understanding the constraints on syntactic generation ...jeboland/bolandblodgett_2001.pdf ·...

Journal of Memory and Language 45,391–411 (2001)doi:10.1006/jmla.2000.2778, available online at http://www.academicpress.com on

Understanding the Constraints on Syntactic Generation: Lexical Bias andDiscourse Congruency Effects on Eye Movements

Julie E. Boland

Institute of Cognitive Science, University of Louisiana at Lafayette

and

Allison Blodgett

Ohio State University

We measured eye movements as people read short stories. The target sentences contained noun/verb homo-nce was

con-regres-

ing times.tacticge syn-

and si-

iguities;

A

T009port970GibmenChrsis.//ww

ABolaCog705

graphs (e.g.,duck) and were preceded by a biasing context sentence. The homograph in the target sentealways disambiguated by a case-marked pronoun, e.g.,She saw his/him duck. Lexical bias effects (reflecting therelative frequency of the noun and verb forms) were found in the initial fixations on the homograph. Intrast, discourse congruency effects were first observed several words downstream in the probability of asive eye movement. Strong discourse congruency effects were also observed in the second pass readWe concluded that the lexical bias effects reflect processing difficulty during the initial generation of synstructure, while the discourse congruency effects reflect later anomaly detection. Thus, the data challentactic processing models in which all relevant and available constraints are brought to bear uniformlymultaneously. © 2001 Academic Press

Key Words:sentence comprehension; parsing; lexical bias; frequency effects; syntactic category ambdiscourse context.

central issue in sentence comprehension re- iohip-c

c

Of course, in a given parsing theory, these com- inonlydi-78,enur-in-tac-tes.al

search is the means by which we structurecoming words into a hierarchical representataccording to the grammar of our language. Tprocess, known as syntactic analysis or parscan be broken down into at least three comnents: (1) generationof syntactic structure, including the identification of alternative strutures at points of ambiguity; (2) selectionof asingle structure; and (3) reanalysisif the struc-ture initially selected turns out to be incorre

39

yn-hattic

e-ture con-am-ner-sesac-on.

his research was sponsored by NSF Grant BC6173/SBR-9729056 to the first author. Additional su was provided by NSF grants SBR-9720473 and IR2331. We are grateful to Suzanne Stevenson, Tson, and two anonymous reviewers for helpful comts on an earlier version of the manuscript. We tha

is Brew and Richard Lewis for advice on corpus ana The stimuli used in this experiment can be found at hw.ucs.louisiana.edu/,jeb9313/webstuff/stimuli1.html.

ddress correspondence and reprint requests to Juliend, University of Louisiana at Lafayette, Institute fonitive Science, P.O. Drawer 43772, Lafayette, L04-3772. E-mail: [email protected].

in-nis

ng,o-

-

t.

ponents are not necessarily implementedthree distinct stages. For example, because one structure is initially generated in the trational garden path model (e.g., Frazier, 191987), the theory makes no distinction betwesyntactic generation and syntactic selection ding the initial parse. In contrast, one can distguish between syntactic generation and syntic selection in any parsing model that activaor generates syntactic alternatives in parallel

The current paper investigates how lexicfrequency and discourse context influence stactic generation. Our central hypothesis is tthese two levels of constraint influence syntacgeneration in quite different ways: lexical frquency influences the ease with which a strucis accessed or generated, whereas discoursestraints can only be used to select among grmatical alternatives once they have been geated. Secondarily, we test several hypotheconcerning the maintenance of multiple synttic alternatives throughout an ambiguous regi

S-p-I-ed-

nkly-ttp:

E.rA

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

All rights of reproduction in any form reserved.

1

D

x

sotv

inct

i

a

o

c

t

,a

da

iar

ns

n-hethe.g.,nec-r,e-

nec-re-

-asureen-ac-na-theu-s ofyn-cti-rsein-de-

borita-n-

s inn-s inord

aintart-notherss-nt

392 BOLAND AN

Our central hypothesis derives from the lecally driven processing theory spelled out Boland (1997a). Syntactic forms are accesfrom the lexicon and these syntactic forms cstitute essential building blocks of syntacstructure. Much of the time, a word will haseveral alternative syntactic forms that compfor selection. The accessibility of the competforms is influenced by their relative frequenin the language. The syntactic forms from lexicon, together with global syntactic knowedge, determine the set of syntactic alternatat each word in a sentence. Thus, we envissyntactic generation as involving both the cess of lexical structures and the integrationthese lexical structures into larger syntactic jects, following the global constraints of thgrammar. Discourse constraints, unlike lexiand syntactic constraints, cannot influence set of syntactic alternatives that is generathough discourse constraints can guide selecof the most likely syntactic structure. Indeedtheory that allowed discourse to restrict synttic generation would make the undesirable pdiction that we cannot construct a syntacanalysis for a sentence in an incongruent course context. As will become clear, our count of sentence comprehension is a typeconstraint-based lexicalist theory, but it diffefrom other constraint-based accounts in drawa distinction between syntactic generation asyntactic selection.

All constraint-based lexicalist models intially activate multiple syntactic alternativespoints of ambiguity (e.g., MacDonald, Peamutter, & Seidenberg, 1994; Trueswell, Tanehaus, & Garnsey, 1994).1 The term “constraint”refers to any pattern from our language expeence that becomes part of our linguistic knowedge; constraints can be based on frequeplausibility, grammaticality, and so forth. Moattempts to specify constraint-based mod
have maintained that all constraints are brougto bear simultaneously during syntactic anal
edeenary

-ate

1 This is not literally true in a model like that of Tabor, Jliano, and Tanenhaus (1997), but one could consider thevolvement of two or more attractors as equivalent to the tivation of syntactic alternatives.

BLODGETT

i-inedn-iceetegy

hel-vesionc- ofb-ealtheed,tionac-

re-ticis-c- ofrsingnd

-t

l-n-

ri-l-cy,telsht

y-

sis. However, different models implement costraint usage in different ways. Most often, tconstraints guide syntactic selection oncesyntactic alternatives have been identified (eSpivey & Tanenhaus, 1998), but in at least omodel, the constraints influence which syntatic structures are initially applicable (TaboJuliano, & Tanenhaus, 1997). To illustrate, wbriefly describe the architecture of two example models.

Spivey and Tanenhaus (1998) used a contionist framework to model the analysis of a duced relative clause ambiguity (e.g.,The horseraced . . .). In their computational model, processing difficulty at the verb was realized competition between a main clause structand a relative clause structure, which were tered into the simulation as unweighted synttic alternatives. Because the syntactic altertives are equally (un)activated at the start of simulation, no constraints are allowed to inflence syntactic generation. Rather, all classeconstraint are used simultaneously during stactic selection. The structures compete for avation from the relevant constraints (discoucontext, lexical frequency, etc.), and the wning syntactic alternative is selected for the veloping syntactic representation.

The Tabor et al. model (1997; see also Taand Tanenhaus, 1999) incorporates a gravtional dynamical systems component, with sytactic alternatives represented as attractormetric space. This model allows multiple costraints to determine the syntactic alternativethe initial candidate set, because at each wposition, the model uses the available constrinformation to situate the system in a new sting position in the space. Although there is actual generation of structure at this point,starting position determines which attracto(i.e., syntactic alternatives) will affect procesing by exerting a gravitational pull. (The amouof pull exerted by each attractor is determinby its frequency.) Intermediate states betwtwo or more attractors map loosely to temporsyntactic ambiguity, but syntactically unambiguous inputs can also result in intermedi

u- in-
states if the grammatical information is inconsistent with other constraints. Syntactic se-
ac-

T

c

opdpin a

am

iv

ese),c

) l. n

eutaleth

8n

tidoe

o

ss

eyts.alin-lesro-

ngtic

ysil-

rennt)lyer

1;.,&4;h

ncyglyn,

,s,

eer,).t-

P,nnt-lec-

u-heric-e-sn-

(3)

UNDERSTANDING CONSTRAIN

lection occurs after each successive word, assystem gravitates to a single attractor/syntastructure.

These two models illustrate some commproperties of constraint-based lexicalist aproaches. Syntactic alternatives are evaluateparallel, using constraints from any level of reresentation. Thus one can, in principle, distguish between the generation of the syntacticternatives and the selection of the optimstructure whenever there is a local syntactic biguity. For example, the fragment They sawher. . ., is consistent with either the accusat(acc) or the possessive (poss) form of her, andeach form can be attached to the matrix vphrase (VP) as part of either an embedded tence (S) or a direct object noun phrase (NPshown in (1). At herat least some of these strutural alternatives would serve as competitorsa model like Spivey and Tanenhaus’s (1998as the attractors in a model like Tabor et a(1997). In either case, the alternative thatmost consistent with the available constraiwould be selected.

a. They saw(heraccsigning)s.

b. They saw((herpossfather) signing)S.

c. They saw(heracc)NP.

d. They saw(herposssignpost)NP.

