden path effects may be observed when verbs with multiple subcategorizations appear with their syntactically more complex subcategorization (cf. Rayner & Frazier 1987, and references therein). Similarly, it appears that syntactic analysis proceeds despite ambiguity in whether a verb is an auxiliary or main verb (see discussion of Dutch hebben in Frazier and Flores d'Arcais 1989) or whether a verb is to be correctly analyzed as inflected or infinitival (Frazier & Flores d'Arcais 1989). However, Frazier and Rayner (1987) suggest that phrasal level syntactic analysis does not begin until the major syntactic category of a word (+/-N, +/-V) is available. They measured eye movements as subjects read sentences like (3) and (4) constrasting A N and N V analyses of ambiguous strings (a, b-forms) and strings with prior disambiguation (c, d-forms). (3)

The a. b. c. d.

warehouse fires... The warehouse fires numerous employees each year. The warehouse fires harm some employees each year. This warehouse fires numerous employees each year. These warehouse fires harm some employees each year.

N-V A-N N-V A-N

Identifying structure under X° (4)

91

... the church pardons... a. Some of us weren't aware that the church pardons very few people. b. Some of us weren't aware that the church pardons are difficult to obtain. c. Some of us weren't award that this church pardons very few people. d. Some of us weren't aware that these church pardons are difficult to obtain.

They found that reading times were long in the disambiguating region of BOTH ambiguous forms (a and b forms). This is expected if category ambiguities lead to delayed analysis, rather than an analysis favoring one particular structure, e.g., one where the parser immediately tries to instantiate a predicted head of phrase before resolving the category ambiguity. Also, in the region of the ambiguous string itself (e.g., warehouse fires or church pardons) reading times were FASTER than in the corresponding region of the disambiguated sentences, as expected if analysis - or at least commitment to one structure - is delayed until the point of disambiguation. Fast reading times in the ambiguous forms also argue against a multiple analysis strategy. Quick processing times in the region of the ambiguity coupled with long processing times, in the disambiguating region of both the AN and NV forms is expected if phrase level syntactic commitments are delayed until category resolution has taken place, as claimed by the Priority of Syntactic Category principle (2a) above. Frazier and Rayner labeled the ambiguities in (3) and (4) AN and NV to avoid problems concerning within word structure, focusing instead on clear category differences of the second target word. The AN ambiguities might be treated by the parser as nominal compounds, as illustrated in (5). Here and throughout I will adopt the convention of eliminating from phrase structure trees any unnecessary preterminals matching a lexical feature specification. Hence, only the N node, as required head of N-max, must be present to dominate the nominal features of church and pardon in the compound church-pardon. (5)

N-max

the church [+Det] [+Noun]

pardon [+Noun]

But this raises a question extremely relevant for present concerns. Why didn't the parser take the first target word (warehouse and church above) to instantiate the head of the NP in the ambiguous forms of (3) and (4), i.e., in all cases except d, where this analysis is already ungrammatical at the point when the first target word is encountered. If the parser had taken the first target word to instantiate the head noun

92 L. Frazier of the noun phrase, then AN (b) forms should have taken longer than NV (a) forms and ambiguous NV (a) forms should have taken no longer than their disambiguated (c) counterparts. Since this pattern of data was not obtained,2 we may conclude that the parser had not yet identified the first target item as the head noun. This argues that feature percolation from the head noun to its maximal projection had not occurred, since otherwise it should have been necessary for the parser to recheck the features of the new head with those of the mother node in the ambiguous AN forms. If it were necessary to revise features that had already percolated from head to mother, an asymmetry in the complexity of the two ambiguous forms should have been observed. In short, assuming the AN forms are really noun-noun compounds, what these data show is that a commitment to the identification of a head of a phrase did not precede syntactic disambiguation of these ambiguous strings. The pattern of equal parsing times on the two ambiguous (a and b) cases can be explained if head identification is an incidental decision (in the sense laid out above). Having attached a word under a predicted node (see Wright & Garrett 1984, Padgett in press, for evidence that obligatory nodes are predicted), the word is head if nothing more is added under X°. However, if a new word is added under X°, the new item becomes the head according to the righthand head rule in (2c) (see Williams 1981). The new head is identified incidentally as a product of an independently required attachment and at no apparent cost, just like adding an additional word or phrase to the current clause even though the clause could have been a complete well-formed clause without the addition. Why, however, would a new word be attached under X°? The answer is simple in the parsing framework assumed here. If the parser has the compound rule in (2c), the compound analysis will result from normal syntactic attachment principles. These require that only the minimal number of nodes be added to the existing phrase marker at each step in parsing. b.

(6)

V-max

N-max Spec I the [+Det]

V-max V°

church [+Noun]

pardons

Having constructed (6) as the minimal attachment analysis of the string the church, attachment of pardons will be delayed until the parser determines whether it is a verb or a noun (by (2a)). Given disambiguation toward a noun, the minimal attachment of a noun into the existing structure is as a sister to church.

Identifying structure under X°

93

2. X°-ATTACHMENT The example in (6) is unambiguous once pardon is identified as a noun. However, given an ambiguous sequence including an additional noun, the Revisions as Last Resort Principle leads us to expect the structure in (7b) to be preferred over the one in (7c). This assumes the parser chooses minimal revisions over more major ones, other things like semantic plausibility being equal. By hypothesis, a minimal revision allows existing structure and interpretation to be carried over from the initial analysis to the revised analysis (see Ferreira & Henderson 1990). Once the processor notes the absence of an inflectional marker on pardon, the word must be labelled as a noun. At this point, the structure in (7a) should be built, as seen above. This structure must be revised on either analysis of committtee, as shown in (7b) and (7c). However, the analysis in (7b) requires only a new syntactic node to be added to the structure - the structure and meaning assigned to the preceding string church pardon may remain intact. By contrast, the revisions of (7a) needed to arrive at (7c) involve not only the addition of a new node (as did (7b)) (7)

a. b. c.

the [church pardon] the [church pardon] committee the [church] [pardon committee]

but also entail that an already structured unit receive a structure and interpretation incompatible with that already assigned. Hence, we expect a preference for (7b), given the interpretation of the Revisions as Last Resort Principle which opts for revisions of a minimal nature as well as the fewest possible revisions. Let's call this Minimal Revisions. In (7) the meanings resulting from the distinct analyses do not differ from each other sharply. In examples like (8), however, it is easier to discriminate between the various interpretations of the ambiguity. In such cases, intuitions about IMMEDIATE processing seem to support the predicted superiority of the left-branching structure corresponding to (7b), (8)

a. b. c.

The corn butter mash impressed me The car cat blanket was torn The river boat people celebrate local holidays

as long as the semantic plausibility and lexical status of the two analyses are balanced. Further, if one tries to force a particular reading by manipulating the relative plausibiltiy of the two structures, as in (9), the structure where the second two nouns constitute the righthand member of the compound (9b and 9d) seems intuitively to be more difficult. Thus, the meaning illustrated in (9a and 9c), though unusual, does not seem intuitively to be difficult to parse. By contrast (9b and 9d),

94 L. Frazier where the meaning strongly favors the presumably unpreferred right branching structure, does seem intuitively to be difficult to parse. (9)

a. b. c. d.

the the the the

[flower bomb] explosion bookshelf [carrier rack] [door paint] factory boat [warehouse sales]

(10)

a. b.

the [wall telephone] cord the nylon [telephone cord]

Examples closer to minimal pairs appear in (10), where again the a-form semantically favors the left-branching structure but the b-form semantically favors a rightbranching structure. Intuitively, (10b) is more difficult to parse.3 Given the option of attaching the third noun to a subsequent noun as in (11) this option (11a) seems preferred to the alternatives (lib) or (11c) - though here the intuitions seem less stable, especially in distinguishing (11a) from (lib). (11)

a. b. c.

the [town map] [stencil drawing] the [[town map] stencil] drawing the [town [map [stencil drawing]]]

If on farther inspection it turns out that a stable general preference exists for the structure in (11a), this would be interesting with respect to the timing of structure assignment. It would suggest the structure revision (insertion of a node dominating N1 and N2) occasioned by N3 in a three noun sequence is accomplished only after the next word has been received. Otherwise (lib) should already have been constructed in long examples like (11) - and presumably be reflected in immediate intuitions as a preference for (lib). Turning to Adjective-Noun compounds, the intuitive preferences are quite different from those seen above for Noun-Noun compounds. In an example like (12) with an Adjective-Noun sequence, the parser will first construct (12a). (12)

a.

N-max

the crispy [+Det] [+Adj]

b.

N-max

crispy [+Adj]

almond [+Noun]

Identifying structure under X° c.

95

N-max

tart [+Noun] [+Noun] (13)a.=(12c)

N-max

b. Spec

A I the crispy [+Det] [+Adj]

almond tart [+Noun] [+Noun]

Principle (2a) specifies that the syntactic parser analyzes only potential X°-s. (14)

a. b. c.

the [crispy] adjective the [crispy ] noun the [pre]

Though crispy is a potential adjectival X°, as illustrated in (14a), it is not a potential nominal X°, as in (14b), anymore than pre- is a potential X°. Hence, assuming each (categorially disambiguated) potential X° is minimally attached into the existing syntactic structure, crispy will be minimally attached as in (12a). Being adjectival, the X° cannot instantiate the predicted head noun. Given its status as an X°, it will not intially be identified as a potential constituent of the head noun unless this results from the lexical lookup process (see below). Once the structure in (12a) is built, the minimal attachment of almond is as shown in (12b) , where it instantiates the predicted N°. Minimal attachment of tart will then result in the structure in (12c). Hence, the structure in (12c) is predicted to be preferred over the left-branching structure in (13b). Immediate intuitions about the ambiguous strings in (15) do seem to favor the analysis in (13a) over that in (13b).

96 L. Frazier (15)

a. b. c.

The crispy almond tarts were delicious The salty butter mash was disgusting The fatty animal sandwiches pleased Marvin

To reiterate, an adjective can be analyzed as an X° only if labelled as an adjective. Hence it may attach in the syntax only where an A° could attach. Unlike a noun, addition of a noun as sister to the adjective will be possible only if the syntactic category of the mother node is revised. Consequently, distinct nominal compounding preferences for ANN vs NNN sequences result in the system proposed here. Perhaps the contrast between the NNN and ANN preferences can be seen best in examples like (16) and (17), directly contrasting pairs which semantically differ from each other only minimally. (16)

a. b. c. d.

farm materials manager city tax control law group restrictions regulation price increase

(17)

a. b. c. d.

rural materials manager urban tax control legal group restrictions regulated price increase

The forms in (16) tend to at first be interpreted with a compound formed of the first two words; by contrast, the ANN forms in (17) intuitively seem more likely to be interpreted with the second and third words forming a compound. The preferences are not absolute, but they are apparent even for these near synonyms. In an informally administered written questionnaire, 16 native English speakers were asked to indicate their first unconsidered interpretation of roughly two dozen ambiguous ANN or NNN forms such as those in (8), (9), (11) and (15)-(17).4 Approximately two thirds of the responses confirmed the predicted preferences. Though the contrasts are far from perfect, the 'matched' ANN-NNN pairs in (16) and (17) perhaps provide the strongest test, since any semantic bias should tend to cancel - rather than potentially cause - distinct preferences for the ANN vs. NNN examples, leading to [12] 3 or 1 [23] responses for both members of a pair. Nevertheless, roughly two thirds of the responses for these pairs also confirmed the prediction, providing initial support for the generality of the intuitive preferences assumed above. One final type of evidence concerns examples like (18), which is multiply ambiguous because of the category ambiguity associated with future.

Identifying structure under X*"O (18) a. b.?? c.

97

the future management crisis [A ][ ] [A ][ ][ [N ][

My own intuitions suggest the preferred reading is the one shown in (18a) where future is interpreted as an adjective modifying a noun-noun compound. The semantically strange analysis in (18c) is also available, where the crisis is current and it concerns management of the future (like leisure time management). What is almost impossible for me to perceive is the reading in (18b), where the crisis concerns management in the future (the managers of tomorrow). These intuitions are readily explained by the present proposal, given that rejecting the preferred analysis (presumably (18a)), the processor still has available to it another preferred structure (18c) and thus, in the absence of evidence requiring the construction of (18b), has no strong reason to construct it.5 To summarize, the results of Frazier and Rayner (1987), suggest that in a determiner-noun sequence, the parser is not yet committed to having the noun as the head of a noun phrase, even though the noun is at the moment the only daughter of N°. Should another noun follow, it may become a sister to the first noun, and thus serve as head to a compound word which heads the noun phrase. Hence, addition of a new head of a word may correspond to the minimal analysis of a new item, and this attachment may occur without any revision of the head-status already assigned, suggesting that in the syntax, identification of the head of a word is an incidental decision. This explains why AN (or noun-noun compound) analyses were no harder than NV strings in the eye-movement experiments. Together with the Minimal Revisions Principle, it explains why a left branching preference exists for NNN sequences. By contrast, ANN sequences do not exhibit a comparable preference because the adjective will be minimally attached into the phrase marker when it is identified; the following noun may not be incorporated as part of an adjectival word, and thus a revision may be avoided only if the noun is analyzed as the head of the NP. Of course, if the word following the adjective were itself an adjective, as in the deep blue hole or a light green coat, then incorporation of the second word as the head of an adjectival compound would be possible, without revision of existing structure. The above preferences are not expected on a model where a lexical processor alone identifies compounds, or on models involving full simultaneous multiple analysis, or delays of analysis under all circumstances, not just when categorial ambiguities are present. In these models, there is no reason to expect the preferences for ANN and NNN sequences to differ from each other.

98 L. Frazier 3. SYNTACTIC COMPOUND FORMATION VS. ANALYSIS AS A LEXICAL COMPOUND OR AS A SYNTACTIC PHRASE

Turning for a moment to lexicalized structures it is apparent that the above preferences disappear if the unpreferred structure involves relations already supplied by the lexicon. Thus the preference for ANN strings to be analyzed as NN compounds is overridden easily if an AN compound (but not an NN compound) already exists in the language, as in the examples in (19). (19)

a. b. c.

the wooden shoe maker the fast food delivery the crispy critter cookies

In the questionnaire study described above, only one quarter (8 out of 33) of the responses to the items in (19) corresponded to the otherwise preferred A[NN] structure; the majority (3/4) of responses identified the familiar compound as a constituent. Similarly, Gleitman and Gleitman (1970) found evidence for a strong role of familiarity in interpreting compounds composed of bird, house and X, where a noun, verb or adjective substitutes for X. They presented subjects with all orders of these items, with either a 132 or 213 stress pattern. Subjects' paraphrases indicated that the familiar compound birdhouse was likely to be correctly identified with unifying stress and more likely to be incorrectly identified with nonunifying stress, than was the plausible but unfamiliar compound housebird. To account for this 'lexical suppletion', where a familiar compound predominates over the otherwise preferred structure in the perception of an ambiguous string, it is necessary to assume that lexical access is not self-terminating. An output corresponding to the individual members of the compound ( e . g f a s t , food) does not preclude an additional output, namely, the compound itself fast-food. But this assumption seems independently required, at least in autonomous multiple access models of the lexicon (Swinney 1979, Seidenberg et al. 1982, Taft & Forster 1976). The only additional assumption needed is that lexicalized compounds become available to the processor before any (semantically equivalent or less plausible) novel compound analysis has been identified and accepted by the processor. Similarly, it seems that a process of semantic attraction takes place whereby highly related elements form a plausible unit taking priority over otherwise preferred analyses - e.g., when one item might instantiate stereotypically an argument of another,6 or supply an obvious dimension along which to construct the analysis as when one word may be taken to supply the material of which potential referents of the other word, an artifact, is made.7 However, intuitions are not helpful in distinguishing between competing accounts of this process - specifically, whether semantic or associative relations guide hypothesis formation or instead only participate in the later evaluation of hypotheses arrived at on independent grounds.

