grammatical search and reanalysis

Journal of Memory and Language 45,308–336 (2001)doi:10.1006/jmla.2001.2777, available online at http://www.academicpress.com on

Grammatical Search and Reanalysis

David Schneider

Cycorp, Inc., Austin, Texas

and

Colin Phillips

Department of Linguistics, University of Maryland

This paper investigates the extent to which existing structural commitments constrain the human parser’ssearch for grammatical analyses of incoming material, specifically whether a Reanalysis As a Last Resort

using a no re- inside aructural are con-e resultst bias to

T

landE-m

(RALR) strategy applies to sentence parsing. Two self-paced reading experiments investigate this issuestructural ambiguity in which a local, easy reanalysis is pitted against a nonlocal attachment requiringanalysis. This ambiguity is created by embedding classic noun phrase/sentential complement ambiguitiesrelative clause modifying a subject NP. The results of both experiments indicate that readers’ existing stcommitments do constrain their subsequent parsing decisions: nonlocal analyses which avoid reanalysissistently favored over local analyses which require an easy reanalysis. This conclusion is confirmed by thof a subcategorization-bias manipulation in Experiment 2, which shows that readers show a consistenavoid reanalysis, rather than a general bias for either local or matrix clause attachments.© 2001 Academic Press

Key Words:parsing; ambiguity; reanalysis; monotonic parsing; subcategorization.

he main objective of this paper is to investi-e

sios

nne

r

n

notonic parsing(Gorrell, 1995; Marcus, Hindle,enelsse of

orral8).llyn- thisghtelythatins ine-na-n-nded-

gate the degree to which commitments madearlier stages in parsing a sentence constrainchoices that are made at later stages in parAlthough it is clear that parsing a sentence volves many cycles of grammatical search flowed by structural commitment, it is much leclear how the commitments made at one cyconstrain the grammatical search undertakethe next cycle. One possibility is that existicommitments strongly constrain subsequsearch, such that the parser only searchesanalyses of incoming material which respectexisting structural commitments. If this is thcase, then reanalysis would only be consideafter all other options have failed. This approaamounts to the strategy sometimes known asRe-analysis As a Last Resort(RALR) (Fodor & Fra-zier, 1980; Frazier, 1990; Frazier & Clifto1998); it also appears under the heading of mo-

Address correspondence and reprint requests to CPhillips, Department of Linguistics, University of Mary

30

-w.

, 1401 Marie Mount Hall, College Park, MD 20742ail: [email protected].

0749-596X/01 $35.008

at theing.n-l-s

cle atgntfor

alleedch

,

& Fleck, 1983; Weinberg, 1993), and it has beimplemented in activation-based parallel modin the form of lateral inhibition mechanism(Vosse & Kempen, 1999). Alternatively, thparser may always select the optimal analysisthe new material, with little regard to whether not this requires reanalysis of existing structucommitments (Gibson, 1991; Stevenson, 199From a computational standpoint, it is generamore efficient to let existing commitments costrain subsequent parsing choices, becausenarrows the search space. However, in the liof evidence that many reanalyses are relativeasy, many psycholinguists have assumed existing commitments only weakly constrasubsequent search processes (see paperFodor & Ferreira, 1998). The experiments rported here attempt to distinguish these altertives. Studies similar to ours have been coducted by Sturt, Pickering, Scheepers, aCrocker (2001). These studies were developindependently of our own, but they yielded compatible results and are discussed further belo

olin

.

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

eo8

sec

r

it

e

v

ot

t

ore

-c-

isen-ier-, iti-

thee-rs,;

-issy.estierer-ri--ine-di-

ise.ginis

theP

erb

d-e,-

l

GRAMMATICAL SEAR

A good deal of recent work has attemptedcharacterize the structural differences betweasier and more costly reanalyses (e.g., FodFerreira, 1998; Gibson, 1991; Pritchett 191992; Sturt & Crocker 1996), in order to explawhy some parsing errors lead to comprehenbreakdown (classic “garden path” sentencwhereas other parsing errors are easily reered from. Our concern here is not to characize the relative difficulty of different reanalyseRather, we are concerned with the logicaprior question of how the parser searchespossible grammatical analyses of new mateIf the parser always searches for new analythat are consistent with existing commitmenthen we make the following strong predictioeven the very easiest reanalyses shouldavoided, because they are not part of the insearch space, and even globally preferred stures will be avoided, if reanalysis is requiredorder to build them.

An answer to our main question would leus to a more detailed understanding of how stences such as (1) are parsed. The sequencThewoman knows the man wrote … is unambigu-ous, and the only possible analysis of the NPtheman is as the subject of the embedded vwrote, but it is less certain how the parser arriat this conclusion when the verb wrote is firstencountered. If grammatical search is cstrained by existing commitments, then parser will search for other analyses beforetermining that the NP the manmust be reanalyzed as the subject of wrote. If search is noconstrained by existing commitments, this analysis may be found immediately.

The woman knows the man wrote …

S

NP VP (1)The woman

V NPknows the man wrote

An answer to the question of how unambigous sequences like (1) are parsed will shed on general questions of how the parser rap
determines the correct analysis of globally u
CH AND REANALYSIS 309

toenr &8,inions),ov-ter-s.llyforial.sests,n: betialruc- in

aden-

erbes

n-hede--

re-

ambiguous sequences, which are probably mcommon than ambiguous sentences.

We will focus on structures similar to (1), because of two useful properties of these strutures. First, the ambiguous NP (the manin (1))is initially misanalyzed, and therefore reanalysis necessary when the embedded verb is countered (Ferreira & Henderson, 1990; Fraz& Rayner, 1982; Rayner & Frazier, 1987). Although this may not be the case in all contextsis certainly true when the NP is a plausible drect object of the main clause verb and when main verb appears frequently with NP complments in corpora (Garnsey, Pearlmutter, Mye& Lotocky, 1997; Juliano & Tanenhaus, 1994Mitchell & Holmes, 1985; Trueswell, Tanenhaus, & Kello, 1993). Second, the reanalysthat is needed in (1) is known to be quite eaRecent experiments by Sturt and colleagu(Sturt, Pickering, & Crocker, 1999) show thareanalysis in sentences like (1) is much easthan the related reanalysis required at the undlined verb in sentences like (2). These expements confirm widely held intuitions (cf. Gibson, 1991; Pritchett, 1988, 1992; papers Fodor & Ferreira, 1998). The fact that the ranalysis in (1) is so easy makes it a good candate for testing the reality of the Reanalysis As aLast Resort(RALR) constraint.

While the woman was reading the magazine fellonto the floor.

Since the reanalysis in (1) is mandatory, it hard to test RALR using exactly this structurHowever, we can test RALR by embeddinstructures like (1) inside a subject NP, as (3a,3b). The effect of the manipulation in (3) that the underlined verb wrote no longer forcesreanalysis; this verb could be analyzed as main clause verb, taking the sentence-initial Nthe creative womanas its subject, as in (4a). Inthis case, no reanalysis of the NP the funny manwould be required. Alternatively, the verb wrotecould be locally attached as an embedded vwithin the relative clause, with the NP the funnymanreanalyzed as its subject, as in (4b). The adition of a gender-specific emphatic reflexivwhich grammatically requires a local an

u-ightidly

(2)

tecedent, allows the two alternative structures ton-

A

t ofe-

ve8, ofen-ts ofntdi-al-ar-algsm-ib-e-ofcalch-

er-ted ofIn ex- at-

(1)uc-n-ionsmi-cru-

es).

eddi-es

e-g.,

n-ralf arel-

it i)bt

saeln

d

rsre l

310 SCHNEIDER

be distinguished. In effect, this ambiguity pagainst one another (i) a reanalysis which isdependently known to be quite easy and (istructure which does not require reanalysis which goes against an independently motivabias, in this instance the local attachment bia(Altmann, van Nice, Garnham, & Henstr1998; Frazier, 1987; Gibson, PearlmuttCanseco-González & Hickok, 1996; Kimba1973; Phillips & Gibson, 1997; Stevenso1994a; Wanner, 1980).

a. The creative woman who knows the funny manwrotesome comedy sketches herself.

b. The creative woman who knows the funny manwrotesome comedy sketches himself thinks he shoulpublish the sketches.

a. S

NP VP

NP S´ wrote some comedy The creativewoman sketches herself

who S (4)

VP

knows NPthe funny man

(3)

PredictionsIf reanalysis is cost-free, we expect the pa

to pursue the local attachment involving analysis, given the independently motivated

b. NP

NP S´The creative woman

who S

VP

knows S´

S

NP VPThe funny man

wrote some comedy sketches himself

cality bias. On the other hand, if the RALR con

ND PHILLIPS

straint applies, then the nonlocal attachmenthe verb wrote should be pursued, avoiding ranalysis but overriding the locality bias.

