Argumenthood and English Prepositional Phrase Attachmenttedlab.mit.edu › tedlab_website ›...

Journal of Memory and Language40, 409–431 (1999)Article ID jmla.1998.2619, available online at http://www.idealibrary.com on

Argumenthood and English Prepositional Phrase Attachment

Carson T. Schu¨tze

Department of Linguistics, UCLA

and

Edward Gibson

Department of Brain and Cognitive Sciences and Department of Linguistics and Philosophy, MIT

Two self-paced reading experiments are presented to assess how temporary ambiguities inprepositional phrase attachment are resolved in English verb–noun-phrase–prepositional-phrasesequences. The hypothesis tested is a preference to maximize argument relations, in contrast to anoverall verb phrase attachment preference (cf. Minimal Attachment). Five syntactic argumenthooddiagnostics were used to construct noun phrase argument and verb phrase modifier completions ofsentences, differing by one word and controlled for frequency. It was found that (1) noun phraseargument completions were read significantly faster in the disambiguating region and (2) unambig-uous verb phrase modifiers were read as quickly as noun phrase arguments and faster than ambiguousverb phrase modifiers. These results suggest that argument relations are maximized in initialcomprehension of the target ambiguity. Alternative potential explanations for the findings areevaluated, including a recency-based account and a lexical-frequency treatment.© 1999 Academic Press

Key Words: parsing; ambiguity; Minimal Attachment; argument; modifier; sentence
comprehension.
m-en-

a) iNPnc

per.

t thei VP)a de-s at-t areu , in( test ands Buts -

ftenPP

ngOnter-

rial5;9;en-oth-erna-

r forteeir

orthn-aud97.d

rcholo

adis-43.

A syntactic ambiguity that has figured proinently in the development of theories of stence comprehension involves the possibletachments of prepositional phrases (PPsverb–noun-phrase–prepositional-phrase (V-PP) sequences, as exemplified by the sentein (1).

(1) (a) The spy saw a cop with a telesco(b) The spy saw a cop with a revolve

Thanks to San Tunstall, Alex Franz, and Dan Grodnetheir assistance with portions of the research repoherein. We are grateful to the following people for thcomments: Lyn Frazier, Janet Fodor, Barbara HemfAlec Marantz, Colin Phillips, Keith Rayner, Mike Tanehaus, John Trueswell, an anonymous reviewer, andences at MIT, LSA 1996, CUNY 1996, and AMLaP 19Standard disclaimers apply. This research was fundepart by the ‘Mind Articulation’ Joint International ReseaProject, sponsored by the Japan Science and TechnCorporation, and by a UCLA Academic Senate grant.

Correspondence concerning this article should bedressed to Carson T. Schu¨tze at the Department of Lingutics, UCLA, Box 951543, Los Angeles, CA 90095-15
[email protected]. Fax: (310) 206-5743.
409

t-n-es

.

In sentence (1a), the PPwith a telescopecan beaken to modify the act of seeing, describingnstrument the spy used (a verb phrase (ttachment reading) or to modify the cop,cribing what the cop was holding (an NPachment reading). Sentences of this formsually not globally ambiguous; for example1b) our knowledge of the real world dictahat revolvers cannot be used for seeing,o the NP attachment reading is forced.ince prepositions likewith can be used in nu

merous ways, an incremental parser ocannot determine which attachment of awill be appropriate until a disambiguatiword (e.g.,revolver) has been encountered.the assumption that a structure and an inpretation is assigned to incoming mateword-by-word (Marslen-Wilson, 1973, 197Tyler & Marslen-Wilson, 1977; Swinney, 197Shillcock, 1982; Garrod & Sanford, 1985; Tanhaus, Carlson, & Seidenberg, 1985; amongers), researchers have proposed several alt

rd

,

i-

in

gy

-

tive answers to the question of how structure is

0749-596X/99 $30.00Copyright © 1999 by Academic Press

All rights of reproduction in any form reserved.

-lt-&

er,y,rksfora,arlrea

ci-in

V-seeathncetaeyins

hmngPro

listeary-n,urere-alex8,renbe

ntatha-

ws

ce.tionence

ofthee-rceen-sey,ent

hatex-edelldi-

ces.7),, p.m-

be-

H,the

’sis athe

esd orenton-h it

hinf

t onep . In( op-e thes le ofg holed is asuist

our

Not t the

410 SCHUTZE AND GIBSON

assigned to an incomplete sentence likeThe spysaw the cop with. . . (Rayner, Carlson, & Frazier, 1983; Taraban & McClelland, 1988; Amann & Steedman, 1988; Clifton, Speer,Abney, 1991; Britt, Perfetti, Garrod, & Rayn1992; Britt, 1994; Spivey-Knowlton & Sediv1995; cf. Konieczny et al., 1997, and wocited there for German; Frazier, 1987,Dutch; Pynte & Prieur, 1997, for French; Igo1995, for Spanish). For example, Rayner, Cson, and Frazier (1983) took an apparent perence for VP attachment to support MinimAttachment (Frazier, 1978), a structural prinple sensitive only to the number of nodessyntactic trees. Using a different range ofNP-PP examples, Abney (1987, 1989) propoan argument preference strategy, which appto a general syntactico-semantic property ofrelationships among constituents of a sentewe discuss the nature of argumenthood in dbelow. However, Clifton, Speer, and Abn(1991) provided experimental evidence agaAbney’s proposal as an initial-choice algoritfor ambiguity resolution, instead reaffirmiMinimal Attachment. More recently, V-NP-Pambiguity resolution has been claimed to pvide evidence for constraint-based/lexicafrequentist approaches (e.g., MacDonald, Pmutter, & Seidenberg, 1994; Sedivy & SpiveKnowlton, 1994; cf. Ford, Bresnan, & Kapla1982), some of which eschew general structbased principles altogether in favor of fquency information about particular lexicitems or lexical classes and “content-basedpectations” (Taraban & McClelland, 1981990). Each of these accounts makes diffepredictions about how the ambiguity shouldresolved across the full range of examples.

In this paper, we present new experimeevidence relevant to Abney’s hypothesisthe processor initially favorsargumentattachments overmodifierattachments,1 in contrast toClifton et al.’s claim that the processor follo

1 We deliberately avoid using the termadjunct in thispaper because it also has a structural sense in lingtheory, under which it might not necessarily be synonymwith nonargument;we usemodifier as a cover term fononarguments regardless of their structural position.
that Clifton, Speer, and Abney (1991) use the termmodifier
-f-l

dlsee;il

t

-/l-

-

-

t

lt

an initial syntactic-category-based preferenWe view Abney’s hypothesis as an elaboraof the constraint-based approach to sentcomprehension, according to which a varietysources of information is brought to bear onresolution of ambiguity, including lexical frquency, semantics, plausibility, and resoucost (e.g., MacDonald, Pearlmutter, & Seidberg, 1994; Trueswell, Tanenhaus, & Garn1994). In particular, the preference for argumattachments is one way of formulating wTaraban and McClelland’s “content-basedpectations” are in terms of the lexically bashypothesis of MacDonald et al. and Trueswet al., using the lexical semantics of the invidual words to guide the parser’s preferenAdapting ideas from Jackendoff (1977, p. 5Marantz (1984, p. 15), Pollard and Sag (1987136), and Grimshaw (1990, p. 108), we sumarize our interpretation of the contrasttween arguments and modifiers in (2).

(2) If a phrase P is an argument of a headP fills a role in the relation described by H,presence of which may be implied by H. Pcontribution to the meaning of the sentencefunction of that role and hence depends onparticular identity of H.

In contrast, if P is a modifier, it predicata separate property of its associated heaphrase. Its semantic contribution is independof other elements and hence is relatively cstant across a range of sentences in whiccombines with different heads.

In (3) we see the contrast manifested witNPs. In (3a), the headstudentimplies the role ohe thing being studied; the sentence tells usroperty of John: that he studies physics3b), from Phoenixpredicates a separate prrty of the student; there is no head inentence that implies the presence of the roeographical origin. The sentence as a wescribes two properties of John: that hetudent and that he is from Phoenix.ic

s

eto cover both arguments and nonarguments; that is no
sense used here.

gu

ix.

toea

theen

le}

n/

thads(6)am

w

e-

rmn.

/nt]

refd oted87

o.

(thnk-

tionerooert

sthetherase

ofbean-tocor-

. Inyn-os-for

s tofol-heisti-rna-ald), ituen-aveit isgu-

thethe

ach-od-

sult,re-

Ourhy-the

aimingingy isereentong

e toet

vent is

411ARGUMENTHOOD AND PP ATTACHMENT

(3) (a) John is a student of physics. [arment]

(b) John is a student from Phoen[modifier]

Now consider VPs. In (4a),on Sundaycan beinterpreted correctly without any referencethe rest of the sentence and has the same ming with each verb. In contrast,on Sandyin (4b)can be interpreted only with reference toverb; the meaning of this phrase is differacross these sentences.

(4) (a) Kim {ate/was sad/broke her ankon Sunday. [modifier]

(b) Kim {depended/blamed the arsodecided} on Sandy. [argument]

As a consequence of (2), arguments onwhole occur with a narrower range of hethan modifiers do, as shown in (5) and(Here and below, an asterisk indicates ungrmaticality.)

(5) (a) a woman/dog/muppet/scarecroandroid with gray hair [modifier]

(b) a member/*dog/*muppet/*scarcrow/*android of Parliament [argument]

(6) (a) John {died/sneezed/broke his asaw Fred/laughed at Bill} in the afternoo[modifier]

(b) John {informed/*saw/*hit/*admired*surprised} his friend of the danger. [argume

Abney (1987, 1989) proposed argument perence as a disambiguation strategy baseintuitions about sentences like (7) (cf. relaproposals by Pritchett, 1988; Gibson, 191991; Konieczny et al., 1997).

(7) I thought about his interest in the Volv

People prefer the interpretation wherein theVolvo describes what he was interested inNP-attached reading), not where I did my thiing (the VP-attached reading).In the Volvois anargument of the nouninterest,but it is not anargument of the verb–preposition combinathought about.We refine Abney’s idea in ordto take account of the fact that argumenthdoes not appear to be an all-or-nothing prop
of PPs; our proposal is given in (8).
-

n-

t

e

.-

/

/

-n

,

e

dy

(8) Argument Preference Strategy:In caseof attachment ambiguity, the parser prefersattachment that maximizes the extent ofargument relation between the attaching phand the attachment site.

While the ideas in (2) are central to the notionargumenthood, in particular cases it mayhard to arrive at clear intuitions on these semtic criteria. As a result, it has proved fruitfulappeal to several syntactic phenomena thatrelate closely with those semantic notionsAppendix A, we have assembled from the stax literature relevant argumenthood diagntics that were used in designing the stimuliour experiments.

