The Preservation of Structure in Language Comprehension: Is Reanalysis the Last Resort?

Journal of Memory and Language 45,283–307 (2001)doi:10.1006/jmla.2001.2776, available online at http://www.academicpress.com on

The Preservation of Structure in Language Comprehension:Is Reanalysis the Last Resort?

Patrick Sturt

Human Communication Research Centre, Department of Psychology, University of Glasgow,Glasgow, Scotland, United Kingdom

Martin J. Pickering

Department of Psychology, University of Edinburgh, Edinburgh, Scotland, United Kingdom

and

Christoph Scheepers and Matthew W. Crocker

Department of Computational Linguistics, Universität des Saarlandes, Germany

During language comprehension, people sometimes have to revise their grammatical analysis of a sentence.Experimental evidence demonstrates that such reanalysis often causes processing difficulty. We might therefore

a choice.rocessor wheneverwhich theeriments

M ounts to

at tKeitcomsupemyPossenPro

ASturof Pland

predict that reanalysis would be disfavored, with the processor preferring not to reanalyze when it had Three experiments on complement-clause ambiguities investigate the conditions under which the pchooses to reanalyze. We contrast two extreme positions, one where the processor avoids reanalysispossible, the other where reanalysis is not disfavored at all. We also consider intermediate positions, in preference to maintain the current analysis is one factor that affects ambiguity resolution. The expdemonstrate that the processor avoids reanalysis even when other factors would support it.© 2001 Academic Press

Key Words:sentence processing; ambiguity resolution; reanalysis; recency.

ost models of human language comprehen-ranks the analyses, then reanalysis am

eeb

changing the ranking.sise-ra-en-ingc-

iffi--s ifisr

sion assume that the processor incorporwords into a grammatical analysis as soonthey are encountered. Sometimes the anainitially chosen by the human language procsor has to be revised because it is incompawith subsequent input. The process of revisthe current analysis and building a new onknown as reanalysis. If the processor instconsiders more than one analysis in parallel,

We thank the members of The Sentence Processing G

28

inherer-lso

pstse

.e-te

he Universities of Glasgow and Edinburgh, as well h Rayner and two anonymous reviewers, for helpfments on earlier drafts of this paper. This research wported by ESRC Grant No. R000222286, British Aca Research Grant No. SG-29789, and a British Acadetdoctoral Fellowship. Portions of the work were prted at the annual CUNY conference on Human Sentecessing in 1999 and in 2000.ddress correspondence and reprint requests to Patt, Human Communication Research Centre, Departmsychology, 58 Hillhead Street, Glascow G12 8QB, Sc, UK. E-mail: [email protected].

0749-596X/01 $35.003

assiss-bleng isadut

There is general agreement that reanalyshould cause processing difficulty, simply bcause it involves additional processing opetions that are not needed otherwise. Experimtal research often uses the finding of processdifficulty as evidence that reanalysis has ocurred. Thus reanalysis is, in some sense, dcult. An implication of this is that one might expect the processor to prefer to avoid reanalysipossible. In order to investigate whether thisthe case, we need to conduct experimentswhich the processor has a choice about whetto reanalyze or not. Below we report three expiments that address this issue. The issue is adiscussed in detail by Schneider and Philli(2001), who report two very similar experimenwhich were conducted independently of thstudies reported here, with compatible results

We first contrast two extreme positions on ranalysis, and compare them with intermedia

oupasulas

d-mye-nce

rickentot-

Copyright © 2001 by Academic PressAll rights of reproduction in any form reserved.

T E

sshyn

eerffecausffeonesheguatre

n, bth

ibsovealysgees

ior(1),

ll

a

ad-theNPandd

-longaly-

ra-ter,r,r-&ed-

y

284 STUR

positions. At one extreme, the Revision-as-LaResort principle proposes that the procesavoids reanalysis whenever possible. In otwords, it changes the currently preferred stactic analysis only when that analysis becomstrictly incompatible with the input. At the othextreme, avoidance of reanalysis does not athe processor’s choice of analysis at all. Wethis the Revision-Irrelevant hypothesis, becait proposes that the processor behaves no diently in its choice of analysis whether revisiis involved or not. In other words, the principlthat determine which analysis is adopted wreanalysis is unnecessary (as in most ambities discussed in the literature) still operwhen at least one alternative does involveanalysis.

We then discuss intermediate positiowhere there is a preference not to reanalyzewhere this preference can be modulated by ofactors. Such factors could include the plausity of the current analysis, with the procespreferring to avoid an implausible analysis eif reanalysis is then necessary. They might include a preference for adopting an analcompatible with the more frequent subcaterization of a verb, even if that analysis involvrevision. Another possibility is that the proc

nc

lv tr en

reen

ent, t-

e

atn-m-9). inen-oneviaerialolicngby

arl-y-ll,

heavy phrases at particular points in a sente

WHAT IS REANALYSIS?

We assume that sentence processing invothe computation of dependencies betweenwords and phrases that are encountered. Foample, in the sentence The troops found thenemy spy, the relations include informatiothat the troopsis the subject of found. Often,words are incorporated directly into the repsentation without breaking existing dependcies. For example, when the main verb foundisencountered, the processor forms the depency between the troopsand found, and does noneed to break any other dependencies (e.g.between theand troops). We will call such nondestructive parsing behavior simple integration.

However, existing dependencies sometim

T AL.

t-orer-s

ctller-

ni-

e-

s,uter

il-rn

soiso-d-

e.

esheex-

--

d-

hat

s

use the term reanalysis(or, equivalently, “revi-sion”) we refer to destructive parsing behavin which a dependency is broken. Consider for example:

The troops found the enemy spy had used up athe supplies.

The verb foundin (1) is ambiguous betweennoun phrase direct object reading (found [NP theenemy spy]) and a sentential complement reing (found [S the enemy spy had used up all supplies]). We call these two readings the analysis and the S analysis, respectively,verbs which allow this ambiguity will be calleNP/S verbs.

Assume that the enemy spyis initially inter-preted as the object of found. A great deal of evidence supports this assumption, at least so as the object analysis is the most frequent ansis for the particular NP/S verb involved (Fzier & Rayner, 1982; Garnsey, PearlmutMyers, & Lotocky, 1997; Rayner & Frazie1987; Pickering & Traxler, 1998; Sturt, Pickeing, & Crocker, 1999; Trueswell, Tanenhaus,Kello, 1993). But this decision must be revokwhen the word had is encountered; the grammatical dependency between found and theenemy spymust be broken, and the enemy spmust be reinterpreted as the subject of had. Ex-perimental evidence demonstrates difficultythis point (in comparison with appropriate cotrols), even though such ambiguities are coparatively easy to revise (Sturt et al., 199Note that, although we discuss reanalysisterms of serial models, any viable model of stence processing must involve reanalysis of kind or another, whether it is implemented backtracking search strategies, as in many smodels, by reordering analyses, as in symbparallel models (Gibson, 1991), or by shiftiactivation patterns, as in models inspired connectionist architectures (MacDonald, Pemutter, & Seidenberg, 1994; McRae, SpiveKnowlton, & Tanenhaus, 1998; TruesweTanenhaus, & Garnsey, 1994).

The Revision-as-Last-Resort Hypothesis

(1)

t-e

In an influential early paper, Fodor and Frazier(1980) proposed theRevision-as-Last-Resort

H

is

s

sr

res c inainn

o (

d

in

:

aS

anis

olev

thatly-a

siib

arl-aticllyon,le,eo-usi-r to

tsncetoesgiveis3;

n-, itn-ngoesto

heis

ealy-notoin-urutor5;-re-

orzey

, or

y-thisiesre-of

IS REANALYSIS T

(RALR) principle, which claimed that reanalysis avoided if possible:

…the partial phrase marker that hasbeen constructed on the basis of previouwords in the sentence is not to be changedin response to subsequent words unlesthere is no other way of proceeding. (Fodo& Frazier, 1980, p.427)

RALR can be seen to follow from a mogeneral principle known as Minimal RevisionThe processor makes as few changes to therent representation (Frazier, 1990; FrazierClifton, 1998) as possible, because this mmizes processing cost. RALR is a special cof Minimal Revisions because reanalysis volves changes to the analysis, but simple igration does not.

To illustrate RALR, consider the effect adding a relative pronoun to the sentenceabove:

The troops who found the enemy spy had useup all the supplies. . .

As before, we assume that the processortially adopts the NP analysis, withthe enemyspy as the direct object offound. When theprocessor encountershad, it has two choiceseither the troops or the enemy spycan beattached as the subject ofhad. The formerinvolves no reanalysis, asfound retains the NPanalysis; but the latter involves reanalysis,the NP analysis offound is replaced by theanalysis. Therefore, RALR predicts thathadshould be preferentially attached withthetroops as its subject. See also SchneiderPhillips (2001) for further discussion of thambiguity in relation to RALR.

Fodor and Frazier’s (1980) definition RALR states that reanalysis is avoided “unthere is no other way of proceeding.” The obous interpretation is to assume that “no oway of proceeding” means “no other grammcal way of proceeding.” In other words, reanasis will only occur if simple integration is impossible. However, this implies that grammatical but wholly implausible analywould never be abandoned. This is incompat

(2)

with most theories, including later develop

E LAST RESORT? 285

:ur-&i-se-

te-

f1)

i-

s

d

fssi-eri--

sle

ments of Frazier’s account. Thus, Rayner, Cson, and Frazier (1983) proposed that a themprocessor checks the plausibility of the initiaadopted analysis (see also Ferreira & Clift1986). If the analysis is sufficiently implausibthen reanalysis ensues. Within alternative thries such as constraint-based accounts, plability can be enough to cause the processochange its ranking of analyses.

A great deal of experimental work suggesthat plausibility can cause reanalysis, and hethat RALR as interpreted above is unlikelybe correct. In particular, a number of studisuggest that the processor can be forced toup its currently favored analysis if this analysbecomes implausible (Rayner et al., 198Tabossi, Spivey-Knowlton, McRae, & Tanehaus, 1994; Trueswell et al., 1994). Howeveris often very difficult to demonstrate that the icreased processing difficulty accompanyisemantic anomaly in studies such as these dactually cause the currently favored analysisbe given up, rather than simply causing tprocessor to weaken its commitment to thanalysis (see Pickering & Traxler, 1998).

A weaker interpretation of RALR allows thprocessor to reanalyze when the current ansis becomes sufficiently implausible. Thus, “other way of proceeding” could be takenmean no way of proceeding with a plausibleterpretation. If so, reanalysis does not occwhen the initial analysis remains plausible, bimplausibility (or perhaps other semanticpragmatic factors (cf. De Vincenzi & Job, 199Kamide & Mitchell, 1997)) may trigger reanalysis. However, the processor would notanalyze as a result of structural preferencesfrequency. For instance, it would not reanalywhen the current analysis involved unwieldconstituents, infrequent argument structuresviolations of recency preferences.

Revision-Irrelevant Hypothesis

The Revision-Irrelevant (henceforth RI) hpothesis represents the other extreme. Onhypothesis, the processor orders its prioritaccording to general parsing preferences,gardless of whether or not the satisfaction

fos

e

ryr

twnwmyl

n

bth

o

r

ecoiten

feu

nn

nhl

itthSof

be-

e isdes, theewting notllyl, ainms ofych-i-iva-de nore-lowures or

eic-8).

istickedhendch-sedc-dss-

reenre-isa

se-quent attachment ofhad is dominated by con-

1

286 STURT

preference which is of particular relevance(2) is recency. In this sentence, the reanalyoption involves choosing a recent attachmsite (the enemy spy) over a less recent one (thetroops). It has always been recognized that acency preference plays an important role in stactic disambiguation (Kimball, 1973; Frazie1978; Stevenson, 1994), at least when thepotential attachment sites are in differeclauses (different results can hold when the tpotential attachment sites form part of a coplex noun phrase, and are not separated bclause boundary (see Cuetos & Mitche1988)). For example, in an eye-movemestudy, Altmann, van Nice, Garnham, and Hestra (1998) contrasted sentences such as thelowing, where there is a clause boundarytween the two potential attachment sites ofadverbyesterdayor tomorrow.

a. He’ll brush the dog he washed yesterday tmake its fur shine again.

b. He’ll brush the dog he washed tomorrow tomake its fur shine again.