To our knowledge, examples like (3) abowere first pointed out by Stevenson (199p.358), who presents them as modificationsexamples in Sturt and Crocker (1996). Stevson recognized these examples as key tesRALR. In Stevenson’s competitive attachmemodel, initial attachments and reanalyses rectly compete with one another, and the locity preference built into Stevenson’s parsing chitecture therefore predicts that the locattachment will be chosen in (3), all other thinbeing equal. Similar predictions apply to a nuber of other parallel parsing models (e.g., Gson, 1991; Jurafsky, 1996). Working within a srial framework, Fodor and Inoue’s model reanalysis (1998, p.105) also predicts that loreanalysis will be chosen over nonlocal attament in (3).

For an interestingly different reason, a prefence for local attachment in (3) is also predicby a number of proposals in the frameworkD(escription)-theory (Marcus et al., 1983). order to account for the ease of reanalysis inamples like (1), it has been suggested thattachment of the underlined verb wrote in directobject/embedded subject ambiguities like does not require retraction of any existing strtural commitments (e.g., Gorrell, 1995; Weiberg, 1993). Phrase-structure representatare assumed to be encoded in terms of donance relations between pairs of nodes (and cially not immediatedominance relations, incontrast to more standard syntactic approachTherefore, when the NP the funny manischanged from a direct object to an embeddsubject, the only structural change is the adtion of new dominance relations among nod(e.g., the new S and S8 both now dominate theNP the funny man). This change requires no rtraction of existing dominance relations (e.main clause VP still dominates NP the funnyman), as illustrated in (5). Due to this nonstadard encoding of phrase-structure, the structu“change” in (1) no longer has the character oreanalysis, and it is therefore predicted to be

sin- auted

,r,l,,

er-

o-
atively easy. For the same reason, approaches-

C

af-m-

ofdy

ex-lowfol-

en-en-

an ande ofe ofnd

nsi-dld,m apre-fter

roup lasteade

ly, a The tock

theted,ext

ae

R8b

n

iln

h

ssr

GRAMMATICAL SEAR

based on D-theory predict the same structuchange to be cost-free when it is embeddedside a relative clause, as is the case in (3).

S S

NP VP NP VPthe woman the woman

V NP V S´knows the knows

funny man (5)

S

NP VPthe wrote

funny man

A number of models predict that reanalyswill be avoided in (3), and hence the nonlocmain clause attachment of the underlined vwill be selected. This prediction is made by srial models which incorporate an explicit RALprinciple (Fodor & Frazier, 1980: Frazier, 197Sturt & Crocker, 1996), and it is also made Vosse and Kempen’s Competitive Inhibitiomodel (Vosse & Kempen, 2000). Vosse aKempen’s model is interesting in the presecontext, both because it is an implemented coputational model and because it captures sosimilar effects to an RALR principle in an actvation-based parallel model, without explicitbuilding in such a principle. The model contaia lateral inhibition mechanism that causes incompatible parses to inhibit one another. Tresults in a reinforcement of highly ranked cadidates and causes reranking to be very coIn this manner, Vosse and Kempen’s model uexisting commitments to constrain the sea

pan-fiedtici-whatandpa-f ae in9, in-

for subsequent analyses, despite its lack of

talipine

b

explicit RALR constraint.

EXPERIMENT 1

This experiment was designed as an initest of whether existing commitments constrgrammatical search. We tested structures those in (3) and (4) above, but we also manlated the probabilistic bias of the embeddverb to take an NP-complement. We includthis manipulation because such probabilistic
ases may affect the strength of the parser’s co
H AND REANALYSIS 311

mitment to an initial analysis, and hence mayfect the parser’s reluctance to give up this comitment.

Participants

Sixty-three members of the University Delaware community participated in the stufor payment. Seven participants (11%) werecluded from the analyses below because of accuracy on the comprehension questions lowing each trial (less than 75% on experimtal sentences or less than 80% on filler stences), leaving a total of 56 participants.

Reading Span Test

All participants took part in a reading sptask based on the test used by DanemanCarpenter (1980). This test was run becausthe complexity of the test sentences, becausthe heterogeneity of the participant pool, abecause of previous reports of differential setivity to probabilistic information in high- anlow-span readers (Pearlmutter & MacDona1995). Participants read sentences aloud frocomputer screen and answered yes/no comhension questions after each sentence. Areading and answering questions about a gof two sentences, their task was to recall theword of each of the sentences that was raloud. If at least four of the initial block of fivpairs of sentences were answered correctblock of five three-sentence sets was tested.procedure continued until a participant failedsuccessfully complete a block (maximum blosize 5 5). The reading-span score is equal tosize of the largest group successfully compleplus 0.2 points per successful trial in the ngroup. In Experiment 1 the mean reading sscore was 2.44 (s.d. 5 0.88). Twenty-five participants who scored 2.6 or higher were classias high-span readers; the remaining 31 parpants were treated as low-span readers. In follows we present the results for high-span low-span readers together, only showing serate results where they differ. The inclusion ocomprehension question after each sentencthis task (following Roberts & Gibson, 199submitted manuscript) forces participants to

ralin-

isl,rbe-

;y

ndntm-me-ys-isn-tly.es

chan

ialinikeu-gdi-
terpret each sentence, and thus reinforces the in-m-

A

as

em

h

mgnb

e

e

l

g

g

ounte

ieaa

en

-

s

san

es,-

omsely

he

P-s of

chli.ri--

t-

312 SCHNEIDER

terference between on-line comprehension rote memorization in this task. This may alexplain why reading-span scores tend to lower in this version of the task than in othversions of the task which do not include coprehension questions.

Materials

Forty-eight sets of four items were used in texperiment, in a 2 3 2 3 2 design, which ma-nipulated the attachment-site (high vs low) the ambiguous verb, the presence or absenctemporary ambiguity at the verb, and the coplement bias of the embedding verb (stronNP-biased vs weakly NP-biased). All conditiobegan with a subject NP which was modified a subject relative clause. The relative claucontained a verb (henceforth the embeddingverb) which allows both NP complements ansentential complements. The embedding vwas followed by an NP (the ambiguousNP)which could be analyzed as the object of the ebedding verb or as the subject of a sentencomplement. The ambiguous NP was followby a transitive verb (the ambiguous verb), whichcould be attached either as a main clause v(high attachment) or as an embedded verb (attachment). Placed four words after the abiguous verb was a gender-marked emphaticflexive that disambiguated toward either a hior a low attachment.

The main clause subject NP and the ambious NP were always animate NPs with contraing gender. This ensured that the disambiguing reflexive was an effective disambiguatand it also ensured that both the main clasubject and the ambiguous NP were semacally good candidates to be the subject of ambiguous verb. The four-word delay betwethe ambiguous verb and the disambiguation wincluded in order to be certain that commments to the attachment of the ambiguous vhad been established. The reflexive was alwfollowed by a four-word PP and additional mterial to ensure a grammatical and natural copletion of the sentence. Emphatic (nonargumtal) reflexives were used rather than argumereflexives, in order to lengthen the delay b
tween the ambiguous verb and the disambigu
ND PHILLIPS

ndober-

e

ofe of

-lysy

se

drb

m-tiald

erbowm-re-h

u-st-at-r,seti-

henast-rbys-

m-n-tale-

tion, and in order to avoid anticipation of a reflexive at the ambiguous verb itself.

The low unambiguous control condition wacreated by addition of the complementizer that.The high unambiguous control condition wacreated by replacing the ambiguous NP with accusative-marked pronoun (him or her). Asample set of stimuli is shown in (7), where thitalicized words are the crucial manipulationand a full list of materials is provided in Appendix B.

a. low ambiguous

The creative woman who knows the funny manwrote

some comedy sketches himselfabout the amusing es-

capades thinks he should publish them.

b. low unambiguous

The creative woman who knows that the funny man

wrote some comedy sketches himselfabout the amus-

ing escapades thinks he should publish them.

c. high ambiguous

The creative woman who knows the funny manwrote

some comedy sketches herselfabout the amusing es-

capades she had seen.

d. high unambiguous

The creative woman who knows him wrote some

comedy sketches herselfabout the amusing escapades

she had seen.

The embedding verbs were divided into twclasses, based on sentence completion norcompiled by Susan Garnsey (p.c.). In 24 of thstimulus sets the embedding verb was strongbiased toward an NP complement (83%1 NP-completions as a proportion of all that-less com-plements). In the remaining 24 stimulus sets tembedding verb was either neutralor weaklybi-ased toward an NP complement (52–78% Ncompletions). Due to the small number of verbin each class, all verbs appeared in a numberexperimental items. Within each class, eaverb was used in the same number of stimuNo S-biased verbs were included in this expement because of the possibility that NPs following strongly S-biased verbs might be initially a

(7)

a-tached as subjects of embedded clauses rather

C

9 ae

thdo

oou am

ou tan

uwa

c tento. o

w

aciomin

aiaivied

ne

ssb-

to foras

in

ionsici-

u- insig-n

n-

hatlowheow

-inhe

meub-Atub- thesesualx-

GRAMMATICAL SEAR

than as direct objects (Garnsey et al., 19Juliano and Tanenhaus, 1994; Trueswell et1993), thereby eliminating the trade-off btween local attachment and reanalysis costthe experiment is based upon. All verbs usethe experiment also satisfied the additional cstraint that they allow animate NPs as directjects. This constraint limited the number verbs that could be used in the experiment, bguaranteed that all candidate subjects of thebiguous verb were plausible. If inanimate abiguous NPs had been used, then this chave created a semantic bias for the verb tothe higher, animate NP as its subject, indepeent of any RALR constraint.