Identifying structure under X°

99

Competition between a compound analysis of an input and a phrasal analysis has not yet been addressed. Let's consider examples like (20). (20)

the house boat and car a. N-max

house boat [+Noun] [+Noun]

and car [+Conj] [+Noun]

N-max Spec house boat [+Noun] [+Noun]

and car [+Conj] [+Noun]

The Minimal Attachment principle predicts that (20a) will be favored over (20b). At the point when boat is encountered, no new nodes need to be added to accomodate the compound analysis. It is difficult to test this prediction because maintaining the semantic plausibility of the compound reading of (20a) while balancing at the same time the parallelism of the conjuncts in the conjoined analysis (20b) is nontrivial. Achieving intonational neutrality or ambiguity also seems especially difficult in this case. Here we will simply note the prediction and point out that testing the prediction is important in showing why the compound analysis would not be systematically overlooked in a syntactic parser operating on X°-s delivered by the lexicon, assuming the grammar permits the structure in (20b) in the first place. If it does not, then there's no potential competition for the compound analysis, and thus no worry that some more minimal structure would always take priority over the compound analysis.

4. REANALYSIS STRUCTURES: AN EXCEPTION TO PRECOMPILATION OF GRAMMATICAL PRINCIPLES?

Next we take up reanalysis (or "coanalysis") structures (see, for example, Hornstein & Weinberg 1981, Fiengo 1974, Di Sciullo & Williams 1987). For purposes of illustration, we will use complex verb formation where a verb and following preposition are reanalyzed as a transitive verb, as indicated by the ability of the object to passivize. If the syntactic parser has access to a rule like (21), then the reanalyzed structure should always be available and preferred to the PP analysis, due to Minimal Attachment. Attaching the preposition as a sister to V does not require

100 L. Frazier any new nodes to be added except the preposition node itself, assuming as before that we omit lexical "preterminal" nodes when lexical features match those of an already postulated syntactic node, e.g., a predicted N° node and a word that is [+noun]. By contrast, analyzing the preposition as part of a Prepositional Phrase does require an additional new node, the maximal projection of the preposition. Hence, the complex verb analysis should be pursued first. Intuitively, this prediction does not seem correct, at least for novel V+P sequences. (21)

V

V -P

For example, in (23) a novel V+P form (or usage) occurs. In (22) this V+P structure should be preferred, and thus a garden path is predicted to occur if (21) is available to the processor given the properties argued for in earlier sections of this paper. The garden path simply does not arise, i.e., there is no indication that the prepositional phrase "after the ceremony" is misanalyzed in (22) due to the parser identifying a reanalyzed form "seen after". (22)

John was seen after the ceremony

(23)

The details were seen after by the caterer

We have an apparent paradox. Novel complex verbs are on occasion comprehended. But if the parser has available to it the rule producing such forms, then general parsing principles validated for phrase-structure and compound word sequences incorrectly predict that postverbal PP sequences should be difficult to process. Assuming that a rule like (21) is available, but only within the lexicon, doesn't solve the problem either. It merely pushes it back a step, since we do not know why the compound rule is available to the syntactic parser but the similar rule (21) is not, and because no substantive theory of processing novel lexical items is available which might have the effect of withholding complex verb analyses initially, but then producing them at some later stage in response to certain types of parsing errors. We might speculate that a novel reanalyzed form is apprehended not by application of a specific rule or template, as in (21), but by extension of the language in line with the underlying syntactic principles of the grammar. Put differently, by hypothesis the reanalyzed structure is not part of the precompiled grammar (see Frazier 1989a, Berwick 1986, Lamontagne 1989), but is instead produced using grammatical principles directly. Reanalysis, on this view, is the formation of a new lexical item on the spot, in order to provide a solution to a problem when parsing the current input. Formation of a complex verb is thus much like coining a new verb form on the spot to satisfy local constraints, e.g., a causative sleep in (24a), or a dative (double object) punch in (24b). Though sleep is an intransitive verb in English and punch is not a double object verb, the parser, nevertheless, copes with the extended uses in (24) without

Identifying structure under X°

101

excessive difficulty, if intuitions are to be trusted on this matter. The novel form may then be stored in the lexicon and in some cases reused by the speaker-hearer. (24)

a. b.

Jimmy slept the bedsheets into a tangled mess Jenny punched me a bunch of net-shots so I could practice my overhand

Though clearly this idea needs to be developed in detail, it leads to an interesting prediction even in its current form. Imagine that other languages are parsed like English, though of course the syntactic processor will use whatever compound (and other) rules are appropriate for the language being parsed. Reanalysis may provide a solution to a breakdown in parsing in other languages as well. The reanalysis will be computationally expensive if it is not endocentric, since revision of existing syntactic decisions will be required - not just sister addition of more material under an existing node. Further, if a reanalyzed structure involves a head (X) and its sister (Y), where X and Y are of distinct syntactic categories, this should be computationally more expensive when the reanalyzed structure is lightheaded than when it is leftheaded. Given an endocentric construction which is leftheaded, a word of any category may be added to its right without structural revision. However, in a lightheaded endocentric construction, if the lefthand member Y is not the same category as X, then a structural revision will be required - mere sister addition will not result in the correct syntactic category. To see this, consider (25). (25)

a.

X

b.

X

c.

X

Structure (25a) will not be difficult to recover, whether it is leftheaded or rightheaded, given incidental identification of word heads. (25b) will also be easy, given that X° is correctly postulated, since Y may simply be added to the existing structure. But (25c) is the difficult case. When Y is parsed it will not be obvious that an X° is being processed. Hence, if Y has already been incorporated into the larger constituent structure representation of the sentence, forming (25c) will necessarily involve a revision of structure, not mere sister addition. It will require the change of a specified (categorial) feature of the existing syntactic representation. Finally, given cases like (25a), we shouldn't be at all surprised if a constituent word became an affix over time (a common process with causatives). This in effect would reflect the addition of a form to the grammar reflecting the now-grammatical structure, without allowing free X° creation in the syntactic components of the grammar.8

102 L. Frazier 5. CONCLUSIONS

I've argued that the syntactic parser operates with the principles in (2). Thus, it has access to (compound rules) with X° on the left, providing they have only X° members on the right. It will create complex words by sister addition. To salvage a parse which violates a grammatical principle, it will create a complex word even if the precompiled grammar doesn't already contain the structure. This presumably reflects the robustness of the parser - its attempt to do its best even with less than perfect input. This view leads to the expectation that reanalyzed constituents will be endocentric. Further, they should tend to be leftheaded or they should involve elements of like syntactic category, when lightheaded structures are formed. It looks as though several principles, those in (2), have been added to our theory of human sentence processing to accommodate the processing of complex words. Is this true? Every theory of syntactic processing must specify what elements are syntactically structured into a phrasal representation. Thus, any theory of parsing must incorporate a principle like (2a), i.e., a principle governing the nature of the units open to syntactic analysis strategies. Hence, while (2a) is a principle of the theory, it or some counterpart would seem to be a necessary component of any theory of parsing. Why the human parser contains (2a) and not some other principle is unclear, however. Principle (2b), on the other hand, is essentially a nonprinciple: it states that the theory has no principle governing the syntactic parser's identification of heads of words. The Sister Attachment Principle is not really an addition to the theory either: given (2c) - the availability of the compound rule within the syntax - sister attachment is simply the result of the already assumed principles for phrase construction. We might ask why the compound rule is available in the syntax? Perhaps all rules with only (potential) X°s on the right are accessible to the syntactic parser. Certainly this would seem to be a good first hypothesis and presumably closely related to principle (2a). The real surprise, then, is the proposal that the syntactic parser will under certain circumstances create a structure not in its precompiled language-specific grammar, if this prevents a breakdown of syntactic analysis and avoids a violation of a general grammatical principle. In effect, this says that the parser favors biology (basic grammatical principles) over convention (staying within the boundaries of already established structures, as determined by the pre-existing lexicon and precompiled rules) accompanied by defeat (breakdown of the syntactic analysis). Recall that this view seemed forced on us by incorrect predictions of the hypothesis which placed productive reanalysis of verb preposition sequences solidly in the (language-specific) grammar. It is also supported by the fact that in general the parser assigns a best fit structure to an illformed input. On the current view, new words (X°) can be formed in the syntax by the parser under highly circumscribed conditions. But these conditions do not necessitate looking into the internal structure of the item already parsed (i.e., the item which was

Identifying

structure

under X°

103

analyzed as a word but which will become a mere constituent of a compound word).9 The morphology-syntax boundary may thus be seen as being permeable, even if morphological and syntactic structures are governed by distinct principles. Structures defined by rules of a morphological nature (e.g., manipulations on argument structures) may create structures identified by the syntactic parser, if the input and output of the rule is a potential X°. This may help explain why languages may exploit morphological means to do the expressive work associated with the phrasal syntax in a language like English - it is not only a lexical processor which can recognize the structures created by such rules.

NOTES * This work was done while I was a fellow at the Netherlands Institute for Advanced Study. I wish to thank the Institute and the NIAS staff for their kind and generous support. I would also like to thank the participants of the Max Planck workshop for their comments, especially Josef Bayer, Greg Carlson and Janet Randall. Finally, this paper has benefitted substantially from the comments and examples contributed by Chuck Clifton, and by two anonymous reviewers. 1. Similar considerations apply with respect to intonation, though the use of written input helps in this case, assuming the role of hyphens to be effectively optional in current English. 2. There was a tendency in experiment 1 for the "AN" forms to take longer to read than the "NV" forms. This tendency was not statistically significant in experiment 1, and was not present even numerically in experiment 2. Experiment 2 used "AN" and "NV" forms which were nearly synonymous as in the church pardons. Whether pardons is a noun or verb, the relation between it and church remains agentive: in both cases, one is most likely to construe the string as indicating that the church gives out pardons. Hence, given the absence of even a numerical tendency for the "AN" forms to be read more quickly than the "NV" forms with such items, it seems most reasonable to construe the numerical disadvantage for "AN" forms in experiment 1 as reflecting the complexity of arriving at an appropriate meaning for them rather than as an indication that the head noun of the noun phrase had already been identified by the parser and now had to be revised. 3. I'd like to thank an anonymous reviewer for this example. Corresponding to (11) in the text, this reviewer also offers (i) wall telephone cord specifications and (ii) nylon telephone cord specifications. My own intuitions suggest (i) is simpler to understand than (ii), as predicted given that a left-branching structure is permissible and semantically acceptable for (i) but not (ii). However, whether (i) is analyzed as a fully left-branching structure is unclear; perhaps it is analyzed as a compound formed of two compounds (wall-telephone and cord specifications). 4. One subject was excluded from the questionnaire data because he clearly did not understand the instructions, e.g., circling only the middle word of a NNN item. The remaining subjects received either 26 items, or 13 items, though many subjects did not respond to every item on the list. Subjects were drawn from secretaries, graduate students, and faculty, mostly from the Linguistics Department at the University of Massachusetts. I am extremely grateful to Kathy Adamczyk for passing out and collecting the questionnaire. 5. It's possible that ANN sequences do not involve compounds; indeed, Josef Bayer (this volume) suggests English may not have productive AN compounds at all. Given this assumption, the preferred reading of an example like rural materials manager would consist of an A that modifies N1 without forming an AN compound. In this case, presumably N2 and N3 do form a compound, since an NP consisting of a specifier and two Ns is not typically allowed. Hence, the suggestion is that adjectives may modify constituents of words, not the word as a whole. Many problems emerge for this suggestion. Why

104 L. Frazier is it possible, even though it is not necessary, to get the stress pattern which would be expected if the A and N form a compound? Why isn't the [AN] reading possible with the stress expected for the [NN] compound analysis? And why can't adjectives modify the constituents of existing English compounds? One might expect, on this view, that a new White House could refer to a freshly painted rendition of the existing White House, a light blackboard could refer to a chalk board of grey color, a deep bluebird to a bluebird of deep blue, and so forth. Moreover, if we use the expected genericity of the meaning of compounds as a test, ANN sequences in which the A modifies N1 again behave as compounds. With stress on Christmas, a phrase like the last Christmas toy simply cannot have the meaning 'the toy from last Christmas'. Further, since last Christmas is not easily construed genetically, it resists compound formation altogether. Hence, even with stress on last, this interpretation is missing. While this behavior is expected on the [AN] compound analysis, it remains entirely unexplained on the alternative view which predicts adjectives may freely go inside words to modify their constituents. Notice that the issue here is distinct from that in well known bracketing paradoxes (e.g., nuclear physicist), where the mental lexicon may supply the processor with a noncompositional meaning. Hence, we will continue to assume that the existence of novel AN compounds in English provides the structure underlying the interpretation where the A modifies Nl. 6. I have not focused on synthetic or verbal compounds in this paper. One aspect of the questionnaire data suggests they exhibit different preferences from the compounds examined here. Cold milk drinkers was the clearest case of a verbal compound in the questionnaire (assuming managers to now be a lexical item that has undergone semantic drift). In cold milk drinkers, ten of the thirteen responses indicated that cold milk was preferably taken as a constituent, in sharp contrast to other ANN forms that did not contain an already familiar AN compound. Perhaps this is due only to the frequent co-occurrence of cold and milk, but I suspect the deverbal status of drinkers is also important. 7. The question of how appropriate meanings are identified for compounds is extremely interesting. The grammar does not dictate the interpretation of compounds, though certain relations seem prohibited. For example, the lid jar does not seem to have a presuppositionally neutral reading on par with the jar lid, where the compound may simply identify a referent using one of its salient properties. Whether this reflects the restricted availability of "part-whole" relations in compounds, or just a pragmatic fact about the relative size of the reference objects used to identify entities is not entirely clear. But in any case, there do seem to be restrictions on the interpretation of compounds, of one sort or the other. Though the grammar does not specify the semantic relation holding within compounds, certain relations seem particularly easy to identify (see note 1, Downing 1977 for adults, Clark 1981 for children, for example). Once a relation is assigned, further compounding seems to obey meaning persistence (see Hellan 1989). 8. For additional examples of reanalysis, see discussion of 'coanalysis' in Di Sciullo and Williams (1987), plausibly the verb cluster data in Haverkort (to appear), and conceivably English dvandacompounds discussed in Lieber (to appear). 9. Williams (1981), Selkirk (1982), and Di Sciullo and Williams (1987) develop a very syntactic view of the structure of words. Following this general approach, lexical structures are defined by rewrite rules, and the geometry of the structures they define participate in feature percolation possibilities, according to, say, the relativized head principle. Baker (1988) provides an interesting critique of the claim that this approach, coupled with the Lexical Integrity Hypothesis, can be maintained as a restrictive view of grammar. He argues that different order of application of affixes can alter interpretation and, therefore, the internal structure of words cannot be opaque to the phrasal syntax. It is unclear, however, whether such examples really show that syntactic rules must be allowed to go inside words. Perhaps interpretative rules alone have access to structure inside X° once an X° has been built An English example like misreattach vs. remisattach (pointed out to me by Tom Roeper) is interesting in this respect. The interpretadon of the forms differs in that only the second presupposes that the first attachment-attempt went astray. However, showing that affix order affects the interpretation of the larger structure does not by itself prove that syntactic rules must register this difference. In any case, it remains unclear at present whether it is only the parser which may form X°s in the syntax, or whether grammatical rules may freely do so.

