Plausibility and the Processing of Unbounded Dependencies:An Eye-Tracking Study

Plausibility and the Processing of Unbounded Dependencies:An Eye-Tracking Study

MATTHEW J. TRAXLER AND MARTIN J. PICKERING

Human Communication Research Centre, University of Glasgow, United Kingdom

Two eye-tracking experiments investigated processing of unbounded dependency constructions.Experiment 1 employed sentences likeThat’s the garage/pistol with which the heartless killer shotthe man yesterday afternoon.Readers experienced greater processing difficulty in implausible sen-tences than in plausible sentences immediately after encountering the verbshot.This demonstratedthat they did not wait until the purported gap location afterman before forming the unboundeddependency. Experiment 2 considered sentences which locally appear to have an unbounded depen-dency that turns out to be incorrect. Data from this experiment demonstrated that readers formed theunbounded dependency immediately, even though they had to reanalyze later. However, there was noevidence that readers formed this unbounded dependency when it was rendered ungrammatical byisland-constraint information. We argue that the processor constructs unbounded dependencies in amanner that is maximally efficient from the point of view of incremental processing.© 1996 Academic

Press, Inc.

This paper is concerned with the processingof unbounded dependencies,as found in (1) be-low:

Which man do you believe Mary loves alot? (1)

Unbounded dependencies also occur in otherconstructions like relative clauses, topicaliza-tions andit-clefts. Sentences like (1) are of greatlinguistic interest, because they present a seri-ous challenge to simple accounts of the syntaxof language. The problem is that the phrasewhich manis a long way from the verbloves,even though they bear a close linguistic rela-tionship. Roughly,which manis the object ar-gument of loves. In most constructions, verbsare very close to their arguments, but in un-

bounded dependencies there is no limit to thenumber of words or clauses that can separatethem. Linguistic theories have to account forthis, preferably without making many additionaltheoretical assumptions.Unbounded dependencies pose an analogous

problem for theories of language comprehen-sion. The processor has to determine that theverb and its argument are related, even thoughthey are separated from each other in the sen-tence. In this paper, we present experiments thatinvestigate how this relationship is constructed.Our first experiment addresses the question ofwhether this relationship can be constructed assoon as the processor encounters the verb. Oursecond experiment asks whether the processordelays forming this relationship if the fragmentis ambiguous and it is locally unclear whetherthe relationship is correct or not. It also consid-ers whether the formation of this relationshipcan be blocked by constraints on possible un-bounded dependencies known as island con-straints.Unbounded dependency constructions con-

tain a phrase (likewhich manin [1]) that can bearbitrarily far from the word with which it isassociated. We call this phrase thefiller (Fodor,1978; Wanner & Maratsos, 1978). The wordwith which the filler is associated is often a verb(like lovesin [1]), but can also be a preposition;

The order of authorship is arbitrary. We gratefully ac-knowledge Melody Terras for her assistance in data collec-tion and analysis. We are particularly indebted to KeithEdwards for extremely capable technical assistance. Wealso acknowledge Simon Garrod, Simon Liversedge, DonMitchell, Keith Rayner, an anonymous reviewer, and themembers of the Human Communication Research CentreSentence Processing Group at the Universities of Edinburghand Glasgow for their advice and encouragement. This re-search was supported by ESRC Grant No. R000234542 anda British Academy Post-Doctoral Fellowship (both awardedto the second author). Address correspondence and reprintrequests to Martin J. Pickering, Human Communication Re-search Centre, Department of Psychology, University ofGlasgow, 56 Hillhead St., Glasgow, G12 9YR, UK.

JOURNAL OF MEMORY AND LANGUAGE 35, 454–475 (1996)ARTICLE NO. 0025

4540749-596X/96 $18.00Copyright © 1996 by Academic Press, Inc.All rights of reproduction in any form reserved.

in our examples, it will always be a verb. It isuncontroversial that the filler and the verb haveto be associated in some manner for the sen-tence as a whole to be interpreted, but it is lessclear how this association actually occurs.Sentence comprehension is essentially an in-

cremental process. Garden-path phenomena(Bever, 1970; Frazier, 1979; Frazier & Rayner,1982) indicate that syntactic processing occurswhile a sentence is encountered. Other experi-ments demonstrate that many aspects of seman-tic interpretation occur over sentence fragments(e.g., Altmann & Steedman, 1988; Garrod,Freudenthal, & Boyle, 1994; Marslen-Wilson,1973, 1975; Trueswell, Tanenhaus, & Garnsey,1994). The processor conducts a considerableamount of linguistic processing as it encounterseach new word. One might therefore expect thatthe processor associates the filler and the verbas soon as it encounters the verb and providesan interpretation for the sentence fragment atthis point.For this model to be correct, two different

components must hold true. First, the processormust be able to construct a simple link betweenthe filler and the verb. This may seem straight-forward, but in fact many theories of unboundeddependencies assume that the filler is only in-directly linked to the verb via an intermediaryelement known as agap.Second, the processorhas to choose to form this link between fillerand verb immediately, even if there is a chancethat this link is in fact incorrect. Many un-bounded dependency constructions are locallyambiguous; it is then impossible to be sure thatforming the link will turn out to be correct. Onthis model, the processor forms the link any-way. Below we discuss these two componentsin turn.

WHEN CAN THE UNBOUNDED DEPENDENCY

BE FORMED?

On the incremental account, the processormust always associate filler and verb as soon asit encounters the verb. Following Pickering(1993), we call thisImmediate Association.Theimmediate association hypothesis has tradition-ally not been adopted within psycholinguistictheory (e.g., Clifton & Frazier, 1989; Fodor,

1978). Most researchers in sentence processinghave assumed that the processor draws uponlinguistic information that is represented in amanner compatible with transformational gram-mar (e.g., Chomsky, 1965, 1981). Transforma-tional grammar assumes a quite complicated ac-count of the grammar of unbounded dependen-cies. It assumes that the filler occurs at its“canonical” location at an underlying level ofrepresentation (D-Structure). This location iswhere the filler is found in other sentence types,near to the verb. In (1), the underlying locationfor which manis immediately afterloves(cf. the“echo” questionYou believe Mary loves whichman a lot?). It is then moved to the beginning ofthe sentence by means of a transformation. Inmodern transformational grammar (Govern-ment-Binding Theory: Chomsky, 1981), thefiller leaves a gap (known as awh-trace) at itsunderlying location. This gap (Ø) is coindexedwith the filler, giving the following representa-tion:

[Which man]i do you believe Maryloves [Ø]i a lot? (2)

Transformational grammar clearly regards un-bounded dependency constructions as special;most other constructions do not involve trans-formations or gaps.Most accounts of the processing of un-

bounded dependencies draw upon transforma-tional grammar and assume that the mental rep-resentation of unbounded dependencies incor-porates gaps. For instance, Fodor (1978, 1989)defined the process of forming an unboundeddependency as the association of filler with gap.The termgap-filling is often treated as a neutraldescription of the process of unbounded depen-dency formation, even though it assumes thatthe transformational description of such con-structions is correct. We return to the relation-ship between processing accounts and linguisticanalysis in the General Discussion below.Experimental evidence is generally inter-

preted in terms of gap-filling. For example,there have been many demonstrations of the so-called filled-gap effect(Boland, Tanenhaus,Garnsey, & Carlson, 1995; Bourdages, 1992;Crain & Fodor, 1985; Frazier & Clifton, 1989;

PLAUSIBILITY AND UNBOUNDED DEPENDENCIES 455

Stowe, 1986). Localized difficulty occurs onusin (3b) compared with (3a) below (Stowe, 1986,Experiment 1):

My brother wanted to know ifRuth will bring us home to Mom atChristmas. (3a)

My brother wanted to know who Ruthwill bring us home to at Christmas. (3b)

In (3b), Stowe suggested that after the processorencountersbring, it inserts a gap right after it,and immediately fills the gap with the fillerwho.The wordus indicates that this analysis isnot correct, and so the processor performs im-mediate reanalysis, causing measurable disrup-tion.Following Fodor (1978), it is generally as-

sumed that gap-filling occurs in the followingway (remember that we are ignoring ambiguityfor the present). The identification of a fillercauses the processor to search for a gap. Whenit finds the location of the gap, it positions thegap and associates the filler with the gap. Onlythen can it interpret the filler as an argument ofthe verb. The unbounded dependency is there-fore not formed until the gap location isreached. Pickering (1993) calls this accountstandard gap-filling.Standard gap-filling contrasts with immedi-

ate association, where the filler is associatedwith the verb directly. Hence, standard gap-filling is incompatible with the simple incre-mental account discussed above. However, forthe examples that we have considered, immedi-ate association and gap-filling make the samepredictions. The reason is that the gap is adja-cent to the verb, and it is assumed that the gapcan be sited as soon as the verb is processed.Thus, previously reported evidence supportinggap-filling is also consistent with the immediateassociation hypothesis.Pickering and Barry (1991) demonstrated

that standard gap-filling and immediate associa-tion make different predictions when the pur-ported gap is not adjacent to the verb. Consider(4) below:

In which tin did you put the cakelast night? (4)

The “canonical” word order for (4) isyou putthe cake in the tin last night.Hence transforma-tional grammar assumes the representation in(5):

[In which tin]i did you put the cake[Ø] i last night? (5)

Standard gap-filling predicts that the filler is notassociated with the verb atput. Instead, the un-bounded dependency can only be formed afterthe processor has reachedcake and sited thegap.In contrast, immediate association predicts

that the processor links the filler and verb atput.Pickering and Barry (1991) provided evidence,based on intuitions about processing difficulty,for immediate association. In particular, theydemonstrated that sentences that contain nestedpatterns of associations according to standardgap-filling, but disjoint patterns according toimmediate association, are not hard to process,and do not behave like the nested constructionsdiscussed by Chomsky (1965). However, theyprovided no experimental evidence for immedi-ate association.To summarize, evidence from sentence pro-

cessing to date is compatible with two accountsof how the processor forms unbounded depen-dencies. Standard gap-filling predicts that theprocessor must wait until it reaches the locationof the gap before it can form the unboundeddependency. Immediate association predictsthat the processor can form the unbounded de-pendency as soon as it encounters the verb. Ifthe gap is not adjacent to the verb (as in [4]above), then the accounts make different pre-dictions about the time course of processing.