Participants had less difficulty readingyester-day (3a), where it can only attach to the morecent verbwashed, than readingtomorrow in(3b), where it can only attach to the less recverb brush. Moreover, this recency preferenproved resistant to all but the most extreme ctextual manipulations. Using another ambiguwhich also involved a clause boundary betwethe two possible attachment sites, Phillips aGibson (1997) showed that the recency preence can also outweigh other preferences, sas referential felicity (Altmann & Steedma1988), the argument-over-adjunct prefere(Liversedge, Pickering, Branigan, & van Gompel, 1998; Schu;autze & Gibson, 1999), aMinimal Attachment (Frazier, 1978). Hence tRI hypothesis should strongly support reanasis in (2).

A recent proposal which is compatible wthe RI hypothesis, at least with respect to ambiguity, was made by Stevenson (1998). proposed a competitive activation model parsing in which there is no clear distinction

(3)

tween simple integration and some types of ranalysis, including the NP/S reanalysis that w

ET AL.

ris

nt

e-n-,oto-a

l,nt-

fol-e-e

e

nten-y,ndr-ch

,ce-dey-

hishe

are considering in this paper. The parse trerepresented by a network of processing noand each new input phrase is attached intonetwork through a process in which the nphrase takes over activation from some exispart of the network. This happens whether orthe attachment of that word would traditionabe seen as involving reanalysis. In the moderecency preference follows from the way which the activation of nodes decays. In terof the processing of (2), the low attachmenthadwould involve activation being shifted awafrom an existing node representing the attament of the enemy spyas an object of the prevous verb. In the case of high attachment, acttion would be shifted away from the norepresenting the attachment of the phrasethetroopsas the root of the tree. Since there isdistinction between simple integration and analysis, recency leads to a preference for attachment, though the network architectalso allows for other factors, such as verb biacontext, to play a part.

Another model which is compatible with thRI hypothesis is the Syntactic Locality Predtion Theory (henceforth SPLT; Gibson, 199SPLT is based on a parsing algorithm whichable to maintain a number of ranked syntacanalyses in parallel. The structures are ranaccording to memory cost (for example, tcost of maintaining a syntactic prediction) aintegration cost (for example, the cost of attaing a phrase to a head which was procesmuch earlier in the input), as well as other fators. The re-ranking of structures corresponto reanalysis in serial models. In (2), both posible attachments ofthe enemy spyare made, soboth the NP analysis [found [NP the enemyspy]], and the S analysis [found [S [NP theenemy spy]]] are maintained in memory. Theis very little difference in memory cost betwethese two analyses at this point; the cost of pdicting the embedded verb in this exampleminimal,1 and there is no cost for predictingmatrix verb. Therefore, the choice of the sub

e-e

Incidentally, this is why SPLT can predict the relativeease of NP/S reanalysis (Sturt et al., 1999).

H

nraosoa

wie

shekeuethfateotiae

th ace

frul8upele

99O inmn

itiw

itiogu

ostnd

d

d

-

ad-ng(4),

e-d byethaten

ly-i-

sing4).forthe

intas-ingrbsbefor-s-thatNP82;Al-er-efer-re-

itialrbs,NP

the

IS REANALYSIS T

siderations of integration cost. As integratioto recent material are less costly than integtions to distant material, Gibson predicts lowtachment in (2). More generally, memory cin Gibson’s framework corresponds to the ntion of embedding complexity. The lack ofdifference in memory cost between the tattachments in (2) suggests that embeddcomplexity is not a factor which affects thattachment preference in this ambiguity.

Avoidance of Reanalysis as a Preference

An intermediate position is that the proceshas a preference to avoid reanalysis, but wthis preference is simply one factor that is tainto account in deciding which analysis to psue. On this account there is a preferencmaintain the currently favored analysis, but is modulated by other factors. One possible tor is plausibility (either in relative or absoluterms). We have suggested above that pealmost certainly do reanalyze when the inianalysis is sufficiently implausible. This papdoes not consider the role of plausibility in decision of whether to maintain or discardanalysis, but focuses on lexical and syntafactors, in particular, verb bias and constituweight.

Verb bias is often operationalized as the quency with which a verb is used on a particanalysis (e.g., Ford, Bresnan, & Kaplan, 19Garnsey et al., 1997; Holmes, Stowe, & Cples, 1989; Mitchell & Holmes, 1985; Trueswet al., 1993). Many studies have shown that vbiases affect processing (Garnsey et al., 1Holmes et al., 1989; Trueswell et al., 1993). some accounts, verb bias is ignored duringtial analysis (Frazier, 1987) but can be eployed during reanalysis (Mitchell, 1989). Oother accounts, verb bias affects the inchoice of analysis (Garnsey et al., 1997), as as the decision about whether to reanalyzere-rank analyses).

One possibility is that, where an ambiguinvolves a choice between simple integratand reanalysis, people are affected by the deof verb bias in deciding whether to contin

oue

E LAST RESORT? 287

sa-t-t-

ong

orren

r- toisc-

ple lren

ticnt

e-ar2;-

lrb7;ni--

alell(or

ynreeeze

adopt the NP analysis if this analysis is the mfrequent). Consider (2), repeated below, acompare it with (4).

The troops who found the enemy spy had useup all the supplies. . .

The troops who discovered the enemy spy haused up all the supplies. . .

The verb found in (2) is strongly biased toward the NP reading, while the verb discoveredin (4) is only weakly biased toward the NP reing (we describe the criteria for determinibias below). The processor might reanalyze because the bias of discoveredis weak enoughto make this option attractive, but not (2), bcause the NP analysis is so strongly preferrethe verb found. A simple claim would assumthat there is a threshold for reanalysis, and the threshold for this construction falls betwethe biases for foundand discovered. An alterna-tive, probabilistic, account predicts that reanasis is more likely to occur with more weakly based verbs. Both versions predict procesdifferences between sentences like (2) and (

It is also possible that there is a thresholdreanalysis, but that this threshold falls belowweakly biased verbdiscovered. This wouldmean that the threshold occurred at the powhen verbs were not NP biased at all. Evensuming a probabilistic mechanism, processdifferences within the class of NP-biased vemight also be less marked than they wouldbetween NP-biased and S-biased verbs. Untunately, it is very difficult to test for these posibilities, because many theories assumethe processor does not strongly activate theanalysis with S-biased verbs (Ford et al., 19Garnsey et al., 1997; Trueswell et al., 1993).though this conclusion is controversial (Picking, Traxler, & Crocker, 2000), it would not bsafe to conclude that a low attachment preence for S-biased verbs was actually due toanalysis in such cases rather than to an inadoption of the S analysis. For NP-biased vehowever, all current theories agree that theanalysis is initially favored.

Structural preferences may also override preference to avoid reanalysis. Fodor and In

(2)

(4)

lly(2000) suggest that the recency preference may

t s be

s

e

na

efuo sa b a4o9).in

d

eace th

yto

si-sre-ithn ife-ctsforle,n-

ntsreported below were designed to distinguish

vencet of beare byral, the isoted9)esultyc-en

xi--

,ngerelt,

2)at-

288 STURT

override RALR. If so, we would expect a low atachment preference in sentences such asFodor and Inoue (2000) argued that readersperience intuitive difficulty with (5), which isimilar in structure to (2), and which can onlysuccessfully parsed as a complete sentencthe high attachment reading:

The lawyer who believed the witness was bi-ased against repeat offenders.

If this intuition is correct, it can only be becaureaders initially attach low. However,believedisbiased toward the sentence-complement ring. Hence, many accounts assume that the wit-nessis not initially attached as an object in acase (Garnsey et al., 1997; Trueswell et 1993).

Another factor that could interact with a prerence against reanalysis relates to constitweight. Cross-linguistic corpus studies shsystematic patterns in the relative orderingsconstituents of different lengths (Hawkin1990, 1995). In English, heavy constituents infrequent before a head, for example, at theginning of a sentence. Main clause subjectsparticularly unlikely to be heavy (Chafe, 199There is also experimental evidence that cstituent length affects parsing (Hirose, 19Thornton, MacDonald, & Arnold, 2000Length might therefore influence the processof (2), repeated below:

The troops who found the enemy spy had useup all the supplies. . .

Low attachment (involving reanalysis) of hadresults in a matrix clause subject which is tremely long and unwieldy, whereas high tachment of hadresults in a main clause subjewhich is considerably shorter. Thus, the procsor may avoid reanalysis in order to avoidheavy main clause subject. We return to issue of constituent weight in Experimentbelow.

Summary

In this section, we have outlined three hpotheses about how the processor chooses

(5)

(2)

analyze. According to RALR, the processodoes not reanalyze unless forced to do so

ET AL.

-(2).ex-

e on

e

grammatical constraints, or possibly by plaubility. According to Revision-Irrelevant, there ino preference to avoid reanalysis per se. Thefore, the processor acts in accordance wstructural preferences, such as recency, evethis means reanalyzing. Finally, avoiding ranalysis may constitute one factor that affeparsing decisions, along with a preference adopting analyses that involve, for exampmore frequent subcategorizations and costituents of appropriate weight. The experime

or

ad-

yl.,

-entwof,ree-re

).n-9;

g

x-t-ts-ae

3

- re-r

among these hypotheses.

EXPERIMENTS

We now report three experiments that involsentences like (2) above. We employ this sentetype because it provides a particularly good tesRALR. In such sentences, reanalysis shouldextremely attractive to the processor. There two reasons for this. First, reanalysis is favoredrecency, which is a particularly strong structupreference, as we have seen above. Secondtype of reanalysis involved in NP/S ambiguitiesparticularly easy, as many researchers have n(Pritchett, 1992; Gorrell, 1995). Sturt et al. (199found that the reanalysis of NP/S ambiguitisuch as (6) causes much less processing difficthan other ambiguities like (7) below, when fators such as plausibility and verb bias had becarefully equated:

The Australian woman saw the famousdoctor had been drinking quite a lot.

Before the woman visited the famousdoctor had been drinking quite a lot.

In summary, testing sentences like (2) mamizes our chances of falsifying RALR. Conversely, if support is found for RALR in (2)then we can be fairly confident that the findishould generalize to other sentence types, whthe reanalysis involved may be more difficuand the attachment preference less strong.

Experiment 1

Experiment 1 employed sentences like (above, but with reflexive pronouns disambiguing the second verb toward either low (recent)

(6)

(7)

byhigh (nonrecent) attachment. We also manipu-

H

d

t

t.

-n

)

hn

rth

y

anch-fer-bi-ard

s.

-re

h

eny

),

,

aster-aln

n-or

-serer

ereore

inrb

n6,i-t-s7;-d-

ber of embedded sentential complement comple-

2 One of the strongly biased verbs,recognized, had nosentences classed as S in the random sample and was thusgiven a bias of 1. However, this is not to say that this verbcannot occur in the S reading. When we looked at the full setof sentences returned from the search, as opposed to thesmaller random sample, a number of examples of S sen-tences were found for this verb. The NP-bias calculated from

IS REANALYSIS T

lated whether the first verb was strongly weakly biased to the NP analysis:

a. weak/high

The troops who discovered the enemy spy hadshot themselves and were later mentioned inthe press report.

b. weak/low

The troops who discovered the enemy spy hashot himself were later mentioned in the pressreport.

c. strong/high

The troops who found the enemy spy had shothemselves and were later mentioned in thepress report.

d. strong/low

The troops who found the enemy spy had shohimself were later mentioned in the press report

RALR predicts thathad is attached high, because this does not require reanalysis. Hethere will be no difficulty in (8a) and (8c), inwhich the pronounthemselvesis consistent withthis analysis, but there will be difficulty in (8band (8d), in which the pronounhimselfis consis-tent with low attachment instead. RI makes topposite predictions, as the recency preferepredicts thathadwill be attached low, which isconsistent with the continuation in (8b) an(8d), but not the continuation in (8a) and (8c).