It should be noted that there are other wayformalize a lexically based parsing proposallowing MacDonald et al. and Trueswell et al. Tapproach we have started with here is a lingucally based lexical semantics proposal. Altetively (and perhaps more in line with MacDonet al.’s and Trueswell et al.’s proposed modelscould be that the processor is sensitive to freqcies of occurrences of similar structures that hbeen encountered in the past. It turns out thatdifficult to separate the lexical-semantics arment-based hypothesis proposed here fromfrequency-based hypothesis with respect toV-NP-PP ambiguity, because argument attments appear to be much more frequent than mifier attachments in this construction; as a reboth theories make the same predictions withspect to most instances of this ambiguity.experiments therefore do not separate thesepotheses, but we suggest ways to do so inGeneral Discussion.

We should emphasize that we do not clthat any single factor determines initial parspreferences; rather, we are simply testwhether the Argument Preference Strategone factor contributing to such decisions. Thwill surely be different degrees of attachmpreference both among arguments and amnonarguments as well, including biases duproperties of particular words (cf. Koniecznyal., 1997, for instrumentals). There could ebe instances where a modifier attachmen
favored over an argument attachment, if other

ce

onncinnd

oreadb)

thae-f thh-

onconthe

the, thesn--at

force-ras). Iichthh

iseby

N:Nd-m

ach

nor aan

trueent

o int 994;C ecto NP( de)w N

nt ofPPo

b ingw reen odesr 84,i ingt ndds

N:G adyn s ana od-i da e isn hati thee er-n h-m asw uldh

t willrF re-f ustr , itm arilya bei VPm yu

ton,S erst thes


factors outweigh the Argument PreferenStrategy.

PREVIOUS FINDINGS

Rayner, Carlson, and Frazier (1983) cducted an eye-tracking study on 12 sentepairs similar to (1), measuring first-pass readtimes in the disambiguating region, and fouthat completions that are pragmatically mlikely to be VP-attached, like (1a), were refaster than NP-attached completions like (1Based on this finding, Rayner et al. arguedinitial VP attachment is always initially prferred, and they take this to be an instance omore general parsing principle Minimal Attacment, stated in (9).

(9) Minimal Attachment:Attach incomingmaterial into the phrase marker being cstructed using the fewest syntactic nodessistent with the well-formedness rules oflanguage. (Frazier & Rayner, 1982, p. 180)

However, because they did not control forargument versus modifier status of their PPsgenerality of this finding remains open to qution. By our criteria, 8 of their 12 items cotrasted VP-attached arguments with NPtached modifiers.

It is important to note that a preferenceVP over NP attachment in V-NP-PP sentenfollows from Minimal Attachment (9) only under certain specific assumptions about phstructure (see Frazier, 1990, for discussionparticular, it is crucial that the node under whthe PP could be attached already exists inVP but must be added to the NP. To see wthis is so, consider the four possible pairwattachment choices.2 The structures assumedClifton et al. are shown in Fig. 1.

1. Argument of V versus argument ofMinimal Attachment would prefer V overattachment only if V attachment involved ading fewer nodes than N attachment. (If the sanumber of nodes were required for each att

2 We assume, along with Frazier (1990) and Clifpeer, and Abney (1991), that arguments must be sist

heir heads, where two nodes in a tree are sisters if
share a common parent node immediately above them.
-eg

.t

e

--

e-

-

s

en

ey

e-

ment, then Minimal Attachment would makeprediction, but Late Closure (Frazier, 1978) orecency/locality preference would predictNP attachment preference.) This would beif N 9 had not yet been built prior to attachm

f the PP, while V9 had been. Under certaheories of phrase structure (e.g., Speas, 1homsky, 1995; cf. Frazier, 1990), the dirbject noun is an immediate daughter ofi.e., it is the highest node below the NP nohen it is first attached (cf. Fig. 1A), and9

must be added in order to attach an argumeN, as in Fig. 1B. In addition, attaching theargument to VP must requireno extra nodes t

e built. This could be true if ternary branchere permitted by the grammar (that is, if a tode were allowed to have three daughter nather than at most two, contra Kayne, 19nter alia), so that the PP argument followhe direct object NP could be a sister of V aaughter of the already-constructed V9, ashown in Fig. 1A.2. Argument of V versus modifier ofiven the ternary branching assumption alreeeded for the previous case, attaching argument of V would be preferred over a m

fier of N if adding an NP modifier involvedding at least one node. This extra nodeeded if N modifiers are adjoined to NP, t

s, attached by adding a new segment ofxisting NP node, as shown in Fig. 1C. Altatively, the preference for V over N attacent would also hold if a modifier had to be

ister of N9, because the requisite N9 nodeould not have been constructed yet and woave to be added.3. Argument of N versus modifier of V:Under

he above assumptions, the NP attachmentequire exactly one new node, namely, N9, as inig. 1B. Thus, for this attachment to be disp

erred, adding a PP as a modifier of VP mequireno new nodes to be constructed. Thusust not be true that VP modifiers are necessdjoined, otherwise a new VP node would

ntroduced (cf. Fig. 2D). Rather, it must be thatodifiers can be sisters to V9, attached directlnder the existing VP, as in Fig. 1D.4. Modifier of N versus modifier of V:It was

ofy
hown in case 2 that NP modifier attachment

FigifieFigre-

reParyarriet bas-tur

asalresa

. 2;rictntsentre

NPod-entbar-ires

er


must require at least one new node, as in1C, and it was shown in case 3 that VP modattachment requires no new nodes, as in1D. Therefore, the VP attachment will be pferred.

Thus, there are three critical assumptionsquired for Minimal Attachment to prefer Vover NP attachment in all situations: ternbranching must be allowed by the grammcertain intermediate nonbranching categomust not be projected, and modifiers muspossible sisters to X9. Suppose instead onesumes a more traditional X-bar phrase struc

FIG. 1. Structures assumed by Clifton et al. (19(C), verb modifier (D).

(cf. Jackendoff, 1977) in which intermediate

.r.

-

,se

e

categories must be projected and all phrmodifiers must be adjoined, as in the structuin Fig. 2. (For simplicity we have omittedphrasal DP projection for determiners in Figalthough this would be required under a stX-bar account, it does not affect the poiunder discussion.) Then Minimal Attachmwould predict that VP arguments (Fig. 2A) apreferred over NP modifiers (Fig. 2C), butarguments (Fig. 2B) are preferred over VP mifiers (Fig. 2D). In both cases, an argumattachment makes use of an existing single-level node, while a modifier attachment requ

): verb argument (A), noun argument (B), noun modifi
91
the addition of an XP. Thus, it is possible to

ci-tureraeren

-PPn,alalsrsthpe

de-

r etfor

onal.2),as

e ofsup-

e.n-

n


accept Minimal Attachment as a parsing prinple but reject the particular phrase strucassumptions adopted by Clifton et al. from Fzier and her colleagues (Frazier & Rayn1982; Frazier, 1990), in which case an argumpreference could be predicted for the V-NPambiguity. However, for clarity of expositioin the remainder of the text we use “MinimAttachment” to refer to the specific proposmade by Frazier and her colleagues. Of couthe trees in Figs. 1 and 2 do not exhaustpossible structures for these four sentence ty

FIG. 2. Possible structures under traditional Xmodifier (C), verb modifier (D).

Because our proposal does not involve countin

-,t

e,es.

nodes in trees, further exploration of thesetails is not relevant.

A number of studies subsequent to Rayneal.’s have found a VP attachment preferencethe V-NP-PP ambiguity: Ferreira and Clift(1986), Clifton and Ferreira (1989), Britt et(1992), Rayner, Garrod, and Perfetti (199and Britt (1994). However, argumenthood wnot controlled for in these studies. Thus, nonthese studies can be taken as unequivocalport for a general VP attachment preferenc

Taraban and McClelland (1988) demo

r theory: verb argument (A), noun argument (B), nou
-ba
gstrated that the VP attachment preference ob-

stebybaenbu

newt isthe

rn

rn-m

T ncea at inp ot fere

or

etsde-

reesete

teglled18witei

wit

tedd thersu

pleder

ple-the

erra-

rn-.

ead-e,ugh-e oranTheyh-s todedforse-

thee6),ye-om-a-ionhes-znylud-

o-ofati-s).edindthethe


tained in the sentences that Rayner et al. tewas not a general preference. In a word-word self-paced reading experiment, Taraand McClelland replicated the VP attachmpreference for Rayner et al.’s sentences,found an NP attachment preference for aset of 18 sentences exemplified in (10); tha(10a) was read more quickly than (10b) indisambiguating region.

(10) (a) The report described the govement’s programs in education.

(b) The report described the goveent’s programs in detail.

hey concluded that PP attachment preferere determined on the basis of all the inform

ion contained in the sentence up to that poerhaps combined with people’s knowledge

he world, and that no general syntactic prence exists:

The results we have reported cannot be accounted fby any syntactic principle of which we are aware—that is, by any principle that does not consider thcontent of the sentence—since our expectation effecoccurred in sentences that differed in the content, annot in the syntactic constituents of the sentencframes (Taraban & McClelland, 1988, p. 611; emphasis in original).

While Taraban and McClelland’s findings aevidence against Minimal Attachment, thfindings do not rule out other syntactic stragies, such as the Argument Preference StraIn particular, argumenthood was not controfor; only two of Taraban and McClelland’sitems contrasted a VP-attached argumentan NP-attached modifier, whereas six of thitems contrasted an NP-attached argumenta VP-attached modifier.

Clifton, Speer, and Abney (1991) conducexperiments that were specifically designetest the argument preference hypothesis. Tcontrasted NP versus VP and argument vemodifier attachments as in (11).

(11)VP argument:(a) The war alienated the young peo

from the social systemand decreasevoter turnout among the young
generation.
d-ntt

,

-

s-t,f-

-y.

hrh

oys

NP modifier:(b) The war alienated the young peo

from our neighborhoodand decreased voter turnout amongyounger generation.

NP argument:(c) The war increased alienationfrom the

social systemand decreased votturnout among the younger genetion.

VP modifier:(d) The war increased alienationfrom the

beginningand decreased voter tuout among the younger generation

Using eye-tracking and phrasal self-paced ring, Clifton et al. found an initial VP preferencas measured by average reading time throout the PP, but on the subsequent region (onmore words following the disambiguation)argument attachment preference appeared.took this result to support the Minimal Attacment claim that the first preference is alwayattach to the VP. Furthermore, they concluthat Abney’s hypothesis about a preferenceargument attachments was true of a later,mantic, stage of parsing.

However, there are several problems withinterpretation of Clifton et al.’s findings. Somof these are pointed out by Konieczny (199who argued that an examination of all the etracking measures reported, especially the cbination of frequency of regression with fixtion durations, strongly suggests a conclusopposite to that of Clifton et al. AlthougClifton et al. did not present all the data necsary for a comprehensive reanalysis, Koniecmade several plausible arguments for concing that Clifton et al. actually found aninitialpreference for arguments over modifiers.