LOCAL AMBIGUITY IN UNBOUNDED

DEPENDENCYCONSTRUCTIONS

On the incremental account, the processorforms unbounded dependencies as soon as theverb is reached. If it is locally ambiguouswhether the filler forms an unbounded depen-dency with the verb, the account predicts thatthe processor will assume that the unboundeddependency is real, and will correct later if nec-essary. In this paper, we limit our investigationto the question of whether the processor ever

TRAXLER AND PICKERING456

assumes the unbounded dependency is real incases of local ambiguity; we do not seek toresolve the question of whether the processorforms the unbounded dependency under all con-ditions.In (1), repeated below, the processor may

consider the possibility that there is an un-bounded dependency betweenwhich manandbelieve:

Which man do you believe Maryloves a lot? (1)

Hence, this sentence contains a local ambiguity.If the processor forms this unbounded depen-dency, it will have to reverse its decision oncethe presence of the wordMary indicates thatwhich manis not an argument ofbelieve.If theprocessor forms unbounded dependencies assoon as possible, then it will be garden pathedon occasions like this. Intuitively, there is nostrong garden path in this instance, but experi-mental evidence (e.g., Frazier & Rayner, 1982;Rayner & Frazier, 1987) demonstrates that syn-tactic misanalysis can lead to processing disrup-tion without readers being aware of it.Fodor (1978) called this strategy thefirst-

resort strategy. In fact, her account assumedstandard gap-filling, but her strategies are alsoconsistent with immediate association. The pro-cessor forms the unbounded dependency assoon as it reaches the verb (or preposition). Incases like (1), the processor will be gardenpathed. The first-resort strategy contrasts withthesecond-resortstrategy, whereby the proces-sor only forms the unbounded dependency if itis not ruled out by the next constituent, and thelast-resortstrategy, whereby the processor doesnot form the unbounded dependency unless ithas no alternative (e.g., when it reaches the endof the sentence).The processor may make use of lexical infor-

mation, with the processor using a more zealousstrategy if the verb preferentially subcategorizesfor a (currently missing, i.e., non-subject) argu-ment of the same category as the filler. In thispaper, we shall not consider the issue of lexicalpreferences in detail, but rather shall ask wheth-er the processor ever uses a first-resort strategy.Experimental work provides suggestive evi-

dence for first-resort processing, at least withtransitive preference verbs. First, there is thefilled-gap effect, as discussed above. A problemis that most demonstrations involve self-pacedreading, which might not reflect the processesof normal reading. Pickering, Barton, andShillcock (1994) did find some evidence for afilled-gap effect in normal reading, but their re-sults may reflect the initial process of formingthe unbounded dependency rather than reanaly-sis (see the discussion of [14] below). Anotherproblem is that all experiments employed con-trol materials containing different constructionsfrom the experimental sentences. For instance,Stowe (1986) contrasted sentences containingan embeddedwh-question like (3b) above withcontrol sentences like (3a) that contain an em-beddedif-question. We simply do not know ifthe materials differ in ways irrelevant to thequestion of interest (e.g., with respect to thefrequency of use of the sentence types).Some evidence for first-resort processing

comes from studies that manipulate plausibility.Tanenhaus, Carlson, and Trueswell (1989) de-scribed a number of experiments that find plau-sibility effects as soon as the verb is reached, ina version of word-by-word self-paced readingwhere subjects monitor when the sentence stopsmaking sense (see also Boland, Tanenhaus, &Garnsey, 1990; Stowe, Tanenhaus, & Carlson,1991; and cf. Garnsey, Tanenhaus, & Chapman,1989, for comparable findings using event-related potentials). However, this method is fur-ther removed from normal reading than stan-dard self-paced reading. Using this task, Bolandet al. (1995) found that plausibility effects onlyoccur on the verb if the filler cannot plausiblyfill any argument associated with the verb. Forexample, subjects often found (6a) implausibleat the verb, but not (6b):

Which prize did the salesman visit . . .(6a)Which movie did your brotherremind . . . (6b)

In (6a),which prizehas to be the object ofvisit,so a reader can be certain that the full sentencewill be implausible. In contrast,remindtakes anNP object and an infinitival complement. If thesentence continuedBill to watch, for instance,


thenwhich moviewould not be the object ofremind,and the sentence would not be implau-sible. Boland et al.’s (1995) results suggest thatthe processor did not form the unbounded de-pendency in (6b). However, the task requiressubjects to determine whether the fragmentcould make sense as part of a complete sen-tence. We cannot be sure that similar effectswould occur in normal reading.This problem is particularly relevant with re-

spect to Boland et al.’s (1995) Experiment 4.They found that subjects often judged (7) belowimplausible before reaching the prepositionto:

Which public library did John donatesome cheap liquor to last week? (7)

The processor must have decided that people donot donate liquor to libraries before encounter-ing to. Hence it must have formed the un-bounded dependency before reaching the prepo-sition. This conclusion is incompatible withboth standard gap-filling and immediate asso-ciation. However, the task may precipitate un-usual processing strategies unconnected withparsing. Subjects may well reason that the frag-ment Which public library did John donatesome cheap liquor . . . is unlikely to be the be-ginning of a plausible sentence, and thus con-clude that it does not make sense. They may, forinstance, readliquor and then anticipate that thenext word will be the prepositionto, and maketheir judgment on this basis. In addition, thisresult depends crucially on subjects’ not havingtreateddonateas taking two NP objects, as inWhich campus party did John donate the cheapliquor? This sentence is marginally acceptablefor some speakers of American English, includ-ing one of the authors (M.T.).Finally, cross-modal priming in speech

(Nicol & Pickering, 1993; Nicol & Swinney,1989) supports a first-resort strategy. Lexicaldecision to an associate of the filler is sped upimmediately after an unbounded dependency isformed. This suggests that the unbounded de-pendency is formed rapidly. However, these re-sults have been questioned by McKoon, Rat-cliff, and Ward (1994; McKoon & Ratcliff,1994; but cf. Nicol, Fodor, & Swinney, 1994).

In conclusion, there is some evidence that theprocessor employs a first-resort strategy withtransitive-preference verbs. The evidence doesnot determine whether this strategy is regularlyemployed during normal reading. Our two ex-periments both address this question. However,Experiment 1 is primarily concerned with thequestion of whether the processor employs im-mediate association or standard gap-filling, andthe sentences employed probably do not containany relevant local ambiguity. Thus, it onlytested the claim that the processor can use afirst-resort strategy when there is no local am-biguity. Experiment 2 tested the claim that afirst-resort strategy is employed when the pos-sible unbounded dependency turns out to be in-correct.Both experiments investigated processing of

unbounded dependency constructions by ma-nipulating plausibility. More specifically, wemanipulated whether the sentence fragment atthe verb would be plausible or implausible if theunbounded dependency had been formed. Thisallows us to monitor processing without havingto compare sentences containing different con-struction types (cf. Pickering & Traxler, 1996).Of course, any effects of plausibility wouldshow that the unbounded dependency has beensemantically interpreted as well as syntacticallyanalyzed.