They saw him sign.

Now consider the situation in (2) at thword sign. There is no syntactic ambiguity, bsign is lexically ambiguous because it isnoun/verb homograph. According to multipaccess theories of word recognition, bothnoun and the verb forms ofsign would be ac-cessed (e.g., Duffy, Morris, & Rayner, 198Swinney, 1979; Tanenhaus, Leiman, & Seideberg, 1979). However, a model of syntacambiguity resolution like that of Spivey anTanenhaus (1998) makes no prediction abprocessing difficulty in this case. Only onsyntactic alternative would be available fevaluation, because only the verb form ofsigncan be attached to the syntactic structure aciated with They saw him. . . . The situation

(1)

(2)

is somewhat different in the Tabor et al. (1997model, because the gravitational system can

S ON SYNTACTIC GENERATION 393

thetic

n- in--

al-l-

e

rbn-

as-inor’sists

e

;-

c

ut

r

o-

influenced by ungrammatical attractors if thare strongly supported by some constrainTherefore, Tabor et al. predict that both lexicand discourse constraints could potentiallyfluence syntactic generation. These exampdemonstrate that unambiguous structures pvide an important test case for determinihow various constraints influence syntacgeneration.

Our hypothesis, that lexical frequency plaa special role in syntactic generation, can belustrated by contrastingsignwith play. Signoc-curs most frequently as a noun, whileplay ismost frequently a verb. The extensive literatuon frequency effects in word recognitiodemonstrates that the dominant (more frequeform of an ambiguous word is more strongavailable during lexical access than weakforms (Binder & Morris, 1995; Binder &Rayner, 1998; Carpenter & Daneman, 198Dopkins, Morris, & Rayner, 1992; Duffy et al1988; Hogaboam & Perfetti, 1975; PachtRayner, 1993; Rayner, Pacht & Duffy, 199Tabossi, Colombo, & Job, 1987), althougsome accounts maintain that these frequeeffects can be negated if the context stronbiases a particular form (Kellas, Paul, Marti& Simpson, 1991; Paul, Kellas, Martin, &Clark, 1992; Martin, Vu, Kellas, & Metcalf1999; Simpson & Krueger, 1991; Vu, KellaMetcalf, & Herman, 2000; responses to somof these claims can be found in Rayner, Bind& Duffy, 1999, and Binder & Rayner, 1999Given the neutral context in (3), all constrainbased approaches would predict thather signwould likely be analyzed as a possessive Nwhile her playwould likely be analyzed as aembedded subject and verb. All constraibased approaches would also agree that setion of the syntactic structure could be inflenced by biasing discourse context and otrelevant constraints. In support of these predtions, Boland (1997b) found that syntactic slection for the ambiguity exemplified in (3) wainfluenced by both lexical and discourse costraints.

a. They saw her sign.
)be b. They saw her play.

D

e

flsc

iv

f

uas

t cife

ce

ppeghlea

teeet

eif-

u-tiveetsf-ofn-

rdsthectsiningye

n-d

gettheis-rd

rsec-nt-his

sess,ss-us

n-re-ndon,tic

394 BOLAND AN

As illustrated above, using Spivey and Tanhaus (1998), most constraint-based modmake clearer predictions about constraint inence during syntactic selection (ambiguity relution) than about how constraints influensyntactic generation in unambiguous structulike (2). We hypothesize that the relatstrength of the noun and verb forms of the hoograph should still influence processing diculty in unambiguous sentences, becauseless frequent form is less strongly available ding word recognition. In contrast, supportiveconflicting discourse context should not inflence the ease of syntactic generation in unbiguous structures. This two-part hypothesiconsistent with exhaustive access modelsword recognition (e.g., Duffy et al., 1988), buwould not follow straightforwardly from seletive access models that allow strongly biascontext (such as unambiguous syntactic inmation) to overwhelm bottom-up frequency fects (e.g., Paul et al., 1992).

To examine the roles of lexical frequenand discourse context, Boland (1997b) embded unambiguous and ambiguous sentencesimple stories. The target sentences wereceded by a biasing context sentence; examfor the item constructed around the noun/vhomographduckare given in (4). The matchintarget sentences are in Table 1, along witsummary of the results. Additional exampare provided in Appendix B. For each set of tget sentences, context type and target type wcrossed to create four conditions, noun conwith noun target, noun context with verb targverb context with noun target, and verb contwith verb target.2 In Table 1, the disambiguaing words in the target sentences are in bo
face type, to highlight the differences in targsentence type.
heis-if-n-born- isn-ev-

2 The point at which the discourse incongruity of mismatched contexts was first apparent differed slightly acroitems. The incongruity was usually apparent at the pronoalthough congruent continuations were sometimes still psible at that point. For example, in the context of (4b), setence (5a) below becomes awkward athis, although a con-gruent continuation likeShe saw his arm twitch as he. . .was possible until it was ruled out by the following wordduck.

BLODGETT

n-elsu-o-e

resem-fi-the

ur-or-m-

isof

it-ngor-f-

yd-

s inre-lesrb

asr-erextt,xt-ld-et

a. Noun context:As they walked around, Katelooked at all of Jimmy’s pets.

b. Verb context:As they walked around, Katewatched everything that Jimmy did.

As shown in Table 1, the lexical bias of thnoun/verb homograph influenced processing dficulty in both unambiguous and (long) ambigous target sentences, consistent with exhausaccess models of word recognition. In targwith long ambiguous regions, the lexical bias efect was found at the first word after the point disambiguation. In unambiguous target setences, the effect was found one to two woafter the homograph, depending on whether discourse was congruent or incongruent. Effeof lexical bias have since been replicated isolated, unambiguous sentences containnoun/verb homographs, though not using etracking (Corley, 1998; Boland et al., 2000, ipublished manuscript). Boland (1997b) foundiscourse congruency effects only in the tarsentences with an ambiguous region. When ambiguous region was short, an effect of dcourse congruency occurred on the first woafter the pronoun, demonstrating that discoucongruency information guided syntactic seletion at the pronoun, as predicted by constraibased models of sentence comprehension. Tfinding provided clear evidence that discourconstraints were computed rapidly. Nonethelediscourse congruency did not influence proceing (syntactic generation) in the unambiguotarget sentences at any point.

The data pattern for the unambiguous setences like those in the top of Table 1 is pdicted by the hypothesis that both lexical adiscourse constraints guide syntactic selectibut only lexical frequency influences syntacgeneration. Boland (1997b) found that only tnoun/verb bias of the homograph—and not dcourse congruency—influenced processing dficulty in unambiguous sentences. Thus, in cotrast to Spivey and Tanenhaus (1998) and Taet al. (1997), Boland concluded that all costraints are not used simultaneously. If Bolandcorrect, she has falsified a central claim of costraint-based models. However, there are s

(4)

-ss

un,os-n-

eral concerns about the results that might qual-ify her conclusions.

,

nnu

ne

a

elfesnindaud

in

ntn-l-

xteds-tyor

n-ss

theiso-

drly

6

on

e

near the barn.

e

The first concern is that the lexical frequeneffects were one to two words later than pdicted. Effects often emerge a word or two dowstream of the predicted location in the self-pacword-by-word reading paradigm that Bola(1997b) used (e.g., see Taraban & McClella1988), but the timing makes it difficult to argdecisively that the relevant effects impacted pcessing during the initial generation of syntacstructure. Likewise, the delay effectively prevethis finding from impacting the exhaustive/seltive access debate in the word recognition liteture. Both Boland et al. (2000, unpublished muscript) and Corley (1998) found effects lexical bias on the noun/verb homograph itsbut those studies used a different measure of cal bias. Boland (1997b) used a contingent quency measure based on sentence complnorms, whereas Boland et al. and Corley ucorpus norms to determine the relative frequeof the noun and verb forms, without considerthe local syntactic and semantic context. This tinction is theoretically important. Tabor et (1997) predict that contingent frequency shopredict processing difficulty because their moincorporates a context-sensitive mechanism

Note. The disambiguating word in each target sentenc

lexical access and syntactic generation, asElman (1990). However, Boland (1997a) predic

cyre-n-ed,dd,ero-tictsc-ra-n-

oflf,

exi-re-tionedcygis-l.ldel

of

that overall bias should predict reading time unambiguous contexts.