Clifton et al.’s materials also suffer from ptential confounds. The relative plausibilitythe different completions was not systemcally controlled (Speer and Clifton, in presAlso, half of the crucial contrast pairs differon the number of words (e.g., (12)) or the kand amount of structure within the PP wherecomparison was conducted (e.g., (13a), with
structures in (13b)). Thus, any differences in

s oeocticrd

iom

ogc

ith

end}.

in

heouNPreeandtheeeith

inveon

oinsige

x-eirn

henceereconthel-le,

betra

re-ch-ner-singep-

ncethesseslly

fectnotA,

p-ssesnce,u-s.

Bc V-N

ri-m fac-t andn ora Wet t thep rat-e ndt ure-b At-t tegya NPm nlyt fierc ast

M

veE s ofM

seo tedf 5),


reading time between PPs in the two versiona sentence could be due to these extrandifferences rather than to the different syntaattachments. Even when the number of wowas matched, the PPs were sometimes idsuch asin record time(14), which might well beread quickly simply because they are recnized as units, in contrast to the nonidiomatiinchurch affairs(Swinney & Cutler, 1979).

(12) John continued the discussion w(persistence/the tired boys).

(13) (a) The teacher encouraged excitemover {the course of the week/learning to rea

(b) [NP the [course [PP of [NP the week]]]vs. [NP PRO [VP learning [IP PRO to [VP read]]]].

(14) Maria increased her involvement{record time/church affairs}.

As for the argument/modifier status of titems, by the criteria used in constructingmaterials, 12 of the 16 items did involve anargument contrasted with a VP modifier; thinvolved NP and VP arguments (items 1, 2,15), and one did not contain a preposition inVP-attached completion (item 16). (SSchutze, 1995 for other potential problems wspecific items.)

In a follow-up study, Speer and Clifton (press) explored the apparent advantage forarguments over VP modifiers found by Cliftet al. and attempted to distinguish effectsargument status versus plausibility on readtimes in V-NP-PP sentences. They found anificant argument advantage in reading timfor low-plausibility items, which cannot be eplained by plausibility as measured by th“sensibleness” ratings, because these diddiffer significantly and were numerically in topposite direction. Unfortunately, the relevaof this study to the construction of concern his questionable, because 7 of the 20 itemstained VP adjuncts that were not PPs, but ratemporal NPs containingaboutused as a quaifier rather than a preposition, for exampabout a minute too late.

One further study in the literature couldtaken as a test of the Argument Preference S
egy, though it was not explicitly designed asd
fus

ss

-

t

r

rb

fg-s

ot

-r,

t-

such. Spivey-Knowlton and Sedivy (1995)ported two experiments involving the attament of PPs headed bywith, one using “actioverbs,” which found a VP attachment prefence modulated by definiteness, and one u“psychological predicates and verbs of perction,” which found an NP attachment preferewhen the direct object was indefinite. Whileauthors acknowledged that these verb cladiffer in that only the former systematicaallows instrumental uses ofwith, they deniedthe relevance of argumenthood to this efbecause they claim that instrumentals arearguments. If, as we argue in Appendixinstrumentals actuallydo have argument proerties, then the contrast between verb clawould be explained by an argument preferewhich would favor VP attachment for instrmental uses ofwith but not for modifier use

ut since all their items containedwith, oneannot extrapolate from their findings toP-PP structures in general.The general picture is that all of the expeents reviewed here confounded various

ors in assessing attachment preferences,one of them constitute clear evidence forgainst the Argument Preference Strategy.

herefore conducted new experiments to tesredictions of the Argument Preference Stgy against those of Minimal Attachment a

hose of theories that claim no general structased preference at all. Because Minimal

achment and the Argument Preference Stragree on their predictions for comparingodifier and VP argument attachments, o

he crucial NP argument versus VP modiomparison, where predictions differ, wested.

EXPERIMENT 1

ethod

Participants.The participants were 33 natinglish speakers, students and other affiliateIT, who were paid for their participation.Materials.We used materials similar to tho

f Clifton et al. (including some items adaprom theirs). Our 15 items, exemplified by (1
iffered on only one word in the NP and VP

rsis

ed

ed

ed

ed

edn’tt.

F msc thd ingt disc t ot wec ifc ugm acm esM inc .2,V P

an-P

2a tions

uet if-f ; 42n ecti edI tenw iti-c e at emp

sid-e

r am

N edt Psa as-s ce,w atedp n-n ral-n d int too ter-s ives eenp in-t ofe eent otsa

if-are.

eder

dter-nal-be

iser,

nfe-htjecton-eeaits

c s ins ean-i like( thea yee


attachment conditions. Slashes demarcategions, which were relevant only for the analy

(15)Sample items(a)NP argument

The company lawyers/consideremployee demands/for araise/butthey/didn’t act until a strike seemimminent.

(b) VP modifierThe company lawyers/consideremployee demands/for amonth/butthey/didn’t act until a strike seemimminent.

(c) no PPThe company lawyers/consideremployee demands/but they/didact until a strike seemed imminen

ive different prepositions were used. All iteontained the same number of words up toisambiguation point. They were tested us

he five relevant argumenthood diagnosticsussed in Appendix A. Given the statemenhe Argument Preference Strategy in (8),ounted an attachment as an argumentlearly passed at least one diagnostic, thoost examples passed several; modifier attents never passed any of the argument tean length of the disambiguating noun

haracters did not differ significantly (NP 6P 5.6,t(14)5 1.17,p . .25). Nouns in the V

condition were marginally more frequent ththose in the NP condition (mean log10 frequencies in Francis & Kucˇera, 1982: NP 1.68, V

.15, t(14) 5 22.01, p 5 .06), which worksgainst our hypothesis that the NP condihould be processed faster.We used a separate paper-and-pencil q

ionnaire in order to check for plausibility derences between the two critical conditionsative English speakers who were not subj

n the on-line reading experiment participatn order to preserve meaning and lexical conhile removing temporary ambiguity, the cral sentences were passivized, so that thachment of the PP was unambiguous, as ex
lified by (16), corresponding to (15) above.
e-.

e

-f

ithh-ts.

s-

s.t

t--

(16) (a) NP argumentEmployee demands for a raise were con

red by the company lawyers.(b) VP modifier

Employee demands were considered foonth by the company lawyers.

P argument PPs always immediately followhe head noun of the subject; VP modifier Pppeared either immediately following the pive participle or at the end of the sentenhichever sounded more natural. Subjects rlausibility on a scale of 1 (natural) to 7 (uatural). They were asked to judge the natuess in the real world of the events describe

he sentences, that is, how likely they wereccur. The 15 stimulus sentences were inpersed with 18 filler items, also all passentences, divided roughly equally betwlausible and implausible according to our

uitions; each subject saw only one versionach experimental item. The difference betw

he NP and VP plausibility ratings was nignificant, means 2.54 (SD5 0.74) for the NPrgument versus 2.63 (SD 5 0.72) for the VP

modifier, Fs , 1. Thus, any reading time dferences between our two critical conditionsunlikely to be due to a plausibility difference

We included a third condition, exemplifiby (15c), in an attempt to rule out anothpossible confound.3 If VP modifiers are reamore slowly than their NP argument counparts, this could be because the argument aysis is initially preferred and must thereforereanalyzed when the disambiguating nounencountered in the VP condition. Howevsuch a slowdown could also be due to an ilicity in the VP condition, namely, that it migsound odd to omit the argument of the obnoun. For example, without a preceding ctext, it sounds strange to sayThey discussed thlikelihood yesterday,because one has no idwhat likelihood refers to in the absence ofomplement. Similarly, because argumentome sense encode an intrinsic part of the mng of the head they attach to, sentences15b) and (15c) might be degraded due tobsence of any indication of what the emplo

3
Thanks to Janet Fodor for suggesting this idea.

ionncmb

msstbe

enen

tesreouerijeud

d-b g-w ol-l terr L)s o rv thw acs ts sino inf ans emt thes prec ighp twl itht ouw oks

R

theon

ctlyvede a

&w

e t

nceses

vinging

tiont re-wenly,thend–6,twounr ofre-iondingttedarenton-

Pster

MPP

c NPa

eson-us-

it-banstsby

VP-there-

if-,

eom-


demands were for. Thus, the third conditomitted the PP entirely, so that its absemight be assessed independently of the aguity in the preposition.

The experiment included another set of itefrom Rayner et al. (1983), in order to tewhether their VP preference result wouldreplicated in our self-paced reading experimWe used the shorter versions of their 12 stence pairs.

Each subject read one version of eachsentence. The 151 12 test sentences weinterspersed with 53 filler sentences of varitypes. These included items for other expments with unrelated hypotheses. Each subencountered the sentences in a different pserandom order.

Procedure.Subjects were timed in a wory-word self-paced noncumulative movinindow reading task (Just, Carpenter, & Wo

ey, 1982) controlled by an IBM PS2 compuunning Micro-Experimental Laboratory (MEoftware. Subjects pressed the space bar teal each subsequent word and cause all oords to revert to dashes. At the end of eentence, a yes/no question appeared oncreen, which subjects answered by presne of two keyboard keys. Subjects were

ormed by a screen message when theywered incorrectly, in order to encourage tho keep paying attention to the content ofentences. The experimental trials wereeded by two screens of instructions and eractice trials. Most sentences spanned

ines on the screen (never more than two), whe disambiguating noun plus at least fords on the first line. The experiment toubjects approximately 20 min to complete.

esults

All subjects answered at least 80% ofquestions in the experiment correctly. Trialswhich the question was answered incorrewere excluded from the analysis; this remo3.6% of the sentences. The data we analyzresidual reading times per word (FerreiraClifton, 1986), derived by subtracting from rareading times each subject’s predicted tim
read words of the same length, calculated by
ei-

,

t.-

t

s-cto-

e-erhheg

--

-to

r

re

o

linear regression equation across all sentein the experiment. Residual reading timgreater than 1000 ms were discarded, remoless than 1% of the remaining data. (Ustighter cutoffs, such as trimming to within 3SDof the mean, separately for each word posiin each condition across subjects, does noveal any hint of effects different from thosereport below.) For purposes of analysis oitems were divided into regions as follows:initial NP (words 1–3, region 1), the verb athe object up to its head noun (words 4region 2), the PP (words 7–9, region 3), thewords following the disambiguating no(words 10–11, region 4), and the remaindethe sentence (region 5). All comparisonsported in this paper were computed by regunless otherwise noted. Mean residual reatimes by region and standard errors are ploin Fig. 3; corresponding raw reading timesgiven in Table 1. There were no significareading time differences among any of the cditions in the first three regions (allFs , 1). Onthe region following disambiguation, the Nargument version was read significantly fathan the VP modifier version,F1(1,32)5 5.55,

Se 5 2552,p , .05,F2(1,14)5 4.96,MSe 51245,p , .05. On that same region, the noondition was read more slowly than thergument condition,F1(1,32) 5 4.79, MSe 5

2428, p , .05, F2(1,14) 5 8.97, MSe 5 736,p , .05, and did not differ significantly from thVP modifier condition,Fs, 1. However, it waconcluded that comparisons involving this cdition were not meaningful—see the Discsion.