Identifying structure under X°

105

REFERENCES Baker, M. 1988. "Morphological and syntactic objects: A review of A.M. Di Sciullo and E. Williams 'On the Definition of Word'". Yearbook of Morphology. Dordrecht: Foris Publications. Berwick, R. 1986. Principle-based Parsing. MIT manuscript. Chomsky, N. 1965. Aspects of the Theory of Syntax. Cambridge, MIT Press. Chomsky, N. 1981. Lectures on Government and Binding: The Pisa Lectures. Dordrecht: Foris Publications. Clark, E. 1981. "Lexical innovation: How children leam to create new words". In V. Deutsch (ed.). The Child's Construction of Language. New York: Academic Press. Crain, S. and J.D. Fodor. 1987. "Sentence matching and overgeneration". Cognition 26, 123-169. Di Sciullo, A.M. and E. Williams. 1987. On the Definition of Word. Cambridge: MIT Press. Downing, P. 1977. "On the creation and use of English compound nouns". Language S3, 810-842. Fiengo, R. 1974. Semantic Conditions on Surface Structure. MIT Doctoral dissertation. Ferreira, F. and J.M. Henderson. 1990. Effects of Sentence Length and Complexity on Syntactic Reanalysis. University of Alberta, manuscript Forster, K. and BJ. Stevenson. 1987. "Sentence matching and well-formedness". Cognition 26,171-186. Frazier, L. 1987. "Sentence processing: A tutorial review". In M. Coltheart (ed.), Attention and Performance XII. Hillsdale, N.J.: Lawrence Erlbaum Associates. Frazier, L. 1989a. "Against lexical generation of syntax". In W. Marslen-Wilson (ed.), Lexical Representation and Process. Cambridge: MIT Press. Frazier, L. 1989b. "Parsing modifiers: Special purpose routines in the H.S.P.M." In D. Balota, G. Flores d'Arcais, and K. Rayner (eds.), Comprehension Processes in Reading. Hillsdale, NJ.: Lawrence Erlbaum Associates. Frazier, L. and G.B. Flores d'Arcais. 1989. "Filler driven parsing: A study of gap filling in Dutch". Journal of Memory and Language 28, 331-344. Frazier, L. and K. Rayner. 1987. "Resolution of syntactic category ambiguities: Eye movements in parsing lexically ambiguous sentences". Journal of Memory and Language 26, 50S-S26. Gleitman, L.R. and H. Gleitman. 1970. Phrase and Paraphrase. New York: Norton. Haverkort, M. 1990. "The definition of morphological and syntactic words". In G. Fanselow (ed.), Parameters and Syntactic Variation. Amsterdam: John Benjamins Publishing Co. Hellan, L. 1989. "On lexical compositionality". In D. Jaspers, W. Klooster, Y. Putseyo, and P. Seuren (eds.), Sentential Complementation and the Lexicon. Dordrecht: Foris Publications. Hornstein, N. and A. Weinberg. 1981. "Case theory and preposition stranding". Linguistic Inquiry 12, 55-91. Lamontagne, G. 1989. On Précompilation. University of Massachusetts, manuscript. Lieber, R. To appear. "Phrasal compounds in English and the morphology-syntax interface". CLS 24-11: Papers from the Parasession on Agreement in Grammatical Theory. D. Brentair, G. Larson, and L. Macleod (eds.). Chicago Linguistic Society. Livant, W.P. 1962. "Productive grammatical operations: The noun compounding of 5-year-olds". Language Learning XII, 15-26. Padgett, J. In press. "Sentential subjects and predictive parsing". To appear in Western Conference on Linguistics, 1989 Session. Seidenberg, M., M. Tanenhaus, J. Leiman and M. Bienkowski. 1982. "Automatic access of the meanings of ambiguous words in context". Psychology 14, 489-537. Selkirk, E.O. 1982. The Syntax of Words. Cambridge: MIT Press. Stowell, T. 1981. Origins of Phrase Structure. Doctoral dissertation, MIT. Swinney, D. 1979. "Lexical access during sentence comprehension: (Re)consideration of context effects". Journal of Verbal Learning and Verbal Behavior 18, 645-660. Taft, M. and K. Forster. 1976. "Lexical storage and retrieval of prefixed words". Journal of Verbal Learning and Verbal Behavior 15, 607-620. Williams, E. 1981. "On the notions 'Lexically Related' and 'Head of a Word'". Linguistic Inquiry 12, 245-274. Wright, B. and M.F. Garrett. 1984. "Lexical decision in sentences". Memory and Cognition 12, 31-45.

Yearbook of Morphology 3 (1990), 107-114

Comments on the paper by Frazier* Josef Bayer

0. INTRODUCTION

It is general wisdom of linguistic theory that the categories N°, A°, V°, P° are the atoms of syntax. These are the traditional heads of X'-syntax. More recent developments (see Chomsky 1986) suggest that "functional" categories such as C° (for complementizers) and 1° (for the verbal inflection) may be included as well. Frazier's proposal here is that the syntax can in principle also go inside X°, thus analyzing word structure, and that this has obvious consequences for the human processor. Frazier proposes the following three parsing principles which I repeat in a slightly simplified form: (I) (II) (III)

Primacy of syntactic category resolution The syntactic processor operates on X°-categories, which are disambiguated for [aN, pV], Incidental head identification Once the head of a word is identified (as [aN, pV]) the head features percolate. Sister attachment A constituent may be added under X° by the principle of Minimal Attachment (see Frazier's article) as long as the result is 'lexically viable'.

Furthermore, the syntactic parser has access to a rule of compounding, X° —> Y°Z°, where Z° is the head of X°. If this rule is used recursively, Minimal Attachment predicts that, all other things being equal, a left branching structure [[XY]Z] will be preferred over a rightbranching structure, [X[YZ]]. The reason is that X can be attached right under the head Y in the former structure, while it has to "wait" in the latter structure until YZ is built. I would like to comment on two aspects of this proposal, both of which are intimately connected to the major claim, namely, that the syntactic processor has access to (Y°)-elements under X°. My first point concerns principles (I) and (II), mainly incidental head identification and feature percolation. The second point Author's address: Max Planck Institut für Psycholinguistik, P.O. Box 310, 6500 AH Nijmegen, The Netherlands.

108 Josef Bayer concerns compounding as a genuinely syntactic mechanism. In the end, I will speculate about a modification to Frazier's proposal.

1. INCIDENTAL HEAD IDENTIFICATION

(I) and (II) jointly predict that once an X° has been identified as, say, an N° (i.e. [+N, -V]), and N° can be unambiguously taken to be the head of an NP, then the features of N° will percolate to the mother node NP. For instance in The warehouse fires numerous..., the noun warehouse will pass its features to the NP as soon as the string is disambiguated, - roughly at numerous. In a locally unambiguous example e.g., this warehouse, the features can percolate immediately. If it turns out that warehouse was just the non-head of a compound, the structure (la) has to be revised to (lb):

the

warehouse

the

N;

N

warehouse

fire

As long as such revisions remain limited, e.g., by adhering to Minimal Attachment, it is predicted that no difficulty for the processor will emerge. I will now show that this may be an artifact of English, and that revisions may become quite dramatic in languages with freer word order and richer inflection. Consider the German NP initiating an embedded clause in (2): (2)

daß [der Hut]... that the hat

Clearly, the parser can rely on the fact that Hut is [+N, —V]. Thus, feature percolation can occur. But notice now that there are a number of features which have effects on the NP. Hut is a masculine, singular noun which is rather underspecified for Case. The article der signals that the Case of the NP is nominative. According to certain syntactic assumptions, the nominative may even be indicative of the grammatical function [NP, S]. Thus, according to (I) and (II), der Hut will presumably be associated with a fairly rich array of features. It will be as in (3): (3)

[mase, sing, nom, [NP,S]]

Comments on the paper by Frazier

109

This analysis is perfectly consistent with a continuation like gestohlen wurde ("was stolen"). But it may lead to problems when the continuation goes as follows: (4)

daß [der Hut + Schachtel] der Deckel abgenommen wurde that the hat box the cover removed was "that the cover of the hat box was removed"

With the exception of the feature for number, nothing is now compatible with (3). The feature matrix of NP changes into (5): (5)

[fem, sing, dat, [NP, VP]]

Under the assumption that the parser makes immediate commitments once an unambiguous head is identified, this change would clearly predict some processing difficulties. As a matter of fact, however, one can be almost sure that there is no difficulty in processing sentences like (4). This situation does not change when we make the compound richer. Since German has an extremely productive system of nominal compounding, we can easily find examples like (6a) or even (6b): (6)

a. b.

daß [den Hut+macher+gewerkschaften] die Luft ausgeht that the hat maker unions the air lacks "that the hat maker unions run out of breath" daß [den Hut+macher+verfolger +gewerkschaften] die Luft ausgeht that the hat maker prosecutor unions the air lacks "that the hat maker prosecutor unions are running out of air"

In both cases, the processor may live with a feature matrix [masc, sing, acc, [NP, VP]] until the occurrence of gewerkschafien. This is a feminine, plural, dative noun. Case syncretism has led to a situation where den is ambiguous between the two matrices. The revision leads to [fem, pi, dat, [NP, V']]. But even in those bizarre cases of compounding, there is no indication of a garden path effect. This is clearly not what one would expect under Frazier's proposal. Given instantaneous assignments, if it is already difficult for the processor to rebracket three X°-constituents, it should be difficult to reassign all the syntactically and semantically relevant features of the NP at a late stage of N°-recursion. Such a conjecture seems at least plausible.

2. SYNTACTIC COMPOUNDING

My second remark concerns the exploitation of a compounding rule by the syntactic parser. Frazier's rule of compound word formation, X° Y°- Z°, suggests unlimited productivity. As a matter of fact, however, (English) compounding is not all that

110 Josef Bayer productive. [NN] is very productive, also [NA], but [VA], [NV], [AV], and [ W ] seem to be very limited, or even non-existent.11 would like to concentrate on [AN]compounds as they play a major role in Frazier's contribution. Frazier observes that [ANN] is obviously preferred with a rightbranching structure i.e., [A[NN]] rather than [[AN]N]. For instance, in the crispy almond tarts, one tends to interpret the adjective as modifying almond tart and not almond. Thus, the preferred bracketing is the opposite of what was found for the [NNN]-sequence. Why is this so? Frazier suggests that upon perception of the A, the parser cannot instantiate an N°. Thus, A is taken as an AP which is attached under a hypothesized NP-node. This leads to (7a). When the first potential head almond is introduced, the A(P) would have to be removed from N' and inserted under N° in order to get a compound structure [No AN0]. Thus, it is easier to build a structure like in (7b), where A stays where it is in (7a) while [NN]-compounding takes place. (7)

a. b.

| N P the [N. crispy [... [up the [N. crispy [N almond tarts]]]

So far, this sounds plausible. There is, however, another explanation which has nothing to do with processing. This explanation involves the claim in (8) which I would like to defend. (8)

In English, [AN]-compounding is not fully productive.

Typically, [AN]-compounds are lexicalized. Examples are blackbird, gentleman, greenhouse, highway, holiday, quicksilver, redneck, smallpox. Marchand (1969) comes to the conclusion that the type blackbird has probably ceased to be productive. In the last 100 years there appear only the words strongpoint (a military term) and strongman (chiefly political). (p. 64)

Frazier's processing explanation for the preference of [A[NN]] can only go through when the grammar of English has productive [AN]-compounding. Let us for the moment assume that this is the case. Adopting an SPE-type compound rule which assigns a righthand head and a lefthand stress would predict a derivation of [[rural materials] manager] as follows in (9) (9)

b.

i l l [N [N [A rural] [N materials]] [N manager]] l 2 [N [A rural] [N materials]]

c.

[N[N [A rural] [N materials]] [N manager]]

a.

l

3

2

Comments on the paper by Frazier

111

The stress in (9c) is clearly absurd. Notice, furthermore, that the overwhelming majority of English [AN]-compounds has a monosyllabic A as the first member (see Bauer 1983: 205f). We expect the 132-stress pattern only in cases like fastfood restaurant, but never in Frazier's examples urban tax control, regulated price increase, and (9). Similar restrictions hold for German [AN]-compounds as Fleischer (1975) and Reis (1985) point out. A, for example, must not be suffixed, and it must not be evaluative, as shown in Reis' examples given under (10). (10)

a. •Farb-ig druck color-y print •Pflanz-lich kost b. plant-ly diet •Bargeld-los verkehr cash-less transfer d. *Feigmann cowardly man e. •Primitivperson f. •Wüstkerl depraved guy

"color print" "vegetarian diet" "cashless money transfer" "coward" "primitive person" "rake"

Since German prenominal adjectives obligatorily inflect for gender, number, and Case, it can be seen more readily than in English that [AN]-compounding is limited and that "on the spot" creations are only possible when A is an inflected syntactic atom. a. b. c. d. e.

Frischluft fresh air Naßzelle wet cabin Neuland new land Kurzschrift short writing Kleinwagen small car

a. •Lauluft mild air b. •Kaltzelle cold cabin c. •Fernland far land

"fresh air" "shower cabin" "virgin territory' "shorthand" "small car" vs. lau-e Luft vs. kalt-e Zelle vs. fern-es Land

112 Josef Bayer d. *Rundschrift round writing e. *Breitwagen broad car

vs. rund-e Schrift vs. breit-er Wagen

The compounds in (11) must be lexicalized because their correspondents in (12) are felt to be ill-formed, while the A makes a perfecdy normal prenominal (inflecting) AP. I venture to say that the situation in English is not different. If this is the case, however, one has to cope with the fact that NPs of the type DET [N. AP [N2 N1 N2]]] can be interpreted as if the AP would modify Nl. These are the less prominent readings which Frazier predicts: rural materials manager, urban tax control and regulated price increase can mean according to this analysis "manager for rural materials", "control of urban taxes", and "increase of regulated prices", respectively. Under our proposal, these readings are clear instances of the well-known bracketing paradox. Bracketing paradoxes are found in structures which require a bracketing A for one level of representation and a non-isomorphic bracketing A' for another. A classical example is unhappier, -er cannot attach to As with a heavy second syllable (e.g. *direct-er). Thus, -er must attach to happy before un- attaches. But what unhappier means is not "not more-happy" but "more not-happy" (see Pesetsky, 1985, among others). The same holds on the phrase-level in cases like generative grammar book. The stress-pattern with the 1-value on grammar indicates that generative is an AP which formally modifies the [NN]-compound. But this gives the wrong semantics. The grammar, not the book is generative } This is not the place to suggest a solution.3 It is enough to become aware of the fact that there is this mismatch between form and content in the language, and that Frazier's "productive [AN]-compounds" are likely to be just cases of this sort. In ambiguous cases in which one reading does not lead to a conceptual clash, it is quite likely that the parser will go for an interpretation whose LF-syntax is isomorphic to its S-structure syntax. One thing should be clear, however. The exclusion of certain [AN]-sequences by the X°-parser is not due to attachment preferences, but to the grammar. The grammar is obviously not in a position to build [AN]-compounds "on the spot". The consequences for syntactic processing are clear: Since syntax proper is fully productive, the generalized compounding schema X° —»Y°- Z° cannot be part of the syntax. If it were, we would expect cross-categorial generality.4 This means that the syntactic parser does not make use of a rule as suggested by Frazier.