Procedure

Sentences were presented on a compscreen using a self-paced reading paradigm a one-word moving-window display (Just, Cpenter, & Woolley, 1982), using the mw-runsoftware for Macintosh developed at MIT. Eatrial began with a series of dashes markinglength and position of the words in the stences. Participants pressed the space bar veal each successive word of the sentenceeach new word appeared, the preceding wdisappeared. The time spent on each wordrecorded. All sentences were followed byyes/no comprehension question. Feedbackaccuracy was given immediately after equestion. Most of the comprehension questquestioned the ambiguity itself, but with sovariation to prevent participants from adoptexperiment-specific strategies.

Up to 100 characters could appear on eline of the display, which allowed all materup to and including the disambiguating reflexto appear on a single line. Following a brpractice session, sentences were presentepseudo-random order in a single block, consing of 48 experimental sentences, 36 sentefrom an unrelated experiment, and 65 filler stences.

Results

Comprehension accuracy and residual reing times at each region were entered into a
peated-measures ANOVA, with ambiguity, at

7; l.,-at

inn-b-ft itm--ldked-

terithr-

hhe- re-Asrdasaonhnseg

chlef in

tachment site (high/low), and verb-cla(strongly vs weakly NP-biased) as within-sujects factors.

Comprehension Accuracy

The overall level of accuracy in responsescomprehension questions was 87%. Resultsall conditions are shown in Table 1. There wno effect of reading span (High versus Low)overall comprehension accuracy (all Fs , 1).Responses to the comprehension questshowed a main effect of attachment site (partpant analysis,F1(1,55) 5 26.61, MSe 5 2.96,p , .0001; item analysis,F2(1,42) 5 29.52,MSe 5 3.02,p , .0001), due to greater accracy in the high-attachment conditions thanthe low-attachment conditions. There was a nificant ambiguity 3 attachment site interactio(F1(1,55) 5 9.72, MSe 5 1.08, p , .01;F2(1,42) 5 10.44, MSe 5 1.07,p , .01), due toa cost of ambiguity in the low-attachment coditions (F1(1,55) 5 7.23, MSe 5 .991,p , .01;F2(1,42) 5 7.36, MSe 5 .929,p , .01), but notin the high-attachment conditions. The fact tcomprehension accuracy was lowest in the ambiguous conditions is consistent with tself-paced reading results (below), which shthe greatest difficulty at the point of disambiguation in this condition. There were no maeffects or interactions involving verb class in tcomprehension questions.

Self-Paced Reading

A regression equation predicting reading tifrom word length was constructed for each sject, using all items (filler and experimental). each word, the reading time predicted by the sject’s regression equation was subtracted fromactual measured reading time, and all analywere performed on these differences (resid

ist-cesn-

ad- re-

reading times). This transformation removes e

TABLE 1

Comprehension Accuracy, Experiment 1

Ambiguous Unambiguous

Low 80% 86%

-High 90% 88%

A

eu

6tes

an

er

y-filor

am rm

ioilas

t w r

ses,ing

ndes

e-ve

-

ithwasex-re-

igu-as

-at-

in-am-

314 SCHNEIDER

traneous variance by subtracting out a baselineach subject, and by controlling for noise dto length effects (Ferreira & Clifton, 198Trueswell & Tanenhaus, 1991). All times reporhere are based on the residual reading timetrials in which the comprehension question wanswered correctly. Reading times greater thstandard deviations from the mean were trimmto 4 s.d. (1108 ms), affecting less than 1%words; reading times greater than 8000 ms wexcluded, affecting less than 0.01% of trials. Tals for the verb fearwere removed from the analsis due to an unusually flat reading-time proAdditionally, the high unambiguous condition item (5) was removed due to a presentation ethat rendered the sentence ungrammatical.

There are two areas of particular interest inconditions. First, reading times at the disabiguating region (region 8 and subsequentgions) can help determine which parsing comitments were made in the ambiguous regSecond, reading times in the ambiguous reg(regions 3–7) provide information on the retive difficulty of making those commitmentBecause the reading times in the ambiguousgions are more easily understood in the lighthe results in the disambiguating regions,take the unorthodox step of presenting the
sults from regions 8–11 before the results fro
publish them 14 /).

ND PHILLIPS

fore

;dfor

as 4

edofrei-

e.fror

ll-

e--

n.on-

.re-ofee-

embedding verbs are combined in all analydue to the absence of interactions involvverb class.

Residual reading times for all conditions aregions are shown in Fig. 1 (raw reading timcan be found in Appendix A). At the word immdiately following the disambiguating reflexithere was a main effect of ambiguity (F1(1,54) 545.94, MSe 5 2340905,p , .0001; F2(1,42) 545.19, MSe 5 2392136,p , .0001), a main effect of attachment site (F1(1,54) 5 10.85, MSe 5552687,p 5 .001; F2(1,42) 5 10.83, MSe 5572958,p 5 .001), and an ambiguity 3 attach-ment site interaction (F1(1,54) 5 25.11, MSe 51279537,p , .0001; F2(1.42) 5 25.21, MSe 51334320,p , .0001). Each of these effects, wthe exception of the effect of attachment site,already significant at the disambiguating reflive itself (region 8), and all of the effects mained significant at regions 10–12.

Pairwise comparisons showed that the ambity 3 attachment site interaction at region 9 wdue to a strong cost of ambiguity in the lowtachment conditions (F1(1,54) 5 58.31, MSe 53407338,p , .0001; F2(1,42) 5 57.41, MSe 53522229,p , .0001), but no cost of ambiguity the high-attachment conditions (p..15). This pattern of results clearly suggests that the dis
mbiguating reflexive causes most difficulty in the
it sug-
regions 3–7. Strongly and weakly NP-biasedlow-attachment conditions. In other words,
FIG. 1. Residual reading times for Experiment 1 (The creative woman who knows 1 / that 2 / the funny man 3 /wrote 4 / some 5 / comedy 6 / sketches 7 / himself 8 / about 9 / the 10 / amusing 11 / escapades 12 / thinks 13 / he should

C

a

echan

atn

gan

n3

isdote nhwh,hu

a

i

i

hbi

ter-

glyusehisgu-m-

i-nsg-

the

on-sisib-n-nt.thetor.is-

asuseu-

ealrs-ue,hating-

owthisx-alllessticbate,NP-

GRAMMATICAL SEAR

gests that participants strongly favor a mclause analysis of the ambiguous verb.

Despite the evidence that high-attachmconditions were much easier than low-attament conditions, analysis of the high-attacment conditions at region 9 reveals a significinteraction of ambiguity with reading-spagroup (F1(54) 5 4.89, MSe 5 216234,p , .05;F2(1,42) 5 5.41, MSe 5 247511,p , .05). Thisinteraction was due to the fact that low-spreaders also showed a cost of ambiguity in high-attachment conditions (high attachmeF1(1,30) 5 4.00, MSe 5 190916, p , .05;F2(1.42)54.80, MSe 5 234637,p , .05), buthigh-span readers did not (all Fs , 1). How-ever, even for the low-span readers, the matude of the slowdown due to ambiguity wmuch greater in the low-attachment conditio(ambig., 146 ms; unambig., 20 ms; slowdow126 ms) than in the high-attachment conditio(ambig., 56 ms; unambig., 25 ms; slowdown,ms). This difference could indicate that incorect low attachments are more difficult to revthan incorrect high attachments, or it could incate that high attachments were more commor a combination of the two. Whereas the cosambiguity in the low-attachment conditions psisted until region 12 and beyond, the costambiguity in the high-attachment conditiowas only observed at region 9. Among higspan readers, the magnitude of the slowdodue to ambiguity was large in the low-attacment conditions (ambig., 121 ms; unambig.ms., slowdown, 113 ms), and very small in thigh-attachment conditions (ambig., 44 ms.,ambig., 42 ms; slowdown, 2 ms).