As for the items from Rayner et al., we omted from the analysis 1 of the 12 that Taraand McClelland (1988) had found in pretewas not given the intended interpretationsubjects. For the remaining 11 items, theattached completions were read faster thanNP-attached completions on the three-wordgion following the disambiguating noun, signicantly by subjects,29.8 vs. 8.8 ms/wordF1(1,33)5 5.21,MSe 5 1139,p , .05,F2(1,10)5 1.73, MSe 5 1286, p 5 .22. (One morsubject was included in this analysis as c
apared to the analysis of our items, in order to

h a

t is

p ana ere adi uat

ef-ms.ise of

nlyantheofri-

ing


balance Rayner et al.’s two conditions.) Wittighter reading time cutoff at 3SD, which re-moves 1.8% of the data, the items effecmarginally significant,F1(1,33)5 6.71,MSe 51435, p , .05, F2(1,10) 5 3.93, MSe 5 871,

5 .08. This result is consistent with Tarabnd McClelland’s (1988) replication of Raynt al. Using the word-by-word self-paced re

ng paradigm they also found no disambig

FIG. 3. Residual read

ion effect on the noun of the PP itself, and on

--

subsequent words the Minimal Attachmentfect was significant by subjects but not by iteSimilarly, when Britt et al. (1992) used thparadigm to present 16 items based on thosRayner et al. in neutral contexts, their osignificant effect was on a region which begafter the disambiguating NP and comprisedwhole remainder of the sentence. By waycomparison, in Rayner et al.’s original expe

times for Experiment 1.

ment, the significant advantage for VP attach-

thain-ria

gute

andt-soofPal.rn,t o

en-d i

NPmanheityarere-bu

ateub.

enceePPliknt

theht

r-l woc cto bea ep e tot di-t h tod con-d (asw is)w 2.

M

veE s ofM n.N t 1.

nts per-i n-t , ase ater al-y

ed

ed

ed

ed

ed

ed

T notb thed lau-s st

s)

5590


ment sentences was found over regionsbegan with the disambiguating word andcluded varying amounts of subsequent mate

Discussion

Our items showed a preference for NP arment over VP modifier attachment, as predicby the Argument Preference Strategy,showed no evidence for an initial Minimal Atachment effect. Our findings thus are connant with the later argumenthood effectClifton et al. (1991), but not with the earlier Vpreference they found. Because Rayner etitems did show a Minimal Attachment patteits absence in our items cannot be an artifacexperimental conditions. However, two pottial confounds remain, which are addresseExperiment 2.

First, the reading time difference betweenand VP completions might reflect inherent coplexity differences between NP argumentsVP modifiers, independent of ambiguity. Tusual method of controlling for this possibilis not available to us: we would like to compeach resolution of the ambiguity to a corsponding unambiguous control sentence,there is no way in English to disambigutoward the NP argument reading without sstantially altering the syntax of the sentence

Second, an anomaly effect due to the absof the argument of the direct object has not bruled out. As can be seen in Fig. 3, the Nocondition appears to have patterned morethe VP condition, which would be consistewith an anomaly-based explanation forNP/VP contrast. However, this is not a straig

TABLE 1

Mean Raw Reading Times per Word (in millisecondfor Experiment 1

Condition

Sentence region

1 2 3 4 5

Noun argument 351 392 345 342 3Verb modifier 353 393 344 371 36No PP 359 385 — 377 37

forward comparison, because the words beinf

t

l.

-d

-

’s

f

n

-d

t

-

en

e

-

compared (e.g.,but theyin (15)) occurred eaier in the no PP condition than in the other tonditions, immediately following the direbject noun. Thus, various confounds couldt play, including differential spillover from threceding word, a possible surprise effect du

he relative shortness of the VP in this conion, etc. For these reasons, we do not wisraw any conclusions based on the no PPition; the anomaly confound hypothesisell as the complexity difference hypothesas tested in a different way in Experiment

EXPERIMENT 2

ethod

Participants.The participants were 82 natinglish speakers, students and other affiliateIT, who were paid for their participatioone of them had participated in ExperimenMaterials. Fifteen NP and VP attachme

entence pairs were identical to those in Exment 1. However, the third condition now coained a PP headed by a different prepositionxemplified in (17c). Again, slashes demarcegions, which were relevant only for the ansis.

(17)Sample items(a)NP argument

The company lawyers/consideremployee demands/for a raise/butthey/didn’t act until a strike seemimminent.

(b) VP modifierThe company lawyers/consideremployee demands/for a month/butthey/didn’t act until a strike seemimminent.

(c) unambiguous PPThe company lawyers/consideremployee demands/after a month/butthey/didn’t act until a strike seemimminent.

his new preposition unambiguously coulde construed as introducing an argument ofirect object noun, and the PP was most pibly a VP modifier. Thus, it allowed us to te

gor effects resulting from the absence of the

mn-ericaener-foin

fersehendthehant

( -em-wlyichis-

talerx-

omri-

hesin

oufo

reaeri

ast itin

sisth

nt th

d 1re-esin

re-andhantheffsomsig-er-, re-

rors4;

undingns

re-thes,

mtoun,nd

-noty

terth

y

x-erees.

VPun-woBe-eadnot

racw in the emf mod


noun’s argument and from the inherent coplexity of the VP modifier structure, indepedent of possible reanalysis effects. Whenevwas feasible, the noun in this PP was identto that in the VP condition (as in (17)), but whthis could not be felicitously achieved, a diffent noun was substituted. (See Appendix Ba complete list of items.) Overall, the nounsthe unambiguous PP condition did not difsignificantly in length or frequency from thoin either of the other conditions. As for tprepositions, those in the unambiguous cotion were marginally longer than those inother conditions (mean 3.80 versus 2.87 cacters,p 5 .068), and significantly less frequelog10 frequency unambiguous5 3.28, ambig

uous5 3.85, t(14) 5 4.07,p , 0.005). Thesdifferences could, if anything, lead the unabiguous PP condition to be read more slothan either of the other two conditions, whwould work against our hypothesis, as dcussed below.

In the interests of maximizing experimenresources, two versions of the experiment wrun on different subjects, in which our 15 eperimental items and filler sentences were cbined with items for two other sets of expements with unrelated hypotheses. We treat tas two subject groups in the analysis, contain50 and 32 subjects, respectively. The first grread a total of 103 sentences, including 48other experiments, and the second group80 sentences, including 32 for other expments.

Procedure.The experimental procedure widentical to that of Experiment 1, except thawas run on a Macintosh Centris computer uscustom software.

Results

Six subjects were omitted from the analy4 had fewer than 80% correct answers onrelevant questions4; 1 noticed an ambiguity ihe practice items similar to the one tested in

4 For two of these subjects, question-answering accuas computed over a subset of 55 of the sentencesxperiment: 15 test items and 40 fillers. We excluded it

or other experiments because they were substantially
ifficult.
-

itl

r

i-

r-

e

-

egprd

-

g

:e

e

experimental items and commented on it; anfeel asleep during the experiment. For themaining 76 subjects, residual reading timwere calculated and data were filtered asExperiment 1: 6.6% of the sentences weremoved due to incorrect question answering,remaining residual reading times greater t1000 ms were removed (less than 1% ofdata). (As in Experiment 1, using tighter cutodoes not reveal any hint of effects different frthose we report below.) As there were nonificant effects of subject group, and no intactions between this and any other factorssults are collapsed across groups.

Residual reading times and standard erfor all three conditions are displayed in Fig.corresponding raw reading times can be foin Table 2. There were no significant readtime differences among any of the conditiopreceding the disambiguating noun (allps $.20). On the two-word post-disambiguationgion, the NP condition was read faster thanVP condition, significantly by subjectF1(1,74) 5 8.67, MSe 5 3734,p , .005, and

arginally by items,F2(1,14) 5 4.04, MSe 52260, p 5 0.06. If this region is expandedthree words including the disambiguating nothe effect is significant by both subjects aitems,F1(1,74)5 8.29,MSe 5 2357,p , .005,F2(1,14)5 4.83,MSe 5 1247,p , .05. Regarding the unambiguous PP condition, it diddiffer significantly from the NP condition in anregion (ps . 0.30). However, it was read fasthan the VP condition on the critical fourregion, significantly by subjects,F1(1,74) 58.88,MSe 5 3697,p , .005, and marginally bitems,F2(1,14)5 4.20,MSe 5 2134,p 5 0.06.

Discussion

Experiment 2 replicates our finding from Eperiment 1 that NP argument sentences wread more quickly than VP modifier sentencAgain, there was no evidence of an initialpreference. In addition, the results for theambiguous condition allow us to rule out talternative explanations for this contrast.cause the unambiguous condition was rfaster than the ambiguous VP condition and

ye

sre
slower than the NP condition, the slowdown in

ribtedm

nylveionnctheis

as

fac-entheforau-

ing


the ambiguous VP condition cannot be attuted merely to the absence of an “expecargument of that noun, nor to the inherent coplexity of a VP modifying the PP structure. Asuch effects would have manifested themsein a slowdown in the unambiguous conditbeginning in the PP region, where the abseof the argument PP became apparent. Rathe slowdown in the ambiguous VP condition

FIG. 4. Residual read

a garden path/reanalysis effect, indicative tha

-”-

s

er,

the ambiguous preposition was initially takenintroducing an argument to that noun.

GENERAL DISCUSSION

We have shown that when several othertors are controlled for, the difference betweargument and modifier attachments affectsinitial analysis constructed by the parserV-NP-PP ambiguities. Thus, in contrast to

times for Experiment 2.

tthors cited earlier, we find that an argument

ionas-ueor,

, ifth

gucortiein-

raNPe turera-8;senth

aseNPthaheefeer-ultsVPar

eaion

tot ihy

ent-onyci-thert isn.

t toa

lo-

o-ncefre-nen-andandhatforac-

oodur-

-i--ts

enic

irst,theen-

, ifasefer-

on-y bi-rtic-NPhisrainbe

s)

N 47V 7U 44


preference is one factor in the initial resolutof this ambiguity and not just a later effect,Clifton et al. (1991) found. Minimal Attachment as formulated by Frazier and colleagcannot therefore be the initial determinantparsing decisions, although, as noted earliestrategy of minimizing tree nodes could becertain syntactic assumptions were made. Inrest of this section, we consider how the arment preference hypothesis relates to othertending accounts, such as statistical propeof lexical items or more general structural prciples.

