Reanalysis in Sentence Processing: Evidence against Current Constraint-Based and Two-Stage Models

Journal of Memory and Language 45,225–258 (2001)doi:10.1006/jmla.2001.2773, available online at http://www.academicpress.com on

Reanalysis in Sentence Processing: Evidence against Current Constraint-Based and Two-Stage Models

Roger P. G. van Gompel and Martin J. Pickering

University of Glasgow, Glasgow, Scotland, United Kingdom

and

Matthew J. Traxler

University of South Carolina

Two eye-tracking experiments investigated processing of VP–NP attachment ambiguities. Experiment 1 testedhen se-attach-

o initialdisam-nalysis.

based lex-nberg,erpret

T ath

WerseChrsugcal his Posgradportackngram

AvanDun229

sentences in which there was an initial bias toward VP attachment. Readers experienced more difficulty wmantic information disambiguated the sentences to NP attachment than when it disambiguated them to VPment or when it was consistent with either analysis. Experiment 2 tested sentences in which there was nbias toward either VP or NP attachment. Readers experienced more difficulty when semantic informationbiguated the sentences to NP attachment or VP attachment than when it was consistent with either aWe argue that these results challenge theories that assume a competition mechanism, such as constraint-icalist accounts (e.g., MacDonald, Pearlmutter, & Seidenberg, 1994; McRae, Spivey-Knowlton, & Seide1998; Spivey-Knowlton & Sedivy, 1995) and fixed-choice two-stage models (e.g., Frazier, 1987). We intthe results in terms of the unrestricted race model (cf. Traxler, Pickering, & Clifton, 1998).© 2001 Academic Press

Key words:sentence processing; syntactic ambiguity resolution; reanalysis; competition.

wo competing and largely incompatibleknown account of this class is thegarden-p
ns
s

nio

e

ialin

es.g.,-&r,

thate-

tt,

classes of account dominate current senteprocessing research. According to one clathe processor computes syntactic analysesally, in two stages. In the first stage, it drawon a restricted range of information sourcesconstruct an initial analysis. During the secostage, it accesses other sources of informatwhich may sometimes cause it to abandoninitial analysis and compute another. The b

s22

thein

n-e-ainld,e,y-;4;an,

e thank Steven Frisson, Simon Garrod, Simon Lidge, Don Mitchell, Robin Morris, Keith Rayner

istoph Scheepers, and an anonymous reviewer for usgestions and discussions and Keith Edwards for techsupport. Special thanks to Chuck Clifton for the use software. Roger van Gompel acknowledges an EStgraduate Studentship and a University of Glasgow Pouate Studentship. Martin Pickering acknowledges su

of ESRC Research Grant R000234542. Matt Traxowledges support from the CLASS and SPAR prs at the University of South Carolina.

ddress correspondence and reprint requests to Ro Gompel, Department of Psychology, University odee, Dundee DD1 4HN, Scotland, UK. Fax:144 1382993. E-mail: [email protected].

0749-596X/01 $35.00Copyright © 2001 by Academic Pres

5

ces,

eri-stodn,

itsst

model, in which the processor makes initdecisions on the basis of strategies definedterms of syntactic information alone and usthematic information in the second stage (eFerreira & Clifton, 1986; Ferreira & Henderson, 1990; Frazier, 1979, 1987; FrazierRayner, 1982; Rayner, Carlson, & Frazie1983). Other two-stage accounts assumethe processor uses thematic information immdiately (Abney, 1989; Crocker, 1995; Pritche1992; see Clifton, Speer, & Abney, 1991).

The second class of model assumes thatprocessor can activate multiple analysesparallel. It employs both syntactic and nonsytactic information in a single stage to forground one analysis, but other analyses remactivated (e.g., MacDonald, 1994; MacDonaPearlmutter, & Seidenberg, 1994; McRaSpivey-Knowlton, & Tanenhaus, 1998; SpiveKnowlton & Sedivy, 1995; Trueswell, 1996Trueswell, Tanenhaus, & Garnsey, 199Trueswell, Tanenhaus, & Kello, 1993; Tarab& McClelland, 1988; Tyler & Marslen-Wilson

v-,efulni-ofRCst-p-

lero-

gerf

All rights of reproduction in any form reserved.

R

s

e soo;

e

sim2

tite;com

r

s

a

scs

nlh b

. r

ible,

et the al-ularWee

r,theol-h-chs is

redm-ayscu-isese-ch

ni-n-r-inr,estsre-re-

ofe,lv-unnguningn-n-n-ns

e-inge-r-

226 VAN GOMPEL, PICKE

1977). Current models of this type are knowasconstraint-based lexicalistmodels.

Current data cannot satisfactorily distinguithese accounts. Most research has attemptedetermine whether all available information used during the very earliest stages of procing. Some studies suggest that the processortially ignores subcategorization preferences,mantic plausibility, and felicity with respect tdiscourse context (e.g., Ferreira & Henders1990; Ferreira & Clifton, 1986; Mitchell, 1987Mitchell, Corley, & Garnham, 1992; Rayner al., 1983; Rayner, Garrod, & Perfetti, 1992However, other studies find evidence that it uone or more of these sources of information mediately (Altmann, Garnham, & Dennis, 199Altmann, Garnham, & Henstra, 1994; Altman& Steedman, 1988; Britt, 1994; Britt, PerfetGarrod, & Rayner, 1992; Garnsey, PearlmutMyers, & Lotocky, 1997; Trueswell, 1996Trueswell et al., 1993, 1994). In order to acount for the latter findings, proponents of twstage accounts claim that the second stage begin extremely quickly and that this latter sof studies tapped processes occurring duringsecond stage. They suggest that currently avable techniques may be insensitive to procesoccurring in the first stage (e.g., Clifton & Fereira, 1989). On the other hand, proponentsone-stage accounts argue that studies that fafind early effects used weak manipulations.

The Source of Processing Difficulty: Reanalyvs Competition

Let us therefore consider whether the counts can be distinguished in other ways. Apfrom strategies and information sources uduring initial processing, the major differenbetween the models concerns the mechaniunderlying ambiguity resolution.

Reanalysis models. Two-stage models are, igeneral, serial models. The processor initiaselects one analysis, say analysis A. If tanalysis subsequently becomes impossibleunlikely), the processor reanalyzes, thereadopting analysis B, and disruption ensues (eFrazier & Rayner, 1982; Rayner et al., 1983)the initial analysis is consistent with later info
mation, there is no need for reanalysis. It do
ING, AND TRAXLER

n

hd toisss-ini-e-

n,

t).es-

;n,r,

--ay

ettheail-ses-of

il to

is

c-artedems

lyis(ory

.g.,If-

not matter whether other analyses are possbecause they have not been considered.

An important distinction within two-stagmodels is between theories that assume thaprocessor (of any speaker of the language)ways adopts the same analysis of a particsyntactic ambiguity and theories that do not. refer to the former as fixed-choice two-stagmodelsand the latter as variable-choice two-stage models. The garden-path theory (Frazie1979, 1987) is the best known example of former. In this model, the processor always flows the structural principles of minimal attacment and late closure in determining whianalysis to adopt. The nonpreferred analysinever adopted.

Several other models can also be considefixed-choice models. These models have in comon that they assume that the processor alwuses the same parsing principles for any partilar structure and that there is no random nothat may cause a nonpreferred analysis to belected. For example, in thematic accounts suas Abney’s (1989), readers should always itially adopt argument attachment analyses. Aother example of a fixed-choice model is refeential theory (Altmann & Steedman, 1988; Cra& Steedman, 1985; Ni, Crain, & Shankweile1996). Although the processor initially proposmultiple analyses in parallel, it rapidly selecthe analysis that requires fewest unsatisfied psuppositions. Analyses that contain more psuppositions are never selected. In the caseambiguities involving a definite noun phrasthis leads to the selection of an analysis invoing a simple noun phrase over a complex nophrase in a null context or a context mentionione entity of the type referred to by the nophrase, and a selection of the analysis involvthe complex noun phrase in a context mentioing more than one such entity. (Note that referetial theory and thematic accounts, unlike gardepath theory, do not necessarily make predictiofor all syntactic ambiguities.)

There are few clearly articulated variablchoice models. The main exception is the tunhypothesis, in which parsing decisions are dtermined by the frequency with which the alte
esnative analyses are used in the language

5

ege

nu

e

l0

e

-eo

s

raicr

tu

n

8 lg

bly

ne

-

-

ng

ndn-form-une-m-

ro-n- nottheyNP1ingm-P2if- to-

t to

cebi-izeui-nsgu-6;

-n,

iern

re-

REANALYSIS IN SEN

(Mitchell, Cuetos, Corley, & Brysbaert, 199see also Brysbaert & Mitchell, 1996; Mitche& Cuetos, 1991). Assuming that speakers mdiffer in their exposure to different analys(Cuetos, Mitchell, & Corley, 1996), the tuninhypothesis predicts that the analysis that diffent people assign to a particular sentence nnot be the same. However, given that frequeis the only type of information readers use ding initial analysis, any given person will havefixed preference for a particular sentence typ

Traxler, Pickering, and Clifton (1998) proposed another variable-choice model, whichshall call the unrestricted race model(see alsoTraxler, Pickering, Clifton, & Van Gompe2000; Van Gompel, Pickering, & Traxler, 200and Lewis, 1999; McRoy & Hirst, 1990, fosimilar proposals). As in constraint-based thries, there is no restriction on the sources offormation that can provide support for the dferent analyses of an ambiguous structuhence it is unrestricted. In the model, the alternative structures of a syntactic ambiguity are gaged in a race, with the structure that is cstructed fastest being adopted. Although itassumed that the processor attempts to consmultiple analyses in parallel, only a singanalysis is adopted. If this analysis is incontent with later information, the processor hasreanalyze, and processing difficulty occuThus, the unrestricted race model is a two-streanalysis model. It is also a variable-chomodel, because the initial analysis of a syntaambiguity is affected both by individual diffeences (as in the tuning hypothesis) and by stactic and nonsyntactic characteristics of sentences. Because the strength of these soof information may differ for each individuasentence, preferences may differ for each stence read by a particular person. Indeed, wesume that any unrestricted two-stage accoua variable-choice model.

In support of this model, Traxler et al. (199conducted three eye-tracking experimentsadjunct attachment ambiguities involving retive clauses. They compared globally ambious sentences with sentences that were disbiguated toward one or other analysis. In o
experiment, they investigated sentences such
TENCE PROCESSING 227

;llays

r-eedcyr-a.-we

,;

ro-in-if-re;

n-n-is

tructleis- tos.ge

cetic-yn-herceslen- as-t is

)ona-u-am-ne

in (1). In (1a) the relative clause can plausibe attached to either NP1 (the son) or NP2 (thedriver), whereas in (1b) and (1c) only oanalysis is plausible:

(1a) The son of the driver that had themoustache was pretty cool. (ambiguous)

(1b) The car of the driver that had themoustache was pretty cool. (NP2 attachment)

(1c) The driver of the car that had themoustache was pretty cool. (NP1 attachment)

Traxler et al. observed that total readitimes for the critical word (moustache) werelonger in the disambiguated conditions (1b) a(1c) than in the ambiguous condition (1a). Aother experiment showed a similar patternrelative clause ambiguities that were disabiguated by the gender of a reflexive prono(himself/herself), but the effect occurred somwhat earlier, on regressions from the words imediately following the disambiguating word.

Because both disambiguated conditions pduced difficulty relative to the ambiguous codition, these results suggest that readers didalways adopt the same analysis. Instead,suggest that the processor adopted the analysis on some proportion of trials, resultin difficulty when the sentence was disabiguated toward NP2 attachment; and the Nanalysis on another proportion, resulting in dficulty when the sentence was disambiguatedward NP1. Hence, these results are difficulreconcile with fixed-choice models.

However, the results only provide evidenfor variable-choice accounts of adjunct amguities. It may not be possible to generalfrom adjunct ambiguities to argument ambigties, as many theories make clear distinctiobetween the processing of adjuncts and arments (Abney, 1989; Frazier & Clifton, 199Liversedge, Pickering, Branigan, & Van Gompel, 1998; Pritchett, 1992; Schütze & Gibso1999). In particular, construal theory (Fraz& Clifton, 1996) makes a distinction betweeprimary and nonprimary phrases, which cor
assponds roughly to the distinction between argu-

oraise

ac

a

le

ntrn

uea

s

rirnllv

-e

f

9

u a

lr5tyn

tionna-e-tlyty”

syn-tol.,y,et

so-nge.,ly

e&.&4)rdsretroldse-ingsso-al.,enodbea-ststyu-heusebe-toll.edr-onaxeirwill


ments and adjuncts. Frazier and Cliftclaimed that ambiguities involving nonprimarelations such as relative clause attachmentbiguities are resolved differently from ambiguties involving primary relations. Ambiguitiethat involve nonprimary phrases are resolvby the principle of construal: Nonprimary reltions are associated with all possible attament sites within the current thematic doma(defined in terms of the most recent thetasigner), and the final choice of attachmentmade between these different sites in an unstricted manner. Therefore, it is a variabchoice model with respect to ambiguities thinvolve nonprimary phrases. However, costrual is a fixed-choice model with respectambiguities that involve at least one primaphrase. If a sentence is ambiguous betweeprimary and non-primary relation, the procesor always adopts the primary relation. Fthermore, if a sentence is ambiguous betwtwo structures that both involve primary reltions, construal predicts that the processorways adopts the structurally simplest analy(determined by minimal attachment and laclosure). Traxler et al.’s (1998) data are constent with a variable-choice model of nonpmary relations, but a fixed-choice model of pmary relations. In order to see if the predictioof the unrestricted race model hold generawe need to investigate ambiguities that invoprimary relations.

Competition models. In most constraintbased lexicalist models, syntactic ambiguity rolution involves a mechanism of competition.Competition is one source of processing diculty in such models (MacDonald et al., 199McRae et al., 1998; Spivey & Tanenhaus, 19Spivey-Knowlton & Sedivy, 1995; Tabor, Juliano, & Tanenhaus, 1997; Tabor & Tanenha1999; Tanenhaus & Trueswell, 1995). Whenternative analyses of an ambiguity have neequal activation, competition is particularstrong and severe processing difficulty is pdicted. Spivey-Knowlton and Sedivy (199proposed “Rather than interpreting elevareading times as evidence that an initial analhas been identified as incorrect, and that a
structure is required, our current approach su
RING, AND TRAXLER

nym--

d-h-ins-isre--

at-oy

as-r-en-al-isteis--i-s

ly,e

s-

fi-4;8;-s,l-

ar-ye-)

edsisew

gests that processing delays are a manifestaof direct competition between opposing altertives [. . .] Near equal activation levels will rsult in lengthy competition, hence greaslowed reading times at the point of ambigui(p. 260).

Constraint-based accounts assume thattactic ambiguity resolution is closely relatedlexical ambiguity resolution (MacDonald et a1994; see also Spivey-Knowlton & Sediv1995; Trueswell et al., 1993, 1994; Garnseyal., 1997). Research on lexical ambiguity relution shows that in neutral contexts, readitime for balanced ambiguous words (i.words with two meanings of approximateequal frequency) is longer than reading timfor unambiguous control words (RaynerDuffy, 1986; Duffy, Morris, & Rayner, 1988)By contrast, Duffy et al. (see also BinderRayner, 1998; Rayner, Pacht, & Duffy, 199showed that when balanced ambiguous woare disambiguated by prior context, they aprocessed as quickly as unambiguous conwords. Biased ambiguous words (i.e., worwith two meanings of greatly unequal frquency) are processed slowly when precedcontext disambiguates the word toward its lefrequently occurring meaning. This is the scalled subordinate-bias effect (Rayner et1994). Thus, lexical processing is slow whneither context nor frequency provides goevidence about which interpretation shouldfavored or when different sources of informtion support different meanings. This suggethat competition occurs in lexical ambiguiresolution when two meanings of an ambigous word are about equally activated by tconstraints (frequency and context). Becasyntactic ambiguity resolution is assumed tosimilar to lexical ambiguity resolution, constraint-based theories predict competitionoccur in syntactic ambiguity resolution as weFor instance, MacDonald et al. (1994) claimthat “[. . .] X-bar structures will each be patially activated to some degree, dependingthe strength of their support from the syntand from other parts of the system (e.g., thassociated argument structures), and they
g-compete with each other just as other represen-

&

c

s

teisfob

ecpt c

a n

d

s

uti

dea

e t

isiscedd,

al.in

r tosi-e

t

ort

)ple99,Inessrd,lylesheon-sh-heeor

ere-nts-hem-ly-hel

ndnd

REANALYSIS IN SEN

tations at the same level do within a single leical entry” (p. 687). In contrast, two-stagmodels draw a clear distinction between sytactic and lexical ambiguity resolution (Frazie1987, 1995; Frazier & Clifton, 1996; RaynerMorris, 1991; Traxler et al., 1998). Proponenof such a distinction usually assume that womeanings are accessed in parallel, whichgive rise to competition effects, whereas sytactic structures are constructed in a serial faion, so that no competition can occur.