If revision is dispreferred, but not a last resothen a high preference may occur when the vestrongly biased toward high attachment, but preference would be reversed, or at least weened, when the bias is weak. We only emploNP-biased verbs so that there would be no conversy that the initial NP analysis is made atherefore that reanalysis is a possibility. We mnipulated whether this bias was weak or strongthe difference in bias is sufficient, we predict interaction between degree of bias and attament, such that there should be more of a preence toward high attachment in the strongly ased conditions, and more of a preference towlow attachment in the weakly biased condition

Method

Participants. Thirty-two participants from the

(8)

University of Glasgow were paid to take par

E LAST RESORT? 289

or

ce

ece

d

rt,b isis

ak-edtro-nda-. If

The data for one further participant were discarded because of a low comprehension sco(below 70% correct, combining scores for botfillers and experimental items).

Items. The experiment included 24 items like(8) (see the Appendix). The items were choswith the aid of a corpus study and a plausibilitpretest, which we describe below.

Corpus study. We determined the NP-bias ofthe experimental verbs, as in Sturt et al. (1999using the 100-million-word British National Cor-pus. Using the Corset tools (Corley, CorleyKeller, Crocker, & Trewin, in press) we obtainedrandom samples of 120 sentences containing ptense tokens of a large number of verbs, aftwhich we excluded sentences in which the complements of the verb appeared in non-canonicpositions. The remaining sentences were theclassified as NP if the verb appeared with a nouphrase direct object and no other complement,S if the verb appeared with a tensed clausal complement without a complementizer. The NP biafor each verb was then calculated as the numbof sentences classed as NP divided by the numbof sentences classed as either NP or S. Verbs wclassed as NP-biased if the NP bias was 0.65greater, as weakly biased if they were in the rang0.65–0.85, and as strongly biased if they werethe range 0.85–0.99. The NP biases for each veare shown in the items in the Appendix.2 Acrossthe items selected for the experiment, the meabias for the strongly biased conditions was 0.9and the mean bias for the weakly biased condtions was 0.73. Note that our method of calculaing verb bias differs somewhat from the methodemployed in other studies (Garnsey et al., 199Trueswell et al., 1993). In an analysis of completion study data Garnsey et al. (1997) comparethe number of NP completions against the num

t.this full sample was 0.98.

an

as

theeotsd

tc

eia

o

t

t

sctos

thef-P

sen-

es

lts

f-nt-

-ge).-totople

e-ble

/

un rel-g-se. ofro--

ed

teds

n-nd ined.st--

cluding fillers, were randomized separately for

3 Essentially identical results were obtained with theWilcoxon signed ranks test for all rating studies in this

290 STURT

tions, including tokens with complementizerswell as tokens without complementizers. In cotrast, we compare NP tokens against sentencomplements including only tokens without complementizers. We calculate the bias in this wbecause, unlike Garnsey et al., who are interein immediate effects of verb argument structupreference, we are dealing with the resolutionan ambiguity several words downstream ofverb, where the potentially useful information rgarding complementizer presence or absencalready available. With the broader definitionS-tokens employed by Garnsey et al. (1997),mean biases were 0.78 for the strongly biaconditions and 0.41 for the weakly biased contions.

Plausibility pretest. The plausibility preteswas conducted to test for plausibility balanbetween the conditions. Four sentences wconstructed for each item. Two of these stences compared the plausibility of the initNP (mis)analysis for the strongly and weakbiased verbs. The remaining two sentences cpared the plausibility of the high and low attacments. For the example item given above,four sentences were as follows:

NP analysis: strong bias

The troops found the enemy spy.

NP analysis: weak bias

The troops discovered the enemy spy.

High attachment

The troops had shot themselves.

Low attachment

The enemy spy had shot himself.

Sentences for 39 candidate items were prinin booklets, randomly interspersed amongstfillers, which were of varying plausibility. Fif-teen participants from the University of Glagow were paid to rate the plausibility of easentence on a scale from 1 (least plausible)(most plausible). From the results of this prete24 items were selected, which combinedhighest overall plausibility with the smallest diference in mean plausibility between the two Nanalysis sentences and the two attachment

ET AL.

s-

tial-y

tedreofe-is

fheedi-

eeren-l

lym-

h-he

ed20

-h

7t,

tical difference in plausibility between thestrong and weak bias misanalysis sentenc(strong5 6.68; weak5 6.58;t1(14)5 1.47,p ..16; t2(23) 5 1.09, p . .28).3 However, theitems did differ in the plausibility of the differ-ent attachments (high5 6.39; low 5 5.79;t1(14) 5 3.25, p , .006; t2(23) 5 3.17, p ,.005). We discuss this in relation to the resubelow.

Procedure. The experiment used the selpaced reading technique, with a segmeby-segment noncumulative moving windowdisplay (Just, Carpenter, & Wooley, 1982), employing the PsyScope experimental packa(Cohen, MacWhinney, Flatt, & Provost, 1993The experiment was run on a Macintosh`compatible computer, which was connecteda button box. Each sentence was divided inseven segments. The segmentation of a samitem is given below, with single slashes corrsponding to segment boundaries and a douslash where there is also a line break.

The troops / who found / the enemy spy / hadshot / themselves // and were later mentioned in the press report.

The first segment consisted of the initial nophrase. The second segment consisted of theative pronoun and the NP/S verb. The third sement consisted of the ambiguous noun phraThe fourth segment consisted of all the wordsthe first verb phrase except the reflexive pnoun, which constituted the critical fifth segment. The sixth, or spillover segment consistedof the first part of the final verb phrase, precedby a conjunction (andor but) in the high attach-ment conditions. The seventh segment consisof the end of the final verb phrase. The itemwere divided into four lists, so that each list icluded exactly one condition of each item, athere were equal numbers of each conditioneach list. Each experimental list was combinwith 72 fillers with similar segmentationTwelve of the fillers included a segment consiing of a single reflexive pronoun. The lists, in

tis-paper.

H

wa

c

nte

i

naw

nnnt

s

itag

u

lee. Apar-nt

lersccu- in- ornde-ould

re-seet ofucedad- the

ede

ergetthetileta

eree--dsreshichat-

f-ultsbles,

aw,atafthntsre-

hee-nt

gs

(strong vs weak) as experimental factors, and

4 “Condition-blind” versions of the same procedure, inwhich quartiles were calculated for the whole data set ofeach segment, regardless of condition, yielded almost identi-cal results for all the experiments reported in this paper.However, we also found that, where there were large differ-ences between conditions, the condition-blind procedurecould sometimes affect slower conditions disproportion-ately, presumably because genuine slow reaction times inthese conditions were being identified as outliers, due to theinclusion of the faster conditions in the overall distribution.

5 Because of an error in the preparing the experimentalscripts, two experimental items omitted the word and in thesixth segment of their high attachment conditions, resultingin globally ungrammatical sentences. These ungrammaticalitems were seen by five participants before the error was dis-covered. The 10 data points corresponding to the sixth andseventh segments of these sentences for these participants

IS REANALYSIS T

each participant, with the constraint that no texperimental items appeared adjacent to eother. Simple yes/no questions followed 50%the experimental and filler trials (e.g.,Were thetroops ignored in the press report?). The correctanswers to the questions was equally balanbetween “yes” and “no.”

The participant was seated in a souproofed booth and was instructed to press middle button with the index finger of his or hdominant hand, in order to see each segmenthe sentence. Each button press revealednext segment of the sentence, and the viewtime for all segments was timed. All segmeexcept the segment currently under view peared as a series of underscore characters,spaces corresponding to the actual spaces itext. The sentences were presented on two liwith the line break located between the fifth asixth segments. The participants were instructo answer the comprehension questions uthe right (“yes”) or left (“no”) buttons. Fivepractice trials were presented initially to famiarize the participants with the experimenprocedure, and the experiment proper bewith 10 filler sentences.

Results and Discussion

All data analyses were conducted on residreading times (Ferreira & Clifton, 1986Trueswell et al., 1994), in order to remove irrevant variance due to differences in segmlength and participant-specific reading speedregression equation was calculated for each ticipant, predicting reading time from segmelength. The regressions used data from all filand experimental items, but to increase the aracy of the regression equations, we did notclude reading times higher than 2000 mslower than 150 ms, or data from the initial afinal segments or the critical fifth segment (bcause it was assumed that these segments whave a large amount of variance and wouldduce the accuracy of the length correction (discussion in Konieczny, 1996)). A second seregression analyses based on all data prodstatistically identical results. The residual reing time for each segment corresponded to

E LAST RESORT? 291

och

of

ed

d-her

t ofthengtsp-ith

thees,d

eding

l-l

an

al;-nt

and the observed reading time. We then trimmoutliers according to the definition of extremvalues in the SPSS function explore; for eachsegment, we calculated the upper and lowquartiles of the distribution of residual readintimes for each condition. Cutoff values were sat three times the interquartile range above upper quartile, and three times the interquarrange below the lower quartile, and any dapoints above or below these cutoff values wreplaced by the relevant cutoff value. This procdure is similar to the percentile-based “trimming” and “winsorizing” procedures discusseby Winer (1971), and it is likely to result in lesdistortion of the data than more usual procedubased on means and standard deviations, wcan be unstable in the presence of outliers (Rcliff, 1993).4 The data-trimming procedure afected 3.6% of the data. Table 1 shows the resfor Experiment 1. For completeness, the tashows not only the residual reading timetrimmed as described above, but also the runtrimmed, reading times in parentheses. Dfor each segment are shown from the critical fisegment onward. Data for the first four segmeare aggregated into a single precritical region,ported as a persegment mean.5

Analyses of variance were calculated on ttrimmed residual reading times for the aggrgated precritical region, and for each segmefrom the critical fifth segment onward, takinattachment site (high vs low) and verb bia

ewere removed before analysis.

a

mh

n

leoia

fi

nen

h

nt

aes

isfectbleIw

tiveall-m-

ofs

ents

low ad-gh,

riti-

s

nsch-stor

ip-r-

ingtherbithis

s)

Strong/low 7 (698) 132 (766) 219 (838) 209 (875)

aggregating over participants (F1) and items(F2). There were no significant effects in thprecritical region (allFs , 1). There was asignificant effect of attachment in the criticfifth segment, corresponding to the disambiguating reflexive pronoun: Reading timewere faster when the reflexive pronoun disabiguated the sentence to high attachment twhen it disambiguated the sentence to loattachment (F1(1,31) 5 7.03, p , .02;F2(1,23) 5 8.30,p , .01). There was no maineffect of verb bias (F1(1,31) 5 2.82, p . .10;F2(1,23) 5 2.393,p . .13), nor was there aninteraction of verb bias and attachme(F1(1,31) 5 2.34, p . .13; F2(1,23) 5 1.12),Hence, participants experienced less difficureading a fragment that disambiguated the stences toward high attachment rather than lattachment. This preference was not signcantly modulated by the degree of verb bi(though numerically, there was a trend towasuch an effect).

The following, sixth segment showed nmain effects of attachment (F1(1,31) 5 3.07,p 5 .09; F2(1,23) 5 1.61) or verb bias (Fs ,1). However, the two factors interacted signicantly (F1(1,31) 5 6.41, p , .02; F2(1,23) 56.25, p , .02). Planned comparisons demostrated that low attached sentences were rmore quickly in the weakly biased conditiothan in the strongly biased condition (F1(1,31) 5 5.36, p , .03; F2(1,23) 5 4.27, p 5.05), but reading times for high attached setences did not differ between strongly anweakly biased verbs (Fs , 2). There were nosignificant effects on the final segment (Fs ,1.14). Mean comprehension accuracy for t

e

l-

s-

anw

t

tyn-w

fi-s

rd

o

-

-ad

n-d

e

tions was 95%, and there were no significadifferences in accuracy between conditions.