3. CONCLUSION

We have seen two problems with Frazier's theory of syntactic processing "under X°". The first was that in a language like German, garden path effects would be expected which are clearly absent. The second was that the syntactic parser should have access to X°-structures which, however, seem to be non-productive. In my

Comments on the paper by Frazier

113

view, these are serious problems as long as we want to maintain that the parsing of compounds is a matter for the syntactic parser. The problems disappear as soon as we claim that there is also a morphological parser which works side by side with the syntactic parser. The two would operate in tandem, the morphological parser looking ahead for a maximal X°-category which is then used as a syntactic atom by the syntactic parser. Minimal Attachment could easily be seen as a strategy that holds for both parsers. Thus, Frazier's explanation of the [[NN]N]-preference could still hold in full generality. An X°-structure look-ahead device can explain why the syntactic processor does not come under strain when an NP grows larger due to X°recursion. And finally, not implementing syntactic rules, the morphological parser can cope with the varying degrees of productivity that are found in compounding and in other word-building processes.

NOTES 1. Some forms (see Selkirk 1982, Ch.2) of this kind, like whitewash and windowshop could be seen as cases of back-formation (see Marchand 1969). 2. In German, the mismatch goes so far that in case of a conflict, AP overtly agrees with the "wrong" head: (i)

ein generative Grammatikiuc/i

Such cases are felt to be a bit awkward, but they occur frequently in the spoken language, and they are clearly better than cases which mirror the semantic relations: (ii)

*ein generativ-e Grammatilcbuch

3. For different accounts see Williams (1981), Sproat (1988), Pesetsky (1985), Di Sciullo and Williams (1987). Pesetsky proposes Quantifier Raising (QR) at the word level. As Di Sciullo and Williams (1987) point out, this proposal cannot account for phrase-level paradoxes because an X°-category (or an affix) would have to attach to an XP. Notice that this would also violate the Head-Movement-Constraint of Chomsky (1986). 4. See Selkirk (1982: 2.1) for relevant discussion of the gaps of X°-rewrite-niles.

REFERENCES Bauer, L. 1983. English word formation. Cambridge: Cambridge University Press. Chomsky, N. 1986. Barriers. Cambridge, MA: MIT Press. Di Sciullo, A.M. and E. Williams. 1987. On the Definition of Word. Cambridge, MA: MIT Press. Fleischer, W. 1975. Wortbildung der deutschen Gegenwartssprache. Tübingen: Niemeyer. Marchand, H. 1969. The Categories and Types of Present-day English Word-formation. München: Beck'sche Verlagsbuchhandlung. Pesetsky, D. 1985. "Morphology and logical form". Linguistic Inquiry 16, 193-246. Reis, M. 1985. "Against Höhle's compositional theory of affixation". In J. Toman (ed.), Studies in German Grammar. Dordrecht: Foris Publications.

114 Josef Bayer Selkirk, L. 1982. The Syntax of Words. Cambridge, MA: MIT Press. Sproat, R. 1988. "Bracketing paradoxes, cliticization and other topics: The mapping between syntactic and phonological structure". In M. Everaert, A. Evers, R. Huybrcgts and M. Trommelen (eds.), Morphology and Modularity. Dordrecht: Foris Publications. Williams, E. 1981. "On the notions 'lexically related' and 'head of a word"*. Linguistic Inquiry 12, 245-274.

Yearbook of Morphology 3 (1990), 115-127

The pronominal clitic [dor] in Dutch: A theoretical and experimental approach Aditi Lahiri, Allard Jongman, and Joan A. Sereno

1. INTRODUCTION

Clitics are studied in a number of areas in linguistics, including phonology, syntax, and semantics. To this, we add the field of psycholinguistics. The present paper investigates the phonology and processing of verb-clitic constructions in Dutch. First, we describe in formal terms the prosodic characterization of the cliticization of cfar ('her') to obstruent-final verb forms. The verb-clitic construction varies phonologically in the phrasal domain; sometimes the host and clitic behave as if they form a single word, and at other times, the host appears to be a separate phonological word and is therefore subject to phonological rules sensitive to word boundaries. Second, we discuss the role of the phonological representation of the verbs in the mental lexicon in processing these host-clitic constructions when they differ in their prosodic structure, having been subject to different phonological processes. We argue that the different prosodic structures do not affect the parsing and processing of the verbclitic constructions; rather, processing time depends on whether the cliticized form corresponds to the underlying representation of the verb stem. To our knowledge, this is the first attempt to combine psycholinguistic research with a formal account of verb-clitic constructions. As such, our experimental results are tentative and should be viewed as the beginning of a research program rather than a definitive conclusion.

2. THE ¿-INITIAL CLITICS

Dutch has a number of ¿¿-initial clitics like dsr 'her', tj dat gedaan had]] who if t this done had

The idea is that I-to-C extends the domain of I, and provides C with an index enabling it to properly govern the subject trace. However, the clear impossibility of (37) shows that the relevant factor cannot be the indexing of Comp (see Reuland 1985 for more extensive discussion). (37)

*Piet vroeg [CP wie; of [IP \ kwam]] Piet asked who if t came

Therefore, these facts do not support I-to-C, neither in its weak, nor in its strong form. Also, the significance of agreement features on Comp is obscure. It is not a general phenomenon; it does not occur in standard Dutch, nor in many dialects. Since the pattern we are discussing also obtains in dialects without overt agreement

146 E. Reuland in Comp, it cannot be an immediate effect of the process licensing I. Where Comp agreement occurs in Frisian and German, it is limited to the second person singular; again, no further contrast correlates with the presence or absence of this phenomenon. Hence, although it could support weak I-to-C, as an argument for strong I-toC it is ineffective. The only further reason in favor of I-to-C seems to be that Tensedness/Finiteness of a clause is marked both on the complementizer and on the verbal inflection. However, this phenomenon also obtains in languages where I-to-C cannot be obligatory, such as French, or English. Hence, again, this is not an argument for strong I-to-C, and dubious for weak I-to-C. The absence of evidence for the structure in (35) is enough to make it implausible. In all cases, agreement and tense features are overtly marked on the verb (see (35)).16 In order to be compatible with (35), this fact requires that the language learner classifies the inflectional morphology as insignificant where it is overtly realized, and assumes that its actual position of realization is in a position where there is no overt marking. V-to-I must be blocked. In the absence of any clues, this would require an independently motivated principle (or parameter). No such principle has been offered in the literature, and none seems conceivable. Let us summarize what has been achieved so far. We started out with a number of observations based on the distribution of adverbs and the targets of extraposition, showing that sequences of the form V-Aux cannot have an S-structure representation • • -V vpl Ii Aux]. We used an argument based on scope facts, showing that.. VP ] [j Vj Aux] is also not a viable option. The upshot of these arguments is that V and Aux/I are not separated by a maximal projection, not at S-structure, but not at Dstructure either. In the present section it was shown that a structure .. .V Aux VP] e] (I is empty) cannot be motivated. Therefore, an alternative is required. In the next section, I will discuss the preliminaries of such an alternative.

4. CONTRASTS BETWEEN GERMANIC SVO AND SOV LANGUAGES

There are a number of striking differences between the Germanic SVO and the SOV languages. They can be summarized in the following table (from Reuland & Kosmeijer 1988). (A)

SOV languages

SVO languages

1.

weak restrictions on VP-internal Subjects

severe restrictions on VP-internal subjects

2.

weak restrictions on scrambling

severe restrictions on scrambling

Head Movement and the Relation between Morphology and Syntax

147

4.1. The domain of I The VP-internal subjects in (Al) agree with the finite verb and have nominative Case. In Dutch, they occur in constructions with Nominative-Dative inversion and in the wide variety of er-constructions. Examples illustrating this are given in (38). (38)

a. b.

Ik denk dat hem deze zaken niet bekend waren I think that him these matters not known were Ik verbaasde me dat er veel mensen oude boeken lazen I wondered (myself) that there many people old books read

The corresponding sentences are ungrammatical in English. The same holds true for Norwegian and Swedish. Under the standard assumption that I licenses nominative Case and exhibits agreement, (Al) indicates a connection between the SOV character and the relative ease for Inflection to entertain a local relation with VP-internal positions. Under the null hypothesis, this is the standard government relation. If so, INFL governs VP-internal positions in SOV languages. (Note that this is inconsistent with the assumption that these have I-to-C, and lack V-to-I.) As is well-known, Dutch, German, and Frisian also allow a considerable amount of scrambling. That is, subcategorized material of the verb, including direct objects, can occur to the left of sentence adverbs. This is illustrated by the examples in (39).17 (39)

a. b.

Ik denk dat Jan waarschijnlijk Marie dat boek geeft I think that John probably Mary that book gives Ik denk dat Jan Marie dat boek waarschijnlijk geeft I think that John Mary that book probably gives

Word order is even less restricted in German, in that objects may move to the left of an (indefinite) subject. Scrambling depends on various factors, such as the nature of the adverb and the nature of the object. We will not address these issues here. We will just use it as a probe into the structure of the IP. Sentence adverbs such as probably and waarschijnlijk are semantically and syntactically associated with I. They cannot move along with VP-proposing.18 This holds true for English and for Dutch. So, (40a) is ungrammatical, in contrast with (40b). (40)

a. *waarschijnlijk een boek lezen zal Jan niet probably a book read will John not "probably read a book, John will not" b. snel een book lezen zal Jan waarschijnlijk niet quickly a book read will John probably not "quickly read a book, John probably will not"

148 E. Reuland This is in accordance with their semantic property of modifying the predication, rather than just the predicate. The association between I and negation is even closer. In French, the negation marker cliticizes onto 1°. Negation adverbials in their canonical position (a in (8)), are governed by 1° (this also holds true for Scandinavian).19 In English, roughly the same pattern obtains, with not behaving as a negation marker and cliticizing onto 1°. The null hypothesis is that negation in Dutch has the same property, with the element functioning as the negation marker staying closer to I than other sentence adverbials. Consider now the following fact. In Modem Dutch, negation is expressed by the element niet. In general, the complements of the verb may occur on either side of sentence adverbs, due to scrambling. Scrambling is generally optional. However, when the element niet expresses sentence negation, it is virtually required to be to the right of any object that can be scrambled.20 This is illustrated in (41). (41)

a. *Ik denk dat Jan niet Marie dat boek geeft I think that John not Mary that book gives b. Ik denk dat Jan Marie dat boek niet geeft I think that John Mary that book not gives

This shows that negation prefers a position 'close* to the verb, and shuns a position 'close* to C. In terms of the structure in (35), one would have to stipulate that niet cannot be adjoined to the VP (or IP), but must remain in the VP. Given the close relation between NEG and I in French and English, this would be quite strange if I is 'just in C', or 'just in I', with no syntactic connection between I and V. Rather, one would expect NEG in Dutch to be close to Comp, or at least external to the minimal VP, respectively. However, if the content of I is actually on the V, where one 'sees' it, the pattern conforms to the null hypothesis. In that case, the verb realizes I, and the position for NEG close to its licenser is actually inside the VP.21 So, the facts indicate a significant difference between NEG (as a negation marker) and other sentential adverbs; whereas the general requirement on S-adverbs is just government by I, NEG stands in a more restrictive relation. A significant relation that is more restrictive is that of sisterhood (Chomsky 1986). The result of NEG cliticized to I in French and English meets the conditions on sisterhood. A more precise version of the null-hypothesis (although necessarily tentative) then, is that a sentential negation marker must be a sister to I.22 Dutch conforms to this hypothesis precisely in the case that I is realized within the VP (and, of course, NEG is too). This view of the relation between NEG and I is consistent with the facts about VPinternal subjects summarized in (Al). So, because of the distribution of VP-internal subjects and negation, one must conclude that the domain of INFL starts out lower in the SOV languages than in the SVO languages. As a consequence, the strong I-to-C approach cannot be a viable alternative to V-to-I. The facts discussed here are consistent with the weak I-to-C

Head Movement and the Relation between Morphology and Syntax

149

approach. But we still face the question of what blocks V-to-I or causes I to be weak, in the SOV languages. 4.2. The domain ofV If the domain of INFL in the SOV languages starts out lower than in the SVO languages, questions arise about the relation between the domains of I and V. Scrambling facts help determine that relation. This is shown by Icelandic. The relevant facts are summarized in (42) and (43). (42)

a.

Hann vissi a8 Skuli mun oft segja Sveini sogu he knew that Skuli will often tell Sveini story b. *Hann vissi a§ Skuli mun Sveini oft segja sogu he knew that Skuli will Sveini often tell story c. '"Hann vissi a5 Skuli mun segja Sveini oft sogu he knew that Skuli will tell Sveini often story

(42) illustrates that in the presence of an auxiliary, scrambling of Sveini over the infinitive segja and/or the adverbial oft is impossible, as is moving segja to the left of oft. When there is no auxiliary, the main verb moves over the adverb, and the indirect object Sveini may too. This is illustrated in (43). (43)

Hann vissi a8 Skuli sagSi Sveini oft sfigu he knew that Skuli told Sveini often story

The relative positions of verb and adverbial indicate that in Icelandic the main verb may move into the INFL position if the latter does not contain an auxiliary (Thrdins son (1986)). That is, Icelandic has V-to-I. In (42) the indirect object may not move over the adverbial, whereas in (43) it may. So, V-to-I correlates with scrambling. This correlation indicates that arguments cannot scramble out of the domain of the verb. Only if V-movement extends the domain is scrambling possible. So, we have the following principle: (PI)

NPs cannot scramble out of the domain of V

That is, a scrambled NP must remain governed by V. As we saw above, there is also a principle (P2) in the grammar. (P2)

NEG must be a sister of I

Given these, the Dutch scrambling data in (39) and (41) lead to the conclusion that the domains of INFL and V may overlap in SOV languages. In the SVO languages without V-to-I in the overt syntax (English, Swedish, Norwegian), these domains are

150 E. Reuland strictly separated. We now face the dilemma that under the analogy with Icelandic, (PI) and (P2) suggest that Dutch has V-to-I as well, whereas we have seen that this cannot be the case. This dilemma is caused, in fact, by the following assumption: (B)

At D-structure, a phrase headed by a finite verb is obligatorily represented by a V-projection contained in an I-projection, i.e., V and I project separately.