Competition models predict that in syntacambiguity resolution, there are two cases whcompetition arises. In the first case, constrafavor analysis A as soon as the ambiguity ariBut constraints that come in at one of the lowing words favor analysis B instead (e.g.,cause they provide disambiguating informtion). Thus, this case involves a changepreferences as a result of constraints that emat different words. This leads to near-equal avation of the two analyses and results in comtition, which causes processing difficulty. Nothat a two-stage model in which analysis Ainitially adopted predicts reanalysis in sucases.

In the second case, constraints support anses A and B approximately equally as soonthe ambiguity arises, and later words do favor either analysis. This also causes comption. In this case, difficulty caused by competion has no equivalent within two-stage modeThey predict no reanalysis and therefore no ficulty in such cases. Hence, investigationthis second case should allow us to distingubetween reanalysis- and competition-bamodels.

An important issue with respect to ambigties that belong to the second case is the course of competition. When multiple analysremain equally supported during several worfor how long will the analyses remain activatin parallel before one analysis is selected competition is resolved? Some researchers hargued that competition continues for a vlong time. A clear example of such a viewfound in MacDonald, Just, and Carpen(1992), who discussed sentences such as
and (2b):

x-en-r,

tsrdann-h-

icre

ntses.l-

e-a-inrgeti-e-eish

ly-asoteti-ti-ls.if-ofished

i-meess,dndaveryiser(2a)

(2a) The experienced soldiers warnedabout the dangers before the midnight raid.

(2b) The experienced soldiers spokeabout the dangers before the midnight raid.

In both sentences, the main clause analysisthe ultimately correct analysis. However, (2a) ambiguous between a main clause and redurelative analysis until the period is reachewhereas (2b) is unambiguous because spokecannot be a past participle. MacDonald et claimed that if both analyses are activated parallel in (2a), this sentence should be harderead than (2b), where only one analysis is posble. Importantly, they claimed that it should bpossible to find such effects in the region beforethe midnight, well downstream from the poinwhere the ambiguity arises (at warned). Hence,they predicted very long-lasting competition.

Some recent computational models also favlong-lasting competition, but they claim thacompetition gradually decreases. The competi-tion-integrationmodel of McRae et al. (1998and Spivey and Tanenhaus (1998) is an examof such a model (see Tabor & Tanenhaus, 19for a model that makes similar predictions). this computational model, the different analysof an ambiguity receive support from variousentence processing constraints. At each wothe analysis that is supported most stronggradually becomes more activated due to cycof forward and backward activation between tanalyses and the constraints. These cycles ctinue until one analysis reaches a certain threold level (determined by a free parameter). Tlonger it takes to reach this threshold, thstronger the competition. Next, the processmoves to the following word. At this word, thfinal activation levels of the analyses at the pvious word are treated as one of the constraifor the current word (the competition-integration model stipulates that this constraint has tsame weight as all the new constraints cobined). If the new constraints support two anases about equally, the activation levels of ttwo analyses will initially become more equaagain. But subsequently, cycles of forward abackward activation between the analyses a
constraints will gradually increase the differ-

R

vs

h

tese

m

ouep t Re7

i2

nemnoioawlen

n

neo

hen-vi-

ity

b-

.

yursa),

ed,

to

use-n-

b-tyine tore- in8)sis:thena-un-be-

cto-m-ti-

ted


ence in activation levels until the threshold leis reached again and the processor movethe next word. In the competition-integratiomodel, it is possible to set the value of tthreshold level such that the difference betwethe final activation levels at a word is greathan at the preceding word. As a result, comption will gradually decrease as the procesmoves from one word to the next. Howevcompetition will never completely disappeaEven the least activated analysis remains (sowhat) activated.

In fact, most constraint-based competitimodels are unclear about the precise time coof competition. However, to our knowledgnone of the these theories claim that the nonferred analysis is abandoned completely. Intheories, multiple analyses remain activatedparallel (though perhaps only weakly) at leaslong as the sentence is ambiguous (e.g., Mcet al., 1998; Spivey & Tanenhaus, 1998; SpivKnowlton & Sedivy, 1995; Tabor et al., 199Tabor & Tanenhaus, 1999). Hence, they assuthat competition is long-lasting.

Evidence for Competition

In the past, a number of studies have claimto find evidence for competition in syntactambiguity resolution. MacDonald et al. (199investigated sentences such as (2a) and (2b)observed that reading times for before the mid-night were longer in (2a) than in (2b). MacDoald et al. took this as evidence that readretained both analyses in memory in the teporarily ambiguous condition, whereas only oanalysis was available in the unambiguous cdition. This can be interpreted as competitbetween the reduced relative and the mclause analysis in the ambiguous condition. Tfurther experiments replicated these resuThese findings clearly pose difficulty for ranalysis models, which claim that only oanalysis is pursued. However, this study wsubsequently criticized by Pearlmutter aMacDonald (1995). They conducted a plausibity study on MacDonald et al.’s materials ashowed that the ambiguous conditions wrated as less plausible than the unambigu
conditions. They argued that this probably e
ING, AND TRAXLER

el to neenerti-orr,r:

e-

nrse,re-allinasaey-;me

edc)and

-rs -

en-nino

ts.-easdil-dreus

plained the relatively long reading times in tambiguous condition. Therefore, the MacDoald et al. study does not provide convincing edence for competition in syntactic ambiguresolution.

In another study, MacDonald (1994) attriuted longer reading times in the region capturedin the coupin (3a) than in fought in the coupin(3b) to competition between constraints.

(3a) The ruthless dictator captured inthe coup was hated throughout the country

(3b) The ruthless dictator fought in thecoup was hated throughout the country.

She suggested that the main clauseanalysis (asin, e.g.,The ruthless dictator captured the cit)in (3a) and (3b) is activated because it occmost frequently or is least complex. In (3captured in the coupis barely consistent withthe main clause analysis, so it is deactivatand the (ultimately correct) reduced relativeanalysis gets more activation. This leadscompetition. But in (3b),fought in the coupiscompatible with the main clause analysis. Ththere is very little competition here. In this rgion, (3a) was harder than (3b) in relation to uambiguous controls.

In another study, McRae et al. (1998) oserved processing difficulty when plausibiliinformation was inconsistent with the maclause analysis and claimed that this was ducompetition between the main clause and duced relative analysis. However, the effectsMacDonald (1994) and McRae et al. (199could equally well have been due to reanalyThe processor may initially have adopted main clause analysis, but then started to realyze when evidence rendered this analysis likely. Hence, these data do not distinguish tween reanalysis and competition.

The Processing of VP–NP AttachmentAmbiguities

Because previous studies have not satisfarily discriminated between reanalysis and copetition, we conducted a new study to invesgate this issue. In this study, we investiga
x-VP–NP attachment ambiguitieslike (4) and (5):

.

t.

.

-

t

otm

b

ulNsn

uleu

n

n

en-

dAtors

is.ong. In

eti-ent

eus bi-

is. in-r, ifigh

uale itishier

m-he arepre-

nghew-ota)ingryli-ue

theands atButoc- this-ne

REANALYSIS IN SEN

(4a) The hunter killed the dangerouspoacher with the rifle not long after sunset(ambiguous)

(4b) The hunter killed the dangerousleopard with the rifle not long after sunset.(VP attachment)

(4c) The hunter killed the dangerousleopard with the scars not long after sunse(NP attachment)

(5a) The hunter killed only thepoacher with the rifle not long after sunset(ambiguous)

(5b) The hunter killed only the leopardwith the rifle not long after sunset. (VP at-tachment)

(5c) The hunter killed only the leopardwith the scars not long after sunset. (NP attachment)

The prepositional phrasewith the rifle/scarsattaches either tokilled (the VP attachmenanalysis), whereby the hunter killed with thrifle or scars, or toleopard/poacher(theNP at-tachment analysis), whereby the leopardpoacher had the rifle or scars. Experimenemployed sentences like (4); Experiment 2 eployed sentences like (5).

These sentences are produced by makingmanipulations. The first concerns the plausiity of the two analyses. In the ambiguous contions [(4a) and (5a)], both analyses are plable; in the VP conditions [(4b) and (5b)], onthe VP attachment is plausible; and in the conditions [(4c) and (5c)], only the NP analyis plausible. We assessed the plausibility maulation in plausibility pretests. The second mnipulation is the initial preference for VP versNP attachment (independent of the plausibimanipulation). We assessed this preferenctwo ways: attachment preferences when plability was carefully controlled [as in (4a) an(5a)], and completion preferences for fragmeending just before the rifle or the scars. Thesepretests demonstrated that the initial preferein (4) was for VP attachment, whereas (5) hno clear initial preference for VP or NP attac
ment. We now outline the predictions of diffe

e

r1-

twoil-di-si-yP

isip-a-s

ity insi-dts

ceadh-

ent theories for the processing of such stences.

Constraint-based theories. Constraint-basetheories predict strong competition in (4c). with the, the VP attachment preference favthe VP analysis. However, at scars, the plausi-bility disambiguation favors the NP analysThis reversal of preferences leads to strcompetition between the VP and NP analysisthe VP attachment condition (4b), no comption is predicted, because the VP attachmpreference at with the is consistent with thplausibility disambiguation. In the ambiguocondition (4a), the VP attachment preferenceases toward the VP analysis at with the, butplausibility does not favor either analysHence, the degree of competition should betermediate between (4b) and (4c). Howevethe activation of the VP analysis is very hwhen the processor reaches rifle, this may out-weigh any competition effects due to eqplausibility of the two analyses, and thereformight, in practice, be impossible to distingu(4a) from (4b) [but (4a) should not be easthan (4b)].

Competition-based theories also predict copetition in (5a), the ambiguous condition for tbalanced ambiguity. Because the materialsbalanced, neither the VP nor NP analysis is ferred when the ambiguity arises (at with). Thisresults in competition at this point. Assumithat competition is relatively long-lasting, tanalyses will continue competing at the folloing words (the rifle) because these words do nprovide any biasing information. Thus, (5should be hard to read at the words followwith. However, if competition decreases verapidly, processing difficulty may become neggible after a few words. We return to this issin Experiment 2.

Competition should be much weaker in VP and NP disambiguated conditions (5b) (5c). Competition in these sentences occurwith thebecause the ambiguity is balanced. of course the same amount of competition curs in (5a), as the sentences do not differ atpoint. At disambiguation, plausibility information provides a strong constraint in favor of o

r-of the analyses (while other constraints do not

tu mnl

lion etr

ie

e

h

th

datou

npro

6in-

retheerorevi-inge-ch

r-

lse

n-ch-anden-eta-in-ionn,n-alin-icc-&

,l-se-isform-un-uc-orc-

fectnced-

re-c-

z,the


favor either analysis). Thus, processing of disambiguated conditions (5b) and (5c) shobe easier than the ambiguous condition (5a).difference is predicted between the disabiguated conditions, because our pretests icated that the plausibility bias was equastrong for the two conditions and there wasinitial preference for one of the analyses.

Fixed-choice two-stage theories. Most two-stage theories predict that the processor alwinitially adopts the VP analysis in sentences (4). According to garden-path theory, this flows from the principle of minimal attachmebecause it is assumed that the VP analysisploys fewer nodes in the phrase structure than the NP analysis (Frazier, 1979; Rayneal., 1983).

Construal theory (Frazier & Clifton, 1996claims that both VP and NP attachment are tially interpreted as primary relations (evthough they may not be primary relations their final interpretation), and therefore, the costrual principle does not apply in such stences. Hence, the processor bases its inanalysis on structural principles only. As in tgarden-path model, the processor should alwinitially adopt the VP analysis as a result of principle of minimal attachment.

Finally, referential theory (Altmann & Steeman, 1988; Crain & Steedman, 1985; Ni et 1996) claims that the processor adopts analysis with the fewest unsatisfied presupptions. In (4), the VP analysis involves a nophrase (e.g.,the dangerous poacher) which re-quires the processor to postulate the existenca single dangerous poacher; whereas theanalysis involves a complex noun phrase (ethe dangerous poacher with the scars), whichrequires the processor to postulate the existeof a set of dangerous poachers. As postulatisingle dangerous poacher requires fewer suppositions than postulating a set of dangepoachers, the VP analysis is adopted.

Most two-stage accounts make the same pdictions for (5) as (4). The exception is the reerential theory as interpreted by Ni et al. (199Ni et al. argued that in referential theory, theclusion of the wordonly makes the NP attach
ment preferred. Ifonly precedes a NP such a
RING, AND TRAXLER

heldNo

-di-lyno

ayskel-tm-

ree et

)ni-ninn-n-itiale

ayse

-l.,hesi-n

e ofNP.g.,

nceg are-us

re-f-).-

poacher in (5), the processor assumes mothan one poacher. As a result, it analyzesfollowing prepositional phrase as a modifiwhich specifies the right poacher and therefadopts the NP analysis. They found some edence from eye-tracking and self-paced readin favor of this account. However, other rsearchers have not found evidence for sustrong effects ofonly on syntactic ambiguityresolution (Clifton, Bock, & Radó, 2000; Pateson, Liversedge, & Underwood, 1999).

Variable-choice two-stage theories. Variable-choice models differ from fixed-choice modewith respect to the proportion of times that thVP or NP analysis is initially chosen. The tuing hypothesis assumes that the initial attament preference for sentences such as (4)(5) is determined by how often readers havecountered that particular structure. Mitchellal. (1995) noted that such frequency informtion can be represented at different levels of lguistic representation. They made a distinctbetween fine-grained frequency informatiowhen information about the frequency of sytactic structures is stored for each individulexical item, and coarse-grained frequencyformation, when it is stored for larger syntactstructures (e.g., for all VP–NP attachment strutures). In a number of articles (e.g., BrysbaertMitchell, 1996; Cuetos, Mitchell, & Corley1996; Mitchell et al., 1995), Mitchell and coleagues have tentatively favored the coargrained variant of the tuning hypothesis. Thvariant predicts that a reader’s preferenceVP or NP attachment is determined by the nuber of times a reader has previously encotered VP and NP attachment in V–NP–PP strtures. The particular verb, noun phrase,preposition that is used is not taken into acount. In particular, asonly constitutes part ofthe noun phrase, its presence should not afpreferences. Therefore, the same prefereshould occur in (4) and (5). Some corpus stuies have found that NP attachment is more fquent than VP attachment in V–NP–PP strutures (Collins & Brooks, 1995; Hindle &Rooth, 1993; Merlo, Crocker, & Berthouzo1997). If these results are representative of

sV–NP–PP structures that our participants en-

hi

nc

ea

cy d

th

hein

e os

na

theric snseth tibsmo k

nimei

,

ri-ren-hea)]the

ictortn-

greeen-tiala-fn-

usm-

lty, as

NPm-onuld

be easier to process than either of the disam-

tsikealic-n-n-lto

esebe

c)]

rtr-e-of

or-

REANALYSIS IN SEN

countered, the tuning hypothesis predicts tmost readers initially adopt the NP analysHowever, it is possible that some participamay have encountered more VP than NP attaments. Consequently, reanalysis may occursome trials in both the VP and NP attachmcondition, and both may be more difficult ththe ambiguous condition.