At the ambiguous verb in region 4 there wno main effect of ambiguity (all Fs , 1), butthere was a main effect of attachment s(F1(1.54) 5 12.34, MSe 5 877088,p , .001;F2(1.42) 5 12.63, MSe 5 915345,p , .001),which persisted until region 5, and an ambigu3 attachment site interaction (F1(1,54) 5 32.21,MSe 5 2288876,p , .0001; F2(1.42) 5 28.95,MSe 5 2098477,p , .0001), which persisteduntil region 7. Pairwise comparisons reveal tthis interaction was due to main effects of amguity for both attachment sites, but with oppos
directions. The low ambiguous condition wa

in

nth--nt

nhet:

ni-ss

n,s1

r-ei-n,

ofr-ofs-n-9en-

s

te

ty

ati-

te

read more slowly than its unambiguous counpart (F1(1,54) 5 22.31, MSe 5 1261511,p ,.0001; F2(1,42) 5 21.8, MSe 5 1287387,p ,.0001), possibly because the NP is more stronpredicted after the complementizer, or becaits attachment is unambiguous and local in tposition. On the other hand, the high unambious condition was read more slowly than its abiguous counterpart (F1(1,52) 5 12.5, MSe 51051882,p , .001; F2(1,42) 5 10.15, MSe 5867926,p , .01). The high unambiguous condtion remained slower than all other conditiountil region 7, the word immediately precedinthe reflexive. In light of the finding that high ambiguous conditions were relatively easy at disambiguation point, we conclude that botham-biguous and unambiguous high-attachment cditions were given a high-attachment analystarting at region 4. Therefore, we cannot attrute the slowdown in the high unambiguous coditions to the cost of finding the high attachmeA more likely cause is the unnaturalness of unanchored pronoun used as a disambiguaWe address this issue further below in the dcussion and in Experiment 2.

Discussion

The clearest finding of this experiment wthat readers consistently chose the main claanalysis (i.e., high-attachment) of the ambigous verb. This finding contrasts with a good dof evidence for a local attachment bias in paing ambiguities where reanalysis is not an issand therefore provides compelling evidence tparsing decisions are constrained by existcommitments, as the Reanalysis As a Last Resort principle predicts.

The preference for high attachments over lattachments was equally clear with bostrongly and weakly NP-biased verbs. Thcould be taken to indicate that the effect of eisting commitments is not graded, and that existing commitments are respected, regardof how well they are supported by probabilisinformation. Alternatively, the lack of a verclass difference could simply reflect the fact thour manipulation was a relatively weak onsince both verb classes were to some extent
sbiased. Ferreira and Henderson (1990) failed to

A

iad

n

hdtft

in th7u

izsr

ioor ao

wdeea

t Ai

a

a

foers.

es-

a- re-an

bi-r &

uredven. Iferusre-l-

wnonof

sn.ityashe

edrst,teu-

at-use

(cf.m-lo-the

asdi-es

ch-sedndh-

aynip-

ulation, and the anomalous behavior of low-

316 SCHNEIDER

find any evidence of verb bias effects in initparsing decisions, but they did notice some ferences in ease of recovery from error basedverb bias. Trueswell et al. (1993) found stroevidence of verb-class effects at the point of dambiguation. It should be noted, however, tthe verb bias differences in the materials usethe Ferreira and Henderson and Trueswell estudies were greater than the verb bias difences in this experiment, and this might be reason for the lack of findings here.

Inferences that can be drawn from readtimes at the disambiguating regions allow usconstrain the possible interpretations of reading times in the ambiguous region (3–The slower reading time for the low ambiguocondition relative to its control is expected, bobecause the disambiguating complementcould facilitate the reading times for the subquent words, and because the ambiguous veattached nonlocally as a main clause verb inambiguous condition, an attachment which mrequire extra time.

The fact that the high unambiguous conditwas read more slowly than the high ambigucondition is more surprising. Based on the sults at the disambiguating region, we canconfident that participants give the same mclause analysis to the ambiguous verb in bhigh ambiguous and high unambiguous contions. Therefore, we cannot attribute the slodown in the unambiguous condition to any adtional cost of the high attachment. We suggthat the slowdown reflects difficulty associatwith the pronoun that is used as a disambiguin this condition. In a string such as The creativewoman who knows him . . . the pronoun is noanchored to any existing discourse referent.though this is fully grammatical and natural some English contexts (as in The people whoknow him best think the President is a genius), itis unnatural in this context and may trigger ultimately fruitless search for an antecedent.

Although high-span readers and low-spreaders showed similar results at most regiothe one striking difference was the evidencea garden path in both high-attachment and lattachment conditions for the low-span read
This suggests that they did not always attach
ND PHILLIPS

lif- ong

is-at in al.er-he

gtoe).s

there-b istheay

nuse-beinth

di--i-stdtor

l-n

n

nns,orw-

ambiguous verb high, but they sometim(probably infrequently) attached it low, in apparent violation of RALR. One possible explnation for the contrast between span groupslates to independent evidence that low-spreaders are less constrained by probabilisticases than are high-span readers (PearlmutteMacDonald, 1995). This could mean that olow-span readers sometimes initially analyzthe ambiguous NP as an embedded subject, ebefore the ambiguous verb was encounteredthis happened, then RALR would no longblock the low attachment of the ambiguoverb. However, the span group difference at gion 9 should be treated with some caution. Athough high-span readers showed no slowdodue to ambiguity in the high conditions at regi9, they showed a marginally significant effect ambiguity at region 8 (F1(1,24) 5 2.49, MSe 5101172, p , .12; F2(1,42) 5 3.14, MSe 5119775,p , .08), due to slower reading timefor the unambiguous high-attachment conditioTherefore, we cannot exclude the possibilthat a small cost of ambiguity at region 9 wmasked by an independent slowdown in thigh unambiguous condition.

There are a number of questions which neto be addressed by a further experiment. Fialthough we took care to make both animaNPs equally plausible subjects for the ambigous verb, it might be objected that the high tachments were made because the main clasubject-verb dependency has special statusGibson, 1998). This can be addressed by ebedding the test structures such that the noncal attachment site does not correspond to main clause.

Second, the use of an unanchored pronoundisambiguator in the high-attachment contions may have introduced spurious differencbetween the high-attachment and low-attament conditions. This problem can be addresby providing an antecedent for the pronoun aby using matched disambiguators in higattachment and low-attachment conditions.

Third, the lack of verb-class differences mbe due to the weakness of our verb-class ma

thespan readers may have been due to their relative

C

nd

tie on

vnoo

),eers

inot

hi

d

N th

gg

dyreowlessesstal

asintheb-in

althoftwobi-rbsixi-eas-d asusd,

ely

GRAMMATICAL SEAR

insensitivity to probabilistic verb argumestructure biases. These can be addresseusing a wider variety of verb classes.

The use of a stronger verb-bias manipulaalso makes it possible to address a couplother accounts of the high attachment biasserved in Experiment 1. The high attachmeobserved there may be due to a simple “aembedding” constraint, which attempts to mimize the degree of embedding at any chpoint in a sentence (and thus necessarily cpetes against local attachment constraintsthey may be due to a constraint which placcost on long subject NPs (the low attachmlengthens the subject NP). It has been obsethat subject NPs tend to be shorter than NPother positions (e.g., Behagel, 1932; Hawk1994). Attention to this statistical property language may lead comprehenders to adopearly closure strategy for subject NPs. If eitof these alternatives is responsible for the hattachment effect, then we should expectacross-the-board high-attachment bias, inpendent of the bias of the embedding verb (bias vs S-bias), and hence independent ofinitial attachment of the ambiguous NP. On other hand, if RALR is responsible for the higattachment bias, then the effect should be hisensitive to the initial attachment of the ambi

r

a

sr,

d

th

at-anm-

ineran-m-f afor

ns.un tons

msstm-nt.n-

ous NP. Experiment 2 was designed to addeach of these issues.

EXPERIMENT 2

This experiment was designed to replicand extend the results of Experiment 1, andaddress the alternative accounts of the resfound there. The basic structures tested weresame as in Experiment 1, but a numberchanges were made in order to address isarising from the first experiment. In this expement we used a wider range of verb classesmore closely matched the high and low unabiguous conditions, and we ensured that all pnouns had an antecedent in the existing course.

Participants

Sixty-four undergraduate students from
University of Delaware, none of whom partic

t by

on ofb-ts

oidi-icem-or

s ant

ved ins,f anerghane-P-thehe-

hlyu-ess

te toults theofuesi-we

m-ro-is-

e

pated in Experiment 1, participated in the stufor payment. Eight participants (12.5%) weexcluded from the analysis based on their laccuracy on the comprehension questions (than 65% on experimental sentences, or lthan 80% on the filler sentences), leaving a toof 56 participants. The exclusion criterion wslightly more liberal in this experiment than Experiment 1, due to the increased length of experimental stimuli, but the percentage of sujects excluded was almost identical to that Experiment 1.