EXPERIMENT 1

Experiment 1 addressed the question ofwhether the processor performs immediate as-sociation or standard gap-filling. We were in-terested in the processes involved in normalreading, and hence tracked eye movements. Wewished to determine whether the processorforms the unbounded dependency on encoun-tering the verb, or whether it waits until the gaplocation, and therefore manipulated the plausi-bility of the experimental materials (see Bolandet al., 1995; Garnsey et al., 1989). We employedmaterials like (8a)–(8d):

That’s the pistol with which theheartless killer shot the haplessman yesterday afternoon. (8a)


That’s the garage with which theheartless killer shot the haplessman yesterday afternoon. (8b)

That’s the garage in which theheartless killer shot the haplessman yesterday afternoon. (8c)

That’s the pistol in which theheartless killer shot the haplessman yesterday afternoon. (8d)

Clearly, (8a) and (8c) make sense, whereas (8b)and (8d) are implausible (or semanticallyanomalous). Sentences (8c) and (8d) controlledfor possible low-level lexical effects due tocombinations ofpistol, garage,andshot.The “canonical” sentence related to (8a) is

The heartless killer shot the hapless man withthe pistol yesterday afternoon.Hence, the gapassociated with the extracted phrasewith whichis located afterman.Sentence (8b) is similar. In(8c) and (8d), the extracted phrase is an adjunct,andshotdoes not subcategorize for it. It is pos-sible that the gap location is afteryesterday af-ternoonin these sentences, as inThe killer shotthe man yesterday afternoon in the garage.However, the gap location is available afterman,as the sentence is incrementally processed,so first-resort standard gap-filling would site thegap right aftermanis reached. In all stimuli, thefinal phrase is a temporal expression, and theextracted adjunct refers to a location. In En-glish, there is a clear preference for locativeadjuncts to precede temporal ones. This sup-ports the assumption that the gap should be be-tweenmanandyesterday.An obvious prediction is that (8b) and (8d)

will take longer to read than (8a) and (8c). Themore interesting issues are when and where theplausibility effect occurs. A plausibility effectcould appear as soon as the verbshot is pro-cessed only if the processor uses immediate as-sociation and a first-resort strategy, interpretsthe sentence fragment immediately, and inte-grates that interpretation with general knowl-edge. Finally, this result would indicate that theinterpretation of sentence fragments has veryrapid effects on eye movements during normalreading. In contrast, standard gap-filling pre-dicts no plausibility effect before the wordman

is encountered. A plausibility effect would ap-pear aroundman yesterdayif the processor usesstandard gap-filling and a first-resort strategy,together with immediate interpretation of thefragment.

Method

Subjects.Twenty-four normally sighted En-glish speakers were paid to participate. Somehad taken part in other eye-tracking experi-ments.Stimuli.We constructed 28 sets of four sen-

tences for Experiment 1 like (8a)–(8d): see Ap-pendix A. A prepositional phrase served as anadjunct or argument of the embedded verb in allmaterials. The plausibility manipulation in-volved two different head nouns (e.g.,pistolandgarage) crossed with two prepositional phrases(e.g.,in whichandwith which). The head nounswere matched for length between plausibilityconditions. Each noun appeared in both a plau-sible and an implausible sentence. Differencesin word frequency, therefore, could not produceany differences between the plausibility condi-tions. Note that all regions of interest were iden-tical across sentences within a material set, andthat the words in the plausible sentences wereidentical to those in the implausible sentences,though they appeared in different combinationsin the individual conditions. One version ofeach material appeared in each of four lists ofitems. All sentences were presented on twolines, with the line break after the wordwhich.Regions.For analysis, we defined one region

as the verb (e.g.,shot), because plausibility ef-fects could occur there if readers performedfirst-resort immediate association. If readersalso interpreted the fragment immediately, thena plausibility effect could appear in a measureof initial processing. We definedman yesterdayas another region, because plausibility effectscould appear there if readers performed stan-dard gap-filling. Defining the region in this wayprovided the best chance of finding an effectaround the purported gap location. The otherregions resulted from the positioning of the linebreak and the two critical regions. Note that allregions included the character spaces before thefirst word in the region.


Norming.To assess the plausibility of the as-sociation between the head noun (e.g.,garage/pistol) and the verb (e.g.,shot), 20 raters read112 questions like (9):

With which pistol did the killershoot the man? (9)

Raters wrote down a number between zero andseven that indicated how much sense each ques-tion made. We eliminated sets of items whenany item produced a mean rating between 2.0and 5.0.Procedure.An SRI Dual Purkinje Generation

5.5 eye-tracker monitored subjects’ eye move-ments. The tracker has angular resolution of 10‘arc. The tracker monitored only the right eye’sgaze location. A PC displayed materials on aVDU 70 cm from subjects’ eyes. The VDU dis-played four characters per degree of visualangle. The tracker monitored subjects’ gaze lo-cation every millisecond and the softwaresampled the tracker’s output to establish the se-quence of eye fixations and their start and finishtimes.Before the experiment started, subjects read

an explanation of eye-tracking and a set of in-structions. The instructions told them to read attheir normal rate and comprehend the sentencesas well as they could. The experimenter thenseated the subject at the eye-tracker and usedbite bars and forehead restraints to immobilizethe subject’s head. Next, subjects completed acalibration procedure. Before each trial, a small“+” symbol appeared near the upper-left-handcorner of the screen. Immediately after subjectsfixated the “+” symbol, the computer dis-played a target sentence, with the first characterof the sentence replacing the “+” on the screen.The “+” symbol also served as an automaticcalibration check, as the computer did not dis-play the text until it detected stable fixation onthe “+” symbol. If readers could not fixate the“+” symbol, the experimenter recalibrated theeye-tracker. When subjects finished readingeach sentence, they pressed a key, and the com-puter either displayed a comprehension ques-tion (e.g.,Did the young man hit the ball?), onabout half of the trials, balanced across condi-tions, or proceeded to the next trial. Half of

these questions had “yes” answers, half had“no.” Subjects received no feedback on theiranswers. Subjects responded to the questions bypressing a button. After subjects completedeach quarter of the experiment, the experi-menter recalibrated the equipment and subjectshad a short break. Thus, the eye-tracker wasrecalibrated a minimum of four times during theexperiment and usually more.The computer displayed each experimental

list in a fixed random order together with 28sentences likeThe doctor explained to the han-dler that the tiger attacked and wounded severalantelope before it was recaptured,and 15 addi-tional fillers of various syntactic types. All sen-tences were displayed on two lines of text.Analyses.An automatic procedure pooled

short contiguous fixations. The procedure incor-porated fixations of less than 80 ms into largerfixations within one character, and then deletedfixations of less than 40 ms that fell within threecharacters of any other fixation. FollowingRayner and Pollatsek (1989), we presume thatreaders do not extract much information duringsuch brief fixations. Before analyzing the eye-movement data, we eliminated the occasionaltrial when the subject failed to read the sentenceor when tracker loss ensued. More specifically,we removed trials where two or more adjacentregions had zero first-pass reading time. Thisprocedure removed 5.2% of the data.First-pass reading timeis the sum of the fixa-

tions occurring within a region before the eyeleft the region. If the eye fixated a point beyondthe end of a region before landing in the regionfor the first time, then the first-pass time for thatregion was zero.Total reading timeis the sumof all fixations in a region. Our main analysesincluded 0-ms fixation times. Subsidiary analy-ses comprised aneye–gazemeasure for first-fixation and first-pass reading time. If the targetword was not directly fixated, then the closestfixation within four character spaces to the leftof the region boundary, but before the first fixa-tion after the target word, was counted as thefixation during which the word was processed.If no such fixation occurred, then the trial waseliminated. Research on perceptual span dem-onstrates that when readers do not fixate on a


target word, they identify it on the prior fixationwhen that fixation lies within three or four char-acters of the target (Rayner & Pollatsek, 1989).1

Results

Table 1 represents subjects’ mean first-fixation times in the critical regions (e.g.,shotandman yesterday). Figure 1 represents sub-jects’ mean first-pass reading times. Figure 2represents subjects’ mean total reading times.The mean first-fixation time on the verb was21.5 ms longer in implausible sentences than inplausible sentences [F1(1,23)4 4.23,p < .05,MSe4 1315;F2(1,27)4 8.24,p < .01,MSe4960]. Mean first-pass reading time on the verbwas 4.3 ms/character longer in implausible sen-tences [F1(1,23)4 4.55,p < .05,MSe 4 48;F2(1,27)4 10.68,p < .01,MSe 4 29]. Meantotal reading time on the verb was 13.1 ms/character longer in the implausible sentences[F1(1,23) 4 15.54, p < .001, MSe 4 132;F2(1,27)4 18.90,p < .001,MSe 4 144]. Thispattern of reaction times suggested that readersformed an association between the filler and theverb as soon as they encountered the verb.The analyses revealed no statistically signifi-

cant differences between plausible and implau-sible sentences on any of the reading time mea-sures in the regionman yesterday,where thegap would be [on first fixation,F1(1,23)4 1.30andF2 < 1; on first pass, bothF < 1; on totaltime, bothp > .25]. No statistically significantdifferences between plausible and implausiblesentences occurred in the region preceding theverb in the first-pass or first-fixation measures.A subsidiary analysis revealed that readers

skipped the verb during their first pass throughthe sentence more often in plausible sentencesthan in implausible sentences. Readers skipped

the verb 3.1 times out of 14 when the sentencewas plausible, and 2.2 times out of 14 when thesentence was implausible [F1(1,23)4 7.27,p <.05,MSe4 1.4;F2(1,27)4 10.00,p < .01,MSe4 .87].Because of this difference in fixation prob-

ability on the verb, we performed a second setof analyses using the eye-gaze measure (seeabove). We removed an additional 1.0% of thedata, for trials when there was no fixation oneither the verb or the four characters before theverb during first pass. Eye–gaze analyses re-vealed an effect of plausibility in the first passeye–gaze measure [F1(1,23)4 3.67, p < .07,MSe 4 1742;F2(1,23)4 18.98,p < .001,MSe4 456]. The effect was only significant byitems for the first-fixation eye–gaze measure[F1(1,23) 4 1.68, p > .20, MSe 4 1188;F2(1,27)4 5.07,p < .05,MSe 4 714].