It is this assumption which either forces a VP, to which nothing can right-adjoin, or an I which has to remain empty. The fact that the V- and I-domains appear to overlap in Dutch shows the way out: (C)

V and I are merged in the morphology and project up simultaneously

(C) can be accommodated in current theory under the following assumption: (D)

head-movement may apply in the lexicon

If a complex head is inserted, its components may project simultaneously, subject to general principles of X'-theory. We will now investigate this option (see also Bayer 1990 for a discussion of the evidence supporting this type of structure).

5. PROJECTING FROM WORDS

5.1. Some preliminaries Suppose we adopt simultaneous projection. How would this fare with current theory? Are there any principles of UG which force splitting off the inflection from the verb, i.e., projecting verb and inflection separately, and assigning them distinct domains? The theory of phrase structure would be a candidate. It could be required that the basis of a projection is always a single X° category. But this cannot be a general well-formedness requirement on trees, since structures resulting from head-movement violate this condition. One could exclude adjunction structures from D-structure, but this cannot be correct either, since no movement source can be motivated for a number of adjunction structures (for instance the adverbials in position a of (8)). So, excluding base-generated adjunction structures of X°s would require an unmotivated decision. Other principles determining D-structures are the projection principle and theta-theory. For instance, given a configuration like [x« N° V°], the projection principle will require an argument position to be projected if the N° is to saturate a theta-role of the V°. However, in the case of a lexical category and a related functional category, no part of the projection principle actually requires a split. This

Head Movement and the Relation between Morphology and Syntax

151

is precisely the case under discussion: Given the structure [Xo V° I°], a separate INFL-position need not be projected in order for (C) to be viable. (If one maintains a distinction between X° and stems, there is, strictly speaking, no adjunction structure, but the other considerations remain valid.) I will now summarize the assumptions needed. (E)

(i) the syntactic categories include V and I (alongside N, A, P, C, etc.) (ii) just like the modal auxiliaries in English, the inflectional affixes of the verb are members of I 23 (iii) syntactic projections are headed by an X° where X has the value of a syntactic category; for a a morpheme which is a member of A, and some X° analyzable as [x ay] (x, y (strings of) members of categories, possibly null), [x a y] is an A" (iv) projection is restricted by the general well-formedness conditions on trees, and the standard X'-schema (see Chomsky 1986).

The implications of (Ei-iv) can be illustrated as follows. Assume fab] is an X°, a is an A and b is a B. (Eiii) allows that it is both true that ab is an A and that ab is a B. The general well-formedness conditions on trees are category-neutral, as is the X'schema. Hence, no general principle prevents that both A and B project up. We will use the notation AIB to express that a node carries two labels. (44) illustrates the case where an X° contains both a V-stem and an inflectional morpheme. (44)

V°/I° V°

P

werk-

-t

Scrambling is illustrated in (45). IP

(45)

Vk/Ik

NP NP

Vj/Ij Vi/Ii

NEG t

V°/I° Vstem

Istem

152 E. Reuland In the area of overlap between the domains of V and I, NEG is governed by 1°, the scrambled NP by V°. We will assume that projection is free. Free projection of [V° I°] will always result in a structure with V ^ (= Vk in (45)) being properly contained in the I-projection, since the V1™" will have to be predicated of the subject (see also Reuland 1986,1988). So, projections not meeting that condition are filtered out. 5.2. The SOV-SVO contrasts In the SVO languages like English and Mainland Scandinavian, there is no overlap between the V- and I-projections. Yet, in these languages too, the inflectional morpheme is realized on the verb. Why, then, is simultaneous projection blocked here? We will show that this follows from an independent parameter: the direction of government. In SVO languages, the VP is head-initial, i.e., the head governs to the right with respect to arguments. The I-projection is not. It is head-medial if one considers the position with respect to both subject and VP. With respect to the subject alone, I governs to the left. So, the contrast SVO/SOV is paralleled by a contrast SIO/SOI. The essential part of the asymmetry is formulated in (F).24 (F)

V and I have uniform government directions in SOV languages but different government directions in SVO languages.

(F) makes it impossible for V-I to project simultaneously in SVO languages because of its interaction with (G). (G)

If a and b are potential heads forming one X" constituent, they can simultaneously project just in case they can form a consistent projection line; i.e. a projection line that can be assigned a consistent position with respect to the elements governed by a-b.

(G) need not be stipulated; it is a corollary of general well-formedness conditions on trees. Its effect can be illustrated as follows. Consider a string of the form (46a) with the analyses (46b) - (46f), using a monostring representation for the phrase marker (free after Lasnik & Kupin 1977). a. b. c. d. e. f.

n n n n n

m m m I"

[v-i] V° F V' I'

Head Movement and the Relation between Morphology and Syntax

153

n and m are lexical items (for the sake of concreteness, we will assume they are nouns), v-i (an inflected verb) is a complex head. Syntactically it counts as one terminal element. Since the clause has two arguments, the verb stem must be transitive, n and m are both thematic arguments of the lexical stem v, and hence of the terminal element v-i. Representations such as (46) only minimally differ from standard tree structures. In contrast to other approaches based on covalency (cf. Huybregts 1985, Haegeman & Van Riemsdijk 1986), all structures are representable as single trees. The only increase in expressive power concerns the labelling of the nodes: a node may bear more than one label. As a consequence of the restricted nature of this extension of the theory, standard notions of government, whether they are based on maximality of projections or on minimality, carry over directly. Order in phrase structure is effected by possibly language-specific ordering principles (Chomsky 1981, Stowell 1981). So, one starts out with unordered objects, and defines an ordering on them on the basis of government properties of the heads involved. Therefore, with the elements of (46), we have (47), without order imposed. (47)

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

{ n { m [v-i]}} s { n {[v-i] m }} etc. { n { m V°}} { n { m 1°}} V'} { n n { I* } I"

The following statements will derive an SOV/I order: (i) if 6 is a V° it is to the right of all of its arguments; (ii) if 6 is an 1° it is to the right of all of its arguments (it is irrelevant for our present purposes whether these statements are primitive or derived from conditions on the assignment of Case or theta-roles). Consider now the statements needed to derive an SV/IO order as illustrated in (48). a. b. c. d. e. f.

n n n n n

[v-i] m V° m 1° m V' I' I"

What one would wish to express is that a verb is to the left of its arguments, and INFL to the right of its arguments (or 'to the right of its nominal argument', and 'to the left of its verbal argument' if the VP also counts as an argument). However, statements (i') and (ii') corresponding to (i) and (ii) above, do not have the required effect: (i') if 8 is a V° it is to the left of all of its arguments; (ii') if 8 is an 1° it is to the right of all of its arguments, v-i as a whole is the only possible value for 8 in these ordering statements. So, v-i should be to the left of all of its arguments by (i')

154 E. Reuland and to the right of its arguments by (ii'). These requirements are inconsistent. Weakening such statements to statements just about government direction (e.g., by having 8 only govern to the left and only govern to the right) leads to a similar result. In fact, no pair of statements about 8 can have the required effect: there is no S in (48a) to which these properties can be consistently ascribed. If n and m do not both count as arguments of v-i for the ordering statements, the inconsistency disappears. But this is tantamount to splitting up the v-i complex. This in fact is what we claim happens in the SVO languages. So, we derive (H). (H)

In SVO languages, a finite verb cannot be represented by a single constituent in the syntax.

Since nothing prevents the finite verb to be represented as a single constituent in SOV languages, we have an explanation for the contrast between these language types. So far, we have found a way to represent a property which the SOV languages turn out to have, and we have an explanation as to why the SVO languages lack that property. We still need to have an explanation as to why the SOV languages have that property. An answer can be given in terms of an economy principle, in line with Chomsky (1988). If a finite verb cannot be assigned a unique head-position in the syntax, an additional head position is projected, either an empty 1°, or an empty V°. (For the sake of concreteness, one may assume a principle Insert a, a being a null element of an arbitrary category; see Reuland & Kosmeijer 1988 for details). In order for both of them to be licensed, the finite verb will have to move through both positions (either via the French or by the English option). If a unique position can be assigned, no movement is needed. Hence, this derivation will always be shorter. Therefore, no special stipulation is needed; the cost associated with splitting up a word reduces to movement being a last resort process.25 So, we get (I). (I)

If a-b is a word, with a and b potential heads, the categories of a and b project separately only if no well-formed tree can be constructed otherwise.

This brings us back to our starting point. Since in the SOV languages, the constituency of V-I is preserved at D-structure, whereas in the SVO languages this is not the case, it cannot be maintained that only thematic properties of S-structure enter into determining D-structure. The notion of a well-formed S-structure, then, is not derivative of the notion of a well-formed D-structure modulo move a, and morphological spell-out. The segmentation of a string into words is an independent factor in determining its D-structure, contra Baker's position outlined in section 2.

Head Movement and the Relation between Morphology and Syntax

155

6. TWO RESIDUAL ISSUES

By way of conlusion I will briefly discuss two issues concerning V-Aux sequences. Our results in section 3 imply that such sequences cannot contain a right VPboundary between the base positions of V and I or V and Aux. Section 5 tells us that the constituent carrying the inflectional properties will be verbal as well, and part of the VP. What does this entail about the structure of sentences such as (10a) (=(49) (and more complex cases)? (49)

dat Cindy een boek gelezen heeft that Cindy a book read has

One could envisage two possibilities. One possibility is that verb and auxiliary are together dominated by a V/I-node. This entails deriving Verb clusters in general by applying Verb-movement in the morphology, and projecting any further structure (including traces where necessary) on that basis, in line with the idea that D-structure is a derivative level (along the lines discussed in the previous section). Independently, a proposal to this effect is made by Bierwisch (this volume). The resulting structure is as illustrated in (50). (50)

CP Spec

C' C

IP

NP

een boek

V/I

gelezen heeft

Such a proposal implies a departure from current conceptions about Verb-raising structures in general, and is probably incompatible with the standard version of the binding theory (see, for instance, Haegeman & Van Riemsdijk 1986 for discussion). However, it is not incompatible with the binding theory of Reinhart and Reuland (to appear).

156 E. Reuland Alternatively, it could be the case that the VP contains (one or more) auxiliary positions in addition to the main verb position. For English, extensive evidence for such an articulation of the VP has been provided by Akmajian, Steele and Wasow (1979). These auxiliary positions could be understood as specifier positions (given a definition of this notion that allows X°s), or as functional head positions; the latter option requires that lexical heads in general do not give rise to a maximal projection, but that only functional heads do; auxiliaries, then, should occupy the lowest functional position; this possibility has been brought to my attention by John Frampton (p.c.). Details aside, under such an analysis the resulting structure would be as illustrated in (51). (51) Spec

een boek gelezen Both analyses satisfy the conditions on structures imposed by our results: There is no maximal projection intervening between verb and auxiliary, hence no position corresponding to a ' in (9) between these two; the inflection is borne by a verbal element within the VP. Both analyses are plausible, and merit serious consideration. At this point we will not offer a choice, as a proper evaluation would carry us beyond the scope of this article. The other question is how the present proposal accommodates VP-preposing. In fact, it needs no special provisos, regardless of whether one takes (50) or (51) as the starting point. Crucial is that it is an economy principle which prevents the emergence of a separate I-projection outside the VP when the VP is in situ. This implies that a separate I-projection will be created whenever this is the only way to derive a grammatical sentence. Now, any theory will have to exclude the structure in (52), i.e., a sentence with fronting of the whole IP. (52)

[ c p IP; [ c . C t, ]]

Head Movement and the Relation between Morphology and Syntax

157

For present purposes, it is irrelevant why this is so, and what principle is responsible. Let us assume now that the rules generating syntactic structure make available both (53a) and (53b). (53)

a.

[cp ... tip ••• tvp/i'--- V/I ]]]

b.

[CP

b

[ V P - V ] I ]]

When V and I are merged, fronting of VP implies fronting of I*. This is excluded since only maximal projections can be moved (Chomsky 1986); the next maximal projection is IP which cannot be fronted, by whatever principle blocks (52). Therefore, the VP cannot be fronted when the V-and I-projections are not separated. Consider now (53b). Derivations starting out from (53b) will not surface, since they violate economy, unless no grammatical sentence can be obtained otherwise. Precisely this is the case with VP-fronting. So, a derivation with VP-fronting starting out from (53b) does not violate economy, hence is allowed. Notice that having two D-structure sources is just an artifact of the directional metaphor employed. If D-structure is viewed as a projection of certain properties of surface structure along the lines discussed earlier, the analysis entails no more than that surface structures with a fronted VP must project a separate Infl-head of IP, whereas surface structures with a VP in situ need not (and hence will not).

NOTES * This paper was presented at the workshop When Words Happen to be Phrases, organized at the Max Planck Institute for Psycholinguistics. I am grateful to the audience, and in particular to Jaklin Komfilt for a stimulating discussion. I would also like to thank Aditi Lahiri and Allard Jongman for organizing such a successful event. Finally, I would like to express my gratitude to Eric Hoekstra, Wim Kosmeijer, Jan Koster, Jan Wouter Zwart, and three anonymous reviewers for their valuable comments. 1. This claim has not passed unchallenged, though. For instance, Pesetsky (1982) claims that categorial selection takes place at LF. 2. In much of the literature, the question of the relation between V and I in Dutch is treated rather implicitly. A relatively early discussion is presented in De Haan (1979), arguing for INFL lowering. Recently, Bennis and Hoekstra (1989) have tentatively opted for V-to-I movement in the syntax. Relevant discussion is also found in the literature on the verb-second phenomenon, for instance, De Haan and Weerman (1986) (based on Frisian), Koster (1986), Weerman (1989), Hoekstra and Marfcz (1989). For German, see for instance Lenerz (1984), Olsen (1984), Haider (1986), etc. For V-movement in Scandinavian, see for instance, Taraldsen (1986), ThrSinsson (1986), Holm berg (1986), Platzack (1986, 1987). For the relation between V and I, see also recent work by Kratzer. This list does not strive for completeness. 3. Note that VP-fronting in these simplex cases involves a Dutch variant of 'do-support'. So, we have (i), but not (ii). (i) (ii)

[boeken lezen] (dat) doet Jan niet [*boeken leest] (dat) Jan niet

So, fronting of the VP leads to a stranded I, which is then supported by the dummy verb doen 'do'.