First, let us consider an alternative structuaccount, namely, that the preference forattachment in our experiments might be dua Recency, Right Association, or Late Closstrategy (cf. Kimball, 1973; Frazier, 1978; Fzier & Fodor, 1978; Gibson, 1991, 199Stevenson, 1994). Because the noun reprea more recent potential attachment site thanverb in a V-NP-PP sequence, a recency-bpreference by itself might lead one to expectattachment to be preferred. However, recallwith different sets of items, we and many otresearchers have found a VP attachment prence for this ambiguity. The Argument Prefence Strategy explains both sets of resgiven our observation that the preferredattachments in previous work were mostlygument attachments (e.g., instrumentalwithphrases). If our NP preference were insttreated as a pure recency effect, some additcompeting factor would have to be positedhandle the examples in which VP attachmenpreferred, and one would have to explain w

TABLE 2

Mean Raw Reading Times per Word (in millisecondfor Experiment 2

Condition

Sentence region

1 2 3 4 5

oun argument 362 393 346 344 3erb modifier 354 390 343 372 34nambiguous PP 357 388 354 345 3

the relative strengths of recency and this com

sfa

e-n-s

l

o

tsed

trr-

,

-

dal

s

peting factor seem to be sensitive to argumhood contrasts. Thus, on grounds of parsimit is preferable to posit a single general prinple, the Argument Preference Strategy, rathan to stipulate a property of a constraint tharelevant only in one particular configuratio(However, see Phillips (1996) for an attempderive some argument preferences frompurely structural principle related to Late Csure.)

Let us now consider a different kind of ptential alternative to the Argument PrefereStrategy, parsing theories based on lexicalquencies such as those of Trueswell and Tahaus (1994) or MacDonald, Pearlmutter,Seidenberg (1994). (Note that, like SpeerClifton (in press), we have found evidence tplausibility differences are not responsiblethe observed disambiguation preference.) MDonald et al. claim that apparent argumentheffects may be reducible to relative co-occrence frequencies:

[The] frequency coding of preferences for differentprepositional phrases (PPs), particularly in a distributed representation described earlier, largely elimnates the argument–adjunct distinction that is common to many discussions of PP attachmenambiguities. . .In our view, an argument is a PP that istrongly (frequently) linked to a word (e.g., the loca-tion role forput) and an adjunct is one that is weakly(infrequently) linked (e.g., the manner role forput). Itis possible that the frequency biases that we havdescribed derive from deeper relationships betweethe semantics of the verbs, nouns, and thematroles. . . (MacDonald et al., 1994, p. 694).

We take up the discussion on three fronts. Fcould frequency biases alone account forresults we have obtained? Second, is argumthood reducible to relative frequency? Thirdit is not, what further experiments could teapart the frequency bias and argument preence hypotheses?

In response to the first question, we cducted a posthoc assessment of frequencases in our 15 items to see whether the paular prepositions more frequently attach torather than VP under certain conditions. Trequired a specific hypothesis as to the gsize of frequencies that the parser might
-sensitive to when resolving a V-NP-PP attach-

thed,

othin

hip

epfirsw-

r

epthe

ep,

ep-

otld,thth

raigfustef thSadto

ons

ge,onn,wasfierhados-ngsedat,er-

het 1

).sede-, theandbeto

o-rm0)–)),n-

heyore,rgu-pro-fun-of

and, weromviceuch

canallyim-

stin-on-igh-twopro-

test


ment (cf. Mitchell, 1994; Gibson, Schu¨tze, &Salomon, 1996; MacDonald, 1997). At leastfollowing possibilities could be considerelisted in order from fine to coarse:

For a given preposition, e.g.,for,

1. Count only occurrences that involve bthe particular noun and the particular verbquestion, in the relevant structural relationsi.e.,

[ VP consider [ NP. . .demand] for. . .] and [VP

consider[ NP. . .demand for. . .]. . .].2. Count only occurrences in which the pr

osition attaches to the particular noun asobject of a verb or to the particular verb folloing a first object, i.e.,

[ VP V [ NP. . .demand for. . .]] and [VP consideNP for . . .].

3. Count only occurrences in which the prosition attaches to the particular noun orparticular verb in question, i.e.,

[ NP. . .demand. . . for. . .] and [VP. . .consider. . . for. . .].

4. Count only occurrences in which the prosition follows the first object of a verb, i.e.

[ VP V [ NP. . .N [PP for. . .]. . .]. . .] and [VP V[NP. . .N] [PP for. . .]. . .].

5. Count all occurrences in which the prosition attaches to an NP or to a VP, i.e.,

[ NP. . . [PP for. . .]. . .] and [VP. . . [PP

for. . .]. . .].

Of these possible grain sizes, 4 and 5 are nthe lexicalist spirit argued for by MacDonaTanenhaus, and their colleagues, becausecount across whole classes of structures rathan ones headed by particular words. Gsizes 1 and 2 occur too sparsely for meaninfrequencies to be computed. We therefore tegrain size 3, based on an automatic search oUPenn Treebank parsed corpora (Marcus,torini, & Marcinkiewicz, 1993), which yieldeproportions such as the following, relevantexample (17) above:

((demand. . . for)/demand) 5 0.390((consider. . . for)/consider) 5 0.015.

Across all 15 items, we found mean proportifor N 1 prep of 0.165 and forV 1 prep of
0.040; theN 1 prepcombinations were propor-
,

-t

-

-

in

eyernlde

n-

tionally significantly more frequent on averat(14) 5 2.87,p , .05, based on the comparisof arcsine proportions. Breaking this dowthere were 11 items where the NP argumentmore frequent, 3 items where the VP modiwas more frequent, and 1 item where bothzero frequency in the corpora. Thus, it is psible that our specific experimental findicould be accounted for by a frequency-baalgorithm. However, it is also worth noting thon an item-by-item basis, the proportion diffences betweenN 1 prep and V 1 prep pairsshowed no hint of a linear correlation with treading time differences in either Experimenor Experiment 2 (p . .5 for each correlation

Turning now to the second question raiabove, it is our contention that if relative frquency were all there was to argumenthoodsyntactic differences between argumentsmodifiers summarized in Appendix A wouldleft unexplained. While one might be ableimagine intuitively why more frequently coccurring pairs of head and PP might confoto the generalizations about ordering (see (2(23) in Appendix A), separation ((35)–(39and pro-form replacement ((30)–(32)), it is uclear why the iterativity ((28)–(29)) andwhextraction ((40)–(50)) tests should work as tdo, instead of in the opposite way. Furthermno frequency-based account of any of the amenthood diagnostics has actually beenposed. Thus, we take argumenthood to bedamentally a syntactico-semantic propertylinguistic expressions. While argumenthoodfrequency surely correlate to a high degreesuggest that frequency differences derive fargument–structure differences, rather thanversa. (MacDonald et al. also acknowledge sa possibility.)

Given this claim about the grammar, werefine the third question: Is the parser actusensitive to grammatical argumenthood or to sple frequencies? These two options are diguishable in principle by using examples that ctrapose a low-frequency argument and a hfrequency modifier. In such cases, thehypotheses make opposite predictions aboutcessing. Our initial attempts to carry out this
indicate that it will be difficult in practice, how-

ablonro-co

semsomexamntatedbute

uisyl-il-

aticsend iarus

sk.al

gewinioncie, aultlike

scet intto

hebetaifor

bil-ace

il-ce

l toin

asedouteyhy-trat-of

rvedncy-f therbi-;

rs).gu-le toanofntfre-lau-allyti-ntal

ofbi-

e ofthe

en-ingfor

icitngitial

s fornounsnly inlish;hus,

s, thensult-bilityo theyrces.


ever. The biggest obstacle is that the availparsed English corpora, consisting of aboutmillion words each, are insufficiently large to pvide reasonable estimates of the actualoccurrence frequencies required, because thevolve the presence of two particular lexical iteOur scrutiny of co-occurrence frequencies cputed from this corpus shows them to betremely sensitive to the particular materials spled from the corpus and intuitively unrepresetive of everyday English. We therefore attempto derive frequency estimates from a largerunparsed corpus, the “Tipster” 1989 AssociaPress Newswire corpus (compiled by the Lingtic Data Consortium at the University of Pennsvania), which consists of approximately 41 mlion words. The problem here is that automsearching cannot do much more than findtences in which the 2 specified words occurrethe requisite order. To determine whether theyin the relevant syntactic relationship, one mfilter all the sentences by hand, an arduous tais therefore impractical to attempt to extractV 1 P and N1 P collocation frequencies from larunparsed corpora. Rather, some way of narrothe search to plausible candidate combinatwas required. We used the relative frequenfrom the Penn Treebank, plus our intuitionsguides to promising choices of words. The resof such preliminary attempts were proportionsthe following: ((study. . .for)/study) 5 0.054 and((possibility. . .for)/possibility) 5 0.018. (While inits intransitive use,study for NPprobably involvea PP argument, the sentences we test will nesarily have a direct object, so the relevant facthat instudy NP for NP,the PP is not an argumeof the verb.) Of course, in order for such pairsbe useful in constructing stimuli for testing tcompeting hypotheses, it must additionallypossible to construct a plausible sentence coning the verb with the noun as its direct object,example, (18).

(18) (a) The UN envoy studied the possiities for a truce before he left on the pemission.

(b) The UN envoy studied the possibities for a while before he left on the pea
mission.
ee

-in-.----

td-

-netIt

l

gssss

s-s

n-

It remains to be seen whether it is practicaconstruct a full set of experimental stimulithis manner.

Frequency-based versus semantically bapproaches differ not only in their claims abhow the parser works, but also in how threlate parsing to language production. Apothesis such as the Argument Preference Segy provides the basis for an explanationboth human parsing behavior and the obseco-occurrence patterns, whereas the frequebased approach does not offer an account oco-occurrence patterns, treating them as atrary (Gibson, Schu¨tze, & Salomon, 1996Stevenson & Merlo, 1997; among otheGiven the semantic characterization of arments versus modifiers, it seems reasonabview the distinction roughly as contrastingintrinsic component of some event or stateaffairs with an incidental property of that eveor state. To the extent that co-occurrencequencies correlate with argumenthood, a psible explanation would be that we are genermore likely to talk about the intrinsic or idenfying properties of events than about accideor nonessential ones.

In conclusion, we have shown that modelsdisambiguation, at least for the V-NP-PP amguity, must take into account some measurthe syntactico-semantic cohesion amongparticular lexical heads that occur in the stence— either argumenthood or somethclosely correlated therewith. The challengefuture work in this area is to develop an expltheory of how all the relevant factors, includiargument status, combine to determine inpreferences in ambiguity resolution.

APPENDIX A: DIAGNOSTICS FORPP ARGUMENTS

In this appendix we summarize six syntactic diagnosticargumenthood of prepositional phrases associated withand verbs. Because our experiments were conducted oEnglish, we discuss only diagnostics that apply in Engmany of them are not applicable in other languages. (Twhen testing for argumenthood effects in other languagerelevant syntax literature for those languages must be coed.) In all cases there is more to be said about the applicaof the diagnostics summarized here than space permits, scannot be safely used without consulting additional sou
Some further details can be found in Schu¨tze, 1995; see Miller,

o nmod

ann th

u-n bnaena) i0a)

theP-Pe nh torely

b uit( um-a etrb -

mul

kenayhersteingific

to

g

dayhen

cathislate(25es a

calt aess tothis

-

-

ennotse itsanand

er inir are

PP isgnoseodi-

entllard

rtar-

by

et-

rful

theotherason.

essesenton,b).

adormrlier,ntsee73;sis,and


1997, for some related discussion. The diagnostics also dall draw the same dividing line between arguments andifiers, and the underlying syntactic mechanisms are in mcases not fully understood; we return to these matters ifinal section.