Discussion

Experiment 1 demonstrated processing diffi-culty around the verb on first-fixation and first-pass measures in the implausible conditioncompared with the plausible condition. This in-dicated that the processor formed the un-bounded dependency immediately after it en-countered the verb. Hence the processor usedthe first-resort immediate association strategy(Pickering, 1993) and immediately interpretedthe fragment on this analysis. Therefore, imme-diate association must be a general characteris-tic of the sentence processor.More generally, the results showed that the

processor can employ a first-resort strategywhen the unbounded dependency is not locallyambiguous. These results do not determine

1 Our main analyses divided the raw fixation durations bythe number of characters in the region to produce a milli-second per character dependent measure. This procedureproduces similar weightings for items of different lengths.Because our critical regions were equally long across con-ditions, we avoid problems with ms/character transforma-tion identified by Trueswell et al. (1994). We did not per-form the ms/character transform for the eye–gaze analyses,because the eye–gaze fixation procedure samples regions ofdifferent lengths for the same item.

TABLE 1EXPERIMENT 1: MEAN-FIRST FIXATION TIME BY SENTENCE

TYPE AND REGION (ms)

Region

Sentence 5 (e.g.,mantype 3 (e.g.,shot) yesterdar)

Plausible 196 237Implausible 217 248

Note.Example sentence: That’s the pistol with which theheartless killer shot the hapless man yesterday afternoon.


whether the processor always uses a first-resortstrategy.The results are incompatible with standard

gap-filling, because standard gap-filling pre-dicts no plausibility effects before the purportedgap location. In addition, the data do not supportany of three other models: (1) The processorperformed association in two stages, at the verband at the gap (Nicol, 1993); (2) the processorsometimes performed association at the verb,and sometimes at the gap; and (3) the processorwaited until a potential end-of-sentence beforeforming an unbounded dependency (cf. Bourd-ages, 1992).Experiment 1 provided good evidence for the

process of incremental interpretation and its ef-

fects on eye movements. The processor detectedthe implausibility of the sentence fragment assoon as it reached the verb. Notice that our ma-nipulation of the preposition rules out any low-level explanation of our effects in terms of lexi-cal associations between the noun in the fillerand the verb.Somewhat surprisingly, Experiment 1 pro-

duced a difference in first-pass fixation prob-ability on the verb between the plausible andimplausible sentences. The skipping effect byitself provided evidence for differences in pro-cessing of plausible and implausible sentencesassociated with the unbounded dependency.The most likely explanation for this effect isthat readers sometimes processed the verb re-

FIG. 2. Experiment 1: Mean total reading time by region and condition. Region 3 corresponds to the wordshotin the example sentence. Region 5 corresponds to the wordsman yesterdayin the example sentence.

FIG. 1. Experiment 1: Mean first-pass reading time by region and condition. Region 3 corresponds to the wordshot in the example sentence. Region 5 corresponds to the wordsman yesterdayin the example sentence.


gion while fixating the end of the precedingword. We therefore predicted that significantdifferences between the conditions wouldemerge on the fixated trials if the regions wereextended to include the end of the precedingword. This prediction was borne out by the eye–gaze analyses.

LOCAL AMBIGUITY AND ISLAND CONSTRAINTS

Our next question is whether the processorcan still employ a first-resort strategy when thepossible unbounded dependency may turn outto be erroneous. Previous research (see above)suggested that the processor can use a first-resort strategy in such cases, but it has not dem-onstrated that this strategy is used in normalreading. A demonstration of this would providestrong evidence for incremental processing insentence comprehension.Consider (9) below:

We like the book that the authorwrote unceasingly and with greatdedication about while waitingfor a contract. (9)

This sentence contains a real unbounded depen-dency between the fillerthe book thatand theprepositionabout.However, at the verbwrote,there is a possible unbounded dependency be-tween the filler and the verb. In Fodor’s (1978)terms, there is a “doubtful gap” afterwrote.Thefirst resort strategy predicts that the processorwill form the possible unbounded dependencybetween the filler and the verb atwrote. In thiscase, this analysis is plausible, because an au-thor is likely to write something (e.g., a book).When the processor reaches the prepositionabout, it realizes that this analysis cannot becorrect (i.e., that it has been garden pathed). InExperiment 2 below, we manipulated the plau-sibility of this possible unbounded dependency.Our final question concerns the processing of

unbounded dependencies that are in fact un-grammatical. Consider (10) below:

*Which fish did the woman whocooked eat rice? (10)

Descriptively, (10) is ungrammatical because it

is impossible to form an unbounded depen-dency between a filler (here,which fish) and averb that is embedded within a relative clausemodifying a subject (here,cooked). The un-grammaticality of (10) is perhaps surprising inthat the sentence would have a clear meaning,as expressed by the grammatical echo questionThe woman who cooked which fish ate rice?Ross (1967) said that subject relative clauseslike who cooked which fishareislands,becauseit is impossible to “move” a filler out of such aclause. Hence the grammar encodesisland con-straint information, which blocks sentences thatwould otherwise be grammatical. Note that sen-tence (10) is astrong island, because it is com-pletely unacceptable. There are alsoweak is-lands, of marginal grammaticality, like (11) be-low (e.g., Szabolsci & Zwarts, 1993):

?Which man did you read no book about? (11)

We focus on strong islands in this paper.What does the processor do when it encoun-

ters a possible unbounded dependency that infact violates an island constraint? Such a situa-tion occurs in (12) below:

We like the book that the authorwho wrote unceasingly and withgreat dedication saw while waitingfor a contract. (12)

If there were no island constraints, then (12)would have an ambiguity like (9) above. Afterwrote, there might be an unbounded depen-dency betweenwrote andbook (as inThe au-thor who wrote the book . . .).It is possible that the use of the island con-

straint information isdelayed.If, in addition,the processor uses a first-resort strategy (withtransitive-preference verbs), then the un-bounded dependency will be formed atwrote.At some later point, the processor would realizethat this was the wrong analysis, and undo theunbounded dependency. It is not completelyclear when this would occur. One possibility isthat it would undo the unbounded dependencywhen it reached the main verbsaw, becausenow the filler might associate with this verb.Alternatively, the processor might undo the un-bounded dependency earlier, while processing


unceasingly and with great dedication.In eithercase, the processor would eventually reach thecorrect analysis, where the filler associates withsaw,andwrote is intransitive.On the other hand, the processor might use

island constraint informationimmediately.If so,it would never form the illegal unbounded de-pendency. If the processor used a first-resortstrategy with immediate use of island constraintinformation, then it would form the legal un-bounded dependency in (9) but not the illegalunbounded dependency in (12).Assuming that the processor attempts to per-

form sentence comprehension in as incrementala manner as possible, we predict that the pro-cessor adopts a first-resort strategy, at least withtransitive-preference verbs, even when the pos-sible unbounded dependency may turn out to bea garden path. Hence we predict immediate for-mation of the unbounded dependency in (9)above. However, it would not be efficient toconsider an ungrammatical analysis, even mo-mentarily, and so we predict that the processoremploys strong island constraint informationimmediately. Similarly, it should simply fail toparse sentences like (10), and never entertainthe possibility that they might be grammatical.On this basis, the processor would never con-

sider the illegal unbounded dependency in (12).But linguistic considerations have suggestedthat the processor might delay use of island con-straint information. It might initially constructillegal unbounded dependencies, only to rulethem out later. In transformational grammar,sentences that violate island constraints are gen-erated by one component of the grammar, butare filtered out by another component (seeChomsky, 1981). As with the issue of gaps, lin-guistic theory suggests the possibility that theprocessor might employ a parsing strategy thatappears inefficient from the point of view ofincremental interpretation.Freedman and Forster (1985) proposed that

the processor initially “overgenerates” sen-tences like (10), and then rules them out byprocesses corresponding to the application ofthe island constraints. They claimed that theprocessor computes a level of representation atwhich it treats sentences that violate island con-

straints as grammatical. They employed a sen-tence-matching task, and found that readers’ re-sponses to sentences that violated island con-straints closely resembled their responses togrammatical sentences. Their account fits wellwith the delayed use of island constraint infor-mation. However, Crain and Fodor (1987) ar-gued that their results could be explained interms of how easy it was to correct the sen-tences into grammatical sentences (see also For-ster & Stevenson, 1987; Stowe, 1992).There is little on-line experimental work on

the processing of island constraint information,and what there is seems to conflict (see Fodor,1989; Pickering et al., 1994). One word-by-word self-paced reading experiment found nofilled-gap effect in (13) below (Stowe, 1986,Experiment 2):

The teacher asked if the silly storyabout Greg’s older brother wassupposed to mean anything. (13a)

The teacher asked what the sillystory about Greg’s older brotherwas supposed to mean. (13b)

In (13b), island-constraint information pre-cludes the unbounded dependency betweenwhat and about which could otherwise beformed when the processor reachesabout (cf.the echo questionThe silly story about whatmeant very little?). No increase in reading timeoccurred atGreg’s in (13b) compared with(13a), though a filled-gap effect did appear incontrol sentences (and in her Experiment 1: See[3] above). This result supported the immediateuse of island constraint information. In contrast,Pickering et al. (1994) found some tentative evi-dence for a filled-gap effect in island environ-ments using eye tracking:

I realize what the artist who paintedthe large mural ate today. (14a)

I realize that the artist who paintedthe large mural ate cakes. (14b)