In the unrestricted race model, preferenmay differ for each sentence that is read bparticular person because preferences arefected both by item characteristics and by invidual factors. Thus, readers may adopt eithe VP or NP analysis initially. In (4), there isstrong initial bias for the VP analysis. Tprocessor takes this bias into account and thfore nearly always adopts the VP analysis tially. Thus, this theory predicts that readers wexperience difficulty when this analysis bcomes implausible, as in (4c), but that littleno difficulty should occur when the NP analybecomes implausible, as in (4b). No processdifficulty should occur in the ambiguous setence (4a), because the initial analysis is alwplausible.

In contrast, because there is no bias for eiVP or NP attachment in (5), the model predthat the processor will adopt each analysisabout half of the trials initially. Hence it makenovel prediction for (5): In the VP conditio(5b) and the NP condition (5c), the proceswill be forced to reanalyze on those trials whit adopted the implausible analysis; but in ambiguous condition (5a), it will never havereanalyze, because either analysis is plausThe ambiguous condition (5a) should be eathan (5b) and (5c) (which should not differ froeach other). Hence, the unrestricted race mmakes essentially the opposite predictionsconstraint-based theories for (5). It also madifferent predictions from fixed-choice modefor sentences like (5).

In all two-stage theories, the processor oreanalyzes if the initial analysis becomes plausible or unlikely. It should not noticwhether the nonselected analysis is implausso long as the selected analysis remains plable. Thus, if it initially adopts the VP analysis
should experience difficulty with (4c). But (4a

ats.tsh-

onntn

es aaf-i-eraere-i-

ill-r

ising-ys

and (4b) should not differ in difficulty. If it ini-tially adopts the NP analysis, it should expeence difficulty with (4b), but not with (4a) o(4c). The same argument applies for the stences in (5). The crucial prediction is that tglobally ambiguous sentences [(4a) and (5should not be more difficult to process than disambiguated sentences.

Summary. Constraint-based theories predcompetition either when two constraints suppdifferent analyses [as in (4c)] or when both costraints support two analyses to an equal de[as in (5a)]. Fixed-choice models (e.g., gardpath theory, construal theory, and referentheory) predict difficulty when the disambigution is inconsistent with an initial choice oanalysis: Either the NP condition or the VP codition should cause difficulty. The ambiguocondition should be as easy as the disabiguated condition that produces least difficubut no easier. A variable-choice account suchthe unrestricted race model predicts that thecondition (4c) should be harder than the abiguous condition (4a) and the VP conditi(4b); and the ambiguous condition (5a) sho

tson a

orree

ole.

ier

deltoesls

ly-

bleusi-it

biguated conditions [(5b)–(5c)].

EXPERIMENT 1

We monitored participants’ eye movemenas they read VP–NP attachment ambiguities l[(4a)–(4c)]. These ambiguities had an initibias toward VP attachment. Note that the predtions of most theories do not differ for these setences. Although the experiment rigorously cotrolled for plausibility and introduced a noveambiguous condition, its main purpose wasdetermine patterns of eye movements for thsentences so that precise predictions couldmade for the more critical sentences [(5a)–(5that were employed in Experiment 2.

Participants

Thirty-six participants were paid to take pain the eye-tracking experiment. All of the paticipants in all parts of the study were nativEnglish-speaking students at the UniversityGlasgow and had normal or corrected-to-n
)mal vision. Some had taken part in other eye-

R

idx

te

.

t.

lamue

A,aa

be

ierhhdda

ne

en

u

ex

an

r

ts.n-hsi-er,ent

s

-n-let

-

ng

f--

soe

-hsetf)


tracking experiments. No one participatedmore than one experiment or pretest reportethis article. Two further participants were ecluded, because they answered more than 2of the statements incorrectly.

Items

We constructed 30 items like (4), repeabelow (see Appendix):

(4a) The hunter killed the dangerouspoacher with the rifle not long after sunset(ambiguous)

(4b) The hunter killed the dangerousleopard with the rifle not long after sunset.(VP attachment)

(4c) The hunter killed the dangerousleopard with the scars not long after sunse(NP attachment)

The VP attachment condition was semanticadisambiguated to the VP analysis; the NPtachment condition was semantically disabiguated to the NP analysis; the Ambiguocondition was semantically ambiguous betwethe two analyses (see plausibility pretest).items consisted of a subject noun phraseverb, an object noun phrase (containing anjective as well as a noun), a prepositionphrase (consisting of the prepositionwith and anoun phrase), and a sentence-final adverphrase. The direct object and the noun phrasthe prepositional phrase contained the defindeterminerthe. On the VP analysis, the prepostional phrase was interpreted as an instrumof the verb; on the NP analysis, it was intepreted as an attribute of the object noun. Tobject noun phrase and the noun within tprepositional phrase differed between contions. The object in the NP attachment contion was the same as the object in the VPtachment condition for half of the items, anthe same as the object in the ambiguous cotion for the other half. The noun within thprepositional phrase was the same in the abiguous and VP attachment conditions, wherthe NP attachment condition had a differenoun. The total length of the adjective pl
noun was identical for each version of a give
ING, AND TRAXLER

nin-5%

d

lyt--snlla

d-l

ial inite-nt-eei-i-t-

ddi-

m-ast

s

item, as was the length of the noun within thprepositional phrase. According to the CeleEnglish database (Baayen, Piepenbrock, & VRijn, 1993), word-form frequency for the nounwithin the prepositional phrase did not diffebetween conditions (F , 1), nor did the loga-rithm of the word-form frequency (F , 1) (seeTable 1).

Pretests

The items were subjected to three pretesThese tests confirmed that the Ambiguous codition was (semantically) plausible on botanalyses, that the other conditions were plauble on one analysis but implausible on the othand that the materials were biased (independof plausibility) toward VP attachment.

Plausibility pretest. We carefully controlledour sentences for plausibility. Previous studiehave often manipulated plausibility by intuitionalone, with the result that VP attachment conditions have often included ambiguous setences where the NP analysis is also plausib(e.g.,The spy saw the cop with binoculars buthe cop didn’t see him, from Rayner et al.,1983). Additionally, even unambiguous sentences can differ in plausibility from eachother, and such differences can impact readitimes (e.g., Traxler & Pickering, 1996). Wetherefore ruled out the possibility that any efect was due to an irrelevant difference in plausibility. We also constructed the sentencesthat plausibility was manipulated on a singlword [rifle or scarsin (4)].

The plausibility pretest investigated the plausibility of the different analyses. It began wit50 sets of items and reduced them to the final of 30. We derived six sentences like (6a)–(6from each item set.

(6a) The hunter killed the dangerouspoacher by using the rifle.

(6b) The hunter killed the dangerousleopard by using the rifle.

(6c) The hunter killed the dangerousleopard by using the scars.

(6d) The dangerous poacher with the
n rifle was killed by the hunter.

r

s

b

n-se

leh2se)t

-re-c-

al,

1edse-

egu-

e-s inre-ini-

Note. Frequency is the number of occurrences of the word form in the Celex database (17.9 million words).

(6e) The dangerous leopard with therifle was killed by the hunter.

(6f) The dangerous leopard with thescars was killed by the hunter.

These sentences were unambiguous paraphof item sets like (4a)–(4c), on the VP or the Nanalysis. For the VP analysis (6a)–(6c), we placed with with by using, by means of (twoitem sets) and with the help of(two item sets).We refer to these conditions as the VP–PP para-phrase conditions. For the NP analysi(6d)–(6f), we converted the original senteninto a passive. We refer to these conditionsthe NP–PP paraphrase conditions. Notice thatthe VP–PP paraphrase of the NP attachmcondition (6c) and the NP–PP paraphrase ofVP attachment condition (6e) derived froanalyses that were constructed to have implable interpretations. All other conditions were drived from analyses that were constructed toplausible. Twenty-two participants rated horealistic they thought that the situations dscribed by (6a)–(6f) were, on a 7-point scawith 7 indicating a very realistic situationand 1indicating a very unrealistic situation. Each par-ticipant rated all 300 sentences, so that all cditions of an item were rated by the same paripants. This enabled us to make a reliacomparisons between the conditions of eachdividual item. At least 20 sentences appea
between two conditions of a single item. Ha
asesP

re-

ce as

entthemusi-e- bewe-le,

on-tic-le

in-red

the participants received the items in one radom order, and half received the exact reverorder of items.

We then selected the final 30 item sets. Tab1 shows the mean plausibility rating for eaccondition for each analysis. We conducted(Paraphrase type: VP–PP vs NP–PP paraphra3 3 (Condition: ambiguous vs VP attachmenvs NP attachment) ANOVAs, treating participants and items as random factors. Thesevealed a significant interaction between the fators: F1(2, 42) 5 436, p , .001, MSe 5 .45;F2(2, 58) 5 1239,p , .001, MSe 5 .22. Forthe purposes of the experiment, it was crucithat conditions (6c) and (6e) were implausiblebut the other conditions were not; Tableshows that this was indeed the case. Planncomparisons showed that the two paraphratypes differed for both the VP attachment condition and the NP attachment condition (Allps, .001). Although the difference between thVP–PP and NP–PP paraphrases in the ambious condition was significant (F1 5 9.71,p 5.003,F2 5 27.6,p , .001), it was numericallyvery small.

We also analyzed the absolute differences btween the scores for the two paraphrase typethe three conditions. Planned comparisons vealed that this difference was much smaller the ambiguous condition than in the VP condtion and the NP condition (all ps , .001). Hence

REANALYSIS IN SENTENCE PROCESSING 235

TABLE 1

Experiment 1: Item Characteristics

Condition

Ambiguous VP attachment NP attachment

Plausibility pretestPlausibility VP-PP paraphrase 6.1 6.1 1.5Plausibility NP-PP paraphrase 5.5 2.1 5.9

Frequency of noun in PP 406 406 479Log frequency of noun in PP 1.88 1.88 2.00

Off-line preference taskPercentage of VP attachment 78.3 95.6 2.2

Completion taskPercentage of VP completions 93.9 94.9 90.1

lfthe plausibility difference was much less in the

t

-dica

ct

htn,utfu

&-r

t

ih

-–ndi-n

nce

m

foresurnt.n-acef-

-he-et

l.,;nof

er

sx-wedts

alldlet

o

fm-

i-a-tenisnd

ntsn-),

but, importantly, both tasks indicated a strongVP attachment preference.

1 Note that due to the way we constructed our materials(see materials section) the NP condition in the completionwas the same as the ambiguous condition for half of the


ambiguous conditions than in the disambiguaconditions.

Off-line preference task. This pretest investigated the bias of the items as they were usethe eye-tracking experiment. Eighteen partpants received exactly the same sentences the eye-tracking experiment, with six particpants seeing each list. Each item was followby the two paraphrases of the item used in plausibility pretest, balanced for order. Partipants indicated whether the VP–PP or NP–PP paraphrase best described the situadescribed by the sentence.

The crucial condition to test the bias of titems is the ambiguous condition. Given that plausibility pretest found that both attachmewere about equally plausible in this conditionpreference for either the VP or NP analysis mbe the result of a bias in the items before critical noun [i.e.,rifle in (4a)], independent oplausibility. Testing a semantically ambiguocondition is a common method to test for sytactic preferences (e.g., Cuetos & Mitche1988; De Vincenzi & Job, 1993; Frazier Clifton, 1996; Gilboy, Sopena, Clifton, & Frazier, 1995; Hemforth, Konieczny, & Scheepe2000).

We predicted that the VP and NP conditiowould normally be given VP and NP interpretions respectively. In other words, participanshould avoid implausible interpretations. Thwould indicate that our plausibility manipulations were strong enough to eventually overrany syntactic bias. Table 1 indicates that twas indeed the case (the two paraphrases alwadded up to 100%). Analyses of variance, treing Condition (ambiguous vs VP vs NP attacment) as a within-participants and within-itemfactor, revealed a significant effect of ConditioF1(2, 34) 5 436,p , .001,MSe 5 102; F2(2,58) 5 314,p , .001,MSe 5 236. Planned comparisons revealed that the percentage of VPattachment responses in the ambiguous cotion was different from both the VP conditioand the NP condition (all ps , .001), which in-dicated that the plausibility manipulatioworked. The 78.3% VP attachment preferenin the ambiguous condition differed fro
chance (both ps , .001), indicating that VP at-
RING, AND TRAXLER

ed

ini-s ini-edthei-hetion

ehetsast

he

sn-ll;

s,

nsa-tsis-deisaysat-h-s

n:

PP

tachment was preferred to NP attachment these items. As plausibility of the two analyswas closely matched, we conclude that oitems displayed a bias toward VP attachmeThis result is crucial for theories such as costraint-based theories and the unrestricted rmodel, which assume that on- and off-line preerences are related.

Completion task. An alternative means of assessing bias employed by constraint-based torists is sentence completion (e.g., Garnseyal. 1997; McRae et al., 1998; Trueswell et a1993; cf. Clifton, Kennison, & Albrecht, 1997Pickering, Traxler, & Crocker, 2000). Eighteeparticipants completed booklets consistingthe 30 experimental items up to the wordwith(e.g.,The hunter killed the dangerous poachwith . . .), together with the 81 fillers used inthe eyetracking experiment, cut off at varioupoints so that they appeared similar to the eperimental items. Again, six participants saeach list, which corresponded to the lists usin the eyetracking experiment. The participanwere instructed to write the first grammaticand meaningful continuation that they couthink of. They were then given a second bookcontaining the experimental fragments and twparaphrase fragments (e.g.,The hunter had. . .andThe dangerous poacher had. . .), balancedfor order, and were told to indicate which othese fragments best corresponded to their copletion.1

We excluded 2.6% of trials, on which particpants said that they did not know which parphrase was correct. Table 1 indicates how ofparticipants employed a VP completion. Thpercentage was very high in all conditions adiffered from chance (both ps , .001). The con-ditions did not differ from each other (both ps ..1). The mean percentage of VP attachme(93.0%) was higher than in the ambiguous codition of the off-line preference task (78.3%

items and the same as the VP condition for the other half.

T

n1el

e4iceaett

clf atorh

ejeer)r

euzrsis

etheae

dlleae thta

an

er-rx-hendor

n.reti-d notx-lsore

e-

jectun

he

en.

hetoaluldcell

,&

gnr ton-

REANALYSIS IN SEN

Design

For the eye-tracking experiment, we costructed three lists of items, consisting of items from each condition, with exactly one vsion of each item appearing in each list. Tweparticipants were randomly assigned to reach list. Each list also contained 81 fillers:sentences employing a range of grammatstructures and 36 items from another expment. Thirty-four items were followed by statement in order to encourage comprehsion—10 statements followed an experimenitem. Participants indicated whether the stament was a correct statement about the preing sentence and received no feedback. Hathe statements required a “correct” responsehalf an “incorrect” response. The experimenitems and fillers were placed in a single randorder, with five fillers preceding the first expemental sentence and two fillers following ttwo breaks in the experiment.

Procedure

Participants’ eye movements were recordwith a Fourward Technologies Dual PurkinGeneration 5.5 eye tracker, which monitorthe right eye (though viewing was binoculaThe tracker had an angular resolution of 10´ aA computer displayed the materials on a scre77 cm from the participants’ eyes. The scredisplayed 3.8 characters per degree of visangle. The tracker monitored participants’ galocation every millisecond and the softwasampled the tracker’s output to establish thequence of eye fixations and their start and fintimes.

Each participant was run individually. Thexperimenter told the participant to read sentences carefully in order to understand thbut to read at his or her normal rate. Bite band head restraints were used to minimize hmovements. Next, the participant completecalibration procedure. Before each item or fia calibration check was performed, and the cbration was repeated if necessary. Each sentwas presented on a single line. After readingsentence, the participant pressed a button wled to the presentation of a comprehension s
ment or the next trial. The experiment, whic
ENCE PROCESSING 237

-0r-vead5alri-

n-ale-ed- ofnd

almi-e

d

d.c.ennale

ee-h

em,rsad

arli-nceheichte-

contained two short breaks, took about half hour.