The data for the fifth, critical segment showhigh attachment advantage, while reading timin the following segment show that this effectmoderated by the degree of verb bias. The efof attachment in the fifth segment is compatiwith the RALR hypothesis, but not with the Rhypothesis, which would have predicted a loattachment advantage. However, an alternaexplanation of our findings is that the smplausibility difference in our items might explain the high attachment preference. To exaine this possibility, we performed an analysisthe critical fifth segment discarding the 12 itemwhich exhibited the greatest high attachmbias. The plausibility of the remaining 12 itemwas equally biased between the high and attachment readings, with a small numerical vantage to the low attachment reading (hi6.12; low, 6.27; both ts , 1). Despite this, theoverall high attachment advantage on the ccal segment remained significant (F1(1,31) 55.48,p , .03; F2(1,11) 5 9.02,p , .02). There-fore, there is no indication that plausibility waaffecting attachment decisions.

There are at least two possible explanatioof the interaction between verb bias and attament in the sixth (spillover) segment. The firis that the choice about whether to reanalyzenot is affected by the degree of verb bias manulated in our experiment. The general prefeence for high attachment indicates that avoidreanalysis is a factor, but the interaction wiverb bias in the sixth segment suggests that vbias has an effect as well. One problem wthis explanation is that the effect of verb bias

292 STURT ET AL.

TABLE 1

Trimmed Residual Reading Times for Experiment 1, with Untrimmed Raw Reading Times in Parentheses (m

Segment 1–4 5 6 7The troops…shot himself (and) were later in the press report

themselves mentioned

Weak/high 3 (680) 38 (642) 150 (799) 160 (821)Weak/low 17 (715) 75 (732) 119 (741) 166 (833)Strong/high 17 (694) 45 (651) 94 (795) 172 (825)

-delayed. Although there is a numerical trend

H

thaekffi

soe tfeh

tibileh

tise ioninl wn2

ddi

dn

y-

d

l

seis

-on

r

nt

he

,ds

r to

isrlyiv-soow

is

The e-

ctlddi-s

ng

byi-en

s

IS REANALYSIS T

toward an interaction in the critical segment,effect is nonsignificant. This account would nurally predict effects of verb bias as soon asfects of attachment site appeared, but the laca reliable interaction may merely reflect insucient power in the design.

An alternative possibility is that the procesfollows RALR, and therefore does not use vbias to inform its decision of whether or notreanalyze. This account explains the main efof attachment site in the critical segment: Tlow disambiguated sentences are incompawith the reflexive pronoun at this point, whthe high disambiguated sentences are not. Tthe processor seeks an alternative, compaanalysis. This is harder in the strongly biacondition because the alternative analysisvolves a highly infrequent subcategorizatiwhereas in the weakly biased condition it volves a less infrequent one. This extra difficufor the strongly biased verbs could manifestself in terms of the relative difficulty of the loattachment condition lasting over two segmeinstead of one. One purpose of Experiment to help distinguish these possibilities.

Experiment 2

Experiment 2 compared low-attached contions with globally unambiguous control contions using the complementizer that:

a. weak/ambiguous

The troops who discovered the enemy spy haused up all the supplies were later mentioned ithe press report.

b. weak/unambiguous

The troops who discovered that the enemy sphad used up all the supplies were later mentioned in the press report.

c. strong/ambiguous

The troops who found the enemy spy had useup all the supplies were later mentioned in thepress report.

d. strong/unambiguous

The troops who found that the enemy spy hadused up all the supplies were later mentioned inthe press report.

The inclusion of unambiguous controls a

E LAST RESORT? 293

et-f-of-

rrboctele

enbled

n-,-

tyit-

ts is

i--

-

takes place. In this experiment, we did not ureflexive pronouns for disambiguation. Thallowed a freer choice of predicates in the ambiguous segment. The global disambiguatiappears at the segment beginningwere. How-ever, the design also allows us to look foreanalysis effects at the wordhad, which arepredicted by the Reanalysis-Irrelevant accou(see below).

Predictions

Figure (1) describes the predictions of tthree models.

RALR. If reanalysis is avoided if possiblethen, in the ambiguous conditions, the worhad used upwill be attached high, so that thetroopswill be treated as the subject of had usedup. In the unambiguous conditions,had used upwill be attached low, so that the enemy spywillbe treated as its subject. Henceforth, we refethe segment had used upas the early criticalsegment. There should be no sign of reanalyseffects in the ambiguous conditions at the eacritical segment; reading times should be equalent to those for unambiguous conditions,long as both attachments are equally easy. Nconsider the segment were later mentioned,which we will refer to henceforth as the latecritical segment. RALR predicts that difficultywill emerge in the ambiguous conditions in thsegment, because the word were can only be attached as the main verb of the sentence. unambiguous conditions would not involve ranalysis because had used upis already attachedlow. Hence, RALR predicts a reanalysis effein this segment (ambiguous conditions shoutake longer to read than unambiguous contions). Note that this predicted effect involvemoving from an NP analysis, with the enemyspy interpreted as the object of the precediverb, to an S analysis, with the enemy spyas thesubject of had. The difficulty of this reanalysisin the late critical segment may be affected verb bias, with greater difficulty when the critcal NP/S verb is strongly NP-biased than whit is weakly NP-biased.

Revision-irrelevant. This hypothesis predictthat had used upis initially attached low, be-

t a

e

n

ec7

on

ntalal., bede-ingng,an-hee-tedriti- itte

heforaliffi-begly

-cean-id

theist,

esd

ana-s intra-d bet-uldrb

he

4 vs

gu-he inm-

asticen. i

hypotheses (see text for details).

spyis treated as its subject. If low attachmeninitially adopted, then the processor has to realyze the NP analysis of the enemy spywhen theearly critical segment is found. Therefore, thshould be reanalysis difficulty at this poinSuch reanalysis effects have been found inumber of previous studies of standard NPambiguities (Frazier & Rayner, 1982; Trueswet al., 1993; Ferreira & Henderson, 1990; Piering & Traxler, 1998; Rayner & Frazier, 198and manifest themselves as extra processingficulty for ambiguous conditions in comparisto unambiguous controls. We have also fou

ET AL.

isn-

ret. a/Sllk-)dif-nd

verb biases and using the same experimetechnique as in this experiment (Sturt et 1999). Therefore, if reanalysis effects are tofound (as predicted by RI), they should be tectable in this segment. In previous work ussimilar segment lengths in self-paced readiwe have found that reanalysis effects for stdard NP/S ambiguities do not spill over into tfollowing segment (Sturt et al., 1999). Therfore, any such increased difficulty associawith the ambiguous sentences in the early ccal segment should be quite localized, andcertainly would not spill over as far as the lacritical segment were later mentioned, where noeffects of ambiguity should be observed. TRevision-Irrelevant hypothesis also allows the difficulty of reanalysis in the early criticsegment to be affected by verb bias; the dculty of the ambiguous conditions may greater when the critical NP/S verb is stronNP biased than when it is weakly NP biased.

Verb bias. It is also possible that our manipulation of verb bias may be enough to influenthe processor’s choice of whether or not to realyze. If so, the processor is more likely to avoreanalysis at the first critical segment whenverb is strongly NP-biased than when itweakly biased. At the early critical segmenextra difficulty with the ambiguous sentencshould occur mainly with the weakly biaseverbs, since these sentences would be relyzed more than the strongly biased sentencethis segment. At the late critical segment, exdifficulty would mainly occur in the strongly biased conditions, since these sentences woulmore likely to have involved an initial high atachment. Hence, both critical segments shoexhibit interactions between ambiguity and vebias. No trials should involve reanalysis in tunambiguous conditions.

Method

Items. As in the previous experiment, 2items were created, treating verb bias (strongweak) and ambiguity (ambiguous vs unambious) as within-items factors. The majority of titems were adapted from Experiment 1. AsExperiment 1, we ran a plausibility pretest co

alcesTheffi--

294 STURT

FIG. 1. Predictions of the RALR, RI, and verb-bibased hypotheses for the early (left) and late (right) crisegments. Black triangles represent ambiguous sentand white triangles represent unambiguous sentencesdotted lines show possible effects of verb bias on the dculty of reanalysis in the RALR and Revision-Irrelevant

ofparing the relative plausibility of the initial NP

H

enthtatFne

s

e

t

e s

at

ter ai

r

e

t

theat-iasr)ceen-bi-200otrb-

ity

ctwoyt-grs.m-dtede-c-ayas

glau-’ses ef-)

d 6)

tn

en-%,c-

ce

IS REANALYSIS T

(mis)analysis for the strongly and weakly biasverbs, and that of the high and low attachmeFrom 28 candidate items we selected 24 combined the highest overall plausibility withe the smallest difference between the high low attachment readings, and between weakly and strongly biased verb readings. the selected 24 items, there was no differebetween the strongly and weakly biased stences (strong, 6.53; weak, 6.54; both ts , 1).Also, the judgments for the high attachmesentences did not differ significantly from thoof the low attachment sentences on the iteanalysis, though the numerical difference wsignificant on the participants analysis (hig6.69; low, 6.49; t1(11) 5 3.06,p , .02; t2(23) 51.71,p . .1). A smaller set of 20 items was slected for which this difference was completeeliminated (high 5 6.70; low 5 6.68; both ts ,1). The experiment employed the larger se24 items, but analyses were also conductedthe smaller set of 20 items (see below).

Results and Discussion

As in Experiment 1, residual reading timwere calculated using regression equationseach participant. Again, the regression analyexcluded short and long times (using the sacriteria), as well as data from the first and lsegments, and from critical regions (segmenand 6). Data trimming was carried out as in Eperiment 1 and affected 1.6% of the daTrimmed residual reading times and untrimmraw reading times are given in Table 2. We port data for each segment separately fromearly critical segment (segment 4) onward. Dfor the first three segments are aggregated one precritical region.

There were no significant effects in the agggated precritical region (all Fs , 1). There wereno effects in the fourth segment (all Fs , 1.6).Thus, there is no reason to conclude thatanalysis is taking place in this segment, withther strongly or weakly biased verbs. Thewere no significant effects in the fifth segmeeither (all Fs , 2.2). The sixth segment (the lacritical segment) showed a significant effectambiguity (F1(1,27) 5 19.86, p , .001;

E LAST RESORT? 295

dts.at

hndheorcen-

nte

msash,

-ly

of on

sfores

mests 4x-a.d

e-thetanto

e-

re-i-

renteof

biguous sentences took longer to read thanunambiguous sentences. Numerically, the ptern of results in this segment resembles the b3 attachment interaction in the sixth (spillovesegment in Experiment 1, with the differenbetween ambiguous and unambiguous stences being slightly greater for the strongly ased verbs than for the weakly biased verbs (ms vs 164 ms). However, this difference did nlead to a significant interaction between vebias and ambiguity (Fs , 1). There was a marginal effect of verb bias (F1(1,27) 5 2.76,p ..1; F2(1,23) 5 4.16,p , .06). The final, seventhsegment also showed an effect of ambigu(F1(1,27) 5 28.78,p , .001; F2(1,23) 5 13.23,p , .002). There was no indication of an effeof verb bias or an interaction between the tfactors (Fs , 1). We also computed three-waANOVAs on the data from the fourth (early criical) and sixth (late critical) segments, takinsegment, verb bias, and ambiguity as factoThis was to evaluate whether the lack of an abiguity effect in the early critical segment anits presence in the late critical segment resulin a significant two-way interaction, as prdicted by RALR, and whether such an interation was modulated by verb bias. The two-winteraction between ambiguity and segment wsignificant (F1(1,27) 5 12.60, p , .002;F2(1,23) 5 28.79,p , .001). The three-way in-teraction was not significant (Fs , 1).