A second concern was the lack of consisteanomaly effects for mismatched discourse coditions in the unambiguous target sentences. Athough Boland (1997b) did not predict conteeffects on syntactic generation, the mismatchtarget sentences were infelicitous in their dicourse context. On any account, this infelicishould have produced an anomaly effect fmismatched conditions.

A third concern is that the semantic and sytactic disambiguation of the homograph was leimmediate for the noun targets compared to verb targets. After a context like (4b)—which biased toward the verb reading of the homgraph—one might initially interpret duck in the“unambiguous” noun target as a nominalizeverb: She saw his duck as the barn door neahit him in the head. In our judgment, this is apossible complication in at least 5 of the 1items (duck, roll, slip, play, bow). There is a syn-tactic complication as well. Suzanne Stevenspointed out to us that the noun targets canactu-ally be analyzed as verbs: e.g.,Jimmy’s dog andMargie’s dog were playing near the barn whil

is highlighted in bold.

UNDERSTANDING CONSTRAINTS ON SYNTACTIC GENERATION 395

TABLE 1

Word Position Where Effects Were First Observed in Boland (1997b)

Observed effects by word position

Target sentence type Lexical bias Discourseof N/V homograph congruency

UnambiguousNoun target: She sawhis duck and chickens congruent context:and,

near the barn incongruent context: no effectVerb target: She sawhim duck and stumble chickens/stumble

near the barnShort ambiguous region

Noun target: She saw herducks and chickensnear the barn. not applicable ducks/bend

Verb target: She saw herbendand stumblenear the barn.

Long ambiguous regionNoun target: She saw her duck andchickens

near the barn. near nearVerb target: She saw her duck andstumble

intsJimmy and Margie watched. The kids saw HISduck under a gate while HERS chased a cat.

D

i

sa

t

h t

oad uh th

eb

rt

eo

r

d

con-

asis-cti-eounsedar-esredn aingnceerfor

rialre-oferts.g.,nheusedh istureInlexhatun

)

b,Experiment 2) materials in an eye movementparadigm, which provides several dependentmeasures reflecting varying amounts of process-

3 Although an unlimited parallel account of syntactic am-biguity resolution is inconsistent with the abundant evidencefor garden path effects, various forms of ranked/limited par-allelism have been considered (e.g., Gibson, 1991; Just &Carpenter, 1992; Pearlmutter & Mendelsohn, 1999, unpub-lished manuscript; Stevenson, 1998). In fact, an account inwhich alternative syntactic representations are ranked ac-

396 BOLAND AN

This analysis is only felicitous in a spoken dcourse with contrastive stress on his. It mightwell be generated under a constraint-baaccount in which all lexical alternatives are ptially activated, but we would not expect the [hisV. . .] analysis to be selected in the contexour experimental texts. No comparable ambigties apply to the verb targets. If the homograpthe noun target sentences is not immediatelyambiguated by the local context as we expeclexical bias effects may be delayed for the ntargets relative to the verb targets. Noun disbiguation would be most delayed when the course context was incongruent, becausecontext would provide some support for the intended form. If correct, this could explain wBoland’s (1997b) lexical bias effects wereword later for incongruent conditions, when target sentences were unambiguous.

The fourth issue is that Boland’s (1997b) “uambiguous” targets illustrated in Table 1 wnot completely devoid of local syntactic amguity. There is a temporary structural ambiguthat is independent of the syntactic class ofhomograph. As shown in (5) and (6), the pnoun can either be (part of) the direct objec(part of) the subject of an embedded clauMore importantly, the noun and verb target stences differ in the point at which these two psibilities were disambiguated. The verb targwere immediately disambiguated as the stture in (6b), e.g.,She saw him duck. . . . Incontrast, the noun targets were not fully disabiguated as the structure in (5a) until the enthe sentence—until the period was encountethe sentence could have continued in a way sistent with (5b), i.e.,She saw[his duck andchickens near the barnNP] eating.

Noun targets

a. She saw[hisNOUNNP]

b. She saw[hisNOUN VPCLAUSE]

Verb targets

a. She saw[himNP]

(5)

(6)

b. She saw[himVERBCLAUSE]

Consider the hypothesis that direct object aclausal structures, corresponding to those in

BLODGETT

s-

ed r-

ofui- in

dis-ed,unm-is-then-yae

n-rei-itytheo- orse.n-s-

and (6), are maintained (in parallel) as long each possibility is syntactically possible, with dcourse congruency influencing the amount of avation for each.3 If so, parallel structures would bsustained over a much longer period for the ntargets than for the verb targets. The increaprocessing load associated with maintaining pallel structures should lead to longer reading timor more re-reading in the noun targets compato the verb targets. This account would explai(nonsignificant) tendency toward slower readtimes for noun targets near the end of the sentein Boland (1997b). We expected to find strongevidence of the increased processing load noun targets in the current experiment.

It must be acknowledged, however, that seaccounts of syntactic ambiguity resolution pdominate in the literature. The best known these is the garden path model, which assthat the simplest analysis is pursued first (eFrazier, 1987), with simplicity being a functioof the number of new nodes required in tphrase structure tree. Our target sentences the structures in (5a) and (6b). No garden patpredicted for noun targets because the strucin (5a) is simpler than the structure in (5b). contrast, the structure in (6b) is more compthan the structure in (6a). Thus, a heuristic tconstructs the minimal structure at the prono

etsuc-

m- of

red,

predicts processing difficulty for verb targets atthe homograph—the point at which the (6aanalysis is ruled out.

EXPERIMENT

The current experiment used Boland’s (1997

nd(5)

cording to the available constraint information, but main-tained to some degree throughout the ambiguous region, re-ceived support from Boland (1997b). Experiments 3 and 4.

seduou llr

m

icie

euisn

fres

sbaih

dt

at

of

r-),ing

y atheon-ple

hichsis.

on-o-o-ive

o- NP

o-use

ne.g.,yysut 7hein-at-he

n-gnt

across the two target conditions; the following

l

Noun target She saw his duck and chickens near the barn.rn

ing, first fixation duration, summed first pareading times, probability of a first pass regrsion, total reading times, etc. The various pendent measures provide a degree of granity not available with tasks that rely on buttpressing to provide the sole dependent measThus, we hoped to better track the influencelexical bias and discourse congruency, as weto examine some predictions of serial and palel accounts of ambiguity resolution.

The most important predictions can be sumarized as follows. First, the hypothesis that leical bias influences syntactic generation predlexical bias effects at the earliest possible po(i.e., the initial fixations on the homograph). Bcause the disambiguation of the homographless straightforward for the noun targets, the lical bias effects might be delayed for the notarget condition when it is in an incongruent dcourse context. Second, discourse congrueshould not influence syntactic generation, so dcourse congruency effects shouldnotbe found atthe earliest possible point (i.e., initial fixationon the pronoun). Instead, the anomaly effectsmismatched context conditions should ariseatively late. And finally, if structural alternativeare generated/evaluated in parallel, and matained (to some degree) as long as they aretactically possible, the processing load willheavier for noun targets compared to verb tgets. Alternatively, if a single structural analysis chosen at the pronoun, using Minimal Attacment (Frazier, 1978), there should be a garpath for verb targets at the homograph, whensimplest analysis is ruled out.

Method

Participants. Twenty-four undergraduates

Verb target She saw him

the Ohio State University participated in the eperiment, for course credit in an introductor

ss-e-lar-nre.of asal-

-x-tsnt-isx-n-cy

is-

sorl-

in-yn-er-s-

enhe

psychology class. All were native speakersEnglish and had normal, uncorrected vision.

Materials. The four stories used in the expeiment were taken intact from Boland (1997bExperiment 2. Thus, target sentences containa noun/verb homograph were preceded bcontext sentence designed to bias either noun reading or the verb reading. Example ctext sentences were given above in (4); examtarget sentences are repeated in Table 2, wdelineates the regions to be used in the analyThe 16 pairs of target sentences were cstructed around 16 different noun/verb homgraphs (see Appendix A for the list of homgraphs and a summary of the normatmeasures described below).