Analyses

On average, the participants made 11% rors. Prior to all analyses, trials with majotracker losses and trials with fewer than four fiations were excluded. This eliminated 3% of ttrials. If a fixation was shorter than 80 ms awithin one character space of the previous next fixation, it was assimilated to this fixatioAll remaining fixations shorter than 80 ms weexcluded. In each of the conditions, this constuted 2% of all fixations. Following Rayner anPollatsek (1989), we assume that readers doextract much information during such short fiations. Fixations longer than 1500 ms were aexcluded. These extremely long fixations apresumably due to a major tracker loss.

We divided experimental items into seven rgions, indicated by slashes in (7):

(7) The hunter/ killed/ the dangerouspoacher/ with the/ rifle/ not long after/ sun-set.

These regions corresponded to (1) the subnoun phrase, (2) the verb, (3) the object nophrase, (4) the words with the, (5) the nounwithin the prepositional phrase (known as tcritical noun region), (6) the next words(known as the postcritical region), and (7) thefinal word of the sentence. The space betwewords was included with the following wordNotice that the postcritical region excluded tfinal word of the sentence. This allowed us interpret effects in this region as due to initiprocessing and hence conclude that they wonot be contaminated by effects of sentenwrap-up (e.g., Just & Carpenter, 1980; Mitche& Green, 1978; Rayner, Kambe, & Duffy2000; Rayner, Sereno, Morris, Schmauder,Clifton, 1989).

We report four eye-tracking measures. First-pass timeis the sum of all fixation times startinwith the reader’s first fixation inside a regiountil the reader’s gaze leaves the region eithethe left or right. For regions consisting of a sigle word, this corresponds to gaze duration
h(Rayner & Duffy, 1986). First-pass regressions

sehhe

o

elie

t ht

cy

plt

ea

ie

n

e

r ettheioneas-on-

ze al.t-iong

-ctsthed re-cted theer-ith

andion.ion,v-entndted

le.ses.

re-

st- re-Ase-

ariti- 4).ionons


include all leftward eye movements that crothe region’s left boundary initiated after a firpass fixation in the region. Regression-path timis the sum of all fixation times starting with treader’s first fixation inside the region until treader’s gaze leaves the region to the right (Brysbaert & Mitchell, 1996; Duffy et al., 1988Konieczny, Hemforth, Scheepers, & Strub1997; Traxler, Bybee, & Pickering, 1997). Fall these measures, if the eye fixated after thegion before it fixated in it, the datum for this rgion was excluded. Total timeis the sum of alfixations in a region. If a reader skipped a regin a measure, resulting in a zero reading timwas excluded from the analyses for that meure. If a reader skipped two or more consecuregions, the data of these regions and thegions following them were excluded from tanalyses on that measure. We assumed thaticipants could not have processed the sentecompletely if this was the case. The latter produre removed less than 1% of the data for anthe analyses.2

Most sentence processing models do not dict in which eye-tracking measures difficushould occur. Previous research has shownfirst-pass times (e.g., Ferreira & Clifton, 198Frazier & Rayner, 1982; Rayner et al., 198Trueswell et al., 1993; 1994), first-pass regrsions (e.g., Altmann et al., 1992; Altmann, vNice, Garnham, & Henstra, 1998; Holmes O’Regan, 1981; Rayner et al., 1983; Picker& Traxler, 1998; Speer & Clifton, 1998), and rgression-path times (e.g., Brysbaert & Mitche1996; Traxler, Bybee, & Pickering, 1997; VaGompel & Pickering, in press) can all refleearly stages of processing difficulty during sytactic ambiguity resolution, and therefore, oearliest effects may occur in any of these meures. Other models such as the competition-igration model (McRae et al., 1998; SpiveyTanenhaus, 1998) have only modeled first-ptimes, but they could, in principle, be extend
to model other eye-tracking measures as wel els,
ach-di-

on

2 Because of an error, the final region of one item wmarginally grammatical. We excluded this complete itefor the total time analyses and the final region for all othanalyses.

RING, AND TRAXLER

sst-

ee.g.,;e,r re--

on, itas-ivere-epar-ncee- of

re-ythat6;3;s-n

&ng-ll,nctn-uras-te-&assd

l.

Using sentences comparable to (5), Traxleal. (1998) found the earliest clear effects in postcritical region, on the first-pass regressmeasure (rather than the first-pass time mure). For instance, participants experienced csiderable difficulty with sentences like The steakwith the sauce that was tough didn’t win a pri,in accord with pretest data and with Gilboy et(1995); but this difficulty first emerged on firspass regressions from the postcritical regafter the critical word tough(see also Pickerin& Traxler, 1998; Pickering et al., 2000, for similar effects). Therefore, we might predict effeto emerge in the postcritical region using first-pass regression measure and the relategression-path time measure. We also predidifferences between conditions to appear ontotal-time measure, for the noun region and phaps other regions around it, in accord wTraxler et al.

Results and Discussion

Table 2 presents the mean reading timespercentage regressions by condition and regFor each eye-tracking measure for each regwe conducted one-way ANOVAs with three leels of Condition (ambiguous vs VP attachmvs NP attachment), treating participants aitems as random factors. Condition was treaas a within-participant and within-item variabTable 3 presents the results of these analyPlanned comparisons between levels areported below.

There were no significant effects on the firpass time measure. The pattern of first-passgression data was more interesting. ANOVrevealed a significant effect of condition in Rgion 3 (the direct object). Most likely, this isspurious effect (note that it was before the ccal region). No effects occurred in Region(with the) or in Region 5 (the critical nounHowever, significant effects emerged in Reg6 (the postcritical region). Planned comparisrevealed that, as predicted by most modthere were more regressions from the NP attment condition than both the ambiguous contion [F1(1, 70) 5 20.1,p , .001; F2(1, 58) 534.6,p , .001] and the VP attachment conditi

asmer
[F1(1, 70) 5 20.9,p , .001; F2(1, 58) 5 33.6,

n

a

oti

n

eobd

thh

Ps

s

ale.e-s-f-sead

ercentageof saccades leaving the region to the left after a first-pass fixation.

p , .001]. The ambiguous and VP conditiodid not differ (Fs , 1). The condition effect inthe final region was also significant.3 Plannedcomparisons revealed that there were fewergressions from the ambiguous condition ththe NP condition [F1(1, 70) 5 6.91,p 5 .011;F2(1, 56) 5 5.76,p 5 . 020]). There were alsfewer regressions from the ambiguous condithan the VP condition [F1(1, 70) 5 5.31,p 5.024; F2(1, 56) 5 4.51,p 5 .038]. The VP andNP conditions did not differ in the final regio(both Fs , 1).

Regression-path times showed similarfects. In Region 3, an (unpredicted) conditieffect was significant by participants, but notitems (p 5 .081). The predicted effect occurrein Region 6. Planned comparisons revealedthe NP condition was read more slowly than tambiguous condition [F1(1, 70) 5 4.22, p 5.044; F2(1, 58)5 9.12,p 5 .004] and the VP
condition [F1(1, 70) 5 7.19, p 5 .009; F2(1,
4:3 The percentage of first-pass regressions from the firegion was high because readers often reread the sentafter they had reached the end.

s

re-n

on

f-ny

ate

58) 5 11.0,p 5 .002]. The ambiguous and VPconditions did not differ (Fs , 1). A similar ef-fect appeared in the final region, with the Ncondition being slower than the ambiguoucondition [F1(1, 70) 5 14.1, p , .001; F2(1,56) 5 6.54, p 5 .013] and the VP condition[F1(1, 70)5 6.21,p 5 .015;F2(1, 56)5 5.73,p 5 .020]. The ambiguous and VP conditiondid not differ (Fs , 1). Hence this measureshowed somewhat different effects in the finregion from the first-pass regressions measur

Total-time analyses showed an effect in Rgion 2 (the verb), presumably due to the regresions pattern in Region 3. Clear total time efects emerged in Regions 4–6. In all theregions, the ambiguous condition was refaster than the NP condition [Region 4:F1(1,70) 5 19.9,p , .001;F2(1, 56)5 11.22,p 5.002. Region 5:F1(1, 70) 5 23.4, p , .001;F2(1, 56) 5 23.9, p , .001. Region 6:F1(1,70) 5 12.2,p , .001; F2(1, 56)5 16.4,p ,.001]. Similarly, the VP condition producedshorter times than the NP condition [RegionF1(1, 70)5 12.8,p , .001;F2(1, 56)5 8.63,

nalence


TABLE 2

Experiment 1: Means

Region

1 2 3 4 5 6 7The hunter killed the dangerous with the rifle not long after sunset

poacher

First-pass reading timesAmbiguous 502 378 672 300 374 500 565VP attachment 491 357 683 305 376 509 553NP attachment 490 370 656 302 373 543 539

First-pass regressionsAmbiguous 5.0 14.0 6.0 8.8 11.3 42.7VP attachment 6.5 6.6 3.9 11.4 11.4 52.1NP attachment 7.8 10.0 3.6 12.2 25.8 51.2

Regression-path timesAmbiguous 502 417 818 354 420 684 1198VP attachment 491 401 748 330 444 663 1252NP attachment 490 408 750 328 431 841 1491

Total timesAmbiguous 622 532 947 379 451 699 698VP attachment 593 476 890 394 468 721 663NP attachment 619 529 922 459 567 829 685

Note. First pass, regression path, and total times are reported in milliseconds and first-pass regressions as the p

p 5 .005. Region 5:F1(1, 70) 5 17.2, p ,

:

n

cohfteas R

ey as

andVPisro-sis- un-VPre-on-

tnote 2).*p significant at the .05 level.

** p significant at the .01 level.

.001; F2(1, 56) 5 17.7, p , .001. Region 6F1(1, 70)5 8.48,p 5 .005;F2(1, 56)5 12.1,p 5 .001]. The ambiguous and VP conditiodid not differ in any of these regions (Fs , 1).

These results demonstrated that the NP dition became more difficult to read than eitthe ambiguous or the VP condition soon athe critical noun. The earliest effects of interoccurred in the postcritical region, on first-pregressions and regression-path times, andsame effect was observed in total times for
gion 4–6. These results are consistent with
s

n-ererststhee-

ther a revision or a competition process. Thare consistent with fixed-choice models suchthe garden-path model, construal theory,referential theory, as they predict that the analysis is always initially adopted. If thanalysis turns out to be implausible, then pcessing becomes difficult. They are also content with variable-choice models such as therestricted race model. Assuming that the analysis is nearly always initially adopted,analysis should hardly ever occur in the VP c

240 VAN GOMPEL, PICKERING, AND TRAXLER

TABLE 3

Experiment 1: ANOVA Results by Region for Effect of Condition

F1(2, 70) MSe F2(2, 58) MSe

Region 1First-pass reading times , 1 , 1Regression-path times , 1 , 1Total reading times , 1 , 1

Region 2First-pass reading times 1.20 5367 1.01 3324First-pass regressions 1.35 62.9 1.05 60.1Regression path times , 1 , 1Total reading times 3.38* 12096 4.55* 6907

Region 3First-pass reading times 1.33 4924 , 1First-pass regressions 7.06** 70.6 4.63* 93.5Regression-path times 3.69* 14061 2.62 20901Total reading times 2.07 16434 1.68 17462

Region 4First-pass reading times , 1 , 1First-pass regressions 1.38 70.5 1.46 33.3Regression-path times , 1 1.85 4079Total reading times 11.17** 5337 6.67** 6962

Region 5First-pass reading times , 1 , 1First-pass regressions , 1 , 1Regression-path times 1.12 7724 , 1Total reading times 13.72** 10271 14.02** 8413

Region 6First-pass reading times 1.63 11677 2.04 6873First-pass regressions 13.67** 173.4 22.74** 98.3Regression-path times 3.93* 81795 6.73** 45994Total reading times 7.00** 25575 9.61** 14639

Region 7First-pass reading times , 1 , 1First-pass regressions 4.11* 283.6 3.45* 199.3Regression-path times 7.30** 117661 4.10* 184031Total reading times , 1 , 1

Note. Degrees of freedom for the items analysis for Region 7 and for all regions on total time are (2, 56) (see foo

ei-dition and nearly always in the NP condition.

T

npranpat

in a iPladl

ioemplte-p t

anduahnalynce ithreo

imltse-cio

993sn

rer-lyil-hethee

hec-ion-ershat

erui-m-&l.,me3), of

e are”

&ndre,P

her inr etn a

ingim-g

mesm-un.-ndses

nlyotgrdad-

REANALYSIS IN SEN

However, if the corpora in which NP attachmeis more common than VP attachment are resentative for the sentences that our participhave been exposed to, the results are incomble with the coarse-grained frequency varianthe tuning hypothesis.

The results are consistent with constrabased theories on the assumption that in thebiguous condition, the VP attachment biascombination with the plausibility of the Vanalysis is strong enough to outweigh the psibility of the NP analysis, so that there is harany competition.

First-pass regressions from the final regsuggested that the VP condition may have bsomewhat more difficult to process than the abiguous condition, though this effect did not apear in any of the other analyses. This resudifficult for fixed-choice models to explain bcause the VP analysis should always be adoand no reanalysis should be necessary in thecondition. It also poses difficulty for constrainbased theories because they predict that, if thing, the ambiguous condition should be harto process than the VP condition. These resfit better with variable-choice models such the unrestricted race model, which predicts ton some occasions, readers may adopt the preferred NP analysis. However, the effect wrelatively weak in that it was obtained in onone analysis and at the end of the senteTherefore, Experiment 2 investigated balanambiguities. If the unrestricted race modelright, we would predict that readers adopt NP analysis more frequently, which should sult in a clear difference between the ambiguand VP condition.

Interestingly, our earliest effects emergedfirst-pass regressions and regression-path tifor the postcritical region. This pattern of resuis different from findings in some other eymovement experiments, which obtained effein first-pass times on the disambiguating reg(e.g., Adams, Clifton, & Mitchell, 1998; Alt-mann et al., 1992, 1998; Garnsey et al., 19Paterson et al., 1999; Trueswell et al., 191994). One reason may be that we employedmantic rather than syntactic disambiguatio
which was employed in many other studies. Se
ENCE PROCESSING 241

te-tsti-

of

t-m-n

u-y

nen-

- is

tedVP-y-

erltssaton-s

ce.dse-

us

nes

tsn

7;,e-,

mantic information may become available moslowly than syntactic information, and it is cetainly less strong in that it usually does not fuldisambiguate. Furthermore, although plausibity was manipulated on the noun in the PP, tactual disambiguation may have been atnext word. For example, (4b) could still bplausibly continued asThe hunter killed theleopard with the rifle wound. This might explainwhy our earliest effects were obtained in tpostcritical region. The fact that our results ocurred in first-pass regressions and regresspath times from this region suggests that readwent back to the parts of the sentence tcaused the disruption.

However, it does not explain why a numbof other studies on VP–NP attachment ambigties found effects in first-pass on the disabiguating noun (Britt et al., 1992; Ferreira Clifton, 1986; Ni et al., 1996; Rayner et a1983). We think there are two reasons. In sostudies (Britt et al., 1992; Rayner et al., 198the disambiguating noun occurred at the enda clause. Reading times at the end of a clausknown to include more integrative “wrap-upprocessing (Just & Carpenter, 1980; MitchellGreen, 1978; Rayner et al., 1989, 2000) atherefore reflect later processing. Furthermothe critical regions in most studies on VP–Nambiguities (and indeed most studies on otsyntactic ambiguities) were much longer thanthe current experiment. For instance, Rayneal.’s (1983) first-pass effects were based ocritical region including the following clause.Therefore, they include much later processthan the first-pass times in our experiment. Silarly, Ni et al. (1996) included words precedinthe critical phrase, and their phrases someticontained adjectives that may have disabiguated the structure before the critical noFinally, Ferreira and Clifton’s (1986) critical region employed syntactic disambiguation aconsisted of at least three words. When one ua short disambiguating region consisting of oone word as in our experiment, it is in fact ntoo surprising to find the effects in the followinregion. First-pass reading times for one-woregions are often very short, and therefore re
-ers will often have planned a saccade before the

terll

sfnc

r

e l

iercee

to-s

ptrded

e

-


processor has evaluated the plausibility of syntactic analysis. This explanation fits wwith current models of eye-movement contsuch as the E-Z reader model (Reichle, Posek, Fisher, & Rayner, 1998), which claims thsaccades to the next word are programmedfore lexical access is complete when accesthe word is slow (e.g., because the word is inquent). Assuming that syntactic and semaprocessing follow lexical access, this predithat the evaluation of the plausibility of a sytactic structure may also lag behind the pgramming of saccades.