To eliminate the possibility that the stronhigh attachment preference was due to the psibility difference found in the participantsanalysis of the pretest, we conducted analybased on the 20 entirely balanced items. Thefect of ambiguity in the sixth (disambiguatingsegment remained significant (F1(1,27) 522.26,p , .001; F2(1,19) 5 43.55,p , .001),as did the interaction between segment (4 anand ambiguity (F1(1,27) 5 14.34, p , .002;F2(1,19) 5 26.10,p , .001). Hence, the effecwas not due to plausibility differences. Meacomprehension accuracy for the 12 experimtal items which included questions was 94and there were no significant differences in acuracy between conditions.

In summary, the high attachment preferen

a

naes

e

c

e

ign

t

i

,ia

bi

wgp

to aer- toun

a

h-nnde-pethetonry

s)

Strong/unamb 23 (714) 25 (740) 36 (748) 221 (642) 15 (717)

the robust effect of ambiguity in the second dambiguating segment. Additionally, there wno clear indication that verb bias was usedmaking the decision of whether or not to realyze (although, numerically, effects of verb biwent the same way as in the previous expment). Thus, any effect of verb bias appearbe very weak.

Experiment 3

It is possible that the high attachment prefence holds for all sentences with this particustructure. Alternatively, the finding might refleproperties of our particular items. Notice ththe string of words preceding the second v(e.g.,the troops who discovered the enemy s)is almost as short as is possible for this ambity. The high attachment preference might hold for longer fragments.

One reason to suspect this follows from notion of a processing window, within which atachments are often assumed to be preferentmade. An early example of such a theory wthe Sausage Machine(Frazier & Fodor, 1978)which assumed that the processor preferentmade attachments within a small window around six words. The reason is that, oncecapacity is exceeded, phrases are “shuntedfor deeper processing. Once a phrase has shunted off, the model predicts difficulty making attachments to this phrase. Other wdow-based theories measure the size of the dow in terms of units other than words (e.number of constituents (Wanner, 1980) or deof embedding of a right branch (Frazier Clifton, 1998). However, they all assume ththe window size is limited, presumably becauit is related to putative short-term memory c

is-s

ina-sri- to

r-lart

atrb

pyu-ot

het-allyas

llyoftheoff”een

nin-in-.,th

&atsea-

Consider (10) below:

a. high/long

The cleaning lady who noticed the fireman ofthe voluntary brigade had badly burned herselfon the arm and was given a warm blanket.

b. high/short

The cleaning lady who noticed the fireman hadbadly burned herself on the arm and was given awarm blanket.

c. low/long

The cleaning lady who noticed the fireman ofthe voluntary brigade had badly burned himselfon the arm and was given a warm blanket.

d. low/short

The cleaning lady who noticed the fireman hadbadly burned himself on the arm was given awarm blanket.

The short conditions are similar in length the items of Experiment 1, for which we foundhigh attachment preference. Given this prefence, any window-based model would haveassume a window that included the high nophrase in the short conditions (i.e.,The cleaninglady who noticed the fireman) at the point wherethe first verb,had, is read. In contrast, the extrphrase (of the voluntary brigade) in the longconditions increases the distance between hadand lady, to such an extent that this high attacment would exceed the window capacity omost window-based accounts. On Frazier aClifton’s account, high attachment would bparticularly difficult in the long conditions because of the difficulty of coming out of “a deeright branch,” while high attachment would beasier in the short conditions because branching is less deep (see Frazier and Clif(1998, pp. 166–167) for a discussion of a ve

(10)

296 STURT ET AL.

TABLE 2

Trimmed Residual Reading Times for Experiment 2, with Untrimmed Raw Reading Times in Parentheses (m

Segment 1–3 4 5 6 7The troops… had used up all… were… in the press

Weak/amb 3 (698) 58 (763) 50 (757) 122 (788) 106 (808)Weak/unamb 3 (720) 51 (759) 34 (749) 241 (618) 211 (687)Strong/amb 8 (701) 44 (758) 2 (706) 179 (844) 109 (803)

similar pair of examples). Because window-

H

r

loi

ov

ij

hrsr

r

i

b

h

m

ch

a

heois

toists,sech

ed

s,h-is-

eri-c-

8)e

er,ea

h-

entob-esebe

ryeher-ch-

d-st

ed”-e,srt

IS REANALYSIS T

based theories predict that the high noun phis less accessible in the long rather than the sconditions, the high attachment preferenshould be reduced or even reversed in the conditions. If so, there should be an interactbetween attachment and length at himself/her-self in (10).

However, length might affect processingour items in other ways. As mentioned abolow attachment in our items results in a maclause subject which is extremely long, whhigh attachment results in a main clause subwhich is shorter. Since long prehead costituents are particularly unusual in languaglike English, this might explain the high attacment preference that we have found in the pceding two experiments. Thus the procesmight include an ambiguity resolution prefeence which minimizes the eventual lengthany prehead constituent. There are varioways in which such a mechanism might woand we will not be able to test all of them. Thsimplest possible mechanism would be onewhich the processor combines each incomhead with the shortest possible dependent cstituent on its left. This simple mechanism oviously is not compatible with the results oExperiments 1 and 2, because at the powhere the first verb is received in the input, tshortest possible dependent constituent islow noun phrase, while Experiments 1 andshow that this attachment is dispreferred.more sophisticated mechanism, which is copatible with the results in Experiments 1 andmight involve a preference against addingthe length of a constituent which is still expecing its licensing head in the input. Alternatively, the preference might simply be to attain such a way as to minimize the length of tmatrix subject. It is possible that a length-basmechanism would lead to a systematic hightachment preference once some thresholdlength has been exceeded in the initial nophrase. The length of the initial noun phraseour previous experiments might be beyond tthreshold. If so, the threshold will be exceedin both the long and the short conditions(10) as well. The threshold-based mechan

E LAST RESORT? 297

asehortceng

on

fe,inleectn-es-e-or-ofusk,einngon--finte

the2A-

2,tot--heedt-of

uninatdfm

using this construction, because it is hardcreate items in which the initial noun phrasemuch shorter than in the previous experimenor in the short conditions of (10). However, it ipossible to test a probabilistic version of thmechanism. Specifically, reluctance to attalow may increase probabilistically with thelength of the initial string of words. If this isso, and if the length manipulation representin (10) is affecting the portion of the lengthscale which is relevant to parsing decisionthen we would expect a stronger high attacment preference in the long conditions. If this the case, then the relative difficulty of disambiguation to low attachment would be greatin the long conditions than in the short condtions; in other words there would be an interation between attachment and length athim-self/herselfin (10).

As we have seen above, SPLT (Gibson, 199predicts a low attachment preference for thshort conditions of this experiment,contra theresults of Experiments 1 and 2, where thequivalent NP was of a similar length. HoweveSPLT makes the opposite prediction for thlong conditions. Recall that SPLT proposesparallel parsing algorithm, and the two attacments of the noun phrasethe fireman (of thevoluntary brigade)are initially made. Thesetwo attachments correspond to the attachmof this noun phrase as, on the one hand, theject of noticed(the NP attachment) and, on thother hand, the subject of a complement clau(the S attachment), whose head verb has topredicted. In the short conditions there is velittle difference in memory cost between thtwo attachments (see above). However, in tlong conditions, a larger memory cost diffeence develops between the NP and S attaments: A new discourse referent (the voluntarybrigade) has to be processed while the embeded verb remains predicted, adding to the cofor this prediction. This difference may benough to remove the S attachment of “noticefrom consideration, resulting in a high attachment preference for the long conditions. Hencthe most straightforward prediction of SPLT ifor a low attachment preference in the sho

ntconditions and a high attachment preference in

d

u

ftr

t h

o

to

te

e

ndnce

is-ive

s.s,the

seti-

nsedi-oti-i-

s:eif-asnt

s

he

298 STURT

the long conditions.6 Given the results of theprevious experiments, we do not expect to finlow attachment preference in the short contions. However, even assuming that SPLT cobe revised to derive a high attachment predtion for the short conditions, we would expethat this high attachment preference wouldstrengthened in the long conditions.

It should be clear that there are many lengbased models, and they make a range of dient predictions. However, they all share same basic prediction that length should intewith attachment preferences in the type of amguity which we discuss in this paper. Thus, olength manipulation in this experiment shouaffect the high attachment preference,strengthening it, weakening it, or reversing There remains the possibility, though, that range of constituent lengths represented bymanipulation may not cover the portion of tlength “scale” that is relevant to parsing desions, so the experiment cannot exhaustivtest all possible length-based explanations.

Method

Participants. Thirty-two participants from theUniversity of Glasgow were paid to participain the experiment. The data for one further pticipant were discarded because of a low coprehension score (less than 70% correct, cbining scores for both fillers and experimenitems).

Items. Twenty-four items were created treaing attachment site (high vs low) and leng(short vs long) as within-items factors, as in (1above. The long conditions involved an exprepositional phrase after the second nphrase. This prepositional phrase could notas a modifier of the verb phrase (becausepreposition was incompatible with the verbBecause of the possibility of finding effects afthe disambiguating word, a spillover segmwas added immediately following the reflexivThis allows us to interpret spillover or delayeffects occuring in a segment in which all con

e6 We thank Ted Gibson for guiding us through the SPpredictions.

ET AL.

adi-ldic-ctbe

th-fer-heactbi-urldbyit.heoure

ci-ely

tear-m-m-

tal

t-th0)raunactthe).ernte.d

di-

The head noun of both the first and seconoun phrases always denoted a human. Heanimacy was controlled in this experiment. Dambiguation involved a gender-marked reflexpronoun.

Plausibility pretest. Fourteen participantsrated 32 candidate items, in five conditionApart from the different number of conditionthe procedure for the pretest was identical to pretests for Experiments 1 and 2.

NP analysis: short

The cleaning lady noticed the fireman.

NP analysis: long

The cleaning lady noticed the fireman of thevoluntary brigade.

Low attachment: short

The fireman had badly burned himself.

Low attachment: long

The fireman of the voluntary brigade had badlyburned himself.

High attachment

The cleaning lady had badly burned herself.

The purpose of the pretest was to find a subof the 32 items which maximized overall plausbility, but which also minimized the plausibilitydifference between the short and long versioof the NP analysis, and minimized all pairwisdifferences among the three attachment contions. The 24 items which were selected did ndiffer statistically in the mean rating of plausbility in the short and long NP analysis condtions (short, 6.11; long, 5.90; t1(13) 5 1.67,p ..1, t2(23) 5 1.90, p . .05). The means of thethree attachment conditions were as followlow/short, 5.92; low/long, 5.76; high, 6.08. Thonly pairwise comparison which reached signicance among the attachment conditions wthat between low/long and high attachme(t1(13) 5 3.03,p , .01; t2(23) 5 2.28,p , .05;all other ts , 1.70). A smaller subset of 20 itemeliminated this difference as well (both ts ,1.25). Although the experiment employed t24-item subset, analyses were run on both ssets (see below).

Procedure. The procedure was identical to thLT

H

ei

i

i

itnbg

e

-

-

b

h

to

s in

he

le,byin

tfto

sts

t

-gerong

tnskat-

es

foro-

ed

ith

wo

cts

IS REANALYSIS T

segmentation was similar to that of Experime1, except for the presence of extra segmentsall conditions there was an extra segment athe disambiguating reflexive pronoun. Thextra segment was also the final segment offirst line. In the long conditions, there was alan extra segment after the second noun phrThe segmentation for the long and short stences is given in (11) (new line marked wdouble slash).

a. long

The cleaning lady / who noticed / the fireman /of the voluntary brigade / had badly burned /herself / on the arm // and was given / a warmblanket.

b. short

The cleaning lady / who noticed / the firemanhad badly burned / herself / on the arm // and wgiven / a warm blanket.