In one member of each target pair, the homgraph was the head noun in a possessivethat began with either his or, in one case,their.In the other member of each pair, the homgraph was the first verb in an embedded clathat began with either him (N 5 15) or them(N5 1). As illustrated in Table 2, the initial regioincluded all words prior to the main verb: thsubject NP plus any preceding phrase (emeanwhile), preverbal adverb, or auxiliarverb. The words in this region were alwaidentical across the two target conditions, bthey varied in length across items from 1 towords, with a mean length of 2.4 words. Tsecond region was always two words long,cluding just the main verb and the disambiguing pronoun. The third region contained thomograph and the word and. The fourth re-gion contained a noun or verb that was cojoined with the homograph, plus the followinword. The conjoined word was always differe

duck and stumble near the ba.


TABLE 2

Example Target Sentences, Delineated by the Regions Used in the Data Analysis

Initial Verb 1 pronoun Homograph Conjoined phrase Fina

x-yword and the two words in the last region wereidentical in 9 out of the 16 items.

ngevearogwoe

c

ni

thar

heeby-s,xt

et,

toomn

herredereionm-nts

usto

es,

f 51

rb-nt-

oredat

erb toex-re-ndesbet,

notto- in

at-tached to an adjustable “bite bar” in order tohold the participant’s head in a fixed position. Astationary forehead-rest and an adjustable chin-

5 Boland first determined that the accusative and nomina-tive forms of her are about equally frequent in this type ofsyntactic context, in both corpus norms and sentence com-pletion norms. Then she collected sentence completions onfragments like these to assess the lexical bias of the homo-

4 Of course, the corpus may not provide an accurate esti-mate of how frequently our experimental participants haveactually encountered the noun and verb forms, because (a) itis limited to edited text, (b) all of the texts predate 1965, and(c) one million words is a fairly small sample. Nevertheless,these disadvantages are balanced by two significant advan-tages. Most importantly, the corpus was compiled by consci-entiously sampling from a broad range of genres, includingnewspapers, novels, nonfiction books and periodicals, gov-ernment documents, and professional journals. To thebest of our knowledge, no other available English corpus

has this property. Second, the corpus is widely used andbroadly accessible among psycholinguists. Therefore, weopted to use the Brown Corpus, despite its disadvantages.

398 BOLAND AND

The target sentences were designed to fit urally into one of four short stories of varyinlengths. The first story contained five target stences, the second had three, the third had seand the fourth story had only one target stence. Target sentences were always separby at least two other sentences, and no tasentence occurred in the first six sentences story. A context sentence preceding the tarsentence was designed to bias the reader toeither an upcoming possessive NP or an upcing embedded clause. Across the set of expmental items, the context sentence operatedther by shifting discourse focus between objeand events (as in the duckexample in (4)) or bymanipulating real-world plausibility, oftebased on spatial relationships within the dcourse model. An example of the latter is item using steer. The male participant is in truck with a dirty rear window. The female paticipant is either inside the truck where she cwatch him steer or behind the truck next to steer. See Appendix B for additional examplNoun-biased and verb-biased contexts wcrossed with noun and verb targets in a two-two factorial design to create four conditionnoun context with noun target, noun contewith verb target, verb context with noun targand verb context with verb target.

Each of the 16 critical items was assignedone of the four conditions in a pseudo-randfashion on the first stimulus list. The conditioof each item was then rotated to form three otexperimental lists, such that each item occuronce in each experimental condition and thwere an equal number of items in each conditon each list. Each story was followed by a coprehension question, to ensure that participaunderstood the stories.

Lexical bias norms. The lexical bias of eachhomograph was estimated in two ways: corpanalysis and sentence completions. Accordingthe Francis and Kucˇera (1982) norms, which arcalculated over the widely used Brown Corputhe homographs had a mean noun usage o
per million and a mean verb usage of 18 per mlion. The log frequencies are provided in Appendix A. This measure provides a rough estima
BLODGETT

at-

n-en,n-ted

getf aetardm-ri-

ei-ts

s-e

-aniss.re

of the overall frequency of the noun and veforms of the homographs.4 As our second measure of lexical bias, we estimated the contingefrequency of the noun and verb forms in ambiguous contexts likeShe saw her. . . [AgentNP, perception verb, ambiguous pronoun]. Fthis, we relied upon completion norms collectby Boland (1997b) on the local population thwas sampled in the current experiment.5

There was a generalized shift toward the vform in the sentence completions comparedthe corpus frequency data. However, upon amining the completion scores and the log fquencies for individual items (both can be fouin Appendix A), it was clear that the differencin the two measures of lexical bias cannot fully explained by a generalized shift. In facthe completions and log frequencies were correlated [r , .10]. The local syntactic contexnot only influenced resolution of the homgraph; it seems to have impacted resolutiondifferent ways for different items.

Apparatus and procedure. A dental impres-sion was made for each participant. This was

il--

te

graph; she found that 54% of the completions were consis-tent with the verb form and 33% were consistent with thenoun form (the remainder could not be disambiguated).

T

anto

m

hathcti

tiemon opo

tere

d

aer ooft

e ta

e

entassIn theassre-

inma-ef-ion,ere

A). re-on-tingy-

malt ofal-r-sists.enf thealhthects re-

ofar-

ura-ch. 1.rse inns to

r-

r-nd

UNDERSTANDING CONSTRAIN

rest also helped secure the head in a comfortfashion. Eye movements were measured usiDr. Bouis monoculor oculometer. The apparacontinuously outputs two voltages, correspding to eye position along the X and Y axes.During the experiment, eye position was sapled every millisecond (ms) and convertedscreen coordinates.

The sensor of the apparatus was first rougaligned by mechanical means. It was further justed to give zero-output voltages when participant looked straight ahead, and balanpositive and negative voltages when the parpant looked at equidistant points along the X andY axes. The experimenter then ran a calibraroutine during which the participant was askto fixate on nine disparate points on the coputer screen in order to establish the relatiship between X/Y voltages and screen positioIf this could not be done with an error rateless than 10 pixels in each dimension, the eximent was aborted. This calibration criteriwas used throughout the experiment.

Participants who were successfully calibraread a practice story, complete with a comphension question. A short break followed, ththe setup was recalibrated before the experimproper began. Each sentence was presentethe screen in its entirety, on a single line, infixed order. The target sentences always peared alone, but some sentences appearpairs, separated by a blank line. For each tasentence, the screen position and durationeach fixation were computed and stored. Mparticipants completed the experiment in ab15 min. The equipment was recalibrated aeach comprehension question to ensure accutracking throughout the experiment. Aside frothe 24 participants whose data are reported hfour participants’ data were omitted becausecalibration error was greater than our accurcriterion during part of the experiment.

Results

We analyzed four dependent measures,
pecting to find lexical bias effects in the earmeasures and discourse congruency effectsthe later measures. We assume that an early

bleg ausn-

-to

lyd-eedci-

ond-

n-.f

er-n

de-n

ent on

ap-

d ingetofstuterratemre,

hecy

x-

fect will be reflected in the first pass dependmeasures, including first fixations and first preading times, in the region of interest. contrast, a late effect might be reflected in second pass reading times or in the first pmeasures—but downstream of the critical gion. Lexical bias was varied continuouslyour stimuli, whereas discourse context was nipulated categorically. Thus, lexical bias fects were evaluated using linear regresswhereas the discourse congruency effects wevaluated using analysis of variance (ANOVAn effect of discourse congruency would bevealed as the interaction of target type with ctext type. We also were interested in contrasthe predictions of our syntactic parallelism hpothesis against the predictions of the miniattachment hypothesis. Both predict an effectarget type in the ANOVAs, but syntactic parlelism predicts greater difficulty for noun tagets, while the minimal attachment hypothepredicts greater difficulty for the verb targeWe present all of the ANOVAs first, and thpresent the regression analyses. In each oprincipal ANOVAs, we included three criticregions, the Verb 1 Pronoun region, in whicthe discourse incongruity is first evident,Homograph region, in which lexical bias effewere predicted, and the Conjoined Phrasegion.