Indeed, a number of eye-movement expments that employed a short critical regioncombination with semantic disambiguation ashowed the earliest effects in regression-bameasures for the postcritical region (Pickeret al., 2000; Pickering & Traxler, 1998, Traxlet al., 1998). The results of Experiment 1 pvide an important indication of the way in whiany effects might reveal themselves in the thretically more critical sentences used in Exp

e

ig

s

he

nstl,e

1,

e

bil-.e:

a

da-P

nthe paraphrases was very small, although the ef-

ment 2.

EXPERIMENT 2

This eye-tracking experiment investigatVP–NP attachment ambiguities in which thewas no initial preference for either VP or Nattachment. It employed items that weclosely related to those in Experiment 1 bincluded the additional wordonly. For thesesentences, current models of syntactic ambity resolution make very different predictionabout which sentences would cause procesdifficulty.

Participants

Twenty-seven participants took part. Six aditional participants were excluded from tanalyses, three because their data reflecthigh percentage of tracker losses and threecause they answered more than 25% of statements incorrectly.

Items

Participants read 30 sentences like (5),peated below (see Appendix):

RING, AND TRAXLER

hellolat-atbe-ingre-ticts

n-o-

ri-insosedngro-ho-ri-

drePreut

u-sing

d-ed abe-the

re-

(5a) The hunter killed only thepoacher with the rifle not long after sunset.(ambiguous)

(5b) The hunter killed only the leopardwith the rifle not long after sunset. (VP at-tachment)

(5c) The hunter killed only the leopardwith the scars not long after sunset. (NP at-tachment)

The items were semantically disambiguatedthe VP or NP analysis or were semantically ambiguous (see plausibility pretest). The itemwere similar to those in Experiment 1, excethat the object noun phrase began with the woonly and did not include an adjective. Thword-form frequency of the noun in the PP dinot differ according to the Celex database [F(1,29) 5 1.15,p 5 .29,MSe 5 110039], althoughthere was a difference in the logarithm of thword-form frequency [F(1, 29)5 5.20,p 5 .03,MSe 5 .63]. The fact that the noun was somewhat more frequent in the NP condition than ithe ambiguous and VP condition works againthe prediction of the unrestricted race modewhich claims that the NP condition should bmore difficult than the ambiguous condition.

Pretests

Pretests were the same as in Experimentexcept as follows.

Plausibility pretest. Eighteen participantsrated the plausibility of 50 sets of items. For thVP analysis, we replaced with with by usingorwith when using(4 item sets). For the final 30sets of items, Table 4 shows the mean plausiity rating for each condition, for each analysis

We again conducted 2 (Paraphrase typVP–PP vs NP–PP paraphrase) 3 3 (Condition:ambiguous vs VP vs NP) ANOVAs and foundsignificant interaction [F1(2, 34) 5 99.7, p ,.001,MSe 5 1.12; F2(2, 58) 5 714,p , .001,MSe 5 .26]. Planned comparisons revealeplausibility differences between the two parphrase conditions for VP attachment and for Nattachment (all ps , .001). In the ambiguousconditions, the numerical difference betwee

b

wa

-ete

e tc

feto

lendas inutn-

si-

r-r-P

cent

en

eg.,dVP

Note. Frequency is the number of occurrences of the word form in the Celex database (17.9 million words).

fect was significant by items [F1(1, 34) 5 1.81,p 5 .19; F2(1, 58) 5 12.9,p , .001].

We analyzed the absolute differencestween the two paraphrases in the three attament conditions. Planned comparisons reveathat the difference between the paraphrasesmuch smaller in the ambiguous condition thin the VP condition and the NP condition (aps , .001).

Off-line preference task. This pretest employed 18 further participants. The mean pcentages of VP–PP paraphrases are presenTable 4.4 ANOVAs revealed a main effect ocondition [F1(2, 34) 5 322, p , .001,MSe 5118; F2(2, 58) 5 174, p , .001,MSe 5 364].Planned comparisons revealed that the percage of VP–PP attachment responses in thebiguous condition differed from those in bothe VP attachment condition and the NP attament condition (all ps , .001), which indicatedthat the plausibility manipulation worked.

Importantly, the 55.0% VP–PP paraphrachoice in the ambiguous condition did not diffrom chance (both Fs , 1.2). This indicates thaplausibility and preference information did n
favor either the VP or NP analysis. Additionall k
Psi-kisd-

4 One additional participant was excluded because chose the NP-PP paraphrase 93% of the time, a patclearly different from any other participant. We presumthat she had failed to understand the task. Including this pticipant would have taken the percentage of VP completiocloser to 50%.

e-ch-ledasn

ll

r-d in

f

nt-am-hh-

ser

ty,

participants hardly ever took the implausibanalysis in the VP and NP conditions: 5.6 a2.8% respectively. We concluded that there wno preference for either VP or NP attachmentthe ambiguous condition for Experiment 2, bthat the final analysis in the disambiguated coditions was nearly always determined by plaubility.

Completion task. Eighteen participants tookpart. We excluded 0.2% of trials, on which paticipants said that they did not know which paaphrase was correct. The overall number of Vcompletions was significantly less than chanon the item analysis, but not on the participaanalysis [F1 5 1.62,p 5 .22,MSe 5 527; F2 511.44, p 5 .002, MSe 5 124] (see Table 4).There were no significant differences betwethe conditions (both ps . .10). The mean per-centage of VP completions, 40.3%, is within thrange normally considered balanced (e.Rayner & Duffy, 1986). We therefore concludethat there was no clear preference for either or NP attachment.

Predictions

The off-line preference and completion tasindicated that there was no bias for either Vor NP attachment for these items. The plaubility pretest and off-line preference tasshowed that plausibility favored one analysover the other in the VP and NP conditions anthat it equally favored both analyses in the am

sheternedar-ns


TABLE 4

Experiment 2: Item Characteristics

Condition

Ambiguous VP attachment NP attachment

Plausibility pretestPlausibility VP-PP paraphrase 5.8 5.8 2.2Plausibility NP-PP paraphrase 5.3 2.5 5.9

Frequency of noun in PP 185 185 277Log frequency of noun in PP 1.72 1.72 2.18Off-line preference task

Percentage of VP attachment 55.0 94.4 2.8Completion task

Percentage of VP completions 37.2 46.1 37.4

biguous condition. Hence, constraint-based the-

R

rcm

)rllsh))

r-

ns

nec

le

aneisn

oif

i

c

a i-

nSn

4tiri

%n-

ndon.on,-ntultseen

ote

ef-he-on

stfi-

on

4hehe

rf-

hatado

ci-


ories predict competition when the items asemantically ambiguous [as in (5a)], but muless competition when the items are disabiguated [as in (5b) and (5c)]; thus (5a) shoucause difficulty compared with (5b) and (5cwhich should not differ. Garden-path theoand construal predict that the parser initiaadopts the VP analysis and only reanalyzethat analysis becomes implausible. Hence, tpredict difficulty with the NP condition (5cversus the other two conditions [(5a) and (5bwhich should not differ. Ni et al.’s (1996) vesion of referential theory predicts that thparser initially adopts the NP analysis and oreanalyzes if that analysis becomes implauble. Hence, it predicts difficulty with the VPcondition (5b) versus the other two conditio[(5a) and (5c)], which should not differ. Thunrestricted race model predicts that eaanalysis is initially adopted about half the timbecause there is no initial bias for either anasis. The initial analysis is implausible half thtime in the disambiguated conditions, butways plausible in the ambiguous conditioHence, processing difficulty will occur in thdisambiguated conditions (due to reanalysbut not in the ambiguous condition. Assumithat Experiment 1 demonstrated the effectsdifficulty on patterns of eye movements fVP/NP ambiguities, we would expect any dferences among the conditions to beginemerge in first-pass regressions and regresspath time for the postcritical region.

Design, Procedure, and Analyses

These were the same as Experiment 1, exas follows. The word only was included in thedirect object region. Nine participants were rdomly assigned to read each list. There werefiller items. Thirty items, including 10 expermental items, were followed by a comprehesion statement with feedback. There was oone programmed break in the experiment. fillers preceded the first experimental senteand three followed the break. On average, pticipants made 8% errors. In each condition,of fixations was excluded because their durawas shorter than 80 ms. One percent of the t
were excluded because they had fewer than f
ING, AND TRAXLER

eh-

ld,yyif

ey

],

elyi-

s

hey-

l-.

),gofr-toon-

ept

n-66

n-lyixcear-%onals

fixations or major tracker losses. Less than 1of the data was removed as a result of two cosecutive regions having zero reading times.

Results and Discussion

Table 5 presents the mean reading times apercentage regressions by condition and regiFor each eye-tracking measure for each regiwe conducted one-way ANOVAs with three levels of Condition (ambiguous vs VP attachmevs NP attachment). Table 6 presents the resof these analyses. Planned comparisons betwlevels are reported below.

As in Experiment 1, first-pass times did nproduce any significant effects. However, thfirst-pass regression measure showed clearfects in Region 6 (the postcritical region), witplanned comparisons showing that fewer rgressions occurred in the ambiguous conditithan in both the VP condition [F1(1, 52) 5 4.27,p 5 .044; F2(1, 58) 5 6.23,p 5 .015] and theNP condition [F1(1, 52) 5 12.1; p , .001;F2(1, 58) 5 13.4, p , .001]. The VP and NPconditions did not differ [F1(1, 52) 5 1.99,p 5 .16; F2(1, 58) 5 1.36,p 5 .25].

The pattern for regression-path times wavery similar. In Region 6, the condition effecwas significant by items and just missed signicance by participants (p 5 .071). Planned com-parisons showed that the ambiguous conditiwas faster than both the VP condition [F1(1,52) 5 4.42, p 5 .041; F(1, 58) 5 10.0, p 5.002] and the NP condition [F1(1, 52)5 3.91,p 5 .053; F(1, 58) 5 9.04,p 5 .004]. The VPand NP conditions did not differ (Fs , 1).

Total times showed an effect in Region (with the). Planned comparisons showed that tambiguous condition was read faster than tVP condition [F1(1, 52) 5 6.24,p 5 .016; F2(1,58) 5 5.70, p 5 .020] and the NP condition[F1(1, 52) 5 5.91,p 5 .019; F2(1, 58) 5 5.40,p 5 .024], which did not differ from each othe(Fs , 1). There was a marginally significant efect in Region 3, the direct object (p , .10 byparticipants and items); the means suggest tthe ambiguous condition may have been refaster than the disambiguated conditions. Nother effects approached significance by parti

ourpants and items.

n a

f

rtee

0 Vis

u Vas

o

h,yd

es

utdf--t

.-s-

t-edi-i-

e,ri--

Note. First pass, regression path, and total times are reported in milliseconds and first-pass regressions as the percentageof saccades leaving the region to the left after a first pass fixation.

The data revealed that the ambiguous cotion was easier to process than both the VPNP attachment conditions. There was no edence for either a VP or NP attachment preence: Both conditions were read equally faThe earliest effects arose in first-pass regsions and regression-path times immediaafter the critical noun. Thus the locus of the fects was identical to that in Experiment 1.

The results fit well with a variable-choicmodel such as the unrestricted race mo(Traxler et al., 1998; Van Gompel et al., 200In this model, readers may adopt either theor NP analysis initially. If the initial analysturns out to be inconsistent with later informtion, reanalysis occurs and processing difficuensues. Given that there was no differencetween the VP and NP condition, our results sgest that reanalysis occurred equally often inVP and NP condition and therefore that the and NP analysis were adopted about equoften during the first stage of processing. Thiconsistent with the off-line and completiopretests, which indicated that the materials fr
Experiment 2 were balanced.
di-nd

vi-er-st.es-lyf-

edel).P

a-ltybe-g-

thePlly isnm

The results may also be compatible witother variable-choice models. For examplesome variants of the tuning hypothesis maclaim that readers are sensitive to fine-grainefrequency information associated withonlyand that VP and NP attachment in sentenccontainingonly are about equally frequent inEnglish. Hence, reanalysis should occur aboequally often in the VP and NP disambiguateconditions. However, the results are more dificult to reconcile with the coarse-grained frequency variant of the tuning hypothesis thawas favored by Mitchell et al. (1995; Brys-baert & Mitchell, 1996; Cuetos et al., 1996)According to this variant, the processor ignores lexical preferences and preferences asociated with particular NPs. Thus, VP/NP atachment ambiguities should always show thsame preference, regardless of whether therect object NP contained an adjective (Experment 1) oronly (Experiment 2). However, Ex-periment 1 showed a strong VP preferencwhereas there was no preference in Expement 2. In order to save this variant of the tun


TABLE 5

Experiment 2: Means

Region

1 2 3 4 5 6 7The hunter killed only the poacher with the rifle not long after sunset

First-pass reading timesAmbiguous 371 331 520 265 310 448 464VP attachment 398 315 555 272 317 463 479NP attachment 405 312 560 286 314 442 491

First-pass regressionsAmbiguous 6.2 9.0 6.8 15.6 13.9 59.1VP attachment 9.8 9.0 7.4 17.1 22.1 56.4NP attachment 6.5 9.5 9.1 14.3 27.1 61.1

Regression-path timesAmbiguous 371 363 620 299 360 562 1063VP attachment 398 374 626 332 379 693 1172NP attachment 405 342 645 342 377 682 1133

Total timesAmbiguous 421 425 706 362 398 633 595VP attachment 469 435 795 409 404 681 597NP attachment 456 412 764 410 432 674 611

ing hypothesis, one would have to assume that

aethmte

e s,on

sora-he at-

-n-ede

i-bi-et

the participants from Experiment 1 and 2 hdifferent preferences, but this seems unlikbecause they were randomly drawn fromsame population. In addition, the results froour pretests, which used different participanwere consistent with the eye-movement expiments.

The results are also inconsistent with fixchoice models, as such models predict thatprocessor always adopts the same analysithat difficulty relative to the ambiguous conditioccurs either for the VP or NP condition, but

** p significant at the .01 level.

for both. The data are incompatible with the ga

dlye

s,r-

d-the

sonot

den-path theory, which predicts that the procesadopts the VP analysis initially and only reanlyzes when this analysis is implausible. Thus, tgarden-path model cannot explain why the VPtachment condition is more difficult than the ambiguous condition. Similarly, the results are icompatible with referential theory, as interpretby Ni et al. (1996), which predicts that thprocessor adopts the NP analysis initially.