Results

As in experiments 1 and 2, residual readtimes were calculated using regression eqtions for each participant. Again, the regressanalyses excluded short and long times (usthe same criteria), as well as data from the fiand last segments, and from the criteria reflexregion (segment 5). Data trimming was carrout as in Experiments 1 and 2, and it affec3.5% of the data. Trimmed residual readitimes and raw reading times are given in Ta3. Data from the first three segments are aggated into a single precritical region.

There were no significant effects in the prcritical region (allFs , 2.3). Reading times inthe extra segment (long conditions only) did ndiffer either (bothFs , 1), nor did they in thefollowing, fourth segment (allFs , 1.7). Thecritical fifth segment (i.e., the reflexive) showea main effect of length (F1(1,31) 5 4.70, p ,.05;F2(1,23)5 6.12,p , .03) and a marginal effect of attachment (F1(1,31) 5 3.37, p , .08;F2(1,23)5 4.27,p 5 .05). These two factors interacted significantly (F1(1,31)5 8.24,p , .01;F2(1,23)5 6.75,p , .02). Contrast analyses revealed that this interaction was driven mainlythe difficulty of the low attached short conditionLow attached sentences took longer to read t

(11)

high attached sentences in the short conditio

E LAST RESORT? 299

nt. Infteristhesoase.n-

th

/as

ngua-oningrstiveededglere-

-

ot

d

-y:an

(F1(1,31)5 7.09,p , .02;F2(1,23)5 7.00,p ,.02), but attachment did not make a differencethe long conditions (bothFs , 1). Short sen-tences took longer to read than long sentencethe low attached conditions (F1(1,31) 5 8.15,p , .01, F2(1,23)5 9.12,p , .01), but lengthdid not make a difference to reading times in thigh attached conditions (bothFs , 1). Thus, atthe point where disambiguation is first availabthe high attachment preference is eliminatedthe addition of the extra prepositional phrasethe long conditions.

In the following sixth (spillover) segmen(e.g., on the arm), there was a main effect olength, with short sentences taking longerread than long sentences (F1(1,31)5 9.18,p ,.01; F2(1,23)5 15.98,p , .001). The main ef-fect of attachment was marginal in the itemanalysis but significant in the participananalysis (F1(1,31)5 5.10,p , .05; F2(1,23)53.29,p , .09). Although the two factors did nointeract significantly (F1(1,31) 5 2.07,p . .1;F2(1,23) 5 2.30,p , .1), contrast analyses revealed that low attached sentences took lonto read than high attached sentences in the lconditions (F1(1,31) 5 10.22, p , .005,F2(1,23) 5 8.49, p , .01), but attachmenmade no difference to the short conditio(bothFs , 1). Additionally, short sentences toolonger to read than long sentences in the hightached conditions (F1(1,31) 5 10.31,p , .005; F2(1,23) 5 11.61, p , .005), butlength did not make a difference to reading timin the low attached conditions (F)(1,31)5 1.11,p . .2, F2(1,23)5 1.83,p . .1). Thus there areclear signs of a high attachment preferencethe long conditions at this delayed point in prcessing.

The seventh (post-spillover) segment showa main effect of attachment (F1(1,31) 5 8.08,p , .01,F2(1, 23) 5 11.47,p , .005) but no ef-fect of length or interaction (all Fs , 1.4). Sim-ilar results were found in the final segment, wa main effect of attachment (F1(1,31) 5 6.33,p , .02; F2(1,23) 5 4.82,p , .05) but no effectof length and no interaction between the tfactors (all Fs , 2.7).

It is clear from these results that length affe

oht

thw

d

thtent.ethrrtcov

s hwng

ta

at-hee-en

ofedrtc-

ent

ede

mss--

ts,as

ernc-

nt

nt

ent

ed

)

Low/short 121 (882) 75 (789)

are parsed. The main effects of length in bthe critical and spillover segments show tprocessing is faster when the preceding conhas included the extra prepositional phrase when it has not. This appears to be at odds visibility-based theories, as well as integratiobased theories such as SPLT, which predictscreased integration difficulty in the long contions in these segments of the sentence (thosee below for a possible explanation of length effect). It also appears that length inacts with the attachment manipulation, as caseen in the results of the critical fifth segmenwe looked only at the results of this critical sment, it would appear that the addition of extra phrase removes the high attachment perence seen in the short conditions. Howevewe also take into account the results from spillover segment, which show a clear attament effect in the long conditions, it seems mlikely that the longer context does not remothe high attachment preference, but rather tematically delays the point in processingwhich the difficulty associated with low attacment manifests itself. To evaluate this claim,ran further analyses of variance on combidata from both the critical and spillover sements, treating segment as a third factor indesign. The combined data replicate the meffects of attachment (F1(1,31) 5 5.50,p , .03;F2(1,23) 5 4.75, p , .05) and of length(F1(1,31) 5 11.76,p , .002; F2(1,23) 5 17.71,

thatextanith

n-in-i-ugher-

be Ifg-eef-, ifheh-ree

ys-at-e

ed-

hein

between these two factors (both Fs , 1). Thisdemonstrates that the strength of the high tachment preference is not affected by tlength of the preceding context, once the dlayed effects in the spillover segment are takinto consideration. The claim that the onsetthe attachment effect is systematically delayin the long conditions in relation to the shoconditions is supported by a three-way interation among length, attachment, and segm(F1(1,31) 5 8.69,p , .01; F2(1,23) 5 9.87,p ,.005).

To eliminate possible confounds associatwith plausibility (see above for discussion), wran an analysis with the reduced set of 20 itefor which length made no difference to the plauibility of the NP misattachment. An analysis conducted on the critical and spillover segmenand taking segment, attachment, and lengthfactors as above, showed that the overall pattof results was identical, with a three-way interation between these factors (F1(1,31) 5 9.46,p , .005;F2(1,19)5 9.24,p , .01). As before,the short conditions showed a high attachmeadvantage in the critical segment (F1(1,31) 57.38,p , .02;F2(1,19)5 5.09,p , .05) but notin the spillover segment (bothFs, 1), while thelong conditions showed no effects of attachmein the critical segment (bothFs , 1) but a highattachment advantage in the spillover segm(F1(1,31)5 7.69,p , .01;F2(1,19)5 6.22,p ,.03). However, probably because of the reduc

300 STURT ET AL.

TABLE 3

Trimmed Residual Reading Times for Experiment 3, with Untrimmed Raw Reading Times in Parentheses (ms

Segment 1…3 Extra 4 5 6The…fireman of…brigade had…burned himself on the arm

High/long 33 (697) 231 (1109) 170 (988) 92 (678) 19 (576)High/short 37 (714) 222 (1026) 90 (697) 108 (722)Low/long 29 (713) 256 (1144) 159 (968) 99 (708) 95 (667)Low/short 62 (767) 197 (1014) 204 (821) 125 (700)

7 8(and) was given a warm blanket

High/long 26 (790) 215 (721)High/short 30 (817) 43 (760)Low/long 102 (821) 73 (828)

npower in the data set, the overall effect of attach-

H

o

e

a

ec

g

r

e

x,

et

r

sgco

at-enn-l.ct3), tosen-ve inre,ess.

d--

ined

edone.e,lye--atce

het-

e-re-er-ewre-el-n are-fer-thee the

IS REANALYSIS T

ment in the two segments was not significa(F1(1,31)5 3.61,p , .07;F2(1,19)5 2.75,p ..1). As before, there was no two-way interactibetween length and attachment (bothFs , 1).Hence, the results reported above are not relato considerations of plausibility.

Mean comprehension accuracy for the 12 perimental items which included questions w91%, and there were no significant differencin accuracy between conditions.

Although the combined data from the criticand spillover segments show no signs of a twway interaction between attachment and lengwe still owe an explanation for two findingarising from the length manipulation. First, ththree-way interaction in the cross-segmanalysis clearly demonstrates that length delay certain aspects of processing. Secomain effects in both critical and spillover sements demonstrate that longer sentences wbeing processed more quickly than their shocounterparts. We will consider the questionthe delay first. One possible reason for the deinvolves the number of discourse referents. Rcall that the long sentences include an exprepositional phrase, which introduces a ndiscourse referent:

The cleaning lady who noticed the fireman [ofthe voluntary brigade] had badly burned herselfon the arm. . .

At the point where the disambiguating refleive pronoun himself/herself is encounteredthere are (in purely configurational terms) twnoun phrases which are possible antecedentsthe reflexive (the cleaning ladyandthe fireman).However, the disambiguation ensures that oof these is ruled out through gender marking.the long conditions, the extra discourse referintroduced by the prepositional phrase is nopossible antecedent for the reflexive evenconfigurational terms. However, the presencethis extra referent in the discourse represention may cause a delay in the computationquired to check the appropriate gender featurNote that this explanation requires a mechaniin which a configurationally defined bindintheory is not used as an initial filter in the searfor antecedents to pronouns and anaph

E LAST RESORT? 301

nt

n

ted

x-ases

lo-th,se

nison, & Albrecht, 1997; Nicol & Swinney,1989).

The main effect of length appears to be odds with the idea that integration difficulty increases with the distance intervening betwethe two items involved in an attachment, a cetral claim of Gibson (1998) in the SPLT modeHowever, one possible reason for this effesuggested by Ferreira and Henderson (199who found similar results, is that readers tendspeed up as they read successive words of a tence. In the long conditions, more words habeen read at the point of disambiguation thanthe corresponding short conditions. Therefoon this account, we would expect reading tim

-

ofor

neInntainofta-e-es.m

hrs

to be shorter at this point in the long condition

GENERAL DISCUSSION

Experiment 1 showed a high attachment avantage using reflexive pronouns for disambiguation, which was moderated by verb biasthe spillover segment. Experiment 2 replicatthe high attachment advantage, and also showthat verb bias did not have any clear effectthe decision about whether or not to reanalyzGiven the robust high attachment preferencand the lack of effects of verb bias in the earcritical segment, it is much more likely that thsmall effects of verb bias which we found in Experiment 1 reflect minor differences in the difficulty of reanalysis. Experiment 3 showed ththe strength of the high attachment preferenwas unaffected by increasing the length of tinitial string of words, though length did affecthe timing with which the effect manifested itself.

In summary, the clear finding which has rpeatedly emerged in the three experiments ported in this paper is a high attachment prefence. The results are compatible with the vithat the processor strongly prefers to avoid analysis if possible. Hence, the Reanalysis-Irrevant hypothesis can be discounted, as camodel in which a general preference to avoid analysis can be overridden by a recency preence, as in Fodor and Inoue (2000). Indeed results are particularly striking given thstrength of the recency preference as well as

-fact that reanalysis should be particularly

e

eteel

bt

a emn

f

eaf

hdm

oeetc ta

iemh

dfor mi

fedin

ingtach-sedomeorallyouse-thellipsse-nsed

ex- usref- thebil-ar-

eruseser.ordochh-ntof

nth-

autd--o-entsand

o

h

302 STURT

straightforward in the sentences studied hHowever, we cannot conclude whether reanasis is truly the last resort of the processor. Thmay be factors which we have not manipulahere, but which will induce reanalysis. And it rmains possible that, given stronger maniputions than we were able to make here, verb and constituent length could also cause processor to reanalyze.

The conclusions of the present paper strengthened by very similar results reportedSchneider and Phillips (2001). In two self-pacreading studies, Schneider and Phillips exained the processing of sentences which begain the fragment in (12):

The creative woman who knows the funny manwrote some comedy sketches herself/himselabout the amusing escapades…

Clearly, (12) is very similar to the sentencconsidered in this paper. Assuming an initial tachment of the funny manas as the object oknows, the attachment of the verb wrotecan ei-ther be low (involving reanalysis) or hig(avoiding reanalysis). As in Experiments 1 anin this paper, this attachment is later disabiguated by a reflexive pronoun (himselfor her-self). Reading times following the reflexive prnoun were consistent with a high attachmpreference, mirroring the results of the prespaper. Schneider and Phillips also manipulaverb bias and found that the basic high-attament preference was unaffected by whetherverb was strongly or weakly NP-biased, agmirroring the results of the present paper.