158 E. Reuland 4. Suppose there were an independent principle, forcing adverbs to be always in the left-periphery of the VP (i.e. position a) in English, French, Dutch, and other languages equally, barring them from position a ' . If such an independent principle would exist (although no such principle has been proposed, and its existence is doubtful), base-generated adverbs would become irrelevant to the argument. The part of the argument based on extraposition would be unaffected, though, and thus, sufficient to support our general claims. However, since no principle limiting adverbs to left-peripheral position has been discovered so far, each step of the argument will have to cover both possibilities: base-generated adverbs in a ' and elements that could have been moved there. 5. The effect of Verb raising in complex clauses is illustrated in (i). (i)

dat Cindy Johan dat lied heeft horen zingen that Cindy John that song has hear song

The D-stracture of such sentences is taken to be as in (ii). (ii)

(dat) [¡p Cindy [yp [p Johan [yp dat lied zingen]] gehoord] heeft]

The complement verb restructures with the matrix verb and the auxiliary, as witnessed by the presence of an infinitive instead of a participle at S-structure (horen instead of the expected gehoord). It is usually assumed that it involves movement of the verb out of its complement, first adjoining to the matrix verb, and then to the I. This construction has been extensively discussed by Evers (1975); for discussion in the framework of Chomsky (1981), see for instance, Reuland (1982, 1983). The rejection of generalized V-to-I will have implications for the analysis of verb raising. Investigating these would lead us too far afield, but see Bierwisch (this volume) for an analysis that seems compatible with our result. 6. I am grateful to Jan Wouter Zwart for the suggestion to consider this type of evidence. 7. I leave open the question of why this is so. 8. Note that the derivation of sequences with this constituent order could have involved scrambling of the direct object to the left of the adverbials; this derivation is irrelevant. Relevant is the question whether this order could have been obtained with one or both adverbs in the position of a ' , or with adjunction of one or two extraposable adverbs to the invisible VP-boundary before or after scrambling. It is the nonexistence of these two options which is relevant to the argument, and brought out by the scope facts. 9. This leaves open the question of the precise relation between the main verb and the auxiliary, which are both VP-internal, see section 6 for some discussion. 10. For instance, if bare adverbs must be governed by a lexical I, they will be admitted in a , but not in a ' , under the assumption discussed here. At this point, we need not pursue this, since I will argue that the alternative under consideration cannot be maintained anyway. 11. Of course, a number of these forms are differentiated in the spelling as in (i). (i)

je travaille; tu travailles; il travaille nous travaillons; vouz travaillez; ils travaillent

For the argument this is irrelevant, however. 12. In note 2 I listed some of the relevant literature. In this section I try to make a clear distinction between two logically possible positions. These positions are often not clearly distinguished in the literature. Nor is it easy to determine who would hold which position, if the distinction is assumed. Therefore, I will not try to ascribe the positions discussed to particular authors. 13. Note, by the way, that the Confl approach would not save Baker's claim discussed in section 2. There is no way in which the presence of Confl as opposed to the presence of distinct I and C can be read off thematic structure and X'-theory. The stipulation of a Confl would represent additional information of the type referred to in section 2. Thus, it would imply a negative answer to Baker's claim. 14. See also section 6.

Head Movement and the Relation between Morphology and Syntax

159

15. Note that Koster's and Pollock's conceptions of strength differ. Unlike Pollock's conception, Koster's does not reduce to properties of AGR. 16. To be fair, I should add that the first statement is not so innocuous as it seems. It could be the starting point of a far-reaching debate on the status of syntactic categories, their canonical realizations, and their canonical interpretations (Reuland 1986 provides some of the ingredients). However, I will not do this here, as the outcome would not affect the correctness of this statement. Hence, I will take the evidence at face value. 17. The assumption that objects may undergo leftward scrambling over sentential adverbs is not incompatible with any previous assumption. The recurrent theme in the previous sections is that in Dutch, sentence adverbials do not occur in the position of a ' , even though no known principle would exclude them from that position if that position existed. Note that, even if no one has claimed this position to exist, it was necessary, for our purpose^, to demonstrate that it does not exist Finally, if it does not exist, one faces the question of why it does not exist. This is the question addressed in section 5. If the position of a' is not available for sentence adverbs, the question comes up what their position is: within the VP proper, or left-adjoined to it? The present section discusses some considerations bearing on this issue. 18. There is no consensus as to what prevents them from being moved along. A requirement that they be governed by I, would have this effect. 19. The position of the negation adverb in French is illustrated in (i). (i)

[u> Je [] n'ai] [yp pas [yp lu ce livre]]] I haven't Adv^Q read that book

Its position in Swedish is illustrated in (ii). (ii)

(att) [¡p han [j 4>] [yp inte [yp kommer ]]] (that) he Adv^o comes

20. The order niet NP is interpreted as constituent negation. When the object is a non-specific indefinite, the order NP niet is impossible. Instead, negation is marked on the NP, in general by the negative determiner geen 'no'. I should mention that one reviewer feels that sentences with pre-complement negation improve when they are assigned a special intonation pattern ('an intonation pattern of their own'). This may be true, but of little significance for the present point, since with sufficient intonational help, there are few orders one cannot get in Dutch. Nevertheless, I personally have not been able to find an intonation contour which makes (41a), or its counterpart with an indefinite direct object, acceptable. Adverbs like waarschijnlijk 'probably', etc., on the other hand, behave differently. Unlike niet they are completely natural in pre-complement position. 21. Middle Dutch had a negation marker enlne in addition to the negation adverbial niet. For Middle Dutch, there is independent evidence that it fits in with the general pattern. Weerman (1989:189) gives a number of reasons to the effect that this negation marker must have been closely associated with the inflection of the verb. 22. Note that in proposals involving the projection of a full NegP, the NegP as a whole is a sister of I (or T). Questions such as why not must cliticize in English, whether or not it projects a NegP, why niet in Dutch would not project, etc., are all very pertinent, but obviously cannot be addressed here. 23. Derivational affixes are not members in that sense. See Reuland and Kosmeijer (1988) for discussion. 24. But cf. Koopman (1984). 25. See Reuland and Kosmeijer (1988) for a related principle, the Preservation Principle. An informal version is given in (i). (i)

PRESERVATION PRINCIPLE (informal): At any level, the representation of a sentence will reflect its observable properties insofar as these properties are representable at that level.

160 E. Reuland REFERENCES Akmajian, A., S. Steele and T. Wasow. 1979. "The category AUX in universal grammar". Linguistic Inquiry 10, 1-64. Baker, M. 1988. Incorporation, Chicago: University of Chicago Press. Bayer, J. 1990. Directionality of Government and Logical Form: A Study of Focusing Particles and Whscope. Habilitationsschrift University of Konstanz/Max Planck Institute Nijmegen. Bennis, H. and T. Hoekstra. 1989. Generatieve Grammatica. Dordrecht* Foris Publications. Bierwisch, M. This volume. "Verb cluster formation as a morphological process". Presented at the Workshop When Words Happen to be Phrases. Max Planck Institute 1989. Bok-Bennema, R. and A. Croughs-Hageman. 1980. Verb-Raising en Struktuurbehoudendheid. Unpublished manuscript University of Amsterdam. Cardinaletti, A., G. Cinque and G. Giusti (eds.). 1988. Constituent Structure. Dordrecht: Foris Publications. Chomsky, N. 1957. Syntactic Structures. The Hague: Mouton. Chomsky, N. 1981. Lectures on Government and Binding. Dordrecht: Foris Publications. Chomsky, N. 1986. Barriers. Cambridge, Massachusetts: MIT Press. Chomsky, N. 1988. Some Notes on Economy of Derivation and Representation. Unpublished manuscript. MIT, Cambridge, Massachusetts. Di Sciullo, A. and E. Williams. 1987. On the Definition of Word. Cambridge, Massachusetts: MIT Press. Evers, A. 197S. The Transformational Cycle in Dutch and German. Distributed by Indiana Linguistics Club. Haan, G. de. 1979. Conditions on Rules. Dordrecht: Foris Publications. Haan, G. de and F. Weerman. 1986. "Finiteness and verb fronting in Frisian". In H. Haider and M. Prinzhorn (eds.), 77-110. Haegeman, L. and H. van Riemsdijk. 1986. "Verb projection raising, scope, and the typology of rules affecting verbs". Linguistic Inquiry 17(3), 417-466. Haider, H. 1986. "V-second in German". In H. Haider and M. Prinzhom (eds.), 49-77. Haider, H. and M. Prinzhom (eds.). 1986. Verb Second Phenomena in Germanic Languages. Dordrecht: Foris Publications. Hellan, L. and K. Koch Christensen (eds.). 1986. Topics in Scandinavian Syntax. Dordrecht: Reidel. Hoekstra, J. and L. Maräcz. 1989. "On the position of inflection in West-Germanic". To appear in Working Papers in Scandinavian Syntax, University of Lund. Holmberg, A. 1986. Word Order and Syntactic Features in the Scandinavian Languages and English. Unpublished Doctoral dissertation. University of Stockholm, Stockholm. Huybregts, M.A.C. 198S. On Reanalysis and Covalency. Unpublished manuscript. Tilburg University. Koopman, H. 1984. The Syntax of Verbs. Dordrecht: Foris Publications. Köster, J. 1986. The Relation between Pro-drop, Srambling and Verb Movements. Unpublished manuscript Groningen University. Köster, J. 1987. Domains and Dynasties: the Radical Autonomy of Syntax. Dordrecht: Foris Publications. Lasnik, H. and J. Kupin. 1977. "A restrictive theory of transformational grammar". Theoretical Linguistics 4, 173-196. Lenerz, J. 1984. "Diachronic syntax: verb position and comp in German". In J. Toman (ed.), 103-132. Muysken, P. and H. van Riemsdijk (eds.). 1986. Features and Projections. Dordrecht: Foris Publications. Olsen, S. 1984. "On deriving V-l and V-2 structures in German". In J. Toman (ed.), 133-164. Pesetsky, D. 1982. Paths and Categories. Unpublished dissertation. MIT, Cambridge, Massachusetts. Platzack, C. 1986. "The position of the finite verb in Swedish". In H. Haider and M. Prinzhom (eds.), 27-49. Platzack, C. 1987. "The Scandinavian languages and the null subject parameter". Natural Language and Linguistic Theory 5(3), 377-402.

Head Movement and the Relation between Morphology and Syntax

161

Pollock, J.-Y. 1989. "Verb movement, universal grammar, and the structure of IP". Linguistic Inquiry 20, 365-424. Reinhart, T. and E J . Reuland. To appear. "Anaphors and logophors: an argument structure perspective". In J. Köster and E J . Reuland (eds.), Long Distance Anaphora. Cambridge: Cambridge University Press. Reuland, EJ. 1982. "Why count your auxiliaries in Dutch?". In J. Pustejovsky and P. Sells (eds.). Proceedings ofNELS 12, GLSA, University of Massachusetts, Amherst, Massachusetts, 221-235. Reuland, E.J. 1983. "On mixing configurational and nonconfigurational properties". In S. Hattorf and K. Inoue (eds.), Proceedings of the Xlllth International Congress of Linguists, Tokyo 1982, 479-484. Reuland, EJ. 1985. "Identifying null subjects in Dutch". Unpublished manuscript. Groningen University. Reuland, E J . 1986. "A feature system for the set of categorial heads". In P. Muysken and H. van Riemsdijk (eds.), 41-88. Reuland, EJ. 1988. "Phrase structure and the theory of levels". In A. Cardinaletti, G. Cinque and G. Giusti (eds.), 263-287. Reuland, EJ. and W. Kosmeijer. 1988. "Projecting inflected verbs". In W. Abraham (ed.), Groninger Arbeiten zur Germanistischen Linguistik 29. University of Groningen, Groningen. Riemsdijk, H. van. and E. Williams. 1981. "NP-structure". The Linguistic Review 1, 171-217. Stowell, T. 1981. Origins of Phrase Structure. Unpublished Doctoral dissertation. MIT, Cambridge, Massachusetts. Taraldsen, T. 1986. "On verb second and the functional content of syntactic categories". In Haider, H. and M. Prinzhorn (eds.), 7-27. Toman, J., ed. 1984. Studies in German Grammar. Dordrecht: Foris Publications. Thräinsson, H. 1986. "On Auxiliaries, Aux and VPs in Icelandic". In L. Hellan and K. Koch Christensen (eds.), 235-267. Weerman, F. 1989. The V2 Conspiracy. Dordrecht: Foris Publications.

Yearbook of Morphology 3 (1990), 163-171

Comments on the paper by Reuland Jaklin Kornfilt

In general, I am sympathetic to the model proposed in this paper; the idea of complex categories formed pre-syntactically and projected according to principles of X'-syntax can lead to interesting insights into a variety of phenomena, as the author shows. In the present article, phenomena concerning typology of word order are discussed; but other kinds of facts, for example the syntax of "restructuring" constructions, can also be treated in insightful ways using complex categories (cf. Bayer & Kornfilt, 1990). Here, I will not attempt to criticize the paper, instead, I shall discuss some consequences of this model for some additional syntactic phenomena. These observations as well as some clarificational questions I shall ask will constitute an invitation to the author, and to other syntacticians interested in the morphologysyntax interface, to spell out and tighten up a model crucially incorporating complex categories, and to look at certain syntactic phenomena not yet considered in this new approach. One basic question to be asked is under what circumstances complex categories are formed pre-syntactically. The approach proposed in the present paper is morphology-based; a complex category is formed in the lexicon (assuming morphology is in the lexicon) where the material it dominates is a word (or is, in some sense, word-like). But this presupposes that we know whether a given morpheme (sequence) forms a word or not. In the examples discussed in Reuland's paper, I take it that the wordstatus of all inflected verbs in question is uncontroversial. But this might not always be the case, and phonological evidence might not always be clear and sometimes even be contradictory. Consider, for example, the Yes/No question morpheme in Turkish, a richly inflected SOV language: (1)

a.

b.

Osman problem-i ?oz-du mill problem-Acc. solve-past Q 'Did Osman solve the problem?' Osman problem-i mi ?oz-du? 'Was it the problem that Osman solved?'

Author's address: Dept. of Foreign Languages and Literatures, Syracuse University, Syracuse, NY 13244,

164 J. Kornfilt c.

Problem-i Osman mi $oz-dii? 'Was it Osman who solved the problem?'