Importantly, the current results provide evdence against fixed-choice models in an amguity that involves primary phrases. Traxler


TABLE 6

Experiment 2: ANOVA Results by Region for Effect of Condition

F1(2, 52) MSe F2(2, 58) MSe

Region 1First-pass reading times 2.47 2769 1.91 4664Regression-path times 2.47 2769 1.91 4664Total reading times 2.47 5451 1.46 10720

Region 2First-pass reading times , 1 1.87 1977First-pass regressions 2.29 53.1 1.52 78.9Regression-path times 1.82 3528 2.14 3999Total reading times , 1 , 1

Region 3First-pass reading times 2.95 4006 1.69 7769First-pass regressions , 1 , 1Regression-path times , 1 , 1Total reading times 3.10 16752 2.72 21394

Region 4First-pass reading times 2.30 1035 1.74 1706First-pass regressions , 1 , 1Regression-path times 1.54 8908 1.20 13745Total reading times 4.05* 5112 3.70* 6567

Region 5First-pass reading times , 1 , 1First-pass regressions , 1 , 1Regression-path times , 1 , 1Total reading times 1.69 4458 1.85 6743

Region 6First-pass reading times , 1 , 1First-pass regressions 6.12** 181.1 7.00** 200.2Regression-path times 2.78 45505 6.37** 27139Total reading times , 1 1.88 10018

Region 7First-pass reading times , 1 , 1First-pass regressions , 1 , 1Regression-path times , 1 , 1Total reading times , 1 , 1

* p significant at the .05 level.

r-al. (1998) showed that in ambiguities that in-

T

usroa

lsr&

oisuuu. ieruoiothtas

s

a&y

uhoean

e

roa

esenateti-al.,&im

ehely

icthereus,ker.te-s of thee-ticalsig-edceassdif-eri-le 7ach

-ceden- ants

ntmsled

s.ac-

cates that the pattern of results was roughly sim-ilar for balanced and biased items.5 Most impor-

5 The NP attachment condition seemed to produce moreregressions than the VP attachment condition in the biasedmaterials, but not in the balanced materials. This was due tothe fact that, overall, there was a slight (nonsignificant) pref-erence for VP attachment. As a consequence of removing allbalanced materials, the difference increased for the biased

REANALYSIS IN SEN

volved nonprimary relations, globally ambigous sentences were easier to read than dibiguated sentences (when there was no stpreference for either analysis). However, it wunclear whether this kind of pattern could abe obtained for ambiguities that involved pmary relations. Construal theory (Frazier Clifton, 1996) predicts a different pattern fthese two types of ambiguities because it variable-choice model with respect to ambigties that involve only nonprimary relations, ba fixed-choice model with respect to ambigties that involve at least one primary relationclaims that the VP/NP attachment ambiguitin the current experiments involve two primarelations and that minimal attachment shoapply. However, the fact that the VP conditiwas more difficult than the ambiguous conditindicates that our readers initially adopted NP analysis on a number of trials. Hence,current experiment suggests that nonprimambiguities do not constitute a special claEven in ambiguities involving primary relations, globally ambiguous sentences are eato process than disambiguated sentences.

Constraint-based theories (e.g., MacDonet al., 1994; McRae et al., 1998; SpiveyTanenhaus, 1998; Spivey-Knowlton & Sediv1995; Tabor et al., 1997; Tabor & Tanenha1999) also cannot satisfactorily account for tdata. Such models predict that the ambigucondition should cause processing difficulty bcause there is no initial preference for eithertachment site, nor does plausibility favor oanalysis over the other. This ought to resultcompetition between the VP and NP analysIn contrast, the VP and NP analyses shouldrelatively easy because plausibility suppoonly one analysis. But in fact the exact oppsite occurred: The ambiguous condition wthe easiest.

In theory, there are a number of ways that thresults might still be consistent with competitiomodels. In the introduction, we mentioned thsome constraint-based theories allow comption to decrease very rapidly (e.g., McRae et1998; Spivey & Tanenhaus, 1998; TaborTanenhaus, 1999). These models might cla
that competition occurred as soon as the ambig
ENCE PROCESSING 247

-am-ngso

i-

r ai-ti-Its

yldnne

herys:

-ier

ld

,s,eus-t-eins.

bets-s

ity arose atwith, but that one analysis becammore activated than its alternative during tnext words, so that competition was relativeweak when readers reached the critical nounriflein the ambiguous condition. Such models predthat, especially if items are not well balanced, tpreferred analysis should quickly become moactivated than the nonpreferred analysis. Thless well-balanced items should produce weacompetition than more closely balanced items

In order to investigate this, we split the marials into balanced and biased items. The biathe materials was determined on the basis ofresults from the eye-tracking study itself. Bcause first-pass regressions from the postcriRegion 6 was the earliest measure to show nificant effects, balanced materials were definas those 15 materials for which the differenbetween the VP and NP attachment in first-pregressions from Region 6 was smallest (a ference of 12.5% or less). The other 15 matals were assigned to the biased group. Tabpresents the means for each condition for egroup of materials.

ANOVAs with Condition as a within-participants and -items factor and Item Bias (balanvs biased) as a within-participants and betweitems factor were conducted. They revealedeffect of Item Bias, significant by participanonly [F1(1, 26) 5 8.44,p 5 .007,MSe 5 309;but F2(1, 28) 5 1.90,p 5 .18, MSe 5 423]. Ifthis is a real effect, it could be due to irrelevadifferences between the two groups of ite(e.g., frequency or length). They also reveaan effect of Condition [F1(2, 52) 5 7.31,p 5.002, MSe 5 389; F(2, 56) 5 6.90, p 5 .002,MSe 5 203], just as in the original analyseThere was no interaction between the two ftors [F1(2, 52) 5 2.39,p 5 .10,MSe 5 290; F2(2, 56) 5 .60,p 5 .55,MSe 5 203], which indi-

u-materials.

R

zig

ghIt

e

mstyn

ni

pA

)

e

edhehe

ands oft of

on,

ofsesheorethebere-rd

heam- Ined aerepe-

y-m-

theandon-n,

dis-m-is-hes ofnt-t 1ses.aimion al.,

1998; Trueswell et al., 1993).6

ro

NP attachment 21.5 32.8 26.2 27.9

6 It is likely that the on-line data provide a better estimateof the item biases. However, we could not use them here,because the ambiguous-disambiguated difference and VP-NP difference are not independent. For example, selectingthose materials where the VP condition is easier than theNP condition necessarily reduces the VP-ambiguous differ-ence. Hence, even if the item biases are caused by random

tantly, when the balanced items were analyseparately, the effect of Condition was still snificant [F1(2, 52) 5 3.52, p 5 .037, MSe 5302; F2(2, 28) 5 6.99,p 5 .003,MSe 5 80.4]Thus, even for the balanced items the ambious condition produced fewer regressions teither the VP or NP attachment condition. fact, the pattern for the balanced items was same to that for the complete set of items. Tresults do not show any evidence that comption is stronger for more balanced items.

Just as items can be biased, so participants have a strong preference for one of the analyHence, we also tested whether participants did not have a strong preference for one analexperienced more competition than participathat did have a strong preference. Participawere split into two groups. Bal-anced participawere defined as those 13 participants for whthe difference between the VP and NP attachmin first-pass regressions from Region 6 was smest (12.2% or less). Biased participants were14 other participants. The means for each grouparticipants are presented in Table 7. ANOVwith Condition as a within-participants and -itemfactor and Participant Bias (balanced vs biaseda between- and within-items factor showedeffect of Condition [F1(2, 50) 5 5.98,p 5 .005,MSe 5 186,F2(2, 58) 5 6.15,p 5 .004,MSe 5383], but no effect of Participant Bias [F1(1, 25)5 .050,p 5 .83,MSe 5 676; F1(1, 29) 5 1.41,p5 .24, MSe 5 208] or interaction between thtwo factors [F(2, 50) 5 .37,p 5 .69,MSe 5 186;F2(2, 58) 5 .53,p 5 .59,MSe 5 383], indicatingthat biased and balanced participants performsimilarly. As shown in Table 7, the pattern for tbalanced participants was very similar to t
overall pattern: The ambiguous condition pro
ING, AND TRAXLER

ed-

u-annheheti-

ayes.

hatsists

ntstschentall-the ofss asan

duced fewer regressions than both the VP NP attachment conditions. Separate analysethe balanced participants showed an effecCondition [F1(2, 58) 5 3.13,p 5 .051,MSe 5294,F2(2, 24) 5 4.97 p 5 .016,MSe 5 97.4].Thus, there was no evidence for competitieven for the most balanced participants.

Finally, we conducted a more direct testmodels that claim that competition decreaquickly. If competition decreases rapidly, tpreferred analysis should quickly become mactivated than its alternative, and therefore activation of the preferred analysis should much higher than the activation of the nonpferred analysis just before the critical worifle/scars. Hence, disambiguation toward tpreferred analysis should be easier than disbiguation toward the nonpreferred analysis.addition, disambiguation toward the preferranalysis should not be more difficult thanglobally ambiguous sentence, where thshould be some (though perhaps weak) comtition due to equal plausibility of the two analses. We tested this by recoding the disabiguated conditions into cases where disambiguation was consistent with the bias into cases where the disambiguation was incsistent with it. Thus, in the consistent conditiothe sentence was either VP biased and VPambiguated, or NP biased and NP disabiguated. In the inconsistent condition, the dambiguation was different from the bias. Tcompletions were used to determine the biathe materials. As the huge difference in perceage of VP completions between Experimenand 2 shows, this task is sensitive to item biaFurthermore, constraint-based theories clthat on-line biases should reflect completdata (e.g., Garnsey et al., 1997; McRae et

m


TABLE 7

Experiment 2: Percentages of First-Pass Regressions fRegion 6 by Condition and Item and Participant Bias

Item bias Participant bias

Balanced Biased Balanced Biased

Ambiguous 11.2 16.5 14.8 13.1VP attachment 21.4 22.8 24.0 20.2

-variability, this would still reduce the difference betweenthe ambiguous and consistent condition.

T

l.og

.3,d

o

m

h hteo

s

h

o

e

yni

rbna

o

ionbly,en

tothe

VP and ambiguous condition. In these condi-

bi- Itdis-me dis-us

usi-ut oren-e of ineureshatm-bi-

eri-on.n-al.,us,r

uses iss an. is forn,ro-nd

en--l-de-esrl-

REANALYSIS IN SEN

Three items were excluded from the anases, because they were perfectly balancedbefore, we analyzed first-pass regressions frthe postcritical region. The mean percentaof regressions for the ambiguous, consisteand inconsistent conditions were 14.5, 26and 23.9% respectively. Analyses with contion (ambiguous vs consistent vs inconsisteas a within-participants and -items factshowed a main effect of condition: [F1(2, 52)5 4.76, p 5 .013, MSe 5 168; F2(2, 52) 55.23,p 5 .008,MSe 5 223]. However, plannedcomparisons showed that the two disabiguated conditions did not differ (Fs , 1).This suggests that the materials were very mogeneous just before the critical noun—novery few materials were strongly biased at tpoint. Further comparisons showed that ambiguous condition produced fewer regrsions than both the inconsistent conditi[F1(1, 52) 5 5.78,p 5 .020; F2(1, 52) 5 5.95,p 5 .018] and the consistent condition [F1(1,52) 5 8.28,p 5 .006; F2(1, 52) 5 9.35,p 5.004]. These results challenge constraint-batheories in which competition is resolved raidly because they predict that, if anything, tambiguous condition should be more difficuthan the consistent condition. No competitishould occur in the consistent condition, bcause the initial bias and plausibility favorthe same analysis. In contrast, some comption should have occurred in the ambiguocondition, because plausibility supported boanalyses. (Note that the nonpreferred analshould not have been completely abandoaccording to current constraint-based theorsee the introduction). Clearly, the data did nsupport this prediction.

Finally, we need to consider whether the sults might be due to a frequency difference tween the NP attachment and ambiguous cotions for the noun in the PP. Although the rfrequency of this noun was controlled for boconditions, there was a difference in the logrithmic frequency: The logarithmic frequencwas higher in the NP than in the ambiguous cdition. If this small difference had an effect, wwould expect the NP condition to be easier th
the ambiguous condition, but this was not th
ENCE PROCESSING 249

y- Asmesnt,

case. The noun frequency of the NP conditwas also higher than the VP condition. Possithis might have neutralized a difference betwethe two conditions. However, it is important note that the crucial contrast was between

i-nt)r

-

o-oris

hes-n

edp-eltn

e-d

eti-usthsised

es;ot

e-e-di-wtha-yn-ean

tions, the noun was identical.

GENERAL DISCUSSION

Experiment 1 used VP–NP attachment amguities with an initial bias for VP attachment.found that sentences that were semantically ambiguated to the NP analysis rapidly becaharder to process than sentences that wereambiguated to the VP analysis or ambiguosentences in which either analysis was plable. Experiment 2 used similar ambiguities, bin this case, there was no bias for either VPNP attachment. It found that ambiguous stences were easier to process than either typdisambiguated sentence (which did not differdifficulty). The effects in Experiment 2 werobtained in the same eye-movement measand regions as in Experiment 1, indicating tsyntactic ambiguity resolution in balanced abiguities is roughly as rapid as in biased amguities.

We have argued that the results from Expment 2 provide evidence against competitiHence, the results are incompatible with costraint-based theories (e.g., MacDonald et 1994; McRae et al., 1998; Spivey & Tanenha1998; Spivey-Knowlton & Sedivy, 1995; Taboet al., 1997; Tabor & Tanenhaus, 1999) becaa fundamental assumption of these theoriethat the processor employs competition amechanism of syntactic ambiguity resolutioGiven that it is such a crucial assumption, ithard to see how such theories can accountour results. For example, without competitiothe competition-integration model that was pposed by McRae et al. (1998) and Spivey aTanenhaus (1998) would have to become antirely different type of model; in fact, it is unclear how the model would work without it. Athough it may be turn out to be possible to velop alternative parallel interactive theoriwithout a competition mechanism (e.g., Pea
emutter & Mendelsohn, 1999, for an initial at-

R

tvc o

oaxie

ig bs,

otinusnvn

m

ci

de

fe5)aime

tnrnhtoni

t arely-thatly-rellyan

heynt

),le--

e noter,

ceiletheatuldne-es-

ri-del

al.,m on

al.haten-con-icu-ngthsche.m-nt, Inneed lit-he


tempt), these theories are very different from current constraint-based theories. Further deopment of these models is required before it be established whether they can account fordata.

Another problem for constraint-based theries is that our results suggest that syntactic biguity resolution cannot be equated with lecal ambiguity resolution, contrary to theclaims (e.g., MacDonald et al., 1994; Trueswet al., 1993, 1994). Research on lexical ambity resolution suggests that balanced wordsneutral contexts are harder to process than anced words in constraining contexts or biawords (Duffy et al., 1988; Rayner & Duffy1986). In lexical ambiguity resolution, it appears likely that the lack of evidence in favorone or other interpretation causes competibetween the meanings of the word and heprocessing difficulty. However, the ambiguocondition in Experiment 2 reproduced theconditions in syntactic ambiguity resolution afound very different results; there was no edence for competition in truly ambiguous setences compared to semantically disambiguasentences. In fact, the ambiguous conditionExperiment 2 was easier than both disabiguated conditions. We conclude that lexiand syntactic ambiguity resolution are very dferent processes.

Our results are also problematic for fixechoice two-stage models such as the gardpath theory (e.g., Frazier, 1979, 1987) and reential theory (Crain & Steedman, 198Altmann & Steedman, 1988; Ni et al., 1996Such models predict that readers should alwinitially adopt the same analysis. Although thprediction is consistent with the findings froExperiment 1, which showed little differencbetween the ambiguous and VP condition, inot consistent with the findings of Experime2: Both the VP and NP conditions were modifficult to read than the ambiguous conditioThe VP–NP attachment ambiguity showed tresults for primary relations are very similarthose obtained by Traxler et al. (1998) for noprimary relations. For both types of ambigu
ties, strongly biased globally ambiguous se
ING, AND TRAXLER

heel-anur

-m-i-rllu-inal-ed

-f

oncesedi--

tedof-

alf-

-n-r-;.yss

iste.at

--

tences are easier than biased sentences thadisambiguated toward their nonpreferred anasis, but about as easy as biased sentencesare disambiguated toward their preferred anasis. This contrasts with ambiguities that aroughly balanced. In such ambiguities, globaambiguous sentences are more difficult thdisambiguated sentences, whichever way tare disambiguated. This finding is inconsistewith construal theory (Frazier & Clifton, 1996which claims that the processor is variabchoice for ambiguities involving nonprimary relations, but fixed choice for ambiguities involving primary relations.

The results fit better with variable-choicmodels, which claim that the processor doesadopt the same analysis on all trials. Howevthey are not compatible with all variable-choimodels. The results are difficult to reconcwith the coarse-grained frequency variant of tuning hypothesis, as it incorrectly predicts ththe sentences in Experiment 1 and 2 shoshow the same preferences, because figrained frequency information such as the prence or absence of only in a NP is not used ininitial parsing.