Optionality

The most common rule of thumb for identifying argments is that arguments to a particular lexical head caobligatory, whereas modifiers are (almost) always optiobut the converse is not true: there are optional argum(Jackendoff, 1977). This test tells us that the PP in (19functioning as an argument of V, while the one in (2could be a modifier:

(19) (a) John put the book in the room.(b) *John put the book.

(20) (a) John saw the book in the room.(b) John saw the book.

This test is listed here for completeness; it is not helpful inpresent context, because if the direct object noun in a V-Nsentence took an obligatory PP argument, there would bambiguity of attachment of the PP—it would have to attacthe NP.Verbsthat take obligatory PP arguments have ra

een used in the experimental literature on this ambigbut see Ferreira & Clifton, 1986 and Britt, 1994), presbly because this would create a confounding asymmetween NP and VP attachment—nounsin most environ

ments do not take obligatory arguments. Our own stitherefore used only optional arguments.

Ordering

Arguments generally must precede modifiers (Jacdoff, 1977; Pollard & Sag, 1987), while modifiers mfollow other modifiers and arguments may follow otarguments, as the following contrasts demonstrate. (Aisks indicate only ungrammaticality of the relevant readof a string; for example, (23b) is grammatical if a specring over Buffalo is under discussion.)

(21) (a) a member of Parliament with gray hair(b) *a member with gray hair of Parliament

(22) (a) a man from Paris with gray hair(b) a man with gray hair from Paris

(23) (a) While we were flying home, I gave the ringmy girlfriend over Buffalo.

(b) *While we were flying home, I gave the rinover Buffalo to my girlfriend.

(24) (a) John saw the mouse in the kitchen on Sun(b) John saw the mouse on Sunday in the kitc

Pollard and Sag (1987) note that constituents in VPoften be reordered so that they superficially violategeneralization, but that such reordering is usually correwith focus on the clause-final constituent. For example,focuses the indirect object, which concomitantly requircertain amount of stress.

(25) Lou handed a book last Sunday to the kids.

ot-ye

el,tss

Po

y

y

i

-

r-

..

n

d)

Hence, if in a given sentence it is impossible to put fostress on somenonfinalconstituent, the word order is nobase order, because only derived word orders force strbe on the final phrase. The following examples makeclearer (italics are used to indicate focal stress).

(26) *Lou handed a book onSundayto the kids.(27) (a) Chris read a book onSundayafter lunch.

(b) Chris read a book afterlunch on Sunday.(28) (a) They complained to thelandlord about the ten

ants.(b) They complained about thetenantsto the land

lord.

Thus, (26) confirms thatto the kidsis an argument—thimpossibility of nonfinal stress indicates that the PP cabe base-generated after a temporal modifier PP. Becaubase position is obligatorily before a modifier, it must beargument. By comparison, the alternative orders in (27)(28) show that neither PP obligatorily precedes the othbase word order. Therefore, the two PPs in each pashown to have the same status. Once the status of oneindependently established, this test can be used to diathe status of another; in this case, (27) involves two mfiers and (28) involves two arguments.

Iterativity

Modifier phrases can usually iterate while argumphrases cannot (cf. Fillmore, 1968; Bresnan, 1982; Po& Sag, 1987):

(29) (a) *Chris rented the gazebo to yuppies, to libeians.

(b) Kim met Sandy in Baltimore in the hotel lobin a corner.

(30) (a) *I met a student of biology, of molecular genics.

(b) I met a student with blue eyes with a wondesmile.

Care must be taken in applying this test, however: ifiterated phrases are semantically incompatible with eachthen this can make the example seem bad for the wrong reFor instance, the badness of*I met a student with blue eywith green eyesis uninformative. In general, good caseem to require modifiers that refer to slightly differproperties or else to a different level of detail (Bruns1992), as with the increasingly specific locations in (29

Pro-form Replacement

If a PP isobligatorily deleted when the noun or verb hewith which it is associated is replaced by a pro-form (a fthat can stand in for an expression that appeared easuch asonein (31) ordo soin (32)), that PP is an argumeof the replaced head; if not, it is a modifier. [For NP,Lakoff, 1970, crediting L. Baker; for VP, see Ross, 19Lakoff & Ross, 1976 (1966); and Klima, 1962; for analysee Jackendoff, 1977.] This is shown for nouns in (31),
for verbs withdo soin (32) and with pseudoclefts in (33).

the

om

o o

so

me

er.

yn b31b

that thNP

ichrm(33eleteentdens n

the

,

o.

byif it60

ut d

byent

h ae.g.,rs

Thebe)ectersh &

r &

ny

so-

n?

as a92)tionalrathersion).(,

b trac-it is

r

t canand

s do

of

utto

entlye for


(31) (a) *The President proposed the solution toforeign crisis, not the one to the domestic crisis.

(b) I know the woman from Peel, not the one frLondon.

(32) (a) *John put a book on the table, and Sue did sthe shelf.

(b) John filled out the form in pen, and Mary didin pencil.

(33) (a) *What the authorities did on Mary was blathe arson.

(b) What Chris did in the backyard was cook dinn

In (31a),do sohas been substituted just forsolution,leavingbehind the following PP headed byto; the ungrammaticalitof the sentence indicates that the PP cannot be left imust be deleted, so it is an argument. By contrast, in (a PP headed byfrom appears withone, just as it did withwoman,so the grammaticality of the sentence indicatesthis PP is not an argument. The facts in (32) show thasame contrast holds for PPs that follow a direct objectIn (33a) we have tried to combine the PPon Marywith thepro-formdid, and this has yielded ungrammaticality, whmeans that this PP must be deleted when the pro-foused, so the PP is an argument. On the other hand, inthe PP is combined with the pro-formdid, and since thsentence is grammatical, that PP is not obligatorily dewhen the pro-form is used, so the PP is not an argum

The noun test is inapplicable if the intended antecenoun is not countable. Thus, the badness of (34) doebear on the argumenthood of the PPs, becausewater anddestructioncannot be antecedents for aoneanaphor.

(34) *The water in the lake is cleaner than the one inriver.

Similarly, the verb replaced bydo so cannot be stativehence the badness of (35).

(35) *Bill knew about our affair, and Harry did so, to

Separation from the Head

If a PP can be separated from its associated nouncopula or a relative clause construction, it is a modifier;cannot, it is an argument (cf. Jackendoff, 1977, p.Grimshaw, 1990).

(36) (a) the man (who) is from Paris(b) the book (that) was by/about Chomsky

(37) (a) *a student (who) was of physics(b) *the solution (that) was to the problem

Certain uses ofwith cannot be paraphrased bybe with forindependent reasons (cf. Freeze, 1992, inter alia), ballow a paraphrase withhave:

(38) (a) *a man (who) was with blue eyes(b) a man (who) has blue eyes

For VP elements, if a preposed PP can be followedquestion it is a modifier, if it cannot it is an argum
(Reinhart, 1983; Emonds, 1976):
n

ut)

te.

isa)

d.t

ot

a

;

o

a

(39) (a) On Tuesday, who drove to the store?(b) At the concert, did you fall asleep?

(40) (a) *On the shelf, who put the book?(b) *To Mary, did John give a ring?

Wh-Extraction

Wh-extraction is a syntactic phenomenon in whicphrase consisting of or containing a question word (who, what, which,etc., most of which begin with the lettewh in English) is moved to the beginning of a clause.position from which thewh-phrase has moved mayindicated with a trace, notated ast. As (41) versus (42shows,wh-extraction of or from a PP that is inside a dirobject is generally possible, but this is not so for modifi(Radford, 1988; cf. Ross, 1967; Jackendoff, 1977; BacHorn, 1976; Chomsky, 1977; Koster, 1978; CulicoveWilkins, 1984; Diesing, 1992):

(41) (a) *Which shelf did you read [a book ont]?(b) *With what kind of sleeves did you buy [ma

sweaterst]?(42) (a) Which problem did the President suggest a

lution to?(b) Of which city did you witness the destructio

There is an orthogonal restriction, traditionally statedban on extraction from “specific” NPs, which Diesing (19suggests is actually a ban on extraction from presupposiNPs, NPs that presuppose the existence of their referent,than asserting it (see Diesing, 1992, for extensive discusShe argues that the complements of experiencer verbslove,like, appreciate, hate) and destruction verbs (destroy, burnan, tear up) are presuppositional. As a consequence, ex

tion is generally impossible from those complements, butpossible from complements to verbs of creation (write, paint,draw) or using (read, play, publish, buy, see). Thus, the latteare the best ones for applying this test.

Further caution is advised because the extraction tesyield grammatical results for the wrong reason. BachHorn (1976) note that the following types of sentencenot unambiguously involve extraction from an NP:

(43) (a) Which country did you explore caves in?(b) Who did you take a picture of?

The reason is thatexplore licenses a PP, independentcaves,as in John explored them in Mexico;similarly, wecan sayJohn took it of Mary(e.g., in response toWhere didthat picture come from?). Thus, (43) tells us nothing abothe possibility of extracting from NPs. The solution isapply the test using a main verb that is shown independnot to license the relevant PP. Thus, (44c) is evidencargumenthood, but (45c) is evidence against it.

(44) (a) I saw a picture of Mary.(b) *I saw it of Mary.(c) Who did you see a picture of?

(45) (a) I know men in overcoats.(b) *I know them in overcoats.
(c) *Which overcoats do you know men in?

tracera

e,nts

rac-

hn

her

hisr

them

mt

god

B thp ofP

orm stui lvel an1 y,1 5;M viewo div& 5;a

thes d (ao thera , ant ker1

nd.

uldh

S thati 996m anS sa -m ent

theargu-s andtheire inrs towhilehapsof a

on ofcticdiag-

mly.e,weenPre-e forbouting.e oflwaysd, forram-

tence)theanticin aan

: theicatef ananyish-

dis-ouble

righters,

If thecloseused

uld

thean-

(6)ctics ofll-or-at is(8)

asseseldsbe

ce of


At the VP level, the analogous pattern emerges in extion from “weak islands,” syntactic constituents of sevtypes (exemplified below) that prohibit certain kinds ofwh-extraction (Rizzi, 1990; Cinque, 1990; Hukari & Levin1995; Szabolcsi & Zwarts, 1993). In these environmeextraction of or from arguments is relatively good, exttion of or from modifiers is much worse.

Wh-islands.

(46) ?To which friend do you wonder [whether Jogave the bookt]?

(47) *On which day last week do you wonder [whetJohn bought the bookt]?

Adversative/factive islands.

(48) To which friend do you deny [that Bob gaveing t]?

(49) *At what time do you regret [that Bob walked toarkett]?