First-pass time onpainted thewas greater in(14a) than (14b). This may have been due to theprocessor forming an illegal unbounded depen-dency betweenwhat andpaintedand then un-doing it when it reachedthe. However, these


results might also simply have demonstratedthat the verbpainted showed that the correctanalysis of the unfolding sentence was complex.Clifton and Frazier (1989) found evidence for

delayed use of island-constraint information,using an end-of-sentence grammatically judg-ment task. Subjects responded to (15b) morequickly than to (15a):

What did John think the girl whoalways won received? (15a)

What did John think the girl whoalways excelled received? (15b)

The verbexcelledis intransitive, and so cannothave an extracted object, butwoncan have onein principle, though in this environment the un-bounded dependency is ruled out by island-constraint information. The finding of difficultywith (15a) suggested that subjects did try totreatwhat as the object ofwon.However, it isalso possible that their finding reflects the factthatexcelledis frequently used transitively, un-like won, and hence that (15b) violates verbpreferences. The results of on-line research areas controversial as those produced by sentence-matching.In conclusion, Experiment 1 and other ex-

perimental work provided some evidence thatthe processor uses a first-resort strategy withtransitive-preference verbs at least. However,we do not know whether the processor uses thisstrategy in normal reading when the unboundeddependency may turn out to be erroneous. It isalso unclear whether the processor employs is-land constraint information immediately duringthe formation of unbounded dependencies, orwhether the use of this information is delayed.Considerations of incrementality and efficiencysuggest that the processor would adopt a first-resort strategy if the unbounded dependencywere possible, but would not construct an un-bounded dependency if it were ruled out bystrong island-constraint information. Experi-ment 2 tested these predictions.

EXPERIMENT 2

Experiment 2 considered the processing ofunbounded dependencies that initially appear to

be possible, but are subsequently shown to bewrong. We contrasted these potential un-bounded dependencies with sentences wherethere was a potential unbounded dependencyonly if the processor initially ignored strong is-land-constraint information. We tested whetherreaders initially formed unbounded dependen-cies if they were possibly erroneous and if theywere ruled out by strong island constraints. Asbefore, we manipulated the plausibility of thepotential unbounded dependency, by using ma-terials like (16):

WAITING FOR APUBLISHING CONTRACTThe big city was a fascinatingsubject for the new book. (16)

We like the book that the authorwrote unceasingly and with greatdedication about while waitingfor a contract. (16a)

We like the city that the authorwrote unceasingly and with greatdedication about while waitingfor a contract. (16b)

We like the book that the authorwho wrote unceasingly and with greatdedication saw while waitingfor a contract. (16c)

We like the city that the authorwho wrote unceasingly and withgreat dedication saw whilewaiting for a contract. (16d)

To improve readability, we introduced a contextsentence and a title. The context produces ashort coherent discourse and satisfies the refer-ential presuppositions of the experimental sen-tence.We manipulated plausibility in thenon-

island conditions (16a and 16b) by varying theextracted noun (e.g.,book/city). We matchedthe two nouns for length and frequency. If read-ers formed an unbounded dependency betweenthe nouns and the verbwrote,then the resultinginterpretation would be plausible for (16a) andimplausible for (16b). The introductory sen-tence mentions both the plausible and implau-sible nouns from the target sentences (e.g.,book


and city). Half of the introductory sentencesmentioned the plausible noun first, and halfmention the implausible noun first.Theislandconditions (16c and 16d) contain a

potential unbounded dependency within an is-land constraint environment (as describedabove). If readers ignore island constraint infor-mation during initial parsing, then (16c) con-tains a more plausible filler–verb associationthan (16d).If we assume that readers take longer to pro-

cess implausible associations than plausible as-sociations (cf. Pickering & Traxler, 1996), thenwe can make predictions about processing ofsentences (16a–16d). Consider first sentences(16a) and (16b), the non-island sentences. Ifreaders use a first-resort strategy, and thereforeform associations between the filler (e.g.,book/city) and the verb (e.g.,wrote) immediately af-ter reaching the verb, then a difference betweensentences with plausible and implausible asso-ciations should appear around the verb in ameasure of initial processing. Next, considersentences (16c) and (16d), which contain island-constraint information. If readers use a first-resort strategy, and ignore island-constraintinformation during the early stages of parsingsentences, then a similar plausibility differenceshould appear around the verb (e.g.,wrote) in ameasure of initial processing. If island-constraint information is used immediately dur-ing parsing, then no plausibility effect shouldappear around the verb, because the island-constraint information rules out any associationbetween the filler (e.g.,book/city) and the verb(e.g.,wrote).Next consider the process of recovery from

misanalysis. Pickering and Traxler (1996) ar-gued that readers have more difficulty abandon-ing a misanalysis if that misanalysis is plausiblethan if it is implausible, because readers seman-tically commit to a plausible analysis more thanan implausible analysis. In the non-island sen-tences (16a) and (16b), the disambiguating re-gion isabout while.If readers misanalyze thesesentences, then they should take longer to pro-cess this region when the misanalysis is plau-sible than when the misanalysis is implausible.If readers misanalyze the island sentences, then

similar effects should occur when the island-constraint information is brought to bear. Wesuggested that this would be most likely to hap-pen soon after the wordsaw is reached. How-ever, reanalysis might occur during the wordsimmediately beforesaw.

Method

Subjects.Thirty-two normally sighted nativeEnglish speakers were paid to participate. Somesubjects had taken part in other eye-trackingstudies.Stimuli. We constructed 28 sets of experi-

mental sentences, together with titles and intro-ductory sentences.2 Only the extracted noun dif-fered between the plausible and implausiblesentences. The island and non-island sentencesdiverged in the final two regions (to differentdegrees in different material sets). We then con-structed four lists of items, with exactly oneversion of each item appearing in each list.Norming. To assess the plausibility of the

ambiguous NP as direct object of the verb in therelative clause, 20 raters read a typewritten listof 104 sentences (e.g.,The author wrote thebook). About half of the sentences related to thisexperiment and half related to a different study.To ensure that the ultimately correct reading ofthe entire experimental sentences was equallyplausible between conditions, some of our sen-tences were “detransformed” versions of the ex-perimental sentences (e.g.,The author wroteabout the book). Raters wrote down a numberbetween zero and seven that indicated howmuch sense each sentence made. We eliminatedpairs of experimental items when either itemproduced a mean rating between 2.0 and 5.0.We found no significant differences betweenconditions in the “detransformed” versions ofthe sentences. Hence, the conditions differed inplausibility on the misanalysis, but did not dif-fer on the ultimately correct analysis.Procedure.The data collection procedures

were identical to those in Experiment 1. All theexperimental sentences were displayed on two

2 All materials can be obtained from the authors on re-quest.


lines, and were split within the region after theverb (e.g.,wrote) and before the disambiguatingpreposition (e.g.,about). At least two wordsseparated the verb from the line break and atleast two words followed the line break beforethe disambiguating preposition or verb.The 28 experimental materials were dis-

played along with 6 two-sentence filler passagesand 28 two-sentence passages [like (17)] fromanother experiment in a pseudo-random order.

The janitor polished the bronzestatues of/for the old maths professorthat the principal hated and thedean of the art school. As thejanitor polished(,) the professorthat the principal hated reviewedthe spring term teaching schedule. (17)

Analyses.In the first stage of analysis, wedetermined which line of text subjects read.This involved some judgment, because sub-jects’ eyes occasionally landed between twolines of text, and because slight head move-ments sometimes caused small, systematicchanges in the recording of fixation location. Inthe vast majority of cases, determining whichline readers actually fixated caused no difficultywhatsoever.The regions that we analyzed,wrote unceas-

ingly, and with great dedication,and aboutwhile/saw while,were identical in the plausible-no island and implausible-no island conditions,and in the plausible-island and implausible-island conditions.3 We removed one item from

the analyses because of an error in the display.We also removed 5.1% of the data, according tothe same criteria as Experiment 1.Our main analyses included 0-ms fixations

that occurred when readers skipped a region.We performed a second set of analyses afterremoving 0-ms fixations. The results of this sec-ond set of analyses matched the results of thefirst set almost exactly, so we do not reportthese analyses.

Results

Table 2 represents subjects’ mean first-fixation times for regions 3, 4, and 5 for bothnon-island and island sentences. Figure 3 rep-resent mean first-pass reading time and Fig. 4mean total reading time for the non-island sen-tences. Figure 5 represents mean first-passreading time and Fig. 6 mean total reading timefor the island sentences. The data were sub-jected to 2 (island: island vs non-island con-straint sentences) × 2 (plausibility: plausible vsimplausible misanalysis) × 2 (region: 3 vs 5)ANOVAs. Statistically significant three-wayinteractions demonstrated that the patterns offixation times differed between island and non-island sentences. Interactions of island con-straint, plausibility, and region appeared in thefirst-fixation [F1(1,31)4 7.37,p < .01,MSe 41615;F2(1,26)4 7.59,p < .01,MSe 4 1225]

3 As in Experiment 1, we employed the ms/character

transform. Note that the regionabout while in the islandconditions differs from the comparable regionsaw whileinthe island conditions, but we make no direct comparisonbetween these regions.