Materials

Materials for this experiment were identicin format to those used in Experiment 1, withe following exceptions. First, four classes embedding verbs were used, instead of the classes used in Experiment 1: strongly NP-ased, weakly NP-biased, S-biased, and vethat only allow NP complements. See AppendC for full details of the verbs used. The condtions with NP-only verbs were included in thexperiment in order to provide a baseline meure of the difficulty of the high attachment, anin order to ensure that the use of pronounsearly disambiguators did not introduce spuriodifferences independent of ambiguity. Seconthe unambiguous conditions were more closmatched across high-attachment and low-tachment conditions than in Experiment 1—accusative pronoun was used in the high unabiguous condition, and a nominative pronounthe low unambiguous condition. Third, in ordto avoid any confounds due to the lack of an tecedent for the pronouns, all items were ebedded as the right-branching complement ohigher clause which provided an antecedent the pronouns in the unambiguous conditioThe antecedent for the disambiguating pronoin the unambiguous conditions correspondedthe ambiguous NP in the ambiguous conditio(e.g.,the funny manin example (8) below). Theadditional clause made the experimental iteslightly longer, but more natural. The higheclause always contained a verb which unabiguously selects a finite sentential complemeFourth, due to the large numbers of stimuli i

i-volved, only 20 sets of items were tested for

t

m

tt

t thee t

a

a

-s p

ri-in-nt,di- ab-uns,ses

acycylybeestion

ity.im-ersin-a

c-un-

ac-lowionthe

if-n

below in the analysis of reading times.

318 SCHNEIDER

each verb class. A sample set of stimuli for experiment is shown in (8). Note that there wno low-attachment conditions for the itemsthe NP-only conditions, since these wouldungrammatical.

a. low ambiguous

The talent agencythinks that the creative woman who

knows the funny manwrote some comedy sketches him-

selfabout the amusing escapades wants to publish the

b. low unambiguous

The funny manthinks that the creative woman who

knows hewrote some comedy sketches himselfabout

the amusing escapades wants to publish them.

c. high ambiguous

The talent agencythinks that the creative woman who

knows the funny manwrote some comedy sketches

herselfabout the amusing escapades she had seen.

d. high unambiguous

The funny manthinks that the creative woman who

knows himwrote some comedy sketches herselfabout

the amusing escapades she had seen.

Procedure

The procedure was identical to that in Expiment 1, except where noted here. Due tolarge number of trials in this experiment,self-paced reading task was divided into blocks of trials each containing 100 senten(35 experimental sentences, 36 sentences unrelated experiments, 29 fillers). The lengththe filler sentences was matched to the lengthe experimental sentences. The order ofblocks was counterbalanced. In this experimall of the comprehension questions involvedambiguity—the question could only be aswered correctly if the ambiguity had been propriately resolved.

Reading Span Task

All participants took part in a reading sptask, as in Experiment 1. The mean reading sscore was 2.21 (s.d. 5 0.84). Twenty-nine participants who scored 2.4 or higher were clafied as high-span readers; the remaining 27

(8)

ticipants were classified as low-span readers

AND PHILLIPS

hisereinbe

.

er-thehewocesfrom ofh of

Results

Analysis procedures were identical to Expement 1, except that the verb class factor cluded two additional levels in this experimeS-bias and NP-only. Since the NP-only contions were always unambiguous, even in thesence of unambiguously case-marked pronothese conditions are not included in the analyunless specifically noted.

Comprehension Accuracy

Table 2 shows mean comprehension accurscores for Experiment 2. The overall accuralevel in this experiment was 80%. The slightlower accuracy level in this experiment canattributed both to the increased length of the tsentences and to the fact that all comprehensquestions specifically tested the ambiguOverall comprehension accuracy was very silar for high-span (80.2%) and low-span read(79.6%), and there were no main effects orteraction involving span group. There wasmain effect of ambiguity on comprehension acuracy, due to higher accuracy scores on theambiguous conditions (F1(1,54) 5 15.62,p ,.0001; F2(1,57) 5 16.42, p , .0001). Therewas also a significant attachment site3 verbclass interaction (F1(2,54)5 12.21,p , .0001;F2(2,57)5 12.41,p , .0001). This interactionwas due to the tendency for comprehensioncuracy to increase as S-bias increased in theattachment conditions, and for comprehensaccuracy to decrease as S-bias increased inhigh-attachment conditions. This pattern of dficulty parallels the patterns of difficulty see

.

nt,hen-p-

npan

si-ar-

TABLE 2

Comprehension Accuracy, Experiment 2

Ambiguous Unambiguous

Low, strong NP-bias 71% 83%Low, weak NP-bias 74% 77%Low, S-bias 78% 85%High, NP-only 91% 86%High, strong NP-bias 82% 87%High, weak NP-bias 79% 84%
High, S-bias 71% 75%

C

rthc-

hn

eeothsh emeont ror

nsoml-a

-n

a

-n

miteTi

th

di-ntialdi-g,heillta-lts

nig.s”reity

- atnsno

m-l-

-ussein-nt.

th3.-in

t

hahet-

GRAMMATICAL SEAR

Self-Paced Reading

As in Experiment 1, all times reported heare residual reading times for trials in which comprehension question was answered rectly. Three trials (. 0.01%) were removed because of residual reading times greater t8000 ms. Residual reading times greater thastandard deviations from the mean wtrimmed to 4 s.d. (1194 ms); this affected lthan 1% of words. Item (44) was removed frthe analysis due to a missing region, with consequence that only 19 weakly NP-biaitems were analyzed, compared to 20 for otverb classes. As in Experiment 1, for ease ofposition we report reading times at the disabiguating region (regions 8 and following) bfore discussing reading times in the ambiguregion (regions 3–7). Region numberimatches the numbering used in ExperimenBecause the complementizer that was no longeused as a disambiguator, there was no regiin this experiment. Raw reading times are ported in Appendix A.

The ANOVA revealed significant interactioinvolving verb class and attachment site, in bthe ambiguous regions and following the disabiguation point. At the word immediately folowing the ambiguous verb (region 5) there wan ambiguity 3 verb class interaction (F1(2,54)5 4.95, MSe 5 265482,p , .01; F2(2,56) 55.24, MSe 5 279262,p , .01) and an ambiguity 3 attachment site 3 verb class interactio(F1(2,54) 5 4.36, MSe 5 233794,p , .05;F2(2,56) 5 4.97, MSe 5 265307,p , .01). Atthe word immediately following the disambiguating reflexive (region 9) there was an tachment site 3 verb class interaction (F1(2,54) 57.80, MSe 5 430588,p , .001; F2(2,56) 58.64, MSe 5 487289,p , .001) and an ambiguity 3 attachment site 3 verb class interactio(F1(2,54) 5 4.18, MSe 5 230867,p , .05;F2(2,56) 5 3.25, MSe 5 183256,p , .05).Based on this evidence that the verb class nipulation modulates the effects of ambiguand attachment site, all further analyses wconducted separately for each verb class. analyses focus on the regions following the dambiguation point (regions 8–12), and on
ambiguous regions (regions 4–7). Reading tim

eeor-

an 4

ressme

ederx--

-usg1.

n 2e-

th-

s

-t-

a-yrehes-e

differences at the NP in region 3 are of lessrect concern here, since there was a substalength difference between the ambiguous contions, in which the NP was three words lonand the unambiguous conditions, in which tNP was a pronoun. As in Experiment 1, we wonly mention span-group differences where stistically justified. In all other cases, the resuwill be for all subjects.

NP-only verbs. Residual reading times iconditions with NP-only verbs are shown in F2. Very few differences between “ambiguouand “unambiguous” NP-only conditions weobserved. There were main effects of ambiguat region 6 (F1(1,54) 5 4.91, MSe 5 227988,p , .05; F2(1,9) 5 5.00, MSe 5 239375,p ,.05) and at regions 10–12 (F1(1,54) 5 7.15,MSe 5 186538,p , .01; F2(1,9) 5 7.08, MSe5 198192,p , 0.01). At region 6 the “ambiguous” conditions were read more quickly, andregions 10–12 the “unambiguous” conditiowere read more quickly. Since there were differences between “ambiguous” and “unabiguous” NP-only conditions immediately folowing the disambiguation (regions 8–9) or immediately following the onset of the ambiguoregion (regions 4–5), we conclude that the uof pronouns as early disambiguators did not troduce spurious differences in this experime

Strongly NP-biased verbs. Disambiguation.Residual reading times in conditions wistrongly NP-biased verbs are shown in Fig.At the word immediately following the disambiguating reflexive (region 9) there was a maeffect of ambiguity (F1(1,54) 5 11.94, MSe5567068,p , 001; F2(1,19) 5 10.15, MSe5504212,p , .01), a main effect of attachmensite (F1(1,54) 5 15.96, MSe5 757975,p ,.0001;F2(1,19)5 15.951, MSe5 792529,p ,.0001), and an ambiguity3 attachment site in-teraction (F1(1,54)5 8.88, MSe5 421673,p ,.01;F2(1,19)5 7.25, MSe5 360176,p , .01).However, pairwise comparisons showed tthe main effects are misleading, and that t

cost of ambiguity was restricted to the low-atachment conditions (high-attachment, allFs ,1; low-attachment,F1(1,54) 5 14.25, MSe5879135,p , .001; F2(1,19) 5 12.27, MSe5

e819816,p , .001). This pattern of results sug-

eveht

Fx

thi-

non

df-

at the

the 10 / fund 11 / drive 12 / to see if he had applied for aid 14 /).

gests that readers commit to the high attachmbefore reaching the disambiguating reflexileading to significant slowdown when the rflexive disambiguates toward the low-attacment parse. For high-span readers this issame pattern observed in Experiment 1.low-span readers this pattern differs from E

telescope 7 / herself 8 / for 9 / the 10 / small observatory 14 /).

nte,--

heor-

showed a significant cost of ambiguity in bohigh-attachment and low-attachment condtions, whereas in this experiment there wasevidence of difficulty at the disambiguation ithe high ambiguous condition.