The question morpheme can have the whole sentence in its scope, as in (la), but it can also form constituent questions, as in (lb) and (lc). Note that the vowel of the question morpheme harmonizes with the last vowel of the preceding "word"; but can we be confident that it forms a word with that preceding sequence? The answer might be yes, since elsewhere in the language, vowel harmony takes a phonological word as its domain. However, this question morpheme does not carry stress; the last syllable of the preceding "word" does. Yet, elsewhere in the language, word-level stress is final (leaving out of consideration some exceptions). Stress, then, seems to argue that the question morpheme is not part of the same word as the preceding morpheme (sequence). One could say, as has been done in traditional grammars, that the question morpheme is a "postclitic" - part of the word, yet not quite. But this is begging the question of the dual nature of such elements. (Note, incidentally, that the standard orthography treats such "particles" as separate entities.) But our syntactic treatment of a sequence of X+Q might depend on our decision on treating such a sequence as a word or not; if not, there wouldn't be any reason to posit a complex category at D-Structure (say, of the form X°/Q°, if we want to treat the question morpheme as a quantifier-like operator); but if we say that the sequence is a word, we might postulate such a complex category and project it. As we see, even a relatively clear-cut ingredient of this model, namely, that of identifying a string as a word, is not always straightforward. Now suppose that we did have a word. What kind of empirical consideration might allow us to project parts of the word separately, and what type of fact might force us to project the complex category of the word simultaneously? Let us look at one case in point from Dutch, discussed in Reuland's paper. In order to argue that the parts of V+I do not project separately in Dutch, Reuland first assumes that, if they did project separately, there would have to be an intervening adverb position (marked as a in Reuland's trees, e.g. his (9)): (2)

CP

Comments on the paper by Reuland

165

Reuland then shows us that in Dutch, such a position is never realized by an element that is not part of the verb (Reuland* s 18): (3)

a. *dat [ [Cindy] [,. [yp [yp een boek gelezen]yp [„ niet] a ]yp [heeft], ],. ]jp that C. a book read not has "That Cindy has not read a book' b. *dat [ [Cindy] [ r [yp [yp een boek gelezen]yp [ (dolcire) + en

—> calcistico + ic' —> addolcire + a- '

232 Book Reviews These examples show that the first step in the formation process creates 'possible but non-existing words' (cf. Scalise 1988), for which synonyms exist: calciatore 'soccerplayer' and dolcificare 'to sweeten', respectively. All of Rainer's possible words are instances of the first interpretation, and must not be confused with 'possible but nonexisting' forms. The last chapter of the book contains a good review of what Rainer calls I'Ipotesi separazionista 'the separation hypothesis' with respect to the relation between form and semantics in word formation. The author analyzes various literature on this topic, and distinguishes between the major and the minor separationists. According to Rainer, major separationists are those linguists whose work is mainly devoted to the separation hypothesis, such as Zwanenburg (1980 and other works in press), Szymanek (1985), and Beard (1981,1986). Minor separationists, on the contrary, are those who deal with this issue only briefly: Rose (1973), Jackendoff (1975), Dressier (1977,1980), Lieber (1981), and Bauer (1983). He concludes, also based on his own research, that the arguments proposed in favor of the separation hypothesis are not convincing. Finally, Rainer applies the 'domain hypothesis' of van Marie (1985, 1986) to Italian names of quality, which is, in Rainer's view, the most promising theory on the problem of blocking. In conclusion, this book offers an excellent description of the formation of Italian quality names. The organization of the adjectives forming the base of WFRs in 'inverse' order is very useful, as is the alphabetical ordering of suffixes introduced by word formation rules. The need to go back from chapter 3 (the list of rules) to chapter 2 (the list of adjectives and corresponding names) to review the data subject to the same rule is slightly inconvenient, but this division serves Rainer's aim to generalize the bases of a WFR. In spite of the above criticisms, Rainer's book is an exhaustive and careful study on Italian word formation. The study of the Italian language lacks works of this type; Rainer's contribution is good and will be very useful to anyone interested in morphology. Antonietta Bisetto

REFERENCES Allen, M. 1978. Morphological Investigations. Ph.D. Dissertation. University of Connecticut Aronoff, M. 1976. Word Formation in Generative Grammar. Cambridge: MIT Press. Bauer, L. 1983. English Word-Formation. Cambridge: Cambridge University Press. Beard, R. 1981. The ¡ndoeuropean Lexicon: A Full Synchronic Theory. Amsterdam: Elsevier. Beard, R. 1986. On the Separation of Derivation from Morphology. Toward a Lexeme!Morpheme Based Morphology. Bloomington: IULC. Dressler, W. 1977. "Elements of a polycentristic theory of word formation". Wiener Linguistische Gazette 15, 13-32.

Book Reviews

233

Dressler, W. 1980. "Universalien von Agens-Wortbildungen". G. Bretschneider and C. Lehmann (eds.), Wege zur Universalienforschung: sprachwissenschaftliche Beiträge zum 60. Geburtstag von Hansjakob Seiler. 110-114. Tubingen: Narr. Jackendoff, R. 1975. "Morphological and semantic regularities in the lexicon". Language 51, 639-671. Lieber, R. 1981. On the Organization of the Lexicon. Ph.D. dissertation. Bloomington: IULC. Marie, J. van. 1985. On the Paradigmatic Dimension of Morphological Creativity. Dordrecht: Foris Publications. Marie, J. van. 1986. "The domain hypothesis: the study of rival morphological processes". Linguistics, 24, 601-627. Rose, J. 1973. "Principled limitations on productivity in denominal verbs". Foundations of Language 10, 509-526. Scalise, S. 1983. Morfologia Lessicale. Padova: Clesp. Scalise, S. 1984. Generative Morphology. Dordrecht: Foris Publications. Scalise, S. 1988. "The notion of 'head' in morphology". Yearbook of Morphology 1, 229-245. Siegel, D. 1977. "The adjacency condition and the theory of morphology". M. Stein (ed.), Proceedings of the 8th Annual Meeting of NELS. 189-197. Amherst. Szymanek, B. 1985. English and Polish Adjectives. A Study in Lexicalist Word Formation. Lublin: Univ. of Lublin. Williams, E. 1981. "On the notions 'lexically related' and 'head of a word'". Linguistic Inquiry 12, 245-274. Zwanenburg, W. 1980. "Form and meaning in morphology". Quaderni di Semantica, 327-337. Zwanenburg, W. To appear. "Derivation constructions and argument structure". Acta Linguistica.

234

Book Reviews

Mervyn F. Lang, Spanish Word Formation. Productive Derivational Morphology in the Modern Lexis. Routledge: London and New York, 1990, xi + 270 p., £ 35,—. Croom Helm Romance Linguistics Series. Romance Philology, which was in the forefront of linguistic theorizing during the diachronico-dialectological period, has undergone a steady process of provincialization in the last thirty years or so, due mainly to its reluctance to cope with the stormy developments experienced by general linguistics and, more specifically, in grammatical theory in the wake of the "Chomskyan revolution". In the light of this regrettable state of affairs, the importance of initiatives such as the foundation of the journal Probus (Foris Publications; vol. 1 (1989)) or the Croom Helm Romance Linguistics Series, whose purpose is to "endorse the view that a General Linguistics that ignores the always rich and often unique data of Romance is as impoverished as a Romance Philology that turns its back on the insight of linguistics [sic] theory" (Editorial statement), cannot be exaggerated. In this spirit, the book under review aims at a mutual fecundation of Spanish word formation1 and the lexicalist theory of morphology. Although both aspects are interwoven in Lang's book, I will separate them here for expository purposes. Lang's ambition is not to describe all word formation devices of modern Spanish, but rather to concentrate on only the most important productive types. These are classified in a traditional manner as rules of compounding (65-90), "emotive" suffixation (91-122), non-emotive suffixation (123-149), adjectivization and adverbialization (150-162), verbalization (163-167), prefixation (168-184), and a remainder of "miscellaneous procedures" (185-201). The description of the single affixes is generally reduced to a presentation of three examples for each formation type plus a short comment on its productivity, as well as its combinatorial, semantic, phonological, or stylistic properties. The domains of the single affixes are not delimited precisely, i.e., Lang's never explicitly stated rules are not generative rules that would enable the reader to decide in all or even most instances whether a particular formation is well-formed or not. This, to some extent, is admittedly a consequence of the very nature of word formation rules, but even more important is the lack of thorough studies on the multivarious restrictions that single word formation rules may be subjected to.2 Lang's descriptions, by the way, might have gained in precision if he had not ignored the, by now, abundant literature on Spanish word formation. Out of approximately 500 books and articles on the subject, Lang's bibliography contains no more than 40, and even such important and readily available books as Gauger (1971), Egea (1979), Faitelson Weiser (1980), Nord (1983),3 Bustos Gisbert (1986), or Laca (1986), have not been included. Unfortunately, Lang's description also contains quite a number of errors of varying magnitude: p. 16: despacito has a morpheme boundary after despac-, not despa-, inversion after invers-, not inver-; p. 25: -izar is not an "infinitive morpheme"; p. 26: the e-epenthesis in tosquedad is implausible; p. 26, passim: the theme vowel should not be analyzed as part of the suffix; p. 27: elefantidsico derives from elefantiasis,

Book Reviews

235

not from elefante; p. 28: perspicuidad derives from perspicuo, not from perspicaz; p. 37: -ez is not "synchronically unproductive"; Nord's corpus contains hipiez and rubiez, my own collection enanez, marcadez, marronez and sordez; p. 56: the bracketings [anti- + [español]A + -ismo]N etc. do not reflect the fact "that these formations are built up in single stages, each derivation containing a previous one and the structure always branching in a binary fashion"; p. 58-59: the statement "Prefixed verbs show a strong preference for nominalizations in -miento" is too general; there is an almost perfect correlation between -miento and a- and en-, but des- and re- behave differently; p. 69: clave etc. are not evolving towards suffixal, but adjectival status; p. 70, 71: the head of radio patrulla is radio, of cabizbajo bajo-, p. 76: V+N-compounds are not "invariably masculine"; names referring to human beings may be masculine or feminine, and with animal and plant names gender is sometimes feminine if the hyperonym is feminine; p. 98-99: it is not true that "current tendencies are to prefer the radical change option" with respect to vowel alternation in Spanish derivation; rather, monophtongizations like bueno > bonísimo are relics, the modern trend being to preserve the diphthong; p. 127: the i of -idad is not "epenthetic"; p. 145: -ido derivations like ladrido do not "suggest prolonged or repeated noise", but on the contrary "das einzelne Element eines zusammengesetzten Geráuschs" (Gauger 1971: 56); p. 170: the statement "an- appends to vowel-commencing bases" is only true for the established formations; in neologistic use we have a- even before vowels: ahistórico, aeuropeo, a-ideológico, etc.; p. 172: de- is not "constrained to established lexical items"; it is also used productively before bases commencing with s: desovietizar, etc.; p. 194: New Zealand, according to all dictionaries I have consulted, is Nueva Zelanda in Spanish, not Neozelandia; p. 238: eurovisivo is derived from Eurovisión, not from visivo; etc. Let us now turn to the appraisal of the general linguistic side of Lang's book, consisting of an introduction to "the theoretical and historical framework" (1-62), a "conclusion" (202-208), as well as several remarks dispersed throughout the book. Lang deliberately adopts an eclectic approach to word formation (62), since in his review of past and present theories he finds them all lacking in some respect. Now, while this may be a defensible position in general terms, I remain unconvinced by most of Lang's arguments against "post-generative4 theorists" (42, passim) like Aronoff, Halle, or Scalise (they are the ones he mentions most often). Among the hypotheses he does not want to subscribe to are the Unitary Base Hypothesis, the Binary Branching Hypothesis, and Blocking. The Unitary Base Hypothesis is certainly the most objectionable of the three, but Lang's case against it (55-6, passim) is weakened by the following: on the one hand, he interprets it in the strictest possible sense, i.e., that a word formation rule may only refer to bases of one syntactic category, which does not correspond to Aronoff's intention, and on the other hand, does not distinguish between productive rules and cases of affix generalization like alcaldable or exceptions like impago. His rejection of the Binary Branching Hypothesis (57, passim) is based on the socalled "parasynthetic" formations of Spanish as well as on suffixations like riachuelo

236

Book Reviews

'streamlet', consisting of a base rio 'river', the pejorative suffix -acho and the diminutive suffix -uelo. The problem with the latter type is that riacho is not an existing Spanish word, and probably not even a possible one either, since -acho is unproductive. Such cases of affix coalescence, which are common in the history of word formation systems, do not, to my mind, invalidate the Binary Branching Hypothesis. What is going on, instead, is that a sequence of two suffixes, one of which has become unproductive, is reinterpreted as one suffix (cluster), in this case -achuelo. The import of parasynthetic formations like aclarar 'to make clear', from claro 'clear', where neither aclaro nor clarar are actual words, moreover, is far from straightforward. Dell (1970: 200-202) proposes the structure [a-[c/aro]A]v, considering -ar an inflectional infinitive ending. The main problem with this proposal is that it does not extend to similar cases like enriquecer 'to make rich', from rico 'rich', where -ec(er) is unequivocally a derivational suffix. Scalise (1983: 202-208) and Alcoba Rueda (1987) try to achieve binarity by postulating possible words as intermediate steps: [a-[[claro]K-ar]v]v (S.), [a-[[claro\\k-ar]w (A.R.). These "overgenerating" solutions are defective in that they systematically introduce nonoccurring and, at least in the case of Alcoba, impossible words, as well. The issue of parasynthetic formations is thus to be considered unresolved at the moment. Lang's critique of the Binary Branching Hypothesis rests on the all but obvious tacit assumption that the structure of parasynthetic verbs is ternary. One possible solution - which I cannot defend here in detail - that allows us to avoid both ternarity and dubious intermediate steps would consist of distinguishing hierarchical and linear structure. Parasynthetic formations would then be binary on the hierarchical level in accordance with their meaning, while the linear realization of 'make' and similar derivational meanings would be carried out by a discontinuous morph, in our case an affix cluster consisting of a prefix and a (zero) suffix, very much in the vein of -achuelo in riachuelo. Blocking is dismissed (61, passim) because of the existence of a considerable number of doublets in Spanish (dictionaries). This in itself, however, is not sufficient to reject a refined version of the theory of blocking, as the one presented in Rainer (1988). Besides synonymy, other factors have to obtain in order for blocking to take place: the two words must be part of the same speaker's lexicon/competence, and the blocking word has to have a certain frequency. Now, doublets in Spanish are often regional or stylistic variants. Furthermore, one of the doublets recorded in dictionaries is often more or less obsolete, and the blocking word may be too infrequent to be able to exert a blocking function. Since Spanish dictionaries are unreliable sources in this respect, we will have to await extensive empirical investigation before we can evaluate the impact of the Spanish data on the theory of blocking. Anyway, it would be an objectionable research strategy to discard theories with a certain amount of empirical support in the face of unclear evidence such as Lang's.

Book Reviews

237

Summing up, one has to say that Lang's book is not a model of the kind of synthesis between Romance and general linguistics that the Croom Helm series advocates. The results of general linguistic interest are rather poor. On the other hand, the general linguistic ingredients will not be very helpful or illuminating to the average hispanistic reader. The most positive feature of Lang's book is the attention paid to current usage, to productivity. For this reason, it could certainly be used as a text-book in undergraduate courses, since there is no comparable book on the market. I think that the book would have made a better impression if the author had restricted himself to such a didactic perspective congenial to him. Franz Rainer

NOTES 1. The subtitle is misleading since the book not only deals with "Derivational Morphology", but with all types of word formation, including compounding, clipping, etc. 2. Cf. Rainer (1989) for an attempt of such a study. 3. Nord's book also concentrates on productive patterns and presents a wealth of neologisms. 4. Lang obviously intended to say "post-transformationalist". This terminological blunder is symptomatic of the author's rather superficial acquaintance with theoretic literature. Another indication of this is the fact that Morris Halle is written with an accent on the final e throughout the book: Hall£\ Still other indicators are his unorthodox use of the term morphonobgical (27, passim), the form paradigmacity for paradigmaticity (195, 199), his partial endorsement of the transformationalist hypothesis (besides the lexicalist one!; p. 8,65-6,82,124,131), as well as strange or totally wrong characterizations of generative phonology (10), allomorphy rules (15), the role of constraints (33), the 'is a' condition (83), the adjacency condition (128), the confusion of Spanish "backformation" of the type cesar > cese (action nominal) with what is normally referred to as back-formation, e.g. editor > edit (149), etc.