TABLE 2EXPERIMENT 2: MEAN FIRST FIXATION TIMES BY SENTENCETYPE AND REGION (ms)

Region

3 (e.g.,wrote 4 (e.g.,and with 5 (e.g.,about while/Sentence type unceasingly) great dedication) saw while)

Non-island, plausible 222 205 238Non-island, implausible 255 219 223Island, plausible 233 210 243Island, implausible 223 209 239

Note.Example of non-island sentence: We like the book that the author wrote unceasingly and with great dedication aboutwhile waiting for a contract. Example of island sentence: We like the book that the author who wrote unceasingly and withgreat dedication saw while waiting for a contract.


and total reading time data [F1(1,31)4 11.17,p < .01,MSe 4 251;F2(1,26)4 8.53,p < .01,MSe 4 268]. This three way interaction wasweaker in the first-pass reading time data[F1(1,31)4 3.29,p < .08,MSe 4 63;F2(1,26)4 2.02, p > .16,MSe 4 74]. The non-islandsentences produced interactions of plausibilityand region on first fixation, first-pass, and totalreading time. In the first-fixation time data, theimplausible sentences’ 32.7-ms disadvantage inthe regionwrote unceasinglybecame a 15.4-msadvantage in the regionabout while[F1(1,31)4 11.81,p < .01,MSe 4 1569; F2(1,26) 412.06,p < .01,MSe 4 1311]. In the first-passreading time data, the implausible sentences’5.3 ms/character disadvantage at the verb be-came a 1.5 ms/character advantage at the prepo-sition [F1(1,31) 4 3.80, p < .06,MSe 4 97;F2(1,26)4 4.77, p < .05,MSe 4 65]. In the

total reading time data, a 15.6 ms/character dis-advantage became a 9.3 ms/character advantage[F1(1,31) 4 16.89, p < .001, MSe 4 294;F2(1,26)4 12.50,p < .01,MSe 4 334]. Datafrom the island sentences produced no such in-teractions on any of the reading time measures(all F < 1).To determine if and when readers formed the

filler–verb association, we subjected data fromregion 3 (e.g.,wrote unceasingly) to 2 (plausi-bility ) × 2 (island constraint) ANOVAs. Theseanalyses produced statistically reliable interac-tions of plausibility and island constraint in thefirst-fixation data [F1(1,31) 4 6.97, p < .01,MSe4 2157;F2(1,26)4 8.87,p < .01,MSe41388]. In the non-island sentences (e.g., [16a] &[16b]), readers’ initial fixations in the verb re-gion (e.g.,wrote unceasingly) were longer whenthe filler–verb association was implausible than

FIG. 4. Experiment 2: Mean total reading time by region and condition for non-island sentences. Region 3corresponds to the wordswrote unceasinglyin the example sentence. Region 5 corresponds to the wordsaboutwhile in the example sentence.

FIG. 3. Experiment 2: Mean first-pass reading time by region and condition for non-island sentences. Region3 corresponds to the wordswrote unceasinglyin the example sentence. Region 5 corresponds to the wordsaboutwhile in the example sentence.


when it was plausible [F1(1,31)4 7.94,p< .01,MSe 4 2157;F2(1,26)4 10.68,p < .01,MSe4 1388]. In sentences where island-constraintinformation ruled out the filler–verb association(e.g., [16c] and [16d]), no plausibility effect ap-peared in the verb region (bothF < 1). Simi-larly, we found no plausibility effect in region 4(e.g., and with great dedication) on first-fixation, first-pass, or total reading time (allF <1). We also found no plausibility effects in re-gion 5 (e.g.,saw while) on first-fixation time(both F < 1), first-pass reading time (bothp >.1), or total reading time (bothp > .2). Ourresults, then, suggest that readers consider thepotential (but inappropriate) filler–verb associa-tion immediately when they encounter the verbin sentences like (16a) and (16b) where island-constraint information does not rule out that as-sociation. By contrast, the data provide no evi-dence that readers consider the potential (but

inappropriate) filler–verb association at anytime during processing of sentences like (16c)and (16d) where island-constraint informationdoes rule out that association.We also found some evidence for a difference

between the plausible and implausible sen-tences in region 5 (e.g.,about while). Here,reading times for the implausible sentenceswere generally faster than reading times for theplausible sentences. This difference approachedstatistical significance in the total-time data,where readers spent 9.3 ms/character longerprocessing the regionabout whilein the plau-sible sentences [F1(1,31)4 5.02,p < .05,MSe4 275;F2(1,26)4 2.14,p > .15,MSe 4 563].

Discussion

The three-way interaction of island con-straint, plausibility, and region demonstratedthat readers’ responses to sentences where the

FIG. 5. Experiment 2: Mean first-pass reading times by region and condition for island sentences. Region 3corresponds to the wordswrote unceasinglyin the example sentence. Region 5 corresponds to the wordssawwhile in the example sentence.

FIG. 6. Experiment 2: Mean total reading times by region and condition for island sentences. Region 3corresponds to the wordswrote unceasinglyin the example sentence. Region 5 corresponds to the wordssawwhile in the example sentence.


unbounded dependency is possible differ fromtheir responses to sentences where the un-bounded dependency is ruled out by strong is-land-constraint information. The interaction ofplausibility and island constraint in region 3(e.g.,wrote unceasingly) on first fixation indi-cated that readers’ responses to island and non-island sentences began to differ as soon as theunbounded dependency could be formed. Thesimple effects demonstrated that readers formedthe unbounded dependency according to thefirst-resort strategy in the non-island sentences.There was no suggestion that readers formed theunbounded dependency in the island sentences.Thus, Experiment 2 provides clear evidence

for first-resort processing of unbounded depen-dencies when the unbounded dependency isgrammatically possible. The plausibility effectin Experiment 2 resembles the correspondingeffect in Experiment 1, but in this experiment,the unbounded dependency eventually turnedout to be incorrect. As subjects would not havelearned to use a first-resort strategy during Ex-periment 2 (since all the unbounded dependen-cies between the head noun and the verb wereeventually ruled out), we can be more certainthat the strategy is an automatic part of humansentence processing.Experiment 2 provides no evidence for mis-

analysis in cases where the doubtful unboundeddependency is ruled out by strong island-constraint information. There was no evidencefor a plausibility effect in the ambiguous region,the following region, at disambiguation, or as across-over between the ambiguous and disam-biguating regions for (16c) and (16d). The in-teraction of island constraint, plausibility, andregion coupled with clear effects of plausibilityin the non-island conditions (16a) and (16b)strongly suggests that lack of sensitivity or sta-tistical power did not produce a Type II Error inthe analyses of the island conditions (16c) and(16d). We conclude, therefore, that no semanticprocessing takes place on the misanalysis if thatmisanalysis requires ignoring the strong island-constraint information implicated in this study.The most parsimonious explanation is that theprocessor employs strong island-constraint in-formation immediately and thus avoids misan-

alysis. These findings are compatible withStowe (1986), but are less compatible with Clif-ton and Frazier (1989) and Pickering et al.(1994).The results of Experiment 2 are consistent

with Pickering and Traxler’s (1996) proposalthat readers have less difficulty adopting, andmore difficulty abandoning, plausible misanaly-ses. The data do not provide conclusive evi-dence for their proposal, however, as the simpleeffect of plausibility in the disambiguating re-gion emerged only marginally in the total timemeasure.

GENERAL DISCUSSION

Experiment 1 showed that the processor per-forms immediate association by forming the un-bounded dependency as soon as it reaches theverb. The processor does not wait until the pur-ported gap location before forming the un-bounded dependency. Indeed, there was no evi-dence that the gap location played any specialrole during processing. Experiment 1 also dem-onstrated that the processor can employ a first-resort strategy, when there is no local ambigu-ity.Experiment 2 showed that the processor can

still employ a first-resort strategy in cases oflocal ambiguity when the possible unboundeddependency may turn out to be erroneous. Itdemonstrated that this strategy is employed innormal reading. It also showed that the proces-sor does not employ the same strategy if theunbounded dependency is ruled out by strongisland-constraint information. Under these con-ditions, there was no evidence that the un-bounded dependency was formed. Hence wesuggest that strong island-constraint informa-tion is employed immediately by the processor.Therefore our results support the claim that

the processor can employ a first-resort immedi-ate-association strategy that is sensitive tostrong island-constraint information during nor-mal reading. We make no claim about whetherthis strategy is always employed, or whether itis restricted to cases where the verb preferen-tially takes an argument of the same category asthe filler. Below, we suggest how these findingscan be modeled in relation to the sentence pro-


cessor and to the way in which linguistic infor-mation is mentally represented. We considerimmediate association, first-resort strategy, andthe role of island-constraint information in turn.