Ambiguous region. At the ambiguous NP anthe two following words (regions 3–5) main e

320 SCHNEIDER AND PHILLIPS

FIG. 2. Residual reading times for NP-only conditions in Experiment 2 (The homeless people think thbrilliant woman who supported 1 / the impoverished man 3 / called 4 / the 5 / large 6 / charity 7 / herself 8 / after 9 /

lower

ere

periment 1: in Experiment 1 low-span readersfects of ambiguity were observed, due to s

FIG. 3. Residual reading times for strongly NP-biased verbs in Experiment 2 (The faculty members wgrateful that the generous benefactor who understood 1 / the persuasive woman 3 / bought 4 / the 5 / powerful 6 /

11 / college 12 / promised 13 / to fund a new

C

s

na

t,

oa

reutteri-at

e

teronse-rs

r

n-

wn-

2)

heh

GRAMMATICAL SEAR

reading times for the unambiguous condition(region 3,F1(1,54) 5 25.51, MSe 5 1245280,p , .0001; F2(1,19) 5 27.83, MSe 5 1336473,p , .0001; region 4,F1(1,54) 5 4.47, MSe 5334924, p , .05; F2(1,19) 5 4.57, MSe 5351057,p , .05; region 5,F1(1,54) 5 5.91,MSe 5 356067,p , .05; F2(1,19) 5 6.60,MSe 5 404011,p , .05).

At the ambiguous NP (region 3) there walso an ambiguity 3 attachment site interactio(F1(1,54) 5 4.87, MSe 5 237853,p , .05;F2(1,19) 5 4.18, MSe 5 201106,p , .05). Thisinteraction was due to a greater slowdownthe nominative pronoun in the low unambigous condition than for the accusative pronouthe high unambiguous condition, although pwise comparisons revealed that the slowdowas significant for both high- and low-attacment conditions (high-attachment,F1(1,54) 53.66, MSe 5 161920,p 5 .055; F2(1,19) 55.02, MSe 5 218523,p , .05; low-attachmenF1(1,54) 5 19.76, MSe 5 1067996,p 5 .0001;F2(1,19) 5 24.08, MSe 5 1275771,p , .0001).The greater slowdown at the low unambiguattachment condition may reflect the fact thnominative pronoun is unexpected followingstrongly NP-biased verb. The ambiguity 3 at-tachment site interaction did not reach signcance at regions 4–5 (all Fs , 1). Note that this
p
the 10 / amusing 11 / escapades 12 / thinks 13 / he should p


asn

foru- inir-wnh-

ust a a

ifi-

where the low unambiguous conditions weread faster than all other conditions throughothe ambiguous region (see Fig. 1). The greasimilarity between the two unambiguous condtions in this experiment may reflect the fact ththe conditions were more closely matched.

At the ambiguous verb (region 4) therwas an ambiguity 3 span-group interaction(F1(1,54) 5 11.41, MSe 5 854597,p , .001;F2(1,19) 5 1.19, MSe 5 781983,p , .01), dueto the fact that the cost of ambiguity was greaamong low-span readers. Pairwise comparisrevealed that the effect of ambiguity in this rgion was significant for low-span reade(F1(1,26) 5 11.80, MSe 5 1109050,p , 0.001;F2(1,19) 5 11.18, MSe 5 1079270,p , .001;ambig.,214 ms; unambig., 46 ms), but not fohigh-span readers (all Fs , 1; ambig., 32 ms;unambig., 9 ms). The interaction with spagroup was not significant at any other region.

Weakly NP-biased verbs. Disambiguation.Results for weakly NP-biased verbs are shoin Fig. 4. At the disambiguating reflexive (region 8) and the following words (regions 9–1there was a significant ambiguity 3 attachmentsite interaction, due to a cost of ambiguity in tlow ambiguous condition, but not in the higambiguous condition (region 8,F1(1,54) 54.03, MSe 5 249324,p , .05; F2(1,18) 5 4.42,

that

attern of results contrasts with Experiment 1,MSe 5 279302,p , .05; region 9,F1(1,54) 5

FIG. 4. Residual reading times for weakly NP-biased verbs in Experiment 2 (The talent agency thinks the creative woman who knows 1 / the funny man 3 / wrote 4 / some 5 / comedy 6 / sketches 7 / himself 8 / about 9 /

ublish them 14 /).

A

:

xlo

t rn

nt

tg

o

nt.he-

inik-. In inlty

ol-ifi-of

oh-

n re-8antity

1

322 SCHNEIDER

4.29, MSe 5 247247,p , .05; F2(1,18) 5 6.30,MSe 5 358144, p , .05; regions 10–12F1(1,54) 5 3.31, MSe 5 106276, p , .07;F2(1,18) 5 3.86, MSe 5 131936,p , .05). Themain effect of attachment site was also signcant, beginning at the word following the refleive, and was due to slower reading times for conditions (region 9,F1(1,54) 5 8.97, MSe 5516875, p , .01; F2(1,18) 5 9.37, MSe 5532526,p , .01; regions 10–12,F1(1,54) 510.81, MSe 5 348488,p 5 .001; F2(1,18) 58.10, MSe 5 276794,p , .01). No interactionsinvolving reading span group were significanany of these regions. At these regions, thesults for weakly NP-biased verbs were similathose for strongly NP-biased verbs. The osignificant costs of ambiguity were observedthe low-attachment conditions, indicating ththe ambiguous verb was consistently givehigh-attachment analysis. Nevertheless,magnitude of the slowdown in the low-attacment conditions was somewhat smaller with weakly NP-biased verbs than with the stronNP-biased verbs.

Ambiguous region. There was a main effect ambiguity at the ambiguous NP (region 3), dto slower reading times for the unambiguoconditions (F1(1,54) 5 7.79, MSe 5 316463,p, .01; F (1,18) 5 8.80, MSe 5 363749,p ,
2
.0

13 / that he had anything to do with the purchase 14 /).

ND PHILLIPS

ifi--w

atre- tolyinat aheh-hely

fueus

tions were significant or marginally significaAt region 6 the only significant effect was tambiguity 3 attachment site 3 span group interaction (F1(1,54) 5 4.26, MSe 5 216490,p ,.05; F2(1,18) 5 3.92, MSe 5 200702,p , .05).

S-biased verbs. Disambiguation. Results forconditions with S-biased verbs are shownFig. 5. Results with this verb class differed stringly from the results with other verb classesparticular, there was a substantial changewhich conditions caused the greatest difficuat the disambiguating regions.

At the disambiguating reflexive and the flowing word (regions 8–9) there were signcant or marginally significant main effects ambiguity (region 8,F1(1,54) 5 3.49, MSe 5249918, p , .07; F2(1,19) 5 3.46, MSe 5253565,p , .07; region 9,F1(1,54) 5 9.14,MSe 5 556718,p , .01; F2(1,19) 5 8.37, MSe5 528453,p , .01). At region 8 there was alsa marginally significant main effect of attacment site (F1(1,54) 5 3.11, MSe 5 222662,p ,.08; F2(1,19) 5 2.75, MSe 5 201270,p , .10).

The ambiguity 3 attachment site interactiodid not reach significance at either of thesegions (all Fs , 1). However, at both region and region 9 there were marginally significinteractions with reading span group (ambigu3 span, region 8,F (1,54) 5 2.93, MSe 5

who

1). At regions 4–5 no main effects or interac-209725, p , .09; F2(1,19) 5 3.02, MSe 5

FIG. 5. Residual reading times for S-biased verbs in Experiment 2 (The family said that the aging uncleclaimed 1 / the teenage girl 3 / purchased 4 / the 5 / new 6 / car 7 / herself 8 / at 9 / the 10 / car 11 / dealership 12 / denied

C

i

e

l

iad

6ve

wi

i

i t

;n

4igh-ch-ers

),

bi-

t

s 4ad

on

in

h be-wo re-nton,the

edend

m-ionr’s

al.ed

GRAMMATICAL SEAR

221071,p , .09; ambiguity 3 attachment site 3span, region 8,F1(1,54) 5 3.72, MSe 5 266697,p , .06; F2(1,19) 5 2.66, MSe 5 194738,p 5.10; ambiguity 3 attachment site 3 span, region9, F1(1,54) 5 1.96, MSe 5 119471,p 5 .16;F2(1,19) 5 3.46, MSe 5 218130,p , .07). Theseinteractions led us to examine the results serately for each reading span group.