ANOVAs. The duration of the first fixation in aregion provides the earliest available index processing difficulty in the region. For each pticipant and item, we computed the mean dtion for the first fixation in each region, for eacondition. These data are summarized in FigTo determine whether there was a discoucongruency effect or an effect of target typethe critical region, participant and item meafor the three critical regions were submitted4(list) 3 3(region) 3 2(target) 3 2(context)ANOVAs. Neither a two-way interaction of taget and context [F1(1,20) 5 1.68, p . .10;F2(1,12) 5 1.58,p . .10] nor a three-way inteaction of target, context, and region was fou

ly in ef-

[F1(2,40) 5 1.17, p . .10; F2(2,24) 5 1.64,p . .10]. Thus, discourse congruency did not af-fect the duration of the initial fixations. There

s

s,

tebe-l-erlly

rech of

FIG. 1. Mean duration (in milliseconds) of the first fixation in each region, for each of the four experimen-

was no main effect of target type [F’s , 1.0],but target interacted with region in the itemanalysis [F1(2,40) 5 2.31,p . .10; F2(2,24) 54.46,p , .05]. Upon examining Fig. 1, the interaction seems to be driven by long fixationsthe verb target conditions during the ConjoinPhrase. Such an effect would be roughly content with the minimal attachment hypotheswhich predicted a garden path for verb targetthe Homograph, in the prior region. However

tal conditions.

4(list) 3 2(context) 3 2(target) ANOVA overthe Co

FIG. 2. Condition means (in ms) for

s

- inedis-

is, ata

target type [F1(1,20) 5 2.30, p . .10; F2(1,12) 5 2.86,p . .10].

The first fixation duration is an incomplemeasure of first pass processing difficulty,cause the initial fixation in a region is often folowed by another fixation before the readmoves on to the next region. This is especialikely when a region includes two or mowords, as our regions did. Therefore, for eaparticipant and item, we computed the mean

400 BOLAND AND BLODGETT

the summed first pass fixation durationsin each data are sum-
njoined Phrase region found no effect ofregion, for each condition. These
the summed first pass fixations in each region.

401

nr

n

b

er-ts

-

is-ldond

-t at

nsAs

s-r

ents-ndy

al

xtd-

marized in Fig. 2. Participant and item meafrom the three critical regions were submitted4(list) 3 3(region) 3 2(target) 3 2(context)ANOVAs. As predicted, neither a two-way interaction of target and context nor a three-winteraction of target, context, and region wfound [F’s , 1.0]. There was no main effect otarget type [F’s , 1.10], nor was there an interaction between target and region [F1(2,40) 51.12,p . .10; F2(1,12) 5 1.61,p . .10].

When a reader experiences difficulty durithe first pass reading, the difficulty can be flected either in longer first pass reading timesthe region or by regressive eye movementsearlier regions. (Information about the landisites for regressions is provided in Appendix CThe probability of a first pass regressionout of aregion was computed by participant and by itefor each condition in each region. These probilities are summarized in Fig. 3, which sugests that discourse congruity effects beganappear several words downstream of the firstcongruous word in the target sentence (the pnoun). Note that the probability of a regressiin the Initial region is zero, by definition. Thprobabilities for the three critical regions wesubmitted to 4(list) 3 3(region) 3 2(target) 32(context) ANOVAs. The main effect of targewas reliable by participants and marginal items [F1(1,20) 5 7.25, p , .05; F2(1,12) 53.19,p , .10]. Note that the direction of the efect is reversed from the trend observed in

FIG. 3. The probability of a first pass r

first fixations: Verb targets exhibited less prcessing difficulty, as predicted by the syntac

ns to

-ayasf-

ge- in tog.)

ma-

g- to

in-ro-onere

tby

f-the

parallelism hypothesis. There was also an intaction of context and target by participan[F1(1,20) 5 6.13, p , .05; F2(1,12) 5 4.71,p , .10] and a fully reliable three-way interaction of region, context, and target [F1(2,40) 56.73,p , .05; F2(2,24) 5 7.38,p , .05]. Theseinteractions are consistent with an effect of dcourse congruity. In ANOVAs at the individuaword positions, the interaction of context antarget was not reliable at the Homograph regi[F’s , 1.0], but it was reliable at the ConjoinePhrase [F1(1,20) 5 19.17,p , .01; F2(1,12) 512.35,p , .01]. Likewise, the main effect of target was absent at the Homograph but presenthe Conjoined Phrase [F1(1,20) 5 5.22, p ,.05; F2(1,12) 5 4.17,p , .10].

Next, we examined thesecond pass readingtimes. This dependent measure sums fixatioon secondary passes through each region.shown in Fig. 4, there was a large effect of dicourse congruency, beginning in the first osecond region. This is to be expected, givour finding that regressive eye movemenwere more likely in the incongruent conditions. The mean summed duration of secopass fixations in a region was computed bparticipants and by items for each condition3region cell. These means for the three criticregions were submitted to 4(list)3 3(region)3 2(target) 3 2(context) ANOVAs. The ex-pected interaction between target and contewas found, confirming that second pass rea

egression out of each region, for each condition.

UNDERSTANDING CONSTRAINTS ON SYNTACTIC GENERATION

o-ticing times were longer in incongruent condi-tions [F1(1,20)5 30.14,p , .01; F2(1,12)5

402

ea

e

er r

hrooeccr

ten

s

et

rathdta

f-de- re-gan

andythem-on-ta-tedxi-ithn-b-un

andn

asch

tedor-forerb

e--ch

r

6.86,p , .05]. In addition, longer times werfound for noun targets compared to verb tgets [F1(1,20)5 15.81,p , .01], but this ef-fect did not generalize across items [F2(1,12)5 1.99;p . .10].

In sum, discourse congruency effects wclearly present, although they were not foundthe most immediate measures of online procing. The discourse congruency effect fiemerged at the Conjoined Phrase region inpercentage of regressive eye movements duthe first pass through that region. Because twere more regressive eye movements for tasentences following incongruent contexts, secpass reading times for the incongruent conditiwere much longer than those for the congruconditions. In contrast, the target type effewere only partially reliable, and they were mumore difficult to interpret. The interaction of taget and region that was observed in the ianalysis of the first fixation data is roughly cosistent with the minimal attachment hypothewhich predicted a garden path for the verb targat the Homograph. However, there was no dence of a garden path for verb targets in other dependent measures. Quite the contsupport for syntactic parallelism was found in participant analyses of the regression data ansecond pass reading times. While we do not

FIG. 4. Condition means (in ms) fo

these effects as unequivocal support for syntacparallelism, they are certainly suggestive.

r-

reinss-sttheingeregetndnsnttsh-m-

is,etsvi-hery,e

theke

Linear regressions. Next, we evaluated the efects of lexical bias on each of the first pass pendent measures, in both the Homographgion and the Conjoined Phrase region. We beby examining the initial fixations. The Homo-graph region consisted of the homograph the word and, so the first fixation was verlikely to have been on the homograph. Thus,fixation duration may reflect access of the hoograph and the addition of the homograph cstituent to the developing syntactic represention for the sentence. Of course, we expecsyntactic integration to be easier when the lecal bias of the homograph was consistent wthe syntactic context. To provide a unidimesional measure of overall lexical bias, we sutracted the log verb frequency from the log nofrequency, as measured in the Francis Kucera (1982) norms derived from the BrowCorpus. A high value on this overall lexical biscore indicates that the noun form was mumore frequent than the verb form. The prediclexical bias effect would result in a negative crelation between bias and fixation duration noun targets and a positive correlation for vtargets.

Figures 5a–5d illustrate the relationships btween the initial fixation duration in the Homograph region and overall lexical bias, for ea

the second pass reading times in each region.

BOLAND AND BLODGETT

ticcontext by target condition. Upon inspecting thescatterplots, we removed two problematic

lo

o

en

fi

h

t

ionsin

inast

nsr-ex-heeedx-

anu-

ion

ion

Note. In all tables, the correlations that are reliable (a 5 .05) are indicated by asterisks. Marginal correlations (p , .10)

items, park and program, which are identifiedby open circles in Fig. 5.6 The remaining lexicabias values were used to predict initial fixatitime in the Homograph region and the Cojoined Phrase region, yielding the correlaticoefficients shown in Table 3. Correlations threliably differ from the null hypothesis arstarred (a 5 .05) and those marginally differeare annotated “mg” (.05, p , .10). All corre-lations are in the predicted direction, exceptthe verb context with noun target conditionthe Homograph region. As noted above, Bola(1997b) found lexical bias effects earlier for tcongruent discourse conditions compared toincongruent conditions. We suspected that

are indicated by “mg.”