Extraposition islands.

(50) Which critic is it time [for Mary to describe her filo t]?

(51) *What country is it a scandal [that the Senatorivorced int]?

aker (1988) points out some apparent exceptions toattern, which might be explained in the frameworkesetsky (1995); see Schu¨tze, 1995, for discussion.

InstrumentalsThe status of instrumentalwith phrases as argumentsodifiers is of particular interest because some previous

es of the V-NP-PP ambiguity have restricted themseargely (Rayner et al., 1983) or entirely (Altmann & Steedm988; Clifton & Ferreira, 1989; Spivey-Knowlton & Sediv995) to the prepositionwith. See Nilsen, 1973; Gruber, 196arantz, 1984; Brunson, 1992; and Larson, 1988, for ref the primary literature concerning instrumentals; see SeSpivey-Knowlton, 1994; Spivey-Knowlton & Sedivy, 199

nd Schu¨tze, 1995, for discussion of other tests.Instrumentals pattern with arguments on three of

yntactic tests. As shown in (52), they cannot be iteratebserved by Lakoff (1968); cf. Bresnan, 1982). Like orguments, they cannot be followed by a question (53)

hey can be extracted from weak islands (54) (cf. Ba988, p. 243):

(52) *John cut the meat with a knife with the sharp e(53) *With the knife, who sliced the salami?(54) With which key do you deny that the butler co

ave opened the door?

ee Schu¨tze, 1995 for additional semantic evidencenstrumentals have argument properties; Konieczny (1

ade this claim based on tests summarized by Pollardag (1987). Given the conclusion thatwith instrumentalre arguments, we did not usewith in any of our experiental items, because our aim was to test VP attachm

hat could only be modifiers.

-l

,

t

is

d-s,

y

s

d,

)d

ts

Interpreting the Diagnostics

It is worth asking what the relationship is betweensyntactic contrasts illustrated here and the nature ofmenthood. One traditional view has been that argumentmodifiers appear in different structural positions and thatcontrasting syntactic behavior follows from the differencposition. Traditionally, arguments were taken to be sistethe syntactic head that assigned their semantic role,modifiers were sisters to an X-bar level category, with perthe further restriction that they must also be daughters(recursive) X-bar level category. Leaving aside the questihow this distinction can be replicated in current syntatheory, we can ask whether the purported argumenthoodnostics appear to reflect this structural dichotomy uniforThe answer seems to be No (Schu¨tze, 1995). Furthermorintuitions on these tests often seem to lie in the range betcomplete grammaticality and strong ungrammaticality.cisely what this means is a deep and unresolved issusyntactic theory, but it is tangential to our hypothesis aparsing. Among the possible answers are the follow(Thanks to an anonymous reviewer for laying out somthese.) First, perhaps arguments and modifiers do not aappear in different syntactic environments, as suggesteexample, by Pesetsky (1995). Second, it is possible that gmaticality judgments are reflecting factors other than senstructure and semantic argumenthood—see Schu¨tze (1996and work cited there. A third possibility is that some ofdiagnostics may reflect an underlying continuous semnotion of argumenthood that is only imperfectly replicatedpositional dichotomy. Fourth, it is possible that there isunderlying dichotomy at the level of semantic compositionrule for combining a head with an argument to form a predmust be different from the rule for predicating a modifier oalready-formed predicate. However, there may well be mcases where the resulting meaning is virtually indistinguable. This is particularly likely in the construction undercussion here, because prepositions can apparently do dduty as either semantically contentful heads in their ownor relatively meaningless words, similar to case markwhose presence is simply dictated by the choice of verb.meaning that a verb assigns to its argument happens to beto the meaning that the preposition itself can assign whento form a modifier, intuitions of marginal acceptability wonot be surprising.

What is relevant to the parsing questions is thatavailable intuitions about sentences, be they “direct” semtic intuitions of the sort described in connection with (3)–or intuitions about the well-formedness of various syntamanipulations, are not binary. As a result, for purposepresentation we assume that argumenthood is not an anothing phenomenon, but that it occurs in degrees. Thwhy in formulating the Argument Preference Strategywe proposed that the parser maximizes theextentof argu-ment relations. This means that an attachment that ponly two or three argument diagnostics, or one that yimarginal results on all of them, is still hypothesized topreferred by the parser over one that shows no eviden
an argument relation.

re thThePPthe

eac

tsus

bigc

m-

(att

form

tionr-

the

er a

/are

the

nting

tanded

nnt

into

til

on-7

an

ith

ns

t-o

ityonal

K.their

-

uli--

bel-theyn-k-

xt.

singre as

n-

e of-

h &

F pe-


APPENDIX B: SELF-PACEDREADING ITEMS

In each item, the two nouns separated by a slash aNP argument and VP modifier versions, in that order.control condition in Experiment 1 omitted the italicizedentirely. The control condition in Experiment 2 replaceditalicized PP with the one given in parentheses aftersentence.

1. The financial administrator announced many cuinthe staff/meetingalthough he knew it would upset numeropeople. (at the meeting)

2. The older campers questioned John’s authorityoverthe group/summerbut they came to respect him after theampfire. (during the summer)3. The community leader withheld his supportfor the

candidate/momentwhile he waited to hear what the incubent would say. (during the debate)

4. The confused suspect admitted his involvementin therobbery/morningbut later he claimed he was coerced.he station)

5. The revised policy increased people’s alienationfromthe system/outsetbut the senate failed to recognize it

any months. (across the country)6. The environmental agency allowed some exemp

from the law/startbut there were still objections from cetain industries. (at the hearing)

7. The science teacher encouraged much excitemenoverthe project/weekby bringing in new experiments for tchildren to try every day. (during the week)

8. The keen shopper expressed his interestin a wallet/hurry just before the store was about to close. (aftmoment)

9. The police detective conducted a searchfor a weaponday but he never looked in the suspect’s car. (at a whouse)

10. The experienced general concealed his surpriseat thedecision/timebut was not sure what to do next. (formoment)

11. The comedian’s friends showed their amusemeatthe story/partybecause they were thoroughly enjoyhemselves. (during the party)

12. The company lawyers considered employee demfor a raise/monthbut they didn’t act until a strike seemimminent. (after a month)

13. The board members discussed Mary’s inclusiointhe process/afternoonafter they had met with the presideover lunch. (at the briefing)

14. The boisterous fans gained some confidencein thecoach/endafter their team came back to force the gameovertime. (by the end)

15. The new conductor noticed Arthur’s absencefromthe rehearsal/beginningbut he didn’t comment on it un
later. (at the rehearsal)
e

h

s

-

s

REFERENCES

Abney, S. P. (1987). Licensing and parsing. In J. McDough & B. Plunkett (Eds.),Proceedings of NELS 1(Vol. 1, pp. 1–15). Amherst, MA: GLSA.

Abney, S. P. (1989). A computational model of humparsing. Journal of Psycholinguistic Research,18,129–144.

Altmann, G., & Steedman, M. (1988). Interaction wcontext during human sentence processing.Cognition,30, 191–238.

Bach, E., & Horn, G. M. (1976). Remarks on “Conditioon transformations.”Linguistic Inquiry,7, 265–299.

Baker, M. C. (1988).Incorporation: A theory of grammaical function changing.Chicago: Univ. of ChicagPress.

Bresnan, J. (1982). Polyadicity. In J. Bresnan (Ed.),Themental representation of grammatical relations(pp.149–172). Cambridge, MA: MIT Press.

Britt, M. A. (1994). The interaction of referential ambiguand argument structure in the parsing of prepositiphrases.Journal of Memory and Language,33, 251–283.

Britt, M. A., Perfetti, C. A., Garrod, S., & Rayner,(1992). Parsing in discourse: Context effects andlimits. Journal of Memory and Language,31, 293–314.

Brunson, B. A. (1992).Thematic discontinuity.Ph.D. dissertation, Univ. of Toronto.

Chomsky, N. (1977). On WH-movement. In P. W. Ccover, T. Wasow, & A. Akmajian (Eds.),Formal syntax (pp. 71–132). New York: Academic Press.

Chomsky, N. (1995). Bare phrase structure. In G. Wehuth (Ed.),Government and binding theory andminimalist program: Principles and parameters in stactic theory(pp. 383–439). Cambridge, MA: Blacwell.

Cinque, G. (1990).Types of A-bar dependencies.Cam-bridge, MA: MIT Press.

Clifton, C., Jr., & Ferreira, F. (1989). Ambiguity in conteLanguage and Cognitive Processes,4, SI 77–103.

Clifton, C., Jr., Speer, S., & Abney, S. P. (1991). Pararguments: Phrase structure and argument structudeterminants of initial parsing decisions.Journal ofMemory and Language,30, 251–271.

Culicover, P., & Wilkins, W. (1984).Locality in linguistictheory.Orlando: Academic Press.

Diesing, M. (1992). Indefinites. Cambridge, MA: MITPress.

Emonds, J. E. (1976).A transformational approach to Eglish syntax.New York: Academic Press.

Ferreira, F., & Clifton, C., Jr. (1986). The independencsyntactic processing.Journal of Memory and Language,25, 348–368.

Fillmore, C. J. (1968). The case for case. In E. W. BacR. T. Harms (Eds.),Universals in linguistic theory(pp.0–88). New York: Holt, Rinehart & Winston.

ord, M., Bresnan, J., & Kaplan, R. M. (1982). A com
tence-based theory of syntactic closure. In J. Bresnan

la-

of-

F cticn-

rom

roum?er

,

e: A.ingovesen

ime

:

with.),

encrt-

n-ingiv.n-

essin-

:

n.

n-ary

ondin.se

2).sion.

ng.

ture

its-an-

se-,

, G.nce

:

is-

pt of

e

llyen-o-at-

Se-

sen--

. S.es-

M s.

M .ish:

M nd

s an

ts:but

et-


(Ed.),The mental representation of grammatical retions (pp. 727–796). Cambridge, MA: MIT Press.

Francis, W. N., & Kucˇera, H. (1982).Frequency analysisEnglish usage: Lexicon and grammar.Boston: Houghton Mifflin.

razier, L. (1978).On comprehending sentences: Syntaparsing strategies.Ph.D. dissertation, Univ. of Conecticut.

Frazier, L. (1987). Syntactic processing: Evidence fDutch. Natural Language and Linguistic Theory,5,519–559.

Frazier, L. (1990). Parsing modifiers: Special purposetines in the human sentence processing mechanisD. A. Balota, G. B. Flores d’Arcais, & K. Rayn(Eds.),Comprehension processes in reading(pp. 303–330). Hillsdale, NJ: Erlbaum.

Frazier, L., & Clifton, C., Jr. (1996).Construal.CambridgeMA: MIT Press.

Frazier, L., & Fodor, J. D. (1978). The Sausage Machinnew two-stage parsing model.Cognition,6, 291–325

Frazier, L., & Rayner, K. (1982). Making and correcterrors during sentence comprehension: Eye mments in the analysis of structurally ambiguoustences.Cognitive Psychology,14, 178–210.