Despite the similar research questions andsults, the two sets of experiments differ innumber of ways. Schneider and Phillips areterested in the impact of individual differencin reading ability, and they show that the sabasic pattern of results is found for both higand low-span readers, though low-span reaappear to be less sensitive to probabilistic inmation. The present paper does not considedividual differences but, in Experiment 3, exaines the issue of constituent weight and influence on parsing processes. Another difence is that Schneider and Phillips examinewider range of verb bias than we did. They

(12)

cluded not only NP-biased verbs but also,

ET AL.

re.ly-red-a-iashe

rebyd-

as

st-

3-

-ntntedh-hein

re-an-se-

ersr-in--tsr- a-

their Experiment 2, S-biased verbs, showthat the use of these verbs can reverse the atment preference. We avoided using S-biaverbs, for reasons given above. There are sminor differences in experimental design. Fexample, the present paper uses semanticneutral complementizers to create unambigucontrol conditions (Experiment 2). This procdure only allows controls to be created for low-attached sentences. Schneider and Phiuse not only complementizers but also camarked pronouns, allowing control conditioto be created for both high- and low-attachconditions.

In summary, the results of the two sets of periments complement each other and allowto conclude that the basic high attachment perence is robust across verb biases withinNP-bias range and across different reading aities, and it is unaffected by lengthening the elier regions.

Modeling the Results

In the model presented in Sturt and Crock(1996), the RALR preference emerges becaof the search strategy employed by the parThe parser can combine the current input wwith the current left context in three ways. Twof these are left and right attachment, whiare used for simple integration. In right attacment, the root of the projection of the curreinput word is attached as a dependent of onethe nodes on the right frontier of the currepartial phrase marker (CPPM). In left attacment, the root of the CPPM is attached asdependent of the projection of the new inpword. The third form of combination is calletree-lowering. It involves inserting the projection of the new input word into the right frontier of the CPPM. In terms of the model, prcessing the sentences in these experimrequires a choice between left attachmenttree-lowering. Figure 2 illustrates these twchoices.

The new word hadprojects to an S node wita noun phrase attachment site to its left (NPz). Ifleft attachment is adopted, then NPz is identifiedwith the root of the CPPM, NPx, and the S node

3

r

h e

eeehoh

FIG. 2. The choice between left attachment and tree-lowering in the model proposed in Sturt and Crocker(1996).

CPPM. If tree-lowering is performed, howeveNPz is identified with NPy on the right frontierof the CPPM, and the S node will be insertedthe daughter of VPx.

Sturt and Crocker (1996) claimed that tparser adopts a serial search, first checkingany ways of combining the projection of thnew node via left or right attachment, and onsubsequently searching for a site to apply trlowering. The result of this is that the modadopts the RALR strategy, because tree-lowing is used for reanalysis, but the left and rigattachment operations operations are nHence, the results are compatible with t

me

IS REANALYSIS THE LAST RESORT? 30

t

rchh

model’s predictions.

APPENDIX: EXPERIMENTAL ITEMS

Segment boundaries are marked with a vertical b(|). Alternatives across the conditions are separawith a forward slash (/).

Items for Experiment 1

In the second segment, the strongly biased verb precethe weakly biased verb, and the NP bias is added in patheses. See text for an explanation of how the bias was culated. In the fifth segment, the reflexive pronoun for tlow disambiguation precedes the reflexive pronoun for thigh disambiguation. The bracketed word in the sixth se

nt (andor but) was used only in the low conditions.

,

as

efor

lye-lr-tt.e

ared

desen-al-ee

g-

1. The councillor | who proposed(0.98)/re-vealed(0.79) | the spending cuts | had not justified |themselves/himself | (and) was rather | apologetic.

2. The politician | who proposed(0.98)/re-vealed(0.79) | the manifesto| had contradicted|itself/himself| (and) was later| criticised.

3. The troops | who found(0.94)/discovered(0.71) |the enemy spy | had shot | himself/themselves | (and)were later mentioned | in the press report.

4. The woman | who heard(0.88)/noticed(0.65) |the hooligan | wasn’t at ease with | himself/herself |(but) told the police | everything.

5. The little girl | who accepted(0.93)/no-ticed(0.65) | her foster parents | couldn’t organise |themselves/herself | (and) was rather confused | by thewhole thing.

6. The detective | who read(0.99)/denied(0.76) | thestatement | had contradicted | itself/himself | (and) wasa bit | confused.

7. The professor | who proposed(0.98)/con-cluded(0.69)| the project | hadn’t really proved|itself/himself | (and) was heavily criticised| in thepress.

8. The weather-man| who guaranteed(0.98)/pre-dicted(0.72)| the good weather| wouldn’t repeat|itself/himself | (and) was| praised.

9. The blind woman | who reported(0.98)/sensed(0.77) | the muggers | had hidden | them-

o

NPbigu-

(0.79) | the new government | had won considerable |

304 STURT

10. The shepherds | who recognised(1)/pre-dicted(0.72) | the cloud formation | would soon dissi-pate | itself/themselves | (and) gathered | their sheep.

11. The sniffer dogs | which found(0.94)/sensed(0.77) | the house fire | had exhausted |itself/themselves | (but) were very | well trained.

12. The manager | who announced(0.91)/esti-mated(0.68) | the redundancies | had justified | them-selves/himself | (and) got promoted | the next month.

13. The biologist | who found(0.94)/discov-ered(0.71) | the ants | had killed | themselves/himself |(and) was only | thirty years old.

14. The rescue workers | who found(0.94)/discov-ered(0.71) | the girl | had injured | herself/themselves |(and) had been | rather nervous.

15. The spice girl | who judged(0.89)/con-cluded(0.69) | the song contest | had distinguished |itself/herself | (and) was going through | a tough divorce.

16. The journalists | who reported(0.98)/denied(0.76) | the story | had contradicted |itself/themselves | (and) were rather | ashamed.

17. The minister | who established(0.99)/revealed(0.79) | the new government | had over-committed |itself/himself | (but) was | very clever.

18. The official | who established(0.99)/con-cluded(0.69) | the investigation | had over-stretched |itself/himself | (and) was feeling | rather sad.

19. The investor | who regretted(0.97)/predicted(0.72) | the economic situation | wouldn’t correct |itself/himself | (but) was on | the news.

20. The bank manager | who guaranteed(0.98)/concluded(0.69) | the purchase | would pay for |itself/himself | (and) got interviewed | by the BBC.

21. The businessman | who announced(0.91)/esti-mated(0.68) | the wholesale prices | had stabilised |themselves/himself | (and) was | rather proud.

22. The jury members | who recognised(1)/doubted(0.83) | the suspect | had prepared |himself/themselves | (but) were bored | with the trial.

23. The woman | who recognised(1)/doubted(0.83)| her ex-boyfriend | had reformed | himself/herself |(but) suffered | from depression.

24. The dictator | who established(0.99)/con-firmed(0.81) | the new regime | couldn’t finance |itself/himself | (and) was | overthrown.

Items for Experiment 2

The following items represent the ambiguous conditi

ET AL.

ns

ased verb precedes the weakly biased verb, and thebias is added in parentheses. To reconstruct the unamous conditions, addthat as the last word of Segment 2.

1. The councillor | who proposed(0.98)/revealed(0.79) | the spending cuts | deserved much | more sup-port | was well known | to the voters.

2. The party members | who proposed(0.98)/re-vealed(0.79) | the manifesto | needed a lot | more pub-licity | were planning | a campaign.

3. The troops | who found(0.94)/discovered(0.71) |the enemy spy | had used up | all the supplies | werelater mentioned | in the press report.

4. The woman | who heard(0.88)/noticed(0.65) |the hooligans | wanted to catch | the last train | wasvery helpful | to the police.

5. The little girl | who accepted(0.93)/noticed(0.65) | her foster parents | could never really | be veryhappy in life | was confused | by everything.

6. The detective | who read(0.99)/denied(0.76) | thestatements | had played an important part | in the courtcase | was scared | of the judge.

7. The professors | who proposed(0.98)/con-cluded(0.69) | the project | had caused a lot of | inter-esting discussion | were discussed | in the press.

8. The weather-man | who guaranteed(0.98)/pre-dicted(0.72) | the high temperatures | could be found |in London at the weekend | was famous | in the city.

9. The blind woman | who reported(0.98)/sensed(0.77) | the muggers | came from | somewherenear Aberdeen | was feeling | very scared.

10. The shepherds | who recognised(1)/pre-dicted(0.72) | the cloud formation | would soon gather| over the hills | were looking | for lost sheep.

11. The sniffer dogs | which found(0.94)/sensed(0.77) | the house fire | had begun to get | veryweak | were well trained | by the police.

12. The biologist | who found(0.94)/discov-ered(0.71) | the tropical ants | could become | ratheraggressive | was respected | by his colleagues.

13. The rescue workers | who found(0.94)/discov-ered(0.71) | the injured girl | had been in the building |for hours | were feeling | rather nervous.

14. The celebrities | who judged(0.89)/con-cluded(0.69) | the song contest | had become unpopu-lar | with the audience | were always arguing | duringthe breaks.

15. The journalists | who reported(0.98)/denied(0.76) | the sensational story | had caused panic |amongst the editors | were ashamed | of their actions.

16. The ministers | who established(0.99)/revealed

bi- public support | were skilled | at politics.

HE

forforlti-

-rths

IS REANALYSIS T

17. The official | who established(0.99)/concluded(0.69) | the investigations | had used up | too muchmoney | was feeling | rather sad.

18. The experts | who regretted(0.97)/predicted(0.72) | the economic situation | couldn’t encourage |new investors | were filmed | on the news.

19. The bank manager | who guaranteed(0.98)/concluded(0.69) | the mortgage agreements | hadhelped | many house buyers | was featured | in a mag-azine.

20. The economist | who announced(0.91)/esti-mated(0.68) | the wholesale prices | had put off | anumber of people | was proud | of his expertise.

21. The jury members | who recognised(1)/doubted(0.83) | the suspect | took the trial | very seri-ously | were interested | in the proceedings.

22. The generals | who established(0.99)/con-firmed(0.81) | the military regime | could provide se-curity | for the people | were bothered | about the army.

23. The announcer | who read(0.99)/confirmed(0.81) | the news reports | had been critical | of the cab-inet | was annoyed | about the policies.

24. The quiz contestants | who accepted(0.93)/no-ticed(0.65) | the silver trophy | looked rather | scruffy |were skillful | at answering questions.

Items for Experiment 3

In the disambiguating segment, the reflexive pronounthe low disambiguation precedes the reflexive pronounthe high disambiguation. The bracketed word in the penumate segment (andor but) was used only in the low conditions. To reconstruct the short conditions, remove the fousegment. The wordshydid not appear in the long conditionin the third segment of item 5.

1. The foster mother | who accepted | the little boy| of the children’s home | didn’t really trust | him-self/herself | at all | (and) was well known | in the com-munity.

2. The boy | who recognised | the princess | of thewealthy tax haven | had cleverly disguised |herself/himself | in a hat | (and) was waiting | down thestreet.

3. The cleaning lady | who noticed | the fireman | ofthe voluntary brigade | had badly burned | himself/her-self | on the arm | (and) was given | a warm blanket.

4. The husband | who accepted | the child bride | ofthe rich family | didn’t really like | herself/himself |very much | (and) was uncertain | about the marriage.

5. The call-girl | who saw | the (shy) monk | of thelong-established abbey | had always undervalued |

LAST RESORT? 305

6. The clergyman | who remembered | the duchess |of the remote principality | had carefully educated |herself/himself | at home | (and) was well respected |in the parish.