difference in timing was due to difficulties in resolving the lexical form of the noun target in in

isblet

ra-asionsom. Intedasasere

-,notherall

6 These items were strongly noun-biased, according tolog frequencies from the Brown Corpus (see Appendixcolumns 1 and 2), but they were verb-biased in Bolan(1997) sentence completions (columns 3 and 4). At first,suspected that the Brown Corpus was outdated; thecreased importance of computers has no doubt increaseuse of programas a verb quite dramatically since the ea1960s, when the Brown Corpus was compiled. Howevthese items were even more strongly noun-biased in a rent textual corpus, the web-based Chicago News Netwarchives. (Details are available from the first author. In brwe analyzed the 100 most recent articles containing word programor park, using only the first token from eacarticle.) Thus, there is a clear discrepancy between the tual corpora and the completion norms. Unfortunately,completion data are problematic because they may be biby the particular sentence fragments that were used incompletion task. (This topic is explored further in the dcussion section.) To be consistent, the programand parkval-ues were excluded from all regressions using the BroCorpus measure of overall lexical bias.

nn-nat

t

ornnde

thehe--

congruent discourse contexts. These suspicwere confirmed by the data pattern reportedTable 3. At the Homograph, the correlationthe noun context with noun target condition wreliably different from that in the verb contexwith noun target condition [z 5 2.11,p , .05],while there was no difference in the correlatiofor the two conditions with verb targets. Unfotunately, the predicted correlations between lical bias and the congruent conditions at tHomograph were not fully reliable. To reducitemwise noise, we followed the strategy usin Boland (1997b). We subtracted the mean fiation duration for the verb target from the mefixation duration for noun target, for the congrent conditions of each item. Analyzed in thmanner, the lexical bias values were a reliapredictor of initial fixation time for congruenconditions, as expected [F(1,12) 5 7.32, p ,.05; r 5 2.62].

We then regressed the initial fixation dutions against contingent lexical bias, which wcomputed as the percentage of verb completminus the percentage of noun completions, frBoland’s (1997b) sentence completion datacontrast to the overall bias data, the predicpattern of positive and negative correlations wobserved only for the incongruent conditions,shown in Table 4. None of the regressions wstatistically reliable.

Next, we examined the summed first pass fixation times. In contrast to the initial fixationsthe predicted pattern of correlations was found in either the Homograph region or tConjoined Phrase region, using either ove

the A,d’swe in-d therlyer,cur-orkief,the

htex-theased theis-


TABLE 3

Correlation Coefficients (r) for Initial Fixation Durations, Predicted by Overall Lexical Bias, with the Predicted Directof the Effect for Each Condition

Condition Homograph Conjoined phrase Predicted direct

Noun context-Noun target 2.55* 2.30 2

Verb context-Verb target .49 mg .30 1

Verb context-Noun target .28 2.72* 2

Noun context-Verb target .49 mg .17 1

lexical bias or contingent frequency as the pre-dictor variable. As shown in Table 5, the data

wn


FIG. 5. The relationship between overall lexical bias (log noun frequency minus log verb frequency) and theduration of the first fixation in the Homograph region. Outliers are shown as open circles and were omitted fromthe analysis. Figure 5a illustrates the correlation for the noun context with noun target condition. Figure 5b illus-trates the marginally reliable correlation for the verb context with verb target condition.

5
FIG. 5—Continued.Figure 5c reveals the lack of correlation in the verb context with noun target condition.Figure 5d shows the marginal correlation for the noun context with verb target condition.

muebl

h ri

oi

emr-nyk

onn

su

ilivid

ectszedve-ns-ng.g

yeim-icalonalyidn-

ofrsethe

alcticunlly

ner-un

ffi-as

on

on

looked a little more coherent when we exained the probability of a first pass regression oof a region. Overall lexical bias did not yield thpredicted pattern in the Homograph region,it did at the Conjoined Phrase. Unfortunateonly one regression was reliable (a 5 .05), andnone of the others approached reliability. Tpredicted pattern was not obtained in eithergion, using contingent lexical bias as the predtor variable.

In sum, the overall lexical bias of the homgraph was a much better predictor of processdifficulty than contingent lexical bias, but evso, the lexical bias effects were almost copletely limited to the initial fixation data. Inteestingly, Boland (1997b) found that contingefrequency was a better predictor of word-bword reading times in her button-pressing tas

Discussion

Noun/verb homographs represent an imptant class of lexical ambiguity that is tightly etwined with phrase structure ambiguities. Costraint-based models of sentence procespredict that the relative frequency of the noand verb forms should predict the accessibof the two syntactic forms, and confirming edence has been reported in self-paced rea

Note. Asterisk indicates reliable correlation (a 5 .05).

-t

uty,

ee-c-

-ngn

-

t-.

r---

ingnty-ing

Corley, 1998; MacDonald, 1993). However, thcurrent study is the first report of such effeusing an eye movement paradigm. We analyfour dependent measures from the eye moment record to obtain detailed informatioabout the timing with which lexical bias and dicourse congruency influenced processiWhile it is always optimal to gather converginevidence from multiple paradigms, the etracking evidence reported here is especially portant, because it demonstrates that the lexbias effects emerged in the very first fixation the noun/verb homograph. In contrast, anomeffects from incongruent discourse context dnot begin to emerge until several words dowstream of the incongruity, in the percentageregressive eye movements. Robust discoucongruency effects were also observed in second pass reading times.

It is important to remember that the lexicbias effects reported here arose during syntageneration, not syntactic selection. The noform of the homograph cannot be grammaticaattached in the context ofShe saw him. . ., sowe assume that no such attachment was geated. Even so, the relative frequency of the noand verb forms influenced the processing diculty of generating the appropriate structure—


TABLE 4

Correlation Coefficients (r) for Initial Fixation Durations, Predicted by Contingent Lexical Bias, with Predicted Directiof the Effect for Each Condition

Condition Homograph Conjoined phrase Predicted directi

Noun context-Noun target .22 2.12 2

Verb context-Verb target .00 2.15 1

Verb context-Noun target 2.22 2.49 mg 2

Noun context-Verb target .41 .43 1

st

ith

on

tasks (Boland, 1997b; Boland & Lewis, 1998;evidenced by a lexical bias effect during the fir

TABLE 5

Correlation Coefficients (r) for Percentage of Regressions Out of the Region, Predicted by Overall Lexical Bias, wPredicted Direction of the Effect for Each Condition

Condition Homograph Conjoined phrase Predicted directi

Noun context-Noun target .08 2.57* 2

Verb context-Verb target .22 .03 1

Verb context-Noun target .23 2.28 2

Noun context-Verb target .36 .37 1


uen

laeic

enr-e

fope

fetebde pezeots

eticee9u u

dr

aaermd

meplyherible,or-

este afterrendthe

land

7b)eyet, tothe

eretheentrallhertheact

c-alceda

ef-ove-7b).evelcede 1for-elyon.igm,ffi-

t isra-asm-textleities

fixations on the homograph in three out of foconditions. The lexical bias effects were delayin the noun target with verb context conditiopresumably because the lexical ambiguity wresolved more slowly.

This pattern of lexical bias effects is simito Boland’s (1997b) findings using self-pacreading, except that she found contingent lexbias to be the best predictor of reading timwhereas we found overall lexical bias to bbetter predictor. Contingent frequency combilocal contextual information with lexical infomation whereas overall frequency is a purlexical constraint. Because the nature of thequency information is an important issue in cstraint-based models, let us consider the imcations of models that rely upon contingfrequency versus overall frequency.

Tabor et al. (1997) investigated several difent types of frequency effects on the deminer/complementizer syntactic category amguity of that. Like Boland (1997b), they founthat syntactically contingent frequencies wbetter than the overall category frequenciespredicting self-paced reading times. For examthat is most often a determiner when it is stence initial but is most often a complementiwhen it follows a verb. These contingent frequcies seemed to guide syntactic ambiguity restion; garden path effects were found when syntactic category of that was disambiguated adifferent from its contingent bias. In the dynamcal systems model of Tabor et al., the contingfrequencies are expected to predict reading in both ambiguous and unambiguous syntacontexts, because the different syntactic contcorrespond to different locations in the represtational space of the system (as in Elman, 19However, Gibson and Tunstall (1999) have sgested that the Tabor et al. results can beplained using overall lexical frequency, withorecourse to contingent frequencies.

In contrast to Tabor et al. (1997), the moassumed in Boland (1997b) predicts that ovelexical biases should be most relevant for unbiguous syntactic contexts. In the Boland count, effects of lexical bias on syntactic genation arise because competing syntactic foare accessed during word recognition, an
word’s more frequent forms are more strong
rd,

as

rdale, aes

lyre-n-li-nt

r-r-i-

reinle,n-ern-lu-he

i-nt

meticxtsn-0).g-ex-t

elallm-c-r-s

a

available. Thus, Boland (1997b) had to assuthat her contingent frequency data were simbetter estimates of overall frequency than corpus data. Such an explanation was plausespecially given our claim that the Brown Cpus underestimated the verb frequencies of parkand program. However, the current data suggthat the Brown Corpus does actually providreasonable estimate of overall frequency afiltering out those two problematic items. (Theis a growing literature comparing corpus acompletion measures as estimates of strength of lexical representations. See Roand Jurafsky, in press, as an example.)