We have proposed a particular variant of vaable-choice models, the unrestricted race mo(see also Traxler et al., 1998; Van Gompel et2000). This model fits well with the results frothe current experiments and the experimentsrelative clause ambiguities in Traxler et (1998). The unrestricted race model claims tthe analyses of a syntactic ambiguity are gaged in a race and that the analysis that is structed fastest is adopted. How often a partlar analysis is adopted depends on the streof the initial bias. If there is no clear initial biafor either analysis, as in Experiment 2, eaanalysis will be adopted about half the timWhen the sentence is semantically disabiguated toward either VP or NP attachmethis results in reanalysis about half the time.contrast, when there is an initial bias for oanalysis, as in Experiment 1, the preferranalysis is usually constructed fastest. Hencetle or no processing difficulty arises when t

n-ambiguity is disambiguated toward this analy-

em

t

lfpo

se(

.t

a

ihe-

nt

n-pt

be-iveiz-usly-ase.na-lly,ionm-therd be

hathisatporthat

ce. Inl-

ier toro-ha-nt-d bywetentle-

choice two-stage model of parsing.

a

a

ly

REANALYSIS IN SEN

sis. In general, the stronger the ambiguity isased, the more often the favored analysis istially selected, the more often reanalysis is quired when the sentence is disambiguatoward the disfavored analysis, and the loften reanalysis is required when it is disabiguated toward the favored analysis. Thus,unrestricted race model claims that gradedfects in syntactic ambiguity resolution are least in part due to the frequency with which initial analysis has to be revised.

In the current study, the off-line and comption pretests were consistent with on-line preences as measured by the eye-tracking exments. However, it is important to stress that conclusions do not depend on the pretests. eye-tracking experiments themselves showthat the materials in Experiment 1 were bia(because the VP and NP conditions differand the materials in Experiment 2 balanced cause the VP and NP conditions did not diffeMost importantly, they showed that the ambigous condition was easier than both disabiguated conditions in balanced ambiguitiesis these eye-tracking results which provide most critical evidence against competition afixed-choice models, not the correspondencetween our on-line data and pretests.

We favor the unrestricted race model overaccount which assumes that readers werelazy and did not resolve the ambiguous contions in the experiments. We think that there a number of reasons why such an explanationot very attractive. Note that it would be incosistent with all current sentence processmodels. All two-stage and constraint-based tories assume that the same processing menisms deal with globally ambiguous and dambiguated sentences. One would havepostulate an additional mechanism which plains why globally ambiguous sentences not resolved, whereas disambiguated senteare. One would also need a mechanism thatermines when a sentence is sufficiently abiguous that processing be suspended. Hesuch a model would clearly be less parsimnious than models that have a single mec
nism to handle both types of sentences. We a

bi-ini-re-tedss-

theef-athe

e-er-eri-ur

Theededd)

be-r).u-m- Ithendbe-

anjustdi-re

n isn-ng

think that this interpretation of our data is ulikely. First, we know that readers did attemto resolve the disambiguated conditions,cause they showed processing difficulty relatto the ambiguous condition. Second, recogning that a sentence is globally ambiguowould require constructing all relevant anases, the VP and NP analyses in the current cThe processor cannot simply have failed to alyze the ambiguous sentences at all. Finathis explanation would predict that suspensof processing also takes place in globally abiguous sentences that are implausible on eianalysis. Thus, sentences such as (8) shouleasy to process as well:

(8) The hunter killed only the leopardwith the rainbow.

Although we did not investigate sentences twere implausible on both readings, we think tprediction is very unlikely to be correct and thdata from sentences such as (8) would supthe unrestricted race model, which claims tthey should be hard.

To summarize, our results provide evidenabout the mechanisms that underlie parsingparticular, the finding of Experiment 2 that baanced ambiguous VP–NP sentences are easprocess than their disambiguated controls pvides clear evidence against both the mecnism of competition assumed by constraibased models and the strategies assumefixed-choice two-stage models. In contrast,have argued that the processing is consiswith the unrestricted race model, a variab

cha-is- toex-areces

de-m-nce,o-

ha-

APPENDIX

Experimental Items from Experiment 1

The farmer injured the young thief with the staff only few days ago.

The farmer injured the black sheep with the staff onlyfew days ago.

The farmer injured the black sheep with the wound ona few days ago.

lsoThe outlaw wounded the tall guard with the crossbow

during the night.

b

e

i

d

t

o

i

a

y

lver

oat

iver

iver

tion

not

not

ars

for

aker

for

rial

erial

ults

ain

ain

ain

tho-

tho-

era-

tol

tol

llar

g

he


The outlaw wounded the grey horse with the crossduring the night.

The outlaw wounded the tall guard with the whiskduring the night.

The knight struck the hated king with the sword durthe dark night.

The knight struck the iron hinge with the sword durthe dark night.

The knight struck the hated king with the beard durthe dark night.

The explorer found the lost tent with the charts aftelong search.

The explorer found the large sea with the charts aftlong search.

The explorer found the large sea with the whales aftlong search.

Patrick transported the black car with the trailer the before yesterday.

Patrick transported the black cow with the trailer the before yesterday.

Patrick transported the black car with the scratch thebefore yesterday.

The policeman touched the dirty beggar with the safter some hesitation.

The policeman touched the dirty stains with the safter some hesitation.

The policeman touched the dirty beggar with the feafter some hesitation.

The hooligan damaged the new shop with the firewolate at night.

The hooligan damaged the new pane with the firewlate at night.

The hooligan damaged the new shop with the discolate at night.

The farm-hand flogged the wicked boy with the brancnumber of times.

The farm-hand flogged the young pony with the brancnumber of times.

The farm-hand flogged the young pony with the bruisnumber of times.

The handyman repaired the enormous toolbox withpliers right after lunch.

The handyman repaired the electric circuit with the plright after lunch.

The handyman repaired the electric circuit with the dage right after lunch.

The walker saw the spying man with the telescope ein the afternoon.

The walker saw the young deer with the telescope ein the afternoon.

The walker saw the young deer with the infection earlthe afternoon.

The farmer deterred the cunning robber with the revojust before dawn.

RING, AND TRAXLER

ow

rs

ing

ng

ing

r a

er a

er a

day

ay

day

tick

ick

ver

rks

rks

unts

h a

h a

e a

the

ers

m-

arly

arly

in

The farmer deterred the dangerous wolf with the revojust before dawn.

The farmer deterred the cunning robber with the raincjust before dawn.

The man repaired the metal toolbox with the screwdrafter one month.

The man repaired the antique watch with the screwdrafter one month.

The man repaired the antique watch with the malfuncafter one month.

The hunter killed the dangerous poacher with the riflelong after sunset.

The hunter killed the dangerous leopard with the rifle long after sunset.

The hunter killed the dangerous leopard with the scnot long after sunset.

Peter yelled at the tall protester with the loudspeakera long time.

Peter yelled at the slow racehorse with the loudspefor a long time.

Peter yelled at the tall protester with the windcheatera long time.

The scoundrel scratched the yellow car with the aeafter the football match.

The scoundrel scratched the red enamel with the aafter the football match.

The scoundrel scratched the yellow car with the faafter the football match.

The angry man hit the nasty boy with the crowbar agand again.

The angry man hit the young cat with the crowbar agand again.

The angry man hit the young cat with the stripes agand again.

The surgeon examined the old doctor with the stescope after half past five.

The surgeon examined the weak heart with the stescope after half past five.

The surgeon examined the old doctor with the tempture after half past five.

The criminal shot the approaching cop with the pisyesterday late at night.

The criminal shot the approaching dog with the pisyesterday late at night.

The criminal shot the approaching dog with the coyesterday late at night.

The doctor wet the brown bottle with the liquid durinthe examination.

The doctor wet the painful skin with the liquid during texamination.

lver The doctor wet the painful skin with the rashes during theexamination.

x

a

a

a

a

tt

tt

tt

t

a

th

t

t

e

e

v

v

g

g

o

o

o

ur-

dur-

vent

vent

bish

r a

r a

r a

a a

a

t a

he

he

les

few

few

ts a

ep

ep

p

ter

fter

fter

rly

arly

REANALYSIS IN SEN

The old workman hurt the filthy miner with the pickaalmost a week ago.

The old workman hurt the nasty poodle with the pickalmost a week ago.

The old workman hurt the filthy miner with the stuttermost a week ago.

The annoying boy damaged the iron shelf with the hmer some time ago.

The annoying boy damaged the wine glass with the hmer some time ago.

The annoying boy damaged the wine glass with cracks some time ago.

The woman dried the iron rack with the towel a liwhile ago.

The woman dried the wet wound with the towel a liwhile ago.

The woman dried the wet wound with the blood a liwhile ago.

The landlady warmed the tiny room with the heawithin a couple of minutes.

The landlady warmed the nice soup with the hewithin a couple of minutes.

The landlady warmed the nice soup with the garlic wia couple of minutes.

The beggar frightened the strange lunatic with penknife late at night.

The beggar frightened the small greyhound with penknife late at night.

The beggar frightened the strange lunatic with the frles late at night.

The teacher poked the funny girl with the umbrella evtime again.

The teacher poked the sandy soil with the umbrella etime again.

The teacher poked the funny girl with the sickness etime again.

The crazy man stabbed the furious knight with the daduring the unrest.

The crazy man stabbed the fierce bulldog with the daduring the unrest.

The crazy man stabbed the fierce bulldog with the cduring the unrest.

The teenager washed the soiled sink with the spearly in the morning.

The teenager washed the injured ear with the spearly in the morning.

The teenager washed the injured ear with the pimearly in the morning.

The hunter aimed at the old poacher with the firearm ding the morning.

rly


e

xe

l-

m-

m-

the

le

le

le

er

ter

in

he

he

ck-

ry

ery

ery

ger

ger

llar

nge

nge

ple

ur-

The hunter aimed at the grey pigeon with the firearm ding the morning.

The hunter aimed at the old poacher with the rabbits ing the morning.

The caretaker cleaned the small bucket with the solafter a while.

The caretaker cleaned the coffee stain with the solafter a while.

The caretaker cleaned the small bucket with the rubafter a while.

The lord protected the disputed field with the fence folong time.

The lord protected the feeding cattle with the fence folong time.

The lord protected the disputed field with the moles folong time.

Experimental Items from Experiment 2

The old man poked only the chap with the umbrellnumber of times.

The old man poked only the soil with the umbrellanumber of times.

The old man poked only the chap with the raincoanumber of times.

The brute frightened only the scoundrel with tpenknife late at night.

The brute frightened only the greyhound with tpenknife late at night.

The brute frightened only the scoundrel with the frecklate at night.

The farmer injured only the shepherd with the staff a days ago.

The farmer injured only the sheepdog with the staff a days ago.

The farmer injured only the sheepdog with the spofew days ago.

The tourists followed only the naturalist with the jeduring the trip.

The tourists followed only the rhinoceros with the jeduring the trip.

The tourists followed only the rhinoceros with the limduring the trip.

The old man pointed at only the girl with the stick afsome time.

The old man pointed at only the star with the stick asome time.

The old man pointed at only the girl with the dress asome time.

The tourist saw only the walker with the telescope eain the afternoon.

The tourist saw only the badger with the telescope ein the afternoon.

The tourist saw only the badger with the infection ea
in the afternoon.

t

th

i

a

a

o

f

f

f

in

n

in

n

n

o

te

e

e

ft

f

ft

e

e

th

s

a

er a

er a

er a

ter

me

fter

bar

last

last

ers

iers

age

ll-

ull-

a-

for

e for

or a

oat

tor-

on-

tile

tile

age


The tribesman knocked down only the enemy with cudgel the other day.

The tribesman knocked down only the tiger with cudgel the other day.

The tribesman knocked down only the tiger with thejury the other day.

The peasant damaged only the shed with the tractor the hard work.

The peasant damaged only the pane with the tractor the hard work.

The peasant damaged only the shed with the windafter the hard work.

The explorer found only the hiker with the compass athe hurricane.

The explorer found only the coast with the compass athe hurricane.

The explorer found only the coast with the harbour athe hurricane.

The intruder struck only the hero with the sword durthe attack.

The intruder struck only the lock with the sword durithe attack.

The intruder struck only the hero with the beard durthe attack.

The hunter killed only the poacher with the rifle not loafter sunset.

The hunter killed only the leopard with the rifle not loafter sunset.

The hunter killed only the leopard with the scars not lafter sunset.

Jonathan spotted only the bag with the binoculars aflong search.

Jonathan spotted only the fox with the binoculars aftlong search.

Jonathan spotted only the bag with the wristwatch aftlong search.

The vandal scratched only the lorry with the aerial athe football match.

The vandal scratched only the paint with the aerial athe football match.

The vandal scratched only the lorry with the faults athe football match.

The African hunter aimed at only the rival with the spduring the pursuit.

The African hunter aimed at only the zebra with the spduring the pursuit.

The African hunter aimed at only the zebra with wound during the pursuit.

The young ruffian kicked only the girl with the trainernumber of times.

The young ruffian kicked only the ball with the trainersnumber of times.

RING, AND TRAXLER

he

e

n-

fter

fter

ws

ter

ter

ter

g

g

g

g

g

ng

r a

r a

r a

er

ter

er

ar

ar

e

a

The young ruffian kicked only the girl with the pigtailsnumber of times.

The adventurer found only the bag with the charts aftlong time.

The adventurer found only the sea with the charts aftlong time.

The adventurer found only the sea with the whales aftlong time.

The man touched only the tramp with the knife afsome hesitation.

The man touched only the stain with the knife after sohesitation.

The man touched only the tramp with the fever asome hesitation.

The angry man battered only the boy with the crowlast week on Sunday.

The angry man battered only the cat with the crowbarweek on Sunday.

The angry man battered only the cat with the stripesweek on Sunday.

The handyman repaired only the toolbox with the pliright after lunch.

The handyman repaired only the lantern with the plright after lunch.

The handyman repaired only the lantern with the damright after lunch.

The police tracked down only the terrorist with the budogs after the chase.

The police tracked down only the narcotics with the bdogs after the chase.

The police tracked down only the terrorist with the dynmite after the chase.

Peter yelled at only the protester with the megaphonea long time.

Peter yelled at only the racehorse with the megaphona long time.

Peter yelled at only the protester with the moustache flong time.

A coastguard rescued only the tourist with the motorbduring the patrol.

A coastguard rescued only the dolphin with the moboat during the patrol.

A coastguard rescued only the dolphin with the poising during the patrol.

The trooper destroyed only the tank with the projecduring the fight.

The trooper destroyed only the wall with the projecduring the fight.

The trooper destroyed only the tank with the camouflduring the fight.

a The criminal shot only the cop with the pistol yesterdaylate at night.

a

a

te

te

te

v

r

io

b

b

n

o

o

e

ri

r

nd

on

-

thvi-

tynal

.eir

nt

he-dy,

t.

).

ure.

ar-e as

REANALYSIS IN SEN

The criminal shot only the dog with the pistol yesterdlate at night.

The criminal shot only the dog with the collar yesterdlate at night.

The caretaker cleaned only the pail with the brush afwhile.

The caretaker cleaned only the suit with the brush afwhile.

The caretaker cleaned only the pail with the holes afwhile.

The man repaired only the toolbox with the screwdrias quickly as possible.

The man repaired only the printer with the screwdrivequickly as possible.

The man repaired only the printer with the malfunctas quickly as possible.

The man deterred only the thug with the shotgun justfore dawn.

The man deterred only the wolf with the shotgun justfore dawn.

The man deterred only the thug with the tattoos justfore dawn.

The teenager washed only the basin with the spoearly in the morning.

The teenager washed only the cheek with the spoearly in the morning.

The teenager washed only the cheek with the pimearly in the morning.

The outlaw wounded only the guard with the crossbduring the night.

The outlaw wounded only the horse with the crossbduring the night.

The outlaw wounded only the guard with the whiskduring the night.

The burglar stabbed only the guy with the dagger duthe night.

The burglar stabbed only the dog with the dagger du
the night.
in

a

).

)

at-kyp:

theicaly-

:e

Cuetos, F., & Mitchell, D. C. (1988). Cross-linguistic differ-

The burglar stabbed only the dog with the collar durthe night.