Immediate Association

Immediate association is compatible withtwo more detailed theories, discussed by Pick-ering (1993). Underpredictive gap-filling,theprocessor makes use of a grammar that containsgaps, such as Government-Binding Theory(Chomsky, 1981). Consider (5) again:

[In which tin]i did you put thecake [Ø]i last night? (5)

Under standard gap-filling, it is impossible forthe processor to form the unbounded depen-dency before it reachescake.But under predic-tive gap-filling, the processor encounters theverb put and then immediately locates the gapdownstream, beyond words that have not beenreached yet (Gibson & Hickok, 1993; cf.Crocker, 1994; Gorrell, 1993). This account isconsistent with immediate association. Picker-ing argued that it is unparsimonious, since itpostulates a theoretical entity (a gap) for whichthere is no processing evidence.Predictive gap-filling might be the preferred

model if linguistic evidence incontrovertiblysupported the existence of gaps. However, out-side the transformational tradition, many well-motivated linguistic theories do not assumegaps (e.g., “Flexible” Categorial Grammars:Moortgat, 1988; Pickering & Barry, 1993;Steedman, 1987; Word Grammar: Hudson,1990; some versions of Head-Driven PhraseStructure Grammar: Pollard & Sag, 1993, chap-ter 9; recent Lexical-Functional Grammar: Ka-plan & Zaenen, 1988). On these accounts, thefiller is associated with the verb directly. Wecan represent the sentence in (5) as follows(Pickering & Barry, 1991):

[In which tin]i did you [put]ithe cake last night? (18)

This representation indicates thedirect associa-tion between filler and verb. It ignores differ-ences between different gap-free linguistictheories. However, we must note that the rep-resentation does not encode any information

about which argument of the verb is taken bythe filler. Gap-free linguistic theories treat thisissue in different ways. See Pickering and Barry(1991) and Pickering (1993) for an approachtaken by flexible categorial grammar.Immediate association is compatible with

both direct association and predictive gap-filling. However, direct association is most par-simonious from a processing point of view, inthat there is no reason to assume the existenceof gaps. Pickering (1993) provided a detaileddescription of how immediate association canbe modeled within an account of incrementalinterpretation. This approach employs a theoryof grammar known as Dependency CategorialGrammar (Pickering & Barry, 1993), in whichthe fundamental linguistic notion is that of unitscalleddependency constituents,which are basedon dependencies between words. Sentence pro-cessing involves the formation of these units inan incremental manner (cf. Ades & Steedman,1982). Within this framework, immediate asso-ciation follows automatically.

First-Resort Strategy

Many accounts that argue for a first-resortstrategy essentially stipulate it as a characteris-tic of the processing of unbounded dependen-cies. For instance, Clifton and Frazier (1989)proposed the Active Filler Strategy (a first-resort strategy applied without reference to lexi-cal preferences), but did not relate it to otherparsing principles like Minimal Attachment orLate Closure. This separation may be a conse-quence of the transformational approach to un-bounded dependencies, which regards them as aspecial construction involving movement, fun-damentally distinct from other constructionsthat do not involve movement.In contrast, Pickering (1994) argued that a

first-resort strategy fell out from a general as-sumption about parsing called thePrinciple ofDependency Formation.This principle holdsequally for local and unbounded dependencies,and is based on the assumption that the proces-sor attempts to form dependency constituents asquickly as possible. Pickering argued that thisstrategy is adopted because it maximizes theamount of incremental interpretation that can


take place during comprehension, as only de-pendency constituents can be given a unifiedinterpretation.The principle of dependency formation pre-

dicts that a number of traditional garden-pathsentences are initially given a particular analy-sis. For example the processor initially resolvesthe subordinate-clause “Late-Closure” ambigu-ity below by attachingthe magazine about fish-ing as the object of the subordinate verbedited:

As the woman edited the magazineabout fishing amused all thereporters. (19)

On this analysis, the processor can form a singledependency constituent, and so it prefers thisanalysis over the alternative where the NP istreated as the subject of the main clause. Similarpredictions are made for some other ambiguoussentences (e.g., reduced complement ambigu-ities). However, when both analyses lead to asingle dependency constituent (as in reducedrelatives and PP-attachment ambiguities, for in-stance), the processor makes the decision withreference to nonsyntactic information like plau-sibility and discourse context. Pickering (1994)discussed the experimental evidence in favor ofthis proposal.The principle of dependency formation pre-

dicts that the processor uses the first-resort strat-egy with unbounded dependencies. In brief, ifthe processor forms an unbounded dependency,it constructs a single dependency constituent,but if it has an unattached filler, then it has toretain two dependency constituents, one con-sisting of the filler, the other consisting of therest of the sentence fragment. Hence the first-resort strategy is adopted, because it allows theformation of a single dependency constituentand therefore maximizes incremental interpre-tation. This account provides a unified accountof the processing of local and unbounded de-pendencies. Unbounded dependencies are notprocessed by some method analogous to trans-formations in linguistic theory, but are treatedlike other constructions.

Island Constraints

Experiment 2 suggested that strong islandconstraint information is accessed during core

parsing processes. This suggests that sentencesthat violate strong island constraints are notgenerated by the grammar. Our results provideno evidence for linguistic theories that “over-generate” strong island-constraint violations inthe basic grammar but rule them out as part ofa special component of the grammar concernedwith constraints on movement (e.g., Chomsky,1981).Many linguistic theories claim that sentences

that violate strong island constraints are simplynot generated by any component of the gram-mar (e.g., Generalized Phrase Structure Gram-mar: Gazdar, Klein, Pullum, & Sag, 1985;“flexible” categorial grammars: e.g., Steedman,1987). Sentences that violate island constraintsare no different from other ungrammatical sen-tences (e.g., *Walks John). However, providinga grammar that rules out all strong islands isvery difficult. Our results tentatively suggestthat such an approach to strong island con-straints might be appropriate. Note that it is im-possible to be sure that the processor treats allstrong islands in the same way. We have onlyconsidered sentences where there is a possibleunbounded dependency into a subject relativeclause.Dependency categorial grammar rules out

this type of strong island-constraint violation.The reason is that there is no dependency con-stituent containingthe book thatandthe authorwho wrote in (16c) (see Pickering & Barry,1993, for details). HenceWe like the book thatthe author who wrotecannot form a dependencyconstituent during incremental processing.Note, however, that dependency categorialgrammar does allow the formation of someweak island violations, like (11) above. Thus,we predict that the processor will delay beforeusing this weak island-constraint information.

APPENDIX A: EXPERIMENT 1 MATERIALS

Each item consists of one of each pair ofwords divided by a slash. The line break oc-curred afterwhich. Material 1 corresponds tosentences (14a)–(14d) in the text.

1. That is the very small pistol/garage with/inwhich the heartless killer shot the hapless manyesterday afternoon.


2. That is the extremely large beam/room with/in which Prince John barred the heavy oakendoor last year.3. That is the wide stick/ocean with/in whichthe scuba diver battered the dangerous sharktwo years ago.4. That is the large reel/park with/in which thefisherman caught the very aggressive trout earlythis year.5. You saw the Australian dogs/farm with/onwhich the rancher herded the stupid and scaredsheep last year.6. That is the builder’s mallet/office with/inwhich the young boy hammered the rusty oldnails this week.7. That is the large knife/hotel with/in whichthe cook sliced the meat last night.8. That is the small tool/shed with/in which theassistant mended the rickety ladder this morn-ing.9. That is the large knife/house with/in whichthe fishmonger gutted the enormous salmon thisafternoon.10. That is the impressive baton/arena with/inwhich the young man hit the spinning ball earlythis evening.11. That is the long chain/train with/into whichthe removers hoisted the grand piano this morn-ing.12. That is the smelly bait/lake with/in whichthe angler hooked the very hungry pike lastsummer.13. That is the very old sling/house with/inwhich the strong woman hurled the smoothblack stone late last night.14. That is the small axe/hut with/in which thepeasant chopped the wood last winter.15. That is the formidable dagger/castle with/inwhich the murderer stabbed the mysterious mantwo weeks ago.16. That is the enormous whip/cage with/inwhich the lion tamer lashed the frightened ani-mal two nights ago.17. That is the modern lighter/mansion with/inwhich the butler lit the roaring fire early thismorning.18. That is the colourful carpet/office with/inwhich the interior decorator covered thewooden floor two days ago.

19. That is the long lance/field with/in whichthe hunter speared the boar early last month.20. That is the Japanese camera/valley with/inwhich the tourist photographed the flaming redflower yesterday.21. That is the tiny tack/room with/in which thesecretary pinned the urgent notice late lastnight.22. That is the rugged cane/cave with/in whichthe stupid biologist poked the hibernating bearlast winter.23. That is the very small brush/attic with/inwhich the famous artist painted the beautifulportrait last year.24. That is the cheery candle/lounge with/inwhich I decorated the beautiful birthday cakelate last night.25. That is the bright green pencil/meadowwith/in which the artist sketched the rural land-scape during the summer.26. That is the large round shovel/island with/on which the fierce pirate dug the very deephole at midnight.27. That is the long pencil/cavern with/onwhich the explorer drew the treasure map latelast century.28. That is the narrow torpedo/harbour with/inwhich the submarine sank the enemy ship dur-ing the war.

REFERENCES

ADES, A., & STEEDMAN, M. J. (1982). On the order ofwords.Linguistics and Philosophy,4, 517–558.

ALTMANN , G. T. M., & STEEDMAN, M. J. (1988). Interactionwith context during human sentence processing.Cog-nition, 30, 191–238.

BEVER, T. G. (1970). The cognitive basis for linguisticstructures. In J. R. Hayes (Ed.),Cognition and the de-velopment of language(pp. 279–362). New York:Wiley.