REFERENCES Bustos G.E. de. 1986. La Composición Nominal en Español. Salamanca: Ed. de la Universidad. Dell, F. 1970. Les Régies Phonologiques Tardives et la Morphologie Dérivationnelle du Français. Ph.D. MIT. Egea, E. 1979. Los Adverbios Terminados en -mente en el Español Contemporáneo. Bogotá: Publicaciones del Instituto Caro y Cuervo. Faitelson-Weiser, S. 1980. Les Suffixes Quantificateurs de l'Espagnol Moderne. Paris: Editions Hispaniques. Gauger, H.-M. 1971. Untersuchungen zur spanischen undfranzözischen Wortbildung. Heidelberg: Winter. Laca, B. 1986. Die Wortbildung als Grammatik des Wortschatzes. Untersuchungen zur spanischen Subjektnominalisierung. Tübingen: Narr. Nord, C. 1983. Neueste Entwicklung im spanischen Wortschatz. Rheinfelden: Schäuble. Rainer, F. 1988. 'Towards a theory of blocking: the case of Italian and German quality nouns". Yearbook of Morphology 1, 155-185. Rainer, F. 1989.1 Nomi di Qualitä nell'Italiano Contemporáneo. Vienna: Braumiiller. Scalise, S. 1983. Morfología Lessicale. Padua: Clesp.

238

Book Reviews

Alfred Holl, Romanische Verbmorphologie und relationentheoretische mathematische Linguistik. Tübingen: Niemeyer 1988, 346 S. DM 132. (Linguist. Arbeiten 216) Any good grammar or any expanded dictionary satisfies the following requirement: If one knows the citation form of a verb and the appropriate paradigm (normally there is an indication given with the citation form) then it is possible to derive all morphologically correct forms. The book under review aims to give a mathematically formalized version of this derivation process. It covers the following languages: Latin, French, Italian, Spanish, Catalan, Portuguese, and Romanian. Although this is done with a great amount of care, several critical remarks are nevertheless necessary. One may ask which new insights this method can give and, furthermore, what linguistics in general, or Romance linguistics, in particular, will gain by it. These questions will be commented upon at the end of my review. First of all, the contents of the book will be reviewed briefly. The first chapter is devoted to some preliminaries on the philosophy of science. The author tries to clarify his own ideas about science in roughly the following words: the justification of a theory can be seen in its application, which gives feedback from reality and human experience (p. 11). The guide lines that are given there for the construction of a scientific theory may be seen as useful landmarks in this context. On pp. 13 and 14 the author distinguishes between the concepts "model equivalent" ("modelläquivalent") and "reality equivalent" ("wirklichkeitsäquivalent"). But the example given from physics is not very illuminating, since a change in coordinates does not change the mathematical model and, therefore, implies the same "reality". The short hint to quantum mechanics on p. 17 needs some clarification: The uncertainty relations of Heisenberg are consequences of the principles of quantum mechanics (more precisely, they are the result of the fact that certain operators do not commute). A common interpretation attributes them to the disturbance caused by the process of measuring. Furthermore, an attempt is made to link the somewhat unsatisfactory state of the art of morphological research with the neglect of the word concept. The second chapter is devoted to the development of a general frame for formalized morphological models. It starts with the models given by Ch.F. Hockett in his paper of 1954: the item-and-arrangement model, the item-and-process model, and the word-and-paradigm model. Although the formalized explanation of another researcher's model is not without problems, this chapter is by far the most interesting of the whole book. It makes fairly clear which decisions must be made if a morphological model is to be built. These involve decisions on the basic syntactical units, on the number of key forms ("Schlüsselformen"; these forms must be known to derive all morphologically correct forms with the help of mechanical rules), as well as decisions on the nature of the rules of synthesis. On p. 47 the author states that it would be logically inconsistent to allow several keyforms and process rules (which, roughly speaking, change the stem of a word like 'take': 'took') in one model. I think it would be economical as well as reasonable (e.g. to deal with

Book Reviews

239

variation in accent). The author decides for a model which is built on the assumption of several keyforms and concatenation rules only: "Unser Ansatz ist eine stark formalisierte und hochgradig abstrahierte Ausprägung des Word-and-ParadigmTypes" ["Our approach is a strongly formalized and abstract version of the wordand-paradigm type of model"] (p. 32). The third chapter serves as a broad description of the model just mentioned. Although it is called a mathematically formalized model, I would emphasize that such a task cannot, strictly speaking, be attributed to mathematics. There are carefully given definitions, as well as a number of symbols, but they remain in the realm of descriptive theory. One misses new results which are deduced from the axioms in a mathematical manner. There is an awkward (and possibly superfluous) long list of neologisms too. Any reader who covers this marathon mile of more than 100 pages can be congratulated on his or her perseverance. Personally, I fear that this is not the right way to gain new friends for the application of mathematical methods. This approach does not show the true fascination and beauty of mathematics. The model is based on seven axioms concerning the following areas: word forms ("Wortformenaxiom"), space of inflexions ("Flexionsraumaxiom"), parts of speech ("Wortartenaxiom"), elimination of synonyms ("Synonymieneliminationsaxiom"), invariance of length ("Längeninvarianzaxiom"), reduction to stems ("Stammreduktionsaxiom"), and reduction to constants ("Konstantenreduktionsaxiom"). If you know that (Ind. Pf. 1. S.) is a "Flexionskategorit" to which the phonographemic word forms and belong, you have already grasped the nature of this kind of modelling. Though some mathematicians include zero in the set of natural numbers, such a procedure is rather strange in a book on linguistics (p. 71)! If you were to count your luggage "zero, one, two, three", I am sure you would think that one piece of your luggage has gotten lost. On p. 104 and p. 112, an ascending chain condition and a maximal condition is compared with the Noetherian condition in the mathematical theory of rings. This is not only quite useless for the average linguist but misleading as well. The set of all objects called "Flexionskategorit" is finite and, therefore, such a condition is automatically satisfied. It is clearly not obvious that the set of all maximal domains of inflections forms a partition of the space of inflections. At least twice (p. 112, p. 134), the equivalence relation to be defined is, in fact, stated incorrectly: Since an equivalence relation is reflexive (p. 69), one must not suppose b* b' in its definition. It might be interesting to point out that the axiom of length reduction (p. 130) is not satisfied for the conjugation in Classical Greek (Ind. Pf. Pass. 3. PI). Furthermore, it is inappropriate to postulate (p. 132) that the space of inflections must be partitioned in so-called n-lengthinvariant subspaces (n > 1) in a non-trivial way: this would exclude the possibility of the inflection of a word class being entirely synthetic. This does not apply to the conjugation in Latin or in the Romance languages, but is true for the declension of Latin nouns. For morphological research, the discussion of the "Stammreduktionsaxiom" and the "Konstantenreduktionsaxiom" is more interesting. Here enter the basic problems of

240

Book Reviews

morphological analysis, namely, the definitions of morphemes, stems, roots, and the like. At this stage, I found some (axiomatically) undefined strings like < / • g> called the "totale Averbo-Konstante" belonging to German (p. 159) or < bin ge-en > called the discontinuous ending of German (p. 140). It is quite clear what the dot means, but such strings have not been defined. At this point, it would also have been appropriate to mention the morphological work on algebraic linguistics by Solomon Marcus and his followers. The concept of an "Analogiebasis" ["analogy base"] is introduced on p. 163, but I, at least, needed the examples on p. 185 to understand it. Strangely enough, up to this point the model is still too complicated. Therefore, a list of exceptions is necessary to reduce it to reasonable size. In chapter 4, the author applies his model to conjugation in the seven languages mentioned at the beginning. On pp. 179-181, some numerical constants are introduced, which serve as a measure of the complexity of verbal morphology. Unfortunately, this interesting topic is not pursued further. There is also a certain number "Zuordnungsformenzahl" (p. 180) which is not entirely clear to me. The author takes more than 120 pages to complete the so-called morphographemics of the verbal complex of the languages. I do not know whether all the data given are correct. But this is not important, since the model could easily be corrected. The treatment of the Latin deponent and semi-deponent verbs appears problematic. The list of key forms does not contain any deponent verb from the a-conjugation. One might therefore think that the model derives morphologically active forms for a deponent verb! Now we may return to the beginning. The presentation of the conjugational patterns in this form does not seem appropriate for educational purposes. A full list of forms is more easily learned than the synthesis rules of the model. The number of listed key forms clearly indicates the difficulty of learning a foreign language. On the other hand, I doubt if the given rules of synthesis contribute to the problem of a possible psychological representation of language rules. Since the presentation is strictly synchronic, it is of no significance for historical linguistics. The aspect of analyzing language by formal modelling is, therefore, the main contribution of the book. The suspicion that a mathematically formal description of the morphology of inflecting languages is inevitably clumsy is demonstrated in a convincing manner. Fritz Schweiger

BOOK NOTICES English Linguistics. Journal of the English Linguistic Society of Japan. Volume 6 of this annual journal contains the following contributions to English morphology: Miho Nishio, 'Feature specification of passive participles'; Tohru Noguchi, 'Morphologically related predicates: ergatives and middles in English'; Wayne P. Lawrence et al., 'English en-, prefix or suffix'. The journal is published by the English Linguistic Society of Japan at Kaitakusha Publishing Co., 5 Kanda-Jinbocho 2-chome, Chiyoda-ku, Tokyo 101, Japan. G.B. W.U. Wurzel (hrsg.). Studien zur Morphologie und Phonologie III. Berlin, DDR: Akademie der Wissenschaften, 1988, 117 p. (Linguistische Studien, Reihe A, Arbeitsberichte 188) This volume contains the following morphological papers: A. Bassarak, 'Morphologisierungstendenzen im Türkischen'; A. Bassarak, 'Zur Natürlichkeit türkischer Tempusformen'; A. Bittner, 'Is anything 'more natural'? Considerations on establishing a hierarchy of naturalness principles'; D. Bittner, 'Motivationsstrukturen im Flexionsverhalten der nhd. Substantive - Vorschlag eines Modells'; W.D. Klimonow, 'Die Rolle des diagrammatischen Ikonismus beim Aufbau von Flexionsformen'; P. Thiele, 'Thesen zum Tempus-Modus-Aspekt-System der portugiesischbasierten Kreolsprachen Westafirikas (unter besonderer Berücksichtigung des Kapverdischen)'; W.U. Wurzel, 'The mechanism of inflection: lexicon representations, rules, and irregularities'; W.U. Wurzel, 'The structuralist heritage in natural morphology'. G.B.

Publications received

[The listing of a publication here does not exclude the possibility of its being reviewed in a next volume] Barry J. Blake, Relational Grammar. London and New York: Routledge, 1990, xiv + 198 p., £ 9.95 (Croom Helm Linguistic Theory Guides) Robin Clark, Thematic Theory in Syntax and Interpretation. London and New York: Routledge, 1990, vii + 264 p., £ 35,— (Croom Helm Linguistics Series) Hartmuth Günther (hrsg.), Experimentelle Studien zur deutschen Flexionsmorphologie. Hamburg: Helmut Buske Verlag, 1989, 210 p., DM 38,— (Beiträge zur Sprachwissenschaft, 2) John T. Jensen, Morphology. Word Structure in Generative Grammar. Amsterdam and Philadelphia: John Benjamins, 1990, x + 210 p., Hfl. 40,— (pb) (Current Issues in Linguistic Theory, 70) Eva Leitzke, (De)nominale Adjektive im heutigen Englisch. Untersuchungen zur Morphologie, Syntax, Semantik und Pragmatik von Adjektive-Nomen-kombinationen der Typen 'atomic energy' und 'criminal lawyer'. Tübingen: Niemeyer, 1989, x + 200 p., DM 78,— (Linguistische Arbeiten, 221) Yakov Malkiel, Diachronic Problems in Phonosymbolism. Edita and Inedita, 1979-1988. Vol. 1. Amsterdam and Philadelphia: John Benjamins, 1990, 274 p., Hfl. 150,— David J.Weber, H.Andrew Black, Stephen R. Mc Connel, Ample: a Tool for Exploring Morphology. Dallas: Summer Institute of Linguistics, 1988, xii + 252 p., U.S. $ 20,—. (Occasional Publications in Academic Computing, 12)

Notes to contributors Contributions should be in English, t h o u g h in certain exceptional cases F r e n c h or G e r m a n m a y b e a c c e p t e d , CONTRIBUTORS WHOSE NATIVE LANGUAGE IS NOT ENGLISH s h o u l d h a v e t h e i r

manuscripts gone over by a native speaker of the language before submission. Three copies of all manuscripts must be submitted to the Editor (address o n inside cover). The only constraint on the length of articles is that they should be no longer t h a n they need to be - a u t h o r s must write concisely Manuscripts must be typed double-spaced with wide margins, o n one side of D I N A4 p a p e r (or similar), a n d should be reasonably divided into sections a n d , if necessary, subsections. The conventions of the present volume should be followed carefully Line drawings (called 'Figures' in the text) and trees must be reproducible originals a n d should be submitted on separate sheets, carefully n u m b e r e d a n d labeled. C a p t i o n s should be typed on a separate sheet and placed at the end of the manuscript. Trees should be n u m b e r e d as if they were part of the linguistic examples, i.e not separately Tables should be n u m b e r e d consecutively a n d titled, and must be referred to in the text. F o o t n o t e s must be kept to an absolute m i n i m u m . They should be supplied on a separate sheet of p a p e r at the end of the manuscript. References should c o n f o r m t o s t a n d a r d scientific practice References are cited in the text by giving the n a m e of the author(s) a n d year of publication (in parentheses), e.g. ' as m e n t i o n e d by C h o m s k y (1957)' ' as has been argued ( C h o m s k y 1957. 18-21)' All citations in the text must be listed fully in a Reference section at the end of the manuscript, by alphabetical o r d e r of a u t h o r s , with complete bibliographical details. J o u r n a l titles must be given in full and underlined, as must b o o k titles. T h e Reference list should be presented as follows C h o m s k y , N 1957 Syntactic Structures The H a g u e M o u t o n C h o m s k y , N and M Halle 1968 The Sound Pattern of English, 2nd edn New York H a r p e r and R o w Fillmore, C . J 1968. " T h e Case for C a s e " In E. Bach a n d R T H a r m s (eds.), Universals in Linguistic Theory New York H o l t , Rinehart and Winston, 1-80 A F o d o r , J and T G Bever 1965 " T h e Psychological Reality of Linguistic S e g m e n t s " Journal of Verbal Learning and Verbal Behavior 4, 414-420. Book Reviews. We will normally invite reviews, but unsolicited reviews m a y be accepted. Reviews will be of three types. ' B o o k Notices' of 500-1,000 w o r d s , ' B o o k Reviews' which deal m o r e extensively with one particular publication (up to 3,000 words), a n d 'Review Articles' which may include reviews of m o r e t h a n one b o o k , a n d which will be treated as a normal article Refereeing. All articles and review articles will be refereed by at least t w o m e m b e r s of the Board of Editors or the Board of C o n s u l t i n g Editors. We will i n f o r m the a u t h o r ( s ) of a decision as soon as possible a f t e r receipt of the article Replies. Articles of special theoretical interest may be sent, with the permission of the author(s), to Consulting Editors or others, for an invited c o m m e n t a r y to a p p e a r with the article We will encourage short replies t o articles already published. Corrections. A u t h o r s are asked to check their manuscripts very carefully before submitting them to prevent delays at the proof stage A u t h o r s will receive p r o o f s for correction which must be returned by dates determined by the publication schedule Offprints. Twenty-five offprints of each p a p e r will be sent free of charge (when there is m o r e than one a u t h o r , the offprints will be sent to the first-named a u t h o r for distribution). C o n t r i b u t o r s to special issues will instead receive six c o m p l e m e n t a r y copies of the entire issue