The discrepancy between the Boland (199self-paced reading data and the currentmovement data may be due, at least in parthe nature of the dependent measures. Incurrent experiment, the lexical bias effects wquite transient, so the lexical bias effects inself-paced study may have reflected a differphenomenon (i.e., some combination of ovebias effects on syntactic generation and otfactors that affect sentence interpretation). Ifcontingent frequencies incorporated the impof local contextual information (such asShe sawher. . .), rather than reflecting pure lexical acessibility, it is not surprising that more globmeasures of processing, such as self-pareading, reflected local context effects tolarger degree.

On the other hand, discourse congruencyfects were much more apparent in the eye mment data reported here than in Boland (199This might raise the concern that discourse-lintegration was hampered by the self-pareading paradigm. However, recall from Tablthat Boland demonstrated that discourse inmation from prior sentences was immediatused to guide syntactic ambiguity resolutiThus, even in the self-paced reading paraddiscourse information was immediately and eciently used to guide syntactic selection. Imore likely that the self-paced reading padigm, with its rigid button-pressing task, wsimply insensitive to the relatively subtle anoaly effects that arose when the discourse conwas incongruent. Eye tracking, with its multipdependent measures, offers more opportun
lyto observe subtle effects.


cayntngrsay

fetocin

onsf-nyr tisdeioru

roamu

s tiot t

dem

aln rnteee

onng

nnga

)ac-sg.,

ryes-ta-d

tics-ticye

tedasur

ep-ete-

ndheylsnt-as

heo-r-ut.nre.

enin

revedingn.ey

entntunen-igu-unat

Despite limitations in our measures of lexibias, our data clearly support the view that stactic generation involves the access and igration of lexical structure, with the dominasyntactic form being accessed more stronthan subordinate forms. In contrast, discoucongruency was reflected in the second ptimes and in the probability of regressive emovements. The size of the second pass efis striking, given that Boland (1997b) failed find consistent discourse congruency effewith these sentences using self-paced readAs predicted, the timing of the discourse cgruity effects in the current experiment contrasharply with the timing of the lexical bias efects, and this suggests that expectations geated from the discourse did not influence stactic generation. Rather, the effects appeabe anomaly detection effects, in response toincongruity of the target sentence with the dcourse context. Thus, these data provide ational evidence for Boland’s distinction betwesyntactic generation and syntactic selectwith lexical bias—but not discourse congency—influencing syntactic generation.

Our data obviously challenge syntactic pcessing models in which all relevant and avable constraints are brought to bear uniforand simultaneously (e.g., Spivey & Tanenha1998; Tabor et al., 1997). Rather, the resultthe experiment presented here, in combinawith the results of Boland (1997b), supporconstraint-based model with at least two cagories of constraints. The first category incluthe relative strength of competing lexical for(i.e., lexical bias) and affects the generationsyntactic alternatives. The second categoryconstraints, which subsumes the first and includes discourse bias, affects the selectioone structure from among generated altetives. Granted, our results might be consiswith the Spivey and Tanenhaus model if thwere to add a syntactic generation componthat is dominated by syntactic and lexical cstraints. However, such a move would chathe character of their model dramatically.

Our comparison of lexical and discourse costraints may bring to mind other investigatioof unambiguous sentences in which investi
tors have contrasted different processing leve
ln-e-tlyessects

tsg.-ts

er-n-tohe-

di-nn,-

-il-lys,ofn

ae-s

sofofsoofa-ntynt-e

-s-

For example, McElree and Griffith (1995found that readers were faster to detect synttic anomalies (e.g.,Some people agree book)compared to thematic role anomalies (e.Some people alarm books) in a paradigm usinga speed–accuracy trade-off. Their primaclaim was that syntactic representations aretablished prior to richer thematic representions. On the other hand, Ni et al. (1998) founthat similar anomalies (e.g.,cats won’t usuallybake) were detected as rapidly as syntacanomalies, but were reflected in different apects of the eye movement record. Syntacanomalies induced immediate regressive emovements while semantic anomalies resulin longer first pass reading times as wellregressive eye movements. Unfortunately, ocurrent data cannot help to resolve the discrancy between McElree and Griffith and Nial., because we did not include similar thmatic/semantic anomalies in our stimuli.

The effects of target sentence type fouhere are also quite interesting, because tprovide data that help distinguish two modeof syntactic ambiguity resolution. The relevasyntactic ambiguity is in whether the constituent beginning at the pronoun is attacheda direct object NP or as part of a clause. Tminimal attachment heuristic predicted prcessing difficulty at the homograph for verb tagets when the minimal structure was ruled oNo such difficulty was predicted for the noutargets, because they used a minimal structuAlthough there was a trend toward a gardpath for verb targets at the Conjoined Phrasethe initial fixation data, no reliable effects wefound. In contrast, we found more regressieye movements and longer second pass reatimes for noun targets in the critical regioThus, there is little support for the view that thminimal attachment analysis was initialladopted.

On the other hand, our data are consistwith a ranked parallel account. This accoupredicted a heavy processing load for the notargets compared to verb targets late in the stence, because the direct object/clause ambity was disambiguated much later for the notargets. It may or may not be meaningful th
ls.the target type effects occurred in the percent-

UNDERSTANDING CONSTRAINT

x

er

o

ih

edse

ad

beelledre

iticeoarneeh

s

Mean 3.04 2.05 33.44 54.06

ow,ver-

ereented

age of regressive eye movements and secpass reading rather than longer first pass fitions. Both behavioral responses have thefect of spending more time in a difficult regiobefore moving on. Although some researchlink regressive eye movements to syntacticanalysis (e.g., Frazier & Rayner, 1982; Ferre& Henderson, 1990; but see Rayner et a1989, for a different view), we believe that ouunderstanding of eye movements is still tlimited to infer the underlying cognitiveprocess on the basis of an eye movement ptern. We must acknowledge that our dawould be consistent with a serial modelwhich garden paths were predicted for tnoun targets. However, such a model wouhave to predict that readers initially attachhis duckas the subject of an embedded clauand then reanalyzed it as a direct object. Wknow of no parsing models that would maksuch a prediction.

The ranked parallel account also explainsotherwise curious trend found by Bolan(1997b), in which verb targets tended tofaster than noun targets near the end of the stence. However, the itemwise variation observin our target type effects suggests that parastructures were not always maintained until dfinitive syntactic information was encountereThis is not surprising, given numerous othersults—including some from our own laboratory—demonstrating that syntactic commments are often made at the point of ambigu(e.g., Boland & Boehm-Jernigan, 1998; MaDonald, 1994; Tabor et al., 1997; Trueswellal., 1994). We suspect that the likelihoodmaintaining parallel representations is probbilistically determined by the amount of suppofor a given analysis, assessed over a wide raof constraints. Such an account is consistwith the noncompetitive, ranked parallel modproposed by Pearlmutter and Mendolso

to

park and point near the mountains.

(1999, unpublished manuscript).

APPENDIX A

This table provides the normative data for our experimetal items. The first two columns list the log of the Kucˇeraand Francis (1982) noun and verb frequencies. The lastcolumns list the percentage of noun and verb completifor ambiguous fragments likeShe saw her duck. . . . Meansfor all measures are provided in the bottom row.


onda-

ef-nrse-iral.,ro

at-taneld

Log(KF frequencies) Sentence completion

Noun Verb % Noun %Verb

duck 0.69 1.95 70 20steer 1.79 1.10 30 70shed 1.39 0.69 20 70roll 2.71 2.89 10 90brush 3.37 2.56 50 50sign 4.29 2.89 65 15park 3.87 0 0 90slip 2.48 1.95 0 90play 4.48 4.70 0 70train 4.20 2.30 80 10check 3.58 3.93 40 50program 5.91 1.61 20 30benefit 3.74 3.00 40 40document 2.40 0.69 60 40bow 2.56 2.56 30 60iron 1.10 0 20 70

ee

n

en-del-.-

--ty-tf-

tgentln

n-

wons

APPENDIX B

An excerpt from an example story is presented belwith both versions of the context sentences and both sions of the target sentences. In the experiment there wno paragraph breaks, because each sentence was presindividually.