At region 9 high-span readers showed signcant main effects of ambiguity (F1(1,28) 55.94, MSe 5 313812,p , .05; F2(1,19) 5 5.45,MSe 5 297727,p , .05) and attachment sit(F1(1,28) 5 4.41, MSe 5 232766,p , .05;F2(1,19) 5 5.68, MSe 5 309987,p , .05), andthe ambiguity 3 attachment site interaction wasignificant in the item analysis and marginasignificant in the participant analysis (F1(1,28)5 3.46, MSe 5 182733,p , .07; F2(1,19) 54.09, MSe 5 223257,p , .05). However, thetwo main effects are misleading, since pairwcomparisons reveal that there was a significcost of ambiguity in the high-attachment contions only (F1(1,28) 5 5.47, MSe 5 426260,p, .05; F2(1,19) 5 7.33, MSe 5 534523, p ,.01; ambig., 121 ms; unambig., 30 ms), andcorresponding effect in the low-attachment coditions (all Fs , 1; ambig., 28 ms; unambig., 1ms). This is the opposite of the pattern obserabove with strongly and weakly NP-biasverbs, and it indicates that the high-span readwere consistently adopting the low-attachmeanalysis for the ambiguous verb.

In contrast to the high-span readers, lospan readers showed clear evidence of dculty in both high and low ambiguous condtions. There was a main effect of ambigu(marginally significant at region 9,F1(1,26) 53.63, MSe 5 252532,p , .06; F2(1,19) 52.58, MSe 5 184748,p , .11; significant atthe p , .01 level when regions 8–9 were combined) and no main effect of attachment siteinteraction of ambiguity and attachment s(all Fs , 1). Pairwise comparisons showedsignificant cost of ambiguity in the high-atachment conditions (regions 8–9,F1(1,26) 54.17, MSe 5 395846,p , .05; F2(1,19) 54.05, MSe 5 372347, p , .05; ambig., 72 msunambig., 30 ms) and a marginally significa
cost of ambiguity in the low-attachment cond

pa-

fi-

sly

senti-

non-

eddersnt

-ffi-i-ty

-ortea-

t

tions (region 9,F1(1,26) 5 2.98, MSe 5187950,p , .09; F2(1,19) 5 3.25, MSe 5203705,p , .08; ambig., 70 ms; unambig.,ms). These results suggest that whereas hspan readers consistently adopt a low attament for the ambiguous verb, low-span readadopt a mix of high and low attachments.

Ambiguous region. At the ambiguous verb inregion 4 and at the following word (region 5there was a significant ambiguity3 attachmentsite interaction (region 4,F1(1,54)5 4.84, MSe5 357346,p , .05; F2(1,19) 5 4.81, MSe5351480,p , .05; region 5,F1(1,54) 5 20.29,MSe5 1129924,p , .0001;F2(1,19)5 22.28,MSe 5 1228903,p , .0001), and at region 5there were also significant main effects of amguity and attachment site (ambiguity,F1(1,54)5 11.45, MSe5 637838,p , .001;F2(1,19)512.63, MSe5 696474,p , .001; attachmensite,F1(1,54)5 6.16, MSe5 342958,p , .05;F2(1,19)5 6.68, MSe5 368616,p , .01). Pair-wise comparisons revealed that at both regionand 5 the high unambiguous condition was remore slowly than the high ambiguous conditi(region 4,F1(1,54)5 5.74, MSe5 418207,p ,.05; F2(1,19)5 4.65, MSe5 355061,p , .05;region 5,F1(1,54) 5 22.24, MSe5 1486158,p , .0001;F2(1,19)5 25.803, MSe5 1731943,p , .0001), with no corresponding slowdownthe low-attachment conditions (allFs, 1).

Discussion

We will first discuss those findings whicwere consistent across reading-span groupsfore considering differences between the tgroups. This experiment replicated the mainsult of Experiment 1, but it showed importavariation due to the verb class manipulatiwhich was not seen in Experiment 1 due to limited range of verb classes tested.

As in Experiment 1, this experiment showthat the ambiguous verb was consistently giva high-attachment analysis when it followestrongly NP-biased or weakly NP-biased ebedding verbs. This again supports the notthat existing commitments constrain the parsesearch for analyses of incoming materiFurthermore, the current experiment show
i-that the high attachments are not due to any spe-

A

rEe

s

re nuo

ohan

thm crbn

s

iderinude

th h

n

rendsswg-

d-x-im-anus

a-nhe towastly

e-hesi-t-

-of

has

athei-i-

atif- ef- 6o-ir

erhely

d-h-ont-

toker

324 SCHNEIDER

cial status of the main clause subject-vedependency, which was one concern aboutperiment 1. In Experiment 2 the additional levof embedding had the effect that neither potetial attachment site involved a main claudependency.

By extending the range of embedding veclasses used in this experiment we were ablsee clear effects of verb class on the parsesigned to the ambiguous verb. In conditiowith S-biased verbs, both reading-span groshowed a tendency to analyze the ambiguverb as an embedded verb. (Low-span readdid not choose this analysis exclusively—wdiscuss this further below.) This clear effect verb class is not surprising if we assume tNPs following S-biased verbs are initially anlyzed as embedded subjects (i.e., not “mimally attached”), as has been argued in a nuber of other studies (Garnsey et al., 199Juliano & Tanenhaus, 1994; Tabor, Juliano,Tanenhaus, 1997; Trueswell et al., 1993). If ambiguous NP is initially analyzed as an ebedded subject, then there is every reasonthe parser to analyze the ambiguous verb aslonging to the same clause, since high attament of the ambiguous verb would require analysis. This verb class effect is predicted RALR, but it is not predicted by an accouwhich attributes the results of Experiment 1 tobias to avoid long subjects or to avoid embeding. Such constraints would lead to an acrothe-board high-attachment bias.

An important concern about Experiment was the difference between the disambiguatused in high- and low-attachment control contions. Two steps were taken in this experimto address this concern: first, the two contconditions were more closely matched, by usa case-marked pronoun in both high and low ambiguous control conditions; second, an adtional opening clause provided an antecedfor the pronouns. These changes appear to hbeen effective, since there was no across-board difficulty associated with the pronounsExperiment 2. This is seen most clearly in tNP-only conditions, which provide a baselinmeasure of the difficulty of the high-attachme
conditions, independent of ambiguity. Howeve
ND PHILLIPS

bx-l

n-e

b toas-spsusersefat-i-m-7;&e-

forbe-h-

e-yt ad-s-

1oni-ntolgn-i-ntavee-

ineet

the pronoun conditions were often read moslowly than their ambiguous counterparts, athere was considerable variability in this acroverb classes. We discuss this variability beloin the context of differences between readinspan groups.

Turning now to differences between reaing-span groups, it was again true in this eperiment that high-span readers showed a spler pattern of results. Whereas high-spreaders consistently analyzed the ambiguoverb in a manner compatible with the probbilistic bias of the embedding verb, low-spareaders were more variable. In particular, in tS-bias conditions, low-span readers appearhave made a combination of both high and loattachments of the ambiguous verb, wherehigh-span readers appear to have consistenmade low attachments. This difference btween span groups is again compatible with tnotion that low-span readers may be less sentive to probabilistic verb biases (cf. Pearlmuter & MacDonald, 1995).

However, it is important to point out that lowspan readers did not show clear evidence variability with the other verb classes. Witstrongly and weakly NP-biased verbs there wdisruption following the disambiguation only inthe low-attachment conditions, suggesting ththe ambiguous verb was consistently given thigh-attachment analysis. We did not find evdence of disruption in high ambiguous condtions, in contrast to Experiment 1.

Turning to the ambiguous regions, starting the ambiguous NP in region 3, we again find dferences between reading-span groups in thefects of the verb class manipulation. Figureshows the slowdown at regions 3–5 in the prnoun (unambiguous) conditions, relative to theambiguous counterparts. The effect of vclass on the high-span readers is striking—t

unambiguous conditions are read more slowthe more they conflict with the verb bias, leaing to opposite trends for high- and low-attacment conditions. For the low-span readers,the other hand, the trends for high- and low-atachment conditions are much more similar one another, again suggesting a much wea
r,effect of verb bias.

GRAMMATICAL SEARCH AND REANALYSIS 325

FIG. 6. Slowdown in unambiguous (pronoun) conditions at regions 3–5, relative to their ambiguous

thnt

a

thtaena

Pre-rb;ed-

ow

onpan in

veoll

an

GENERAL DISCUSSION

Comparison of Experiments

In both experiments we found support for claim that existing parsing commitments costrain the search for analyses of subsequencoming material. This confirms the existencea Reanalysis As a Last Resortprinciple, or anactivation-based analog of this principle suchVosse and Kempen’s lateral inhibition mechnism (Vosse & Kempen, 1999).

Experiment 2 showed that the parsing of temporarily ambiguous structure was substially affected by the probabilistic bias of thembedding verb. However, this finding is etirely compatible with RALR, if we assume th

counterparts, Experiment 2.

S-bias 92*

e-

in-of

as-

en-

-t

ing of the ambiguous NP. If the ambiguous Nwas parsed as a direct object, then RALR pdicts a high attachment of the ambiguous veif the ambiguous NP was parsed as an embded clause subject, then RALR predicts a lattachment for the ambiguous verb.