7. The woman | who recognised | the rapist | of theinnocent teenager | had clearly contradicted | him-self/herself | on the tape | (and) was interrogated | atthe trial.

8. The handyman | who mentioned | the dinnerlady | of the school canteen | had actually confused |herself/himself | a lot | (and) was offered | early retire-ment.

9. The salesgirl | who mentioned | the workman |from the unreliable company | was now blaming |himself/herself | for the trouble | (and) sent a letter | ofexplanation.

10. The movie man | who discovered | the main ac-tress | of the historic drama | was always injecting |herself/himself | with drugs | (and) appeared | on adocumentary.

11. The barmaid | who reported | the conman | ofthe criminal gang | had already betrayed | himself/her-self | to the press | (but) was later questioned | by theinspector.

12. The policeman | who reported | the prostitute |of the high-class brothel | had badly hurt | herself/him-self | in the fight | (and) was interviewed | by a re-porter.

13. The countess | who heard | the choirboy | of theold cathedral | had really enjoyed | himself/herself | atthe concert | (and) organised a charity event | after-wards.

14. The foreman | who doubted | the chairwoman |of the management group | had fully committed | her-self/himself | to the deal | (and) was supported | by thetrade union.

15. The spokeswoman | who proposed | the chair-man | of the finance committee | needed to promote |himself/herself | carefully | (and) was seeking | expertadvice.

16. The Irishman | who heard | the chorus girl | ofthe popular cabaret | had badly misbehaved |herself/himself | in Paris | (and) almost missed | hisflight.

17. The ambulanceman | who found | the trappedgirl | from the neighbouring village | had seriously cut| herself/himself | on a nail | (and) contacted the hospi-tal | immediately.

18. The policewoman | who found | the madman |from the mental institution | had somehow hidden |

306 STURT ET AL.

--

es

e

t--

-an-

ticn,

n

19. TheArab prince|who revealed| the mistress|ofthe middle eastern dictator| had deeply criticised| her-self/himself| on TV | (and) was very sincere| about it.

20. The manageress | who acknowledged | the er-rand boy | of the courier firm | should not assert | him-self/herself | so much | (but) would always encourage |the workforce.

21. The old man | who recalled | the priestess | ofthe local chapel | couldn’t really organise |herself/himself | well | (and) was unpopular | with theneighbours.

22. The gypsy girl | who revealed | the Mafia god-father | of the Sicilian brotherhood | had fatally shot |himself/herself | in Rome | (and) was in all | thetabloids.

23. The businessman | who proposed | the womancandidate | of the political party | needed to assert |herself/himself | more | (and) became very serious |about it.

24. The newsgirl|who recalled| the cameraman|ofthe silly sitcom|had really distinguished|himself/her-self | at work| (and) was always helpful| at the studio.

d

:

-tss.

.-s

n-,

-

REFERENCES

Altmann, G., & Steedman, M. (1988). Interaction with cotext during human sentence processing. Cognition, 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.

Chafe, W. (1994). Discourse, consciousness and tim.Chicago: Univ. of Chicago Press.

Clifton, C., Kennison, S. M., & Albrecht, J. E. (1997). Reaing the words him and her: Implications for parsingprinciples based on frequency and on structure. Journalof Memory and Language, 36,276–292.

Cohen, J. D., MacWhinney, B., Flatt, M., & Provost,(1993). Psyscope: A new graphic interactive enviroment for designing psychological experiments. Behav-ioral Research Methods, Instruments, & Compute,25,257–271.

Corley, S., Corley, M. M. B., Keller, F., Crocker, M. W., &Trewin, S. (in press). Finding syntactic structure in uparsed corpora: The gsearch corpus query sysComputers in the Humanities.

Cuetos, F., & Mitchell, D. C. (1988). Cross-linguistic diffeences in parsing: Restrictions on the use of the late sure strategy in Spanish. Cognition, 30,72–105.

De Vincenzi, M., & Job, R. (1995). An investigation of laclosure: The role of syntax, thematic structure apragmatics in initial and final interpretation. Journal ofExperimental Psychology: Learning, Memory, anCognition, 21,1303–1321.

Ferreira, F., & Clifton, C. (1986). The independence of stactic processing. Journal of Memory and Language,25,348–367.

n-

-

e

d-

J.n-

rs

n-tem.

r-clo-

tend

d

yn-

Ferreira, F., & Henderson, J. M. (1990). Use of verb information in syntactic parsing: Evidence from eye movements and word-by-word self-paced reading. Journalof Experimental Psychology: Learning, Memory, andCognition, 16,725–745.

Ferreira, F., & Henderson, J. M. (1993). Reading processduring syntactic analysis and reanalysis. CanadianJournal of Psychology, 47,247–275.

Fodor, J. D., & Frazier, L. (1980). Is the human sentencparsing mechanism an ATN? Cognition, 8, 417–459.

Fodor, J. D., & Inoue, A. (2000). Garden path reanalysis: Atach (anyway) and revision as last resort. In M. D. Vincenzi & V. Lombardo (Eds.),Cross linguistic perspec-tives on sentence processing(pp. 21–61). Dordrecht:Kluwer.

Ford, M., Bresnan, J., & Kaplan, R. M. (1982). A competence-based theory of syntactic closure. In J. Bresn(Ed.), The mental representation of grammatical relations(pp. 727–796). Cambridge, MA: MIT Press.

Frazier, L. (1978). On comprehending sentences: Syntacparsing strategies. Unpublished doctoral dissertatioUniversity of Connecticut, Storrs.

Frazier, L. (1987). Sentence processing: A tutorial review. IM. Coltheart (Ed.),Attention and performance(Vol.12, pp. 559–586). Hillsdale, NJ: Erlbaum.

Frazier, L. (1990). Identifying structure under X0. Yearbookof morphology, 3, 87–105.

Frazier, L., & Clifton, C. (1998). Sentence reanalysis, anvisibility. In J. D. Fodor & F. Ferreira (Eds.),Reanaly-sis in sentence processing(pp. 143–175). Dordrecht:Kluwer.

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

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

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

Gibson, E. A. F. (1991). A computational theory of humalinguistic processing: Memory limitations and processing breakdown. Unpublished doctoral dissertationCarnegie Mellon University, Pittsburgh, PA.

Gibson, E. A. F. (1998). Linguistic complexity: Locality ofsyntactic dependencies. Cognition, 68,1–76.

Gorrell, P. (1995). Syntax and parsing. Cambridge, UK:Cambridge Univ. Press.

Hawkins, J. A. (1990). A parsing theory of word order universals. Linguistic Inquiry, 21,223–261.

Hawkins, J. A. (1995).A performance theory of orderand constituency. Cambridge, UK: Cambridge Univ.Press.

Hirose, Y. (1999). Resolving reanalysis ambiguity in Japan-ese relative clauses. Unpublished doctoral dissertation,The City University of New York, New York, NY.

Holmes, V. M., Stowe, L., & Cupples, L. (1989). Lexical ex-

IS REANALYSIS THE LAST RESORT? 307

pectations in parsing complement verb sentences. Jour-nal of Memory and Language, 28,668–689.

Just, M., Carpenter, P., & Wooley, J. (1982). Paradigms

a

t

s

b-

.

Kn

ec

t-

r

n

r

a

ing: Eye movements in the analysis of semantically bi-ased sentences. Journal of Verbal Learning and Verbal

-

d

rid

ng.

nted-

-e-

ln-

s,yn

a-

.-

.pa-s.-

e:

processes in reading comprehension. Journal of Exper-imental Psychology: General, 111,228–238.

Kamide, Y., & Mitchell, D. C. (1997). Relative clause tachment: Nondeterminism in Japanese parsing. Jour-nal of Psycholinguistic Research, 26,247–254.

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

Konieczny, L. (1996). Human sentence processing: A mantics-oriented approach. Unpublished doctoral dissertation, University of Freiburg, Germany.

Liversedge, S. P., Pickering, M. J., Branigan, H. P., & vGompel, R. P. G. (1998). Processing argumentsadjuncts in isolation and context: The case ofphrase ambiguities in passives.Journal of Experimental Psychology: Learning, Memory, and Cognition, 24,461–475.

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

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

Mitchell, D. C. (1989). Verb guidance and other lexical fects in parsing. Language and Cognitive Processes, 4,123–154.

Mitchell, D. C., & Holmes, V. M. (1985). The role of spcific information about the verb in parsing sentenwith local ambiguity. Journal of Memory and Language, 24,542–559.

Nicol, J., & Swinney, D. (1989). The role of structurecoreference assignment during sentence compresion.Journal of Psycholinguistic Research, 18, 5–19.

Phillips, C., & Gibson, E. (1997). On the strength of local attachment preference. Journal of Psycholinguistic Research, 23,323–346.

Pickering, M. J., & Traxler, M. J. (1998). Plausibility and covery 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: Evideagainst frequency-based accounts. Journal of Memoryand Language, 43.

Pritchett, B. L. (1992). Grammatical competence and paing performance. Chicago: Univ. of Chicago Press.

Ratcliff, R. (1993). Methods for dealing with reaction timoutliers. Psychological Bulletin, 114,510–532.

Rayner, K., Carlson, M., & Frazier, L. (1983). The intertion of syntax and semantics during sentence proc

and

t-

ure

e--

anandy-

S.lu-

.dion.

ef-

-es

-

inhen-

he

e-

ry,

).ce

s-

e

c-ess-

Behavior, 22,358–374.Rayner, K., & Frazier, L. (1987). Parsing temporarily am

biguous complements. Quarterly Journal of Psychol-ogy, 39A, 657–673.

Schneider, D., & Phillips, C. (2001). Grammatical searchand reanalysis. Journal of Memory and Language.

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

Stevenson, S. (1994). Competition and recency in a hybnetwork model of syntactic disambiguation. Journal ofPsycholinguistic Research, 23,295–321.

Stevenson, S. (1998). Parsing as incremental restructuriIn J. D. Fodor & F. Ferreira (Eds.),Reanalysis in sen-tence processing(pp. 327–363). Dordrecht: Kluwer.

Straub, K., & Badecker, W. (1994). The role of grammar iprocessing pronouns and reflexives. Paper presento the Psychonomic Society, St. Louis, MO, November.

Sturt, P., & Crocker, M. W. (1996). Monotonic syntactic processing: A cross-linguistic study of attachment and ranalysis. Language and Cognitive Processes, 11,449–494.

Sturt, P., Pickering, M. J., & Crocker, M. W. (1999). Structurachange and reanalysis difficulty in language comprehesion. Journal of Memory and Language, 40,136–150.

Tabossi, P., Spivey-Knowlton, M. J., McRae, K., & TanenhauM. K. (1994). Semantic effects on syntactic ambiguitresolution—evidence for a constraint-based resolutioprocess. Attention and Performance, 15,589–615.

Thornton, R., MacDonald, M. C., & Arnold, J. E. (2000).The concomitant effects of phrase length and informtional content on sentence processing. Journal of Psy-cholinguistic Research, 29,195–203.

Trueswell, J. C., Tanenhaus, M. K., & Garnsey, S. M(1994). Semantic influences on parsing: Use of thematic role information in syntactic ambiguity resolu-tion. Journal of Memory and Language, 33, 285–318.

Trueswell, J. C., Tanenhaus, M. K., & Kello, C. (1993)Verb-specific constraints on sentence processing: Serating effects of lexical preference from garden pathJournal of Experimental Psychology: Learning, Memory, and Cognition, 19,528–553.

Wanner, E. (1980). The ATN and the Sausage MachinWhich one is baloney? Cognition, 8, 209–225.

Winer, B. J. (1971). Statistical principles in experimentaldesign(2nd ed.). New York: McGraw–Hill.

(Received March 28, 2000)(Revision received October 31, 2000)Published Online June 21, 2001