This is a story about a man named Wilbur. Wilburis the mayor of a small town called White Bluffs. Oneof his best friends is his neighbor, Ann. Ann workedas Wilbur’s campaign manager. Before Wilbur be-came a politician, he felt strongly about the environ-ment.Noun context:And when a clean water petition camearound he decided to run for mayor.Verb context: But when a clean water petition camearound he didn’t know what to do.Noun target: Ann made his sign and posters beforethe election.Verb target: Ann made him sign and talk with pol-luters about the problem.

As the mayor, Wilbur convinced the town councilto develop a nature preserve. He found a wooded areain the foothills that would be perfect. They were ableto purchase it within a month. Wilbur was very proudof it, and drove his two daughters out to the site.Noun context: The girls liked to ride in the car andthey were also anxious to see the spot.Verb context: The girls didn’t like to ride in the car,but they agreed to go along.Noun target: They seemed relieved when they saw hispark and campgrounds in the mountains.Verb target: They seemed relieved when they saw him

e of ours analyses. the in-

es- the Con-e sentence


APPENDIX C

The table below presents the probability of a regressive eye movement landing in a given region. Although nonpredictions hinged upon these data, they are consistent with the patterns found in the regression and second pasThe Initial region was more likely to be refixated for noun targets, compared to verb targets. This is consistent withcreased processing load attributed to the noun targets. The Verb 1 Pronoun region was most likely to be the goal of a regrsive eye movement in the incongruent conditions. Thus, when participants made regressive eye movements fromjoined Phrase and Final regions because of the discourse incongruity, they were often regressing to the region of thwhere the incongruity originated.

ConjoinedInitial Verb 1 pronoun Homograph Phrase Final

Noun context-Noun target .39 .36 .29 .28 .08Noun context-Verb target .28 .42 .27 .26 .08

t

t

cs

n

e

Verb context-Noun target .46 .54 .43 .20 .05Verb context-Verb target .29 .32 .21 .28 .10

l

or-e-

nd

tic,

. L.n

r-entses.

ns-,

u-

ofg

REFERENCES

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

Binder, K. S., & Morris, R. K. (1995). Eye movemenand lexical ambiguity resolution: Effects of prior encounter and discourse topic. Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 21,1186–1196.

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

Binder, K. S., & Rayner, K. (1999). Does contextual strengmodulate the subordinate bias effect? A reply to Kelland Vu.Psychonomic Bulletin & Review, 6, 518–522.

Boland, J. E. (1997a). The relationship between syntaand semantic processes in sentence comprehenLanguage and Cognitive Processes, 12, 423–484.

Boland, J. E. (1997b). Resolving syntactic category ambigties in discourse context: Probabilistic and discourse cstraints. Journal of Memory and Language, 36,588–615.

Boland, J. E., & Boehm-Jernigan, H. (1998). Lexical costraints and prepositional phrase attachment. Journal ofMemory and Language, 39,684–719.

Boland, J. E., & Lewis, R. (1998). The lexical generationsyntax and its implications. The 39th Annual Meetiof the Psychonomics Society, Dallas.

Carpenter, P. A., & Daneman, M. (1981). Lexical retrievand error recovery in reading: A model based on fixation. Journal of Verbal Learning and Verbal Behavior, 20,137–160.

Corley, S. (1998). A statistical model of human lexical category disambiguation. Unpublished doctoral dissertation, University of Edinburgh.

Dopkins, S., Morris, R. K., & Rayner, K. (1992). Lexicaambiguity and eye fixation in reading: A test of com

h

s -

thvi-

thas

ticion.

ui-on-

n-

ofg

alye-

--

peting models of lexical ambiguity resolution.Journalof Memory and Language, 31, 461–476.

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

Elman, J. L. (1990). Finding structure in time. CognitiveScience, 14,179–211.

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

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

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

Frazier, L. (1987). Theories of sentence processing. In JGarfield (Ed.),Modularity in knowledge representatioand natural language understanding. Cambridge, MA:MIT Press.

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

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

Hogaboam, T. W., & Perfetti, C. A. (1975). Lexical ambigity and sentence comprehension. Journal of VerbalLearning and Verbal Behavior, 14,265–274.

Just, M. A., & Carpenter, P. A. (1992). A capacity theorycomprehension: Individual differences in workinmemory. Psychological Review, 99,122–149.

l-Kellas, G., Paul, S. T., Martin, M., & Simpson, G. B. (1991).

Contextual feature activation and meaning access. In


G. B. Simpson (Ed.),Understanding word and sen-tence(pp. 47–71). Amsterdam: Elsevier.

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

MacDonald, M. C. (1994). Probabilistic constraints and syn-

.n

i

-

e-

) l

opo

o

S

Stevenson (Eds.),Sentence processing and the lexicon:Formal, computational, and experimental perspectives,John Benjamins: Amsterdam.

Simpson, G. B., & Krueger, M. A. (1991). Selective accessof homograph meanings in sentence context. Journal ofMemory and Language, 30,627–643.

bi-of

g-

ring.en-

om-

ingunt ts

of

S.ing

ch-ess-

xi-

M.he-u-

eo-

tactic ambiguity resolution. Language and CognitiveProcesses, 9, 157–201.

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

Martin, C., Vu, H., Kellas, G., & Metcalf, K. (1999)Strength of context as a determinant of the subordibias effect. Quarterly Journal of Experimental Psychology, 52A, 813-–839.

McElree, B., & Griffith, T. (1995). Syntactic and thematprocessing in sentence comprehension: Evidencea temporal dissociation.Journal of Experimental Psychology: Learning, Memory, & Cognition, 21,134–157.

Ni, W., Fodor, J. D., Crain, S., & Shankweiler, D. (1998Anomaly detection: Eye movement patterns. Journal ofPsycholinguistic Research, 27,515–539.

Pacht, J. M., & Rayner, K. (1993). The processing of homphonic homographs during reading: Evidence from movement studies. Journal of Psycholinguistic Research, 22,251–271.

Paul, S. T., Kellas, G., Martin, M., & Clark, M. B. (1992Influence of contextual features on the activationambiguous word meanings. Journal of ExperimentaPsychology: Learning, Memory, & Cognition, 18,703–717.

Rayner, K., Binder, K. S., & Duffy, S. A. (1999). Contextual strength and the subordinate bias effeQuarterly Journal of Experimental Psychology, 42,841–852.

Rayner, K., Pacht, J. M., & Duffy, S. A. (1994). Effects prior encounter and global discourse bias on the cessing of lexically ambiguous words: Evidence freye fixations. Journal of Memory and Language, 33,527–544.

Rayner, K., Sereno, S. C., Morris, R. K., & Schmauder,R. (1989). Eye movements and on-line language cprehension processes. Language and CognitiveProcesses, 4, 121–149.

Roland, D., & Jurafsky, D. (in press). Verb sense and vsubcategorization probabilities. In P. Merlo &

S.n.

ate-

cfor

).

o-ye

.of

-ct.

fro-m

A.m-

erb.

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

Stevenson, S. (1998). Parsing as incremental restructuIn J. D. Fodor & F. Ferreira (Eds.), Reanalysis in stence processing, pp. 327–363. Kluwer: Boston.

Swinney, D. A. (1979). Lexical access during sentence cprehension: (Re)consideration of context effects. Jour-nal of Verbal Learning and Verbal Behavior, 18,645–659.

Tabor, W., Juliano, C., & Tanenhaus, M. K. (1997). Parsin a dynamical system: An attractor-based accoof the interaction of lexical and structural constrainin sentence processing. Language and CognitiveProcesses, 12,211–272.

Tabor, W., & Tanenhaus, M. K. (1999). Dynamical modelssentence processing.Cognitive Science, 23,491–550.

Tanenhaus, M. K., Leiman, J. M., & Seidenberg, M.(1979). Evidence for multiple stages in the processof ambiguous words in syntactic contexts.Journal ofVerbal Learning and Verbal Behavior, 18,427–440.

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

Tabossi, P., Colombo, L., & Job, R. (1987). Accessing lecal ambiguity: Effects of context and dominance. Psy-chological Research, 49,161–167.

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

Vu, H., Kellas, G., Metclaf, K., & Herman, R. (2000). Thinfluence of global discourse on lexical ambiguity reslution. Memory and Cognition, 28,236–252.

(Received November 5, 1999)(Revision received November 6, 2000)Published online August 22, 2001

understanding the constraints on syntactic generation ...jeboland/bolandblodgett_2001.pdf ·...

Documents