Freeze, R. (1992). Existentials and other locatives.Lan-guage,68, 553–595.

Garrod, S. C., & Sanford, A. J. (1985). On the real tcharacter of interpretation during reading.Languageand Cognitive Processes,1, 43–59.

Gazdar, G., Klein, E., Pullum, G., & Sag, I. (1985).Gen-eralized phrase structure grammar.Cambridge, MAHarvard Univ. Press.

Gibson, E. (1987). Garden-path effects in a parserparallel architecture. In A. Miller & J. Powers (EdsProceedings of the Fourth Eastern States Conferon Linguistics(pp. 88–99). Columbus, OH: Depament of Linguistics, Ohio State Univ.

Gibson, E. (1991).A computational theory of human liguistic processing: Memory limitations and processbreakdown.Ph.D. dissertation, Carnegie Mellon Un

Gibson, E. (1998). Linguistic complexity: Locality of sytactic dependencies.Cognition,68, 1–76.

Gibson, E., Schu¨tze, C. T., & Salomon, A. (1996). Threlationship between the frequency and the procecomplexity of linguistic structure.Journal of Psycholinguistic Research,25, 59–92.

Grimshaw, J. (1990).Argument structure.Cambridge, MAMIT Press.

Gruber, J. S. (1965).Studies in lexical relations.Ph.D.dissertation. MIT.

Hukari, T. E., & Levine, R. D. (1995). Adjunct extractioJournal of Linguistics,31, 195–226.

Igoa, J. M. (1995, April).Parsing decisions and the costrual hypothesis: Attachment preferences in primphrases in Spanish.Paper presented at the SecSymposium on Psycholinguistics, Tarragona, Spa

Jackendoff, R. (1977).X-bar syntax: A study of phra
structure.Cambridge, MA: MIT Press.
-In

--

e

g

Just, M. A., Carpenter, P. A., & Woolley, J. D. (198Paradigms and processes in reading comprehenJournal of Experimental Psychology: General,111,228–238.

Kayne, R. S. (1984).Connectedness and binary branchiDordrecht: Foris.

Kimball, J. (1973). Seven principles of surface strucparsing in natural language.Cognition,2, 15–47.

Klima, E. S. (1962). Structure at the lexical level andimplication for transfer grammar. In1961 International Conference on Machine Translation of Lguages and Applied Language Analysis(Vol. 1, pp.97–108). London: HMSO.

Konieczny, L. (1996).Human sentence processing: Amantics-oriented parsing approach.Ph.D. dissertationAlbert-Ludwigs-Univ. Freiburg.

Konieczny, L., Hemforth, B., Scheepers, C., & Strube(1997). The role of lexical heads in parsing: Evidefrom German.Language and Cognitive Processes,12,307–348.

Koster, J. (1978).Locality principles in syntax.DordrechtForis.

Kuno, S., & Takami, K. (1997). Remarks on negativelands.Linguistic Inquiry,28, 553–576.

Lakoff, G. (1968). Instrumental adverbs and the concedeep structure.Foundations of Language,4, 4–29.

Lakoff, G. (1970). Global rules.Language,46, 627–639.Lakoff, G., & Ross, J. R. (1976). Why you can’tdo sointo

the sink. In J. D. McCawley (Ed.),Notes from thlinguistic underground, Syntax and semantics(Vol. 7,pp. 101–111). New York: Academic Press. [Originapublished as “A criterion for verb phrase constitucy,” Report NSF-17 of the Aiken Computation Labratory of Harvard University to the NSF on Mathemical Linguistics and Automatic Translation, 1966.]

Larson, R. K. (1988). Implicit arguments in Situationmantics.Linguistics and Philosophy,11, 169–201.

MacDonald, M. C. (1997). Lexical representations andtence processing: An introduction.Language and Cognitive Processes,12, 121–136.

MacDonald, M. C., Pearlmutter, N. J., & Seidenberg, M(1994). The lexical nature of syntactic ambiguity rolution. Psychological Review,101,676–703.

arantz, A. (1984).On the nature of grammatical relationCambridge, MA: MIT Press.

arcus, M. P., Santorini, B., & Marcinkiewicz, M. A(1993). Building a large annotated corpus of EnglThe Penn Treebank.Computational Linguistics,19,313–330.

arslen-Wilson, W. D. (1973). Linguistic structure aspeech shadowing at very short latencies.Nature,244,522–523.

Marslen-Wilson, W. D. (1975). Sentence perception ainteractive parallel process.Science,189,226–228.

Miller, P. (1997, September).Complements vs. adjuncSome tests which are supposed to distinguish themdon’t really do so.Paper presented at the LAGB me
ing, University of Hertfordshire.

ns-

h::

as-

P ,

P dSLra-

P ram

P tachd

R st

R ac-cesallyr-

R rse

R n.

R :

R

R t,

S d. I.),MIT).s.

S s:gy.

S ofrs-ra-ce

S nal

S y o

:

S nd

S ng

S ybrid

S ando-

S ncects.

S ro-l

S an

T . S.nta-yal,

T at-encetions.

T om-A..),.

T d aam-.

sing

T M.the-on.

T eing.

V si-

(


Mitchell, D. C. (1994). Sentence parsing. In M. A. Gerbacher (Ed.),Handbook of psycholinguistics(pp. 375–409). New York: Academic Press.

Nilsen, D. L. F. (1973).The instrumental case in EnglisSyntactic and semantic considerations.The HagueMouton.

Pesetsky, D. (1995).Zero syntax: Experiencers and ccades.Cambridge, MA: MIT Press.

hillips, C. (1996).Order and structure.Ph.D. dissertationMIT.

ollard, C., & Sag, I. A. (1987).An information-basesyntax and semantics. Vol. 1: Fundamentals. CLecture Notes Number 13.Leland Stanford JunioUniv.: Center for the Study of Language and Informtion.

ritchett, B. (1988). Garden path phenomena and the gmatical basis of language processing.Language,64,539–576.

ynte, J., & Prieur, B. (1996). Prosodic breaks and atment decisions in sentence parsing.Language anCognitive Processes,11, 165–191.

adford, A. (1988).Transformational grammar: A fircourse.Cambridge: Cambridge Univ. Press.

ayner, K., Carlson, M., & Frazier, L. (1983). The intertion of syntax and semantics during sentence proing: Eye movements in the analysis of semanticbiased sentences.Journal of Verbal Learning and Vebal Behavior,22, 358–374.

ayner, K., Garrod, S., & Perfetti, C. A. (1992). Discouinfluences during parsing are delayed.Cognition,45,109–139.

einhart, T. (1983).Anaphora and semantic interpretatioLondon: Croom Helm.

izzi, L. (1990).Relativized Minimality.Cambridge, MAMIT Press.

oss, J. R. (1967).Constraints on variables in syntax.Ph.D.dissertation, MIT.

oss, J. R. (1973).Nearer to Vee.Mimeographed handouMIT.

chutze, C. T. (1995). PP attachment and argumenthooC. T. Schu¨tze, J. B. Ganger, & K. Broihier (EdsPapers on language processing and acquisition,Working Papers in Linguistics(Vol. 26, pp. 95–151Cambridge, MA: MIT Working Papers in Linguistic

chutze, C. T. (1996).The empirical base of linguisticGrammaticality judgments and linguistic methodoloChicago: Univ. of Chicago Press.

edivy, J., & Spivey-Knowlton, M. (1994). The usestructural, lexical, and pragmatic information in paing attachment ambiguities. In C. Clifton, Jr., L. Fzier, & K. Rayner (Eds.),Perspectives on sentenprocessing(pp. 389–413). Hillsdale, NJ: Erlbaum.

hillcock, R. (1982). The on-line resolution of pronomianaphora.Language and Speech,25, 385–402.

peas, M. (1994). Null arguments in a theory of economprojection. In E. Benedicto & J. Runner (Eds.),Func-
tional projections, University of Massachusetts Occa-(
I

-

-

s-

n

f

sional Papers(Vol. 17, pp. 179–208). Amherst, MAGLSA.

peer, S. R., & Clifton, C., Jr. (in press). Plausibility aargument structure in sentence comprehension.Mem-ory and Cognition,26.

pivey-Knowlton, M., & Sedivy, J. C. (1995). Resolviattachment ambiguities with multiple constraints.Cog-nition, 55, 227–267.

tevenson, S. (1994). Competition and recency in a hnetwork model of syntactic disambiguation.Journal ofPsycholinguistic Research,23, 295–322.

tevenson, S., & Merlo, P. (1997). Lexical structureparsing complexity.Language and Cognitive Prcesses,12, 349–399.

winney, D. A. (1979). Lexical access during sentecomprehension: (Re)consideration of context effeJournal of Verbal Learning and Verbal Behavior,18,645–659.

winney, D. A., & Cutler, A. (1979). The access and pcessing of idiomatic expressions.Journal of VerbaLearning and Verbal Behavior,18, 523–534.

zabolcsi, A., & Zwarts, F. (1993). Weak islands andalgebraic semantics for scope taking.Natural Lan-guage Semantics,1, 235–284.

anenhaus, M. K., Carlson, G. N., & Seidenberg, M(1985). Do listeners compute linguistic represetions? In D. Dowty, L. Karttunen, & A. M. Zwick(Eds.), Natural language processing: Psychologiccomputational, and theoretical perspectives(pp. 359–408). Cambridge: Cambridge Univ. Press.

araban, R., & McClelland, J. L. (1988). Constituenttachment and thematic role assignment in sentprocessing: Influences of content-based expectaJournal of Memory and Language,27, 597–632.

araban, R., & McClelland, J. L. (1990). Parsing and cprehension: A multiple-constraint view. In D.Balota, G. B. Flores d’Arcais, & K. Rayner (EdsComprehension processes in reading(pp. 231–263)Hillsdale, NJ: Erlbaum.

rueswell, J. C., & Tanenhaus, M. K. (1994). Towarlexicalist framework for constraint-based syntacticbiguity resolution. In C. Clifton, Jr., L. Frazier, & KRayner (Eds.),Perspectives on sentence proces(pp. 155–179). Hillsdale, NJ: Erlbaum.

rueswell, J. C., Tanenhaus, M. K., & Garnsey, S.(1994). Semantic influences on parsing: Use ofmatic role information in syntactic disambiguatiJournal of Memory and Language,33, 285–318.

yler, L. K., & Marslen-Wilson, W. D. (1977). The on-lineffects of semantic context on syntactic processJournal of Verbal Learning and Verbal Behavior,16,683–692.

estergaard, T. (1977).Prepositional phrases and prepotional verbs: A study in grammatical function.TheHague: Mouton.

Received January 16, 1998)
Revision received September 25, 1998)

Argumenthood and English Prepositional Phrase Attachmenttedlab.mit.edu › tedlab_website ›...

Documents

Transcript of Argumenthood and English Prepositional Phrase Attachmenttedlab.mit.edu › tedlab_website ›...