REFERENCES

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

Adams, B. C., Clifton, C., Jr., & Mitchell, D. C. (1998Lexical guidance in sentence processing? PsychonomicBulletin & Review, 5, 265–270.

Altmann, G. T. M., Garnham, A., & Dennis, Y. I. L. (1992Avoiding the garden path: Eye-movements in conteJournal of Memory and Language, 31,685–712.

Altmann, G. T. M., Garnham, A., & Henstra, J. A. (1994Effects of syntax in human sentence parsing: Eviden


y

y

r a

r a

r a

er

as

n

be-

e-

e-

ge

nge

ple

w

w

rs

ng

ing

g

n

.xt.

against a structure-based parsing mechanism. Journalof Experimental Psychology: Learning, Memory, aCognition, 20,209–216.

Altmann, G. T. M., & Steedman, M. J. (1988). Interactiwith context during human sentence processing. Cog-nition, 30,191–238.

Altmann, G. T. M., van Nice, K. Y., Garnham, A., & Henstra, J. A. (1998). Late closure in context. Journal ofMemory and Language, 38,459–484.

Baayen, H. R., Piepenbrock, R., & Van Rijn, H. (1993). TheCELEX lexical database (CDROM). Philadelphia, PA:Linguistic Data Consortium, Univ. of Pennsylvania.

Binder, K. S., & Rayner, K. (1998). Contextual strengdoes not modulate the subordinate bias effect: Edence from eye fixations and self-paced reading. Psy-chonomic Bulletin & Review, 5, 271–276.

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

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

Brysbaert, M., & Mitchell, D. C. (1996). Modifier attachmein sentence parsing: Evidence from Dutch. QuarterlyJournal of Experimental Psychology, 49A,664–695.

Clifton, C., Jr, Bock, J., & Radó J. (2000). Effects of tfocus particle only and intrinsic contrast on comprehension of reduced relative clauses. In A. KenneR. Radach, D. Heller, & J. Pynte (Eds.),Reading asa perceptual process(pp. 591–619). Oxford, UK:Elsevier.

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

Clifton, C., Jr., Kennison, S. M., & Albrecht, J. E. (1997Reading the words her, his, him: Implications for pars-ing principles based on frequency and on structJournal of Memory and Language, 36,276–292.

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

Collins, M., & Brooks, J. (1995). Prepositional phrase tachment through a backed-off model. In D. Yarows& K. Church (Eds.),Proceedings of the third workshoon very large corpora(pp. 27–38). Cambridge, MAMIT Press.

Crain, S., & Steedman, M. J. (1985). On not being led upgarden-path: The use of context by the psychologprocessor. In D. Dowty, L. Kartunnen, & A. Zwick(Eds.),Natural language parsing(pp. 320–358). Cambridge, UK: Cambridge Univ. Press.

Crocker, M. W. (1995). Computational psycholinguisticsAn interdisciplinary approach to the study of languag.Dordrecht, The Netherlands: Kluwer.

).ce

ences in parsing: Restrictions on the use of the late clo-sure strategy in Spanish. Cognition, 30,73– 105.

nen

c

l

-

v

n

cf

.

s-

eec

Mbe

Ln

ys

y

c

G.ce

-edual

anad-

-

nd

yn-

).ng

S.so-

.ndion.

nd

).al-.

n-e

g:M.

.nce

ry,

nglr

t,ing:sta-h


Cuetos, F., Mitchell, D. C., & Corley, M. M. B. (1996). Parsiin different languages. In M. Carreiras, J. Garcia-Alb& N. Sabastian-Galles (Eds.),Language processing iSpanish(pp. 145–187). Hillsdale, NJ: Erlbaum.

De Vincenzi, M., & Job, R. (1993). Some observationsthe universality of the late closure strategy: Evidenfrom Italian. Journal of Psycholinguistic Research, 22,189–206.

Duffy, S. A., Morris, R. K., & Rayner, K. (1988). Lexicaambiguity and fixation times in reading. Journal ofMemory and Language, 27,429–446.

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

Ferreira, F., & Henderson, J. (1990). Use of verb informtion in syntactic parsing: Evidence from eye moments and word-by-word self-paced reading. Journalof Experimental Psychology: Learning, Memory, aCognition, 16,555–568.

Frazier, L. (1979). On comprehending sentences: Syntaparsing strategies. Ph.D. dissertation, University oConnecticut. Indiana University Linguistics Club.

Frazier, L. (1987). Sentence processing: A tutorial reviewM. Coltheart (Ed.),Attention and performance XII(pp.559–586). Hillsdale, NJ: Erlbaum.

Frazier, L. (1995). Constraint satisfaction as a theory of tence processing. Journal of Psycholinguistic Research, 24,437–468.

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

Frazier, L., & Rayner, K. (1982). Making and correcting rors during sentence comprehension: Eye movemin the analysis of structurally ambiguous sentenCognitive Psychology, 14,178–210.

Garnsey, S. M., Pearlmutter, N. J., Myers, E., & Lotocky,A. (1997). The contributions of verb bias and plausiity to the comprehension of temporarily ambiguous stences. Journal of Memory and Language, 37,58–93.

Gilboy, E., Sopena, J.-M., Clifton, C., Jr., & Frazier,(1995). Argument structure and association preferein Spanish and English complex NPs. Cognition, 54,131–167.

Hemforth, B., Konieczny, L., & Scheepers, C. (2000). Stactic attachment and anaphor resolution: Two siderelative clause attachment. In M. W. Crocker, M.Pickering, & C. Clifton, Jr. (Eds.),Architectures andmechanisms for language processing(pp. 259–281).Cambridge, UK: Cambridge Univ. Press.

Hindle, D., & Rooth, M. (1993). Structural ambiguitin lexical relations.Computational Linguistics, 19,103–120.

Holmes, V. M., & O’Regan, J. K. (1981). Eye fixation paterns during the reading of relative-clause sentenJournal of Verbal Learning and Verbal Behavior, 20,417–430.

Just, M. A., & Carpenter, P. A. (1980). A theory of readinFrom eye fixations to comprehension. PsychologicalReview, 87,329–354.

RING, AND TRAXLER

ga,

ones

of

a-e-

d

tic

In

en-

r-ntses.

.il-n-

.ces

n- ofJ.

t-es.

g:

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

Lewis, R. (1999). Attachment without competition: A computational model of race-based parsing in a limitworking memory. Paper presented at the 12th annCUNY conference, New York, NY.

Liversedge, S. P., Pickering, M. J., Branigan, H. P., & VGompel, R. P. G. (1998). Processing arguments andjuncts in isolation and context: The case ofby-phrase am-biguities in passives.Journal of Experimental Psychology: Learning, Memory, and Cognition, 24,461–475.

MacDonald, M. C. (1993). The interaction of lexical asyntactic ambiguity. Journal of Memory and Language,32,692–715.

MacDonald, M. C. (1994). Probabilistic constraints and stactic ambiguity resolution. Language and CognitiveProcesses, 9, 157–201.

MacDonald, M. C., Just, M. A., & Carpenter, P. A. (1992Working memory constraints on the processiof syntactic ambiguity.Cognitive Psychology, 24,56– 98.

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

McRae, K., Spivey-Knowlton, M. J., & Tanenhaus, M. K(1998). Modeling the influence of thematic fit (aother constraints) in on-line sentence comprehensJournal of Memory and Language, 38,283–312.

McRoy, S. W., & Hirst, G. (1990). Race-based parsing asyntactic disambiguation. Cognitive Psychology, 14,313–353.

Merlo, P., Crocker, M. W., & Berthouzoz, C. (1997Attaching multiple prepositional phrases: Generized backed-off estimation. In C. Cardie & RWeischedel (Eds.),Proceedings of the Second Coference on Empirical Methods in Natural LanguagProcessing. (p. 149–155). Providence, RI.

Mitchell, D. C. (1987). Lexical guidance in human parsinLocus and processing characteristics. In Coltheart,(Ed.), Attention and performance XII(pp. 601–618).Hillsdale, NJ: Erlbaum.

Mitchell, D. C., Corley, M. M. B., & Garnham, A. (1992)Effects of context in human sentence parsing: Evideagainst a discourse-based proposal mechanism. Jour-nal of Experimental Psychology: Learning, Memoand Cognition, 18,69–88.

Mitchell, D. C., & Cuetos, F. (1991). The origins of parsistrategies. In C. Smith (Ed.),Current issues in naturalanguage processing(pp. 1–12). Austin: Center foCognitive Science, Univ. of Texas.

Mitchell, D. C., Cuetos, F., Corley, M. M. B., & BrysbaerM. (1995). Exposure-based models of human parsEvidence for the use of coarse-grained (nonlexical) tistical records. Journal of Psycholinguistic Researc,24,469–488.

Mitchell, D. C., & Green, D. W. (1978). The effect of con-text and content on immediate processing in reading.

n

9s

a

-

e

r

)

ce bl

d

s

f-ay

t

o

n

-

R

.n

d

-

i-f

-

gofn

f

m-

e,

h-ss-

n:.

e-

-

,n-al

g

c-

e-

REANALYSIS IN SEN

Quarterly Journal of Experimental Psychology, 30,609–636.

Ni, W., Crain, S., & Shankweiler, D. (1996). Sidesteppigarden paths: Assessing the contributions of syntax,mantics and plausibility in resolving ambiguities.Lan-guage and Cognitive Processes, 11,283–334.

Paterson, K. B., Liversedge, S. P., & Underwood, G. (199The influence of focus operators on syntactic proceing of “short” relative clause sentences.QuarterlyJournal of Experimental Psychology, 52A, 717–737.

Pearlmutter, N. J., & MacDonald, M. C. (1995). Individudifferences and probabilistic constraints in syntacambiguity resolution. Journal of Memory and Language, 34,521–542.

Pearlmutter, N. J., & Mendelsohn, A. A. (1999). Serial vs.parallel sentence comprehension. Unpublished manu-script.

Pickering, M. J., & Traxler, M. J. (1998). Plausibility and rcovery from garden paths: An eye-tracking study. Jour-nal of Experimental Psychology: Learning, Memoand Cognition, 24,940–961.

Pickering, M. J., Traxler, M. J., & Crocker, M. W. (2000Ambiguity resolution in sentence processing: Evidenagainst likelihood. Journal of Memory and Language,43,447–475.

Pritchett, B. (1992). Grammatical competence and parsinperformance. Chicago: Univ. of Chicago Press.

Rayner, K., Carlson, M., & Frazier, L. (1983). The interation of syntax and semantics during sentence procing: Eye movements in the analysis of semanticallyased sentences. Journal of Verbal Learning and VerbaBehavior, 22,358–374.

Rayner, K., & Duffy, S. A. (1986). Lexical complexity anfixation times in reading: Effects of word frequencverb complexity, and lexical ambiguity. Memory &Cognition, 14,191–201.

Rayner, K., Garrod, S. C., & Perfetti, C. A. (1992). Dicourse influences in parsing are delayed. Cognition, 45,109–139.

Rayner, K., Kambe, G. A., & Duffy, S. A. (2000). The efect of clause wrap-up on eye movements during reing. Quarterly Journal of Experimental Psycholog53A, 1061–1080.

Rayner, K., & Morris, R. K. (1991). Comprehensioprocesses in reading ambiguous sentences: Reflecfrom eye movements. In G.B. Simpson (Ed.),Under-standing word and sentence(pp. 175–198). Amster-dam: Elsevier.

Rayner, K., Pacht, J. M., & Duffy, S. A. (1994). Effects of priencounter and global discourse bias on the processinlexically ambiguous words: Evidence from eye fixatioJournal of Memory and Language, 33,527–544.

Rayner, K., & Pollatsek, A. (1989). The psychology of reading. Englewood Cliffs, NJ: Prentice-Hall.

Rayner, K., Sereno, S. C., Morris, R. K., Schmauder, A.
& Clifton, C., Jr. (1989). Eye movements and on-linlanguage comprehension processes. Language andCognitive Processes, 4, SI 21–49.

gse-

).s-

ltic

-

y,

.ce

g

-ss-i-

y,

-

d-,

nions

rg ofs.

.,

Reichle, E. D., Pollatsek, A., Fisher, D. L., & Rayner, K(1998). Toward a model of eye movement control ireading. Psychological Review, 105,125–157.

Schütze, C. T., & Gibson, E. (1999). Argumenthood anEnglish prepositional phrase attachment. Journal ofMemory and Language, 40,409–431.

Speer, S. R., & Clifton, C., Jr. (1998). Plausibility and argument structure in sentence comprehension. Memory &Cognition, 26,965–978.

Spivey, M. J., & Tanenhaus, M. K. (1998). Syntactic ambguity resolution in discourse: Modeling the effects oreferential context and lexical frequency. Journal ofExperimental Psychology: Learning, Memory, andCognition, 24,1521–1543.

Spivey-Knowlton, M., & Sedivy, J. (1995). Resolving attachment ambiguities with multiple constraints. Cogni-tion, 55,227–267.

Tabor, W., Juliano, C., & Tanenhaus, M. K. (1997). Parsinin a dynamical system: An attractor-based account the interaction of lexical and structural constraints isentence processing. Language and CognitiveProcesses, 12,211–271.

Tabor, W., & Tanenhaus, M. K. (1999). Dynamical models osentence processing. Cognitive Science, 23,491–515.

Tanenhaus, M. K., & Trueswell, J. C. (1995). Sentence coprehension. In J. Miller & P. Eimas (Eds.),The hand-book of perception and cognition: Speech, languagand communication(Vol. 11, pp. 217–262). San DiegoCA: Academic Press.

Taraban, R., & McClelland, J. R. (1988). Constituent attacment and thematic role assignment in sentence proceing: Influence of content-based expectations. Journalof Memory and Language, 27,597–632.

Traxler, M. J., Bybee, M. D., & Pickering, M. J. (1997). In-fluence of connectives on language comprehensioEye-tracking evidence for incremental interpretationQuarterly Journal of Experimental Psychology, 50A,481–497.

Traxler, M. J., & Pickering, M. J. (1996). Plausibility andthe processing of unbounded dependencies: An eytracking study. Journal of Memory and Language, 35,454–475.

Traxler, M. J., Pickering, M. J., & Clifton, C., Jr. (1998). Ad-junct attachment is not a form of lexical ambiguity resolution. Journal of Memory and Language, 39,558–592.

Traxler, M. J., Pickering, M. J., Clifton, C., Jr., & Van GompelR. P. G. (2000). Architectures and mechanisms for setence processing: Is syntactic parsing a form of lexicambiguity resolution? In M. De Vincenzi & V. Lombardo(Eds.), Cross-linguistic studies of sentence processin(pp. 149–174). Dordrecht, The Netherlands: Kluwer.

Trueswell, J. (1996). The role of lexical frequency in syntatic ambiguity resolution. Journal of Memory and Lan-guage, 35,566–585.

Trueswell, J., Tanenhaus, M. K., & Garnsey, S. (1994). S
e mantic influences on parsing: Use of thematic role in-
formation in syntactic disambiguation. Journal ofMemory and Language, 33,285–318.


Trueswell, J., Tanenhaus, M. K., & Kello, C. (1993). Verb-specific constraints in sentence processing: Separating

r

ei

Adams, Clifton, and Mitchell (1998). PsychonomicBulletin & Review.

.c.

effects of lexical preference from garden-paths. Jour-nal of Experimental Psychology: Learning, Memoand Cognition, 19,528–553.

Tyler, 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.

Van Gompel, R. P. G., & Pickering, M. J. (in press). Lexi-cal guidance in sentence processing: A note

y,

ng.

on

Van Gompel, R. P. G., Pickering, M. J., & Traxler, M. J(2000). Unrestricted race: A new model of syntactiambiguity resolution. In A. Kennedy, R. Radach, DHeller, & J. Pynte (Eds.),Reading as a perceptualprocess(pp. 621–648). Oxford, UK: Elsevier.

(Received December 23, 1998)(Revision received October 16, 2000)Published Online June 21, 2001

Reanalysis in Sentence Processing: Evidence against Current Constraint-Based and Two-Stage Models

Documents

Transcript of Reanalysis in Sentence Processing: Evidence against Current Constraint-Based and Two-Stage Models