BOLAND, J. E., TANENHAUS, M. K., & GARNSEY, S. M.(1990). Evidence for the immediate use of verb controlinformation in sentence processing.Journal of Memoryand Language,29, 413–432.

BOLAND, J. E., TANENHAUS, M. K., GARNSEY, S. M., &CARLSON, G. N. (1995). Verb argument structure inparsing and interpretation: Evidence from wh-questions.Journal of Memory and Language,34,774–806.

BOURDAGES, J. (1992). Parsing complex NPs in French. InH. Goodluck, & M. Rochemont (Eds.),Island con-


straints: Theory, acquisition and processing(pp. 61–87). Dordrecht, The Netherlands: Kluwer.

CHOMSKY, N. (1965).Aspects of the theory of syntax.Cam-bridge, MA: MIT Press.

CHOMSKY, N. (1981).Lectures on government and binding.Dordrecht, The Netherlands: Foris.

CLIFTON, C., JR., & FRAZIER, L. (1989). Comprehendingsentences with long distance dependencies. In G. Carl-son & M. Tanenhaus (Eds.),Linguistic structure inlanguage processing(pp. 273–317). Dordrecht, TheNetherlands: Kluwer.

CRAIN, S., & FODOR, J. D. (1985). How can grammars helpparsers? In D. R. Dowty, L. Karttunen, & A. Zwicky(Eds.),Natural language parsing(pp. 94–128). Cam-bridge, UK: Cambridge University Press.

CRAIN, S., & FODOR, J. D. (1987). Sentence matching andovergeneration.Cognition,26, 123–169.

CROCKER, M. W. (1994). On the nature of the principle-based sentence processor. In C. Clifton, L. Frazier, &K. Rayner (Eds.),Perspectives on sentence processing(pp. 245–266). London: Erlbaum.

FODOR, J. D. (1978). Parsing strategies and constraints ontransformations.Linguistic Inquiry,9, 427–473.

FODOR, J. D. (1989). Empty categories in sentence process-ing. Language and Cognitive Processes,4, 155–209.

FORSTER, K. I., & STEVENSON, B. (1987). Sentence matchingand well-formedness.Cognition,26, 171–186.

FRAZIER, L. (1979).On comprehending sentences: Syntacticparsing strategies.West Bend, IN: Indiana UniversityLinguistics Club.

FRAZIER, L., & CLIFTON, C. (1989). Successive cyclicity inthe grammar and parser.Language and Cognitive Pro-cesses,4, 93–126.

FRAZIER, L., & RAYNER, K. (1982). Making and correctingerrors during sentence comprehension: Eye movementsin the analysis of structurally ambiguous sentences.Cognitive Psychology,14, 178–210.

FREEDMAN, S., & FORSTER, K. I. (1985). The psychologicalstatus of overgenerated sentences.Cognition,19,101–131.

GARNSEY, S. M., TANENHAUS, M. K., & CHAPMAN, R.(1989). Evoked potentials and the study of sentencecomprehension.Journal of Psycholinguistic Research,18, 51–60.

GARROD, S. C., FREUDENTHAL, D., & BOYLE, E. (1994). Therole of different types of anaphor in the on-line reso-lution of sentences in a discourse.Journal of Memoryand Language,33, 39–68.

GAZDAR, G., KLEIN, E., PULLUM, G. K., & SAG, I. A. (1985).Generalized phrase structure grammar.Cambridge,MA: Harvard University Press.

GIBSON, E., & HICKOK, G. (1993). Sentence processing withempty categories.Language and Cognitive Processes,8, 147–161.

GORRELL, P. (1993). Evaluating the direct association hy-pothesis: A reply to Pickering and Barry (1991).Lan-guage and Cognitive Processes,8, 199–146.

HUDSON, R. A. (1990).English word grammar.Oxford,UK: Blackwell.

KAPLAN, R. M., & ZAENEN, A. (1988). Long-distance de-pendencies as a case of functional uncertainty. In M.Baltin & A. Kroch (Eds.),Alternative conceptions ofphrase structure(pp. 17–42). Chicago: University ofChicago Press.

MCKOON, G., & RATCLIFF, R. (1994). Sentential context andon-line lexical decision.Journal of Experimental Psy-chology: Learning, Memory, and Cognition,20,1239–1243.

MCKOON, G., RATCLIFF, R., & WARD, G. (1994). Testingtheories of language processing: An empirical investi-gation of the on-line lexical decision task.Journal ofExperimental Psychology: Learning, Memory, andCognition,20, 1219–1228.

MARSLEN-WILSON, W. D. (1973). Linguistic structure andspeech shadowing at very short latencies.Nature,244,522–523.

MARSLEN-WILSON, W. D. (1975). Sentence perception as aninteractive parallel process.Science,189,226–228.

MOORTGAT, M. (1988).Categorial investigations: Logicaland linguistic aspects of the Lambek Calculus.Dor-drecht, The Netherlands: Foris.

NICOL, J. (1993). Reconsidering reactivation. In G. T. M.Altmann & R. C. Shillcock (Eds.),Cognitive models ofspeech processing: Psycholinguistic and computa-tional perspectives II(pp. 321–347). Cambridge, MA:MIT Press.

NICOL, J., FODER, J. D., & SWINNEY, D. A. (1994). Usingcross-modal lexical decision tasks to investigate sen-tence processing.Journal of Experimental Psychology:Learning, Memory, and Cognition,20, 1229–1238.

NICOL, J., & PICKERING, M. J. (1993). Processing syntacti-cally ambiguous sentences: Evidence from semanticpriming. Journal of Psycholinguistic Research,22,207–237.

NICOL, J., & SWINNEY, D. (1989). The role of structure incoreference assignment during sentence comprehen-sion.Journal of Psycholinguistic Research,18, 5–19.

PICKERING, M. J. (1993). Direct association and sentenceprocessing: A reply to Gorrell and to Gibson andHickok. Language and Cognitive Processes,8, 163–196.

PICKERING, M. J. (1994). Processing local and unboundeddependencies: A unified account.Journal of Psycho-linguistic Research,23,

PICKERING, M. J., & BARRY, G. (1991). Sentence processingwithout empty categories.Language and CognitiveProcesses,6, 229–259.

PICKERING, M. J., & BARRY, G. (1993). Dependency catego-rial grammar and coordination.Linguistics,31, 855–902.

PICKERING, M. J., BARTON, S., & SHILLCOCK, R. C. (1994).Unbounded dependencies, island constraints and pro-cessing complexity. In C. Clifton, L. Frazier, & K.Rayner (Eds.),Perspectives on sentence processing(pp. 199–224). London: Erlbaum.


PICKERING, M. J., & TRAXLER, M. J. (1996).Plausibility andrecovery from garden paths: An eye-tracking study.Unpublished manuscript.

POLLARD, C., & SAG, I. A. (1993). Head-driven phrasestructure grammar.Stanford, CA. and Chicago, IL:CSLI and University of Chicago Press.

RAYNER, K., & FRAZIER, L. (1987). Parsing temporarily am-biguous complements.Quarterly Journal of Experi-mental Psychology,39A, 657–673.

RAYNER, K., & POLLATSEK, A. (1989).The psychology ofreading.Englewood Cliffs, NJ: Prentice Hall.

ROSS, J. R. (1967).Constraints on variables in syntax.Bloomington, IN: Indiana University Linguistics Club.

STEEDMAN, M. J. (1987). Combinatory grammars and para-sitic gaps.Natural Language and Linguistic Theory,5,403–439.

STOWE, L. A. (1986). Parsing WH-constructions: Evidencefor on-line gap location.Language and Cognitive Pro-cesses,1, 227–245.

STOWE, L. A. (1992). The processing implementation ofsyntactic constraints: The sentence matching debate. InH. Goodluck & M. Rochemont (Eds.),Island con-straints: Theory, acquisition and processing(pp. 419–443). Dordrecht, The Netherlands: Kluwer.

STOWE, L. A., TANENHAUS, M. K., & CARLSON, G. M.

(1991). Filling gaps on-line: Use of lexical and seman-tic information in sentence processing.Language andSpeech,34, 319–340.

SZABOLSCI, A., & ZWARTS, F. (1993). Weak islands and analgebraic semantics for scope taking.Natural Lan-guage Semantics,1, 235–284.

TANENHAUS, M. K., CARLSON, G., & TRUESWELL, J. C.(1989). The role of thematic structures in interpretationand parsing.Language and Cognitive Processes,4,211–234.

TANENHAUS, M. K., STOWE, L., & CARLSON, G. (1985). Theinteraction of lexical expectations and pragmatics inparsing filler-gap constructions. InProceedings of theSeventh Annual Cognitive Science Society Meetings(pp. 361–365). Hillsdale, NJ: Erlbaum.

TRUESWELL, J., TANENHAUS, M., & GARNSEY, S. (1994).Semantic influences on parsing: Use of thematic roleinformation in syntactic disambiguation.Journal ofMemory and Language,33, 285–318.

WANNER, E., & MARATSOS, M. (1978). An ATN approach tocomprehension. In M. Halle, J. Bresnan, & G. A.Miller (Eds.), Linguistic theory and psychological re-ality. Cambridge, MA: MIT Press.

(Received January 18, 1995)(Revision received May 17, 1995)


Documents

Plausibility and the Processing of Unbounded Dependencies:An Eye-Tracking Study