The costs of ambiguity at the disambiguatipoint for each verb class and each reading-sgroup in the two experiments are summarizedTable 3.

Across the two experiments, there are fidifferent ambiguous verb-class conditions (twin Experiment 1; three in Experiment 2). For afive of these verb-class conditions, high-spreaders show disruption due to ambiguity in ei-ther the high-attachment or the low-attachment

e,

nt

the verb bias had immediate effects on the pars-conditions, but never in both. In each instanc

TABLE 3

Summary of Cost of Ambiguity (Ambiguous Minus Unambiguous Residual Reading Times) at Region 9 in Experiments 1 and 2

High-span readers Low-span readers

High-attachment Low-attachment High-attachment Low-attachme

Exp. 1 Exp. 2 Exp. 1 Exp. 2 Exp. 1 Exp. 2 Exp. 1 Exp. 2

NP-only 16 222Strong NP-bias 2 22 107* 65* 49* 213 168* 110*Weak NP-bias 0 244* 97* 53*r9–10 63* 217 102* 55*r8–9

13 42*r8–9 66**

A

ch

hw

e 2 ehmnc

oan

a

rd

rv

rfoaeRe

l

ext isly-erbw

ntlyse it isch-the ei-lveuldltsces

wo am-m-tic

on-ec-d, itxist-re-dplussertil

hee- ofing

onrter

Be-ite toressi-

at-le-w

326 SCHNEIDER

the verb bias straightforwardly predicts whiattachment site will cause difficulty. On tother hand, in three of the five verb-class contions, low-span readers show disruption dueambiguity in both high-attachment and low-at-tachment conditions. This difference betwethe span-groups can be explained under thesumption that low-span readers are less setive to probabilistic verb biases, although tdoes not explain why low-span readers shoa slowdown in the high ambiguous conditionsExperiment 1 but did not show a parallel effwith the NP-biased conditions in Experiment

Sturt et al. (2001) show results very similarthose shown here. In particular, their resultsNP-biased sentences show the same preferfor high attachments that we have shown. Talso show that increasing the length of the trix subject has no effect on high attachmecountering a possible objection that high attament might only hold within a certain proximity. Likewise, our studies provide strong suppfor the result that the easy reanalysis is not min NP-biased contexts. To that we add evideinvolving a wider selection of verbs, whicdemonstrate that when the parser’s initial comitments are reversed, its treatment of the biguous verb is reversed. This finding suppothe existence of an RALR constraint and pvides evidence that the results are not relateany bias toward making matrix attachments.

Relevance to Other Theoretical Assumptions

There is a tendency for advocates of separsing models to assume RALR, and for adcates of parallel parsing models to not assuRALR. It is true that it is more straightforwato implement RALR in a serial model (cSchneider, 1999), but we cannot claim that results provide evidence in favor of a serial proach to parsing, as we have seen that therparallel models which incorporate RAL(Vosse & Kempen, 1999) and serial modwhich do not (Fodor & Inoue, 1998).

However, our results are relevant to accouof why reanalysis is so easy in sentences (1), repeated below.

The woman knows the man wrote… (1)

ND PHILLIPS

hedi- to

en as-nsi-ised

inct.

tofornceeya-ts,h--rtdece

hm-m-rtso- to

ialo-med.urp- are

ls

ntsike

An interesting proposal made in the contof D(escription)-theory (Marcus et al., 1983)that the reanalysis in (1) is not really a reanasis at all, since the change at the underlined vin (1) only requires monotonic addition of nedominance relations. This approach elegapredicts that the change in (1) is easy, becaudoes not involve reanalysis. However, if thisthe correct analysis of (1), then the low attament in our test sentences, which requires same change, should not involve reanalysisther. But if the low attachment does not invoreanalysis, then it is unclear why readers shoconsistently avoid it. For this reason, our resucast doubt on the D-theory claim that sentenlike (1) do not involve reanalysis.

Alternative Accounts of the Results

Our account of the results consists of tcomponents. First, a parse is assigned to thebiguous NP, either as a direct object or an ebedded subject, according to the probabilisbias of the embedding verb or any other cstraints that may be relevant at this point. Sond, once the ambiguous verb is encountereis attached in a manner that preserves the eing structure. Under this interpretation, the sults follow from independently motivateproperties of how noun phrases are parsed,a general architectural property of the parthat it respects existing commitments unforced to do otherwise (i.e.,Reanalysis As aLast Resort).

Two alternative possibilities substitute tvery general RALR constraint with more spcific constraints that either restrict the lengthsubject NPs or restrict the depth of embeddthat the parser will allow.

A number of studies of language productihave noted that subject NPs tend to be shothan NPs occurring later in a clause (e.g.,hagel, 1932; Hawkins, 1994). Therefore,might be suggested that the parser is sensitivthis distributional generalization and therefoaims to complete subject NPs as soon as poble. In our experimental materials, the high tachment of the ambiguous verb allows comption of the subject NP, whereas the lo
attachment commits the parser to an even longer

C

hhidtthe wr n

tiemitRtreiclant idcts

r

le

e9

oyawi i

u

e

o

am-ded

fac-nlyit isr tohe

nt

ci- for

uc-r inhe

ofc-

likeen-r at-

forgnd-re-

uchthe

lesraleduc-ingbse-ngavelts isinne

in

ef-rmmal

GRAMMATICAL SEAR

subject NP. A constraint which restricts tlength of subjects can account fairly well for tresults of Experiment 1. However, the avoance of long subjects should be an across-board effect, which applies regardless of properties of the embedding verb. For this rson, the clear verb class differences whichobserved in Experiment 2 pose a problem foaccount based on subject length restrictioThe parser may be sensitive to subject lengthstrictions, but this cannot be the sole explanaof our results. In order to account for the effof the verb-class manipulation, we must assuthat revisions of existing structural commments carry some nontrivial cost. The RALconstraint is one instance of such a constrain

Similarly, the results of Experiment 1 aconsistent with an account based on constraon the depth of embedding that the parser tolerate, but such a constraint does not expthe greater variability observed when additioverb classes were introduced in ExperimenAgain, we do not wish to claim that the avoance of embedding (or other kinds of syntacomplexity) is irrelevant to parsing decisionwe simply claim that it cannot substitute foconstraint on the avoidance of reanalysis.

A third alternative account is more promisinIn a number of principle-based parsers, parsdecisions are driven by the need to compgrammatical dependencies involving themarelations, case relations, and agreement rtions (e.g., Abney, 1987, 1989; Crocker, 19Gibson, 1991; Pritchett, 1988, 1992). In theframeworks, our results could be attributed tneed to complete a subject-verb dependencan existing subject NP, for reasons of themrole assignment, case licensing, or both. If assume that the ambiguous NP is parsed manner consistent with the probabilistic biasthe embedding verb, then in conditions wNP-biased verbs, the only incomplete subjeverb dependency involves the main clause sject. This dependency can be completedparsing the ambiguous verb as a main claverb. In conditions with S-biased verbs, whthe ambiguous NP may have been parsed aembedded clause subject, there is an additi
embedded subject-verb dependency whi

ee-

he-ea-e

ans.re-onct

e-

.

ntsanin

al2.-ic;a

needs to be completed, and therefore the biguous verb will be parsed as an embedclause verb.

Notice also that the account based on satistion of grammatical dependencies not omatches our account in empirical coverage—also based on a theoretical claim very similathe RALR account. Whereas RALR forces tparser to maintain existing structural commit-ments, the dependency-completion accouforces the parser to honor existing structural pre-dictions. Under both accounts, prior parsing desions impose strong constraints on the searchanalyses for new material. With regard to strtures like (3), the two accounts are very similaspirit and make almost identical predictions. Ttwo accounts differ in regard to the parsingsimpler structures like (1). Under the RALR acount, lessons from the parsing of sentences(3) can be applied to the parsing of simple stences like (1), and they suggest that the parsetempts to find an alternative attachment sitethe incoming verb before ultimately identifyinthe parse involving reanalysis. Under the depeency-completion account, on the other hand,sults about the parsing of (3) do not tell us mabout how the parser goes about identifying

g.ingte

ticla-6;se a fortice

n aofthct-ub-byse

res annal

correct parse of sentences like (1).

CONCLUSION

Incremental parsing involves repeated cycof grammatical search leading to structucommitments. The two experiments reporthere investigated the extent to which the strtural commitments made at one cycle of parsguide the grammatical search process at suquent cycles. Most existing models of parsitake some stand on this issue, but there hbeen very few empirical tests of this. The resuof our experiments indicate that the parserstrongly guided by its existing commitments searching for analyses of new material. Ostraightforward way of implementing this is the form of the Reanalysis As a Last Resortcon-straint, but as we have emphasized, similar fects can be captured in other ways, in the foof a lateral inhibition mechanism, or in the forof a commitment to satisfy existing structur
chpredictions.

A

c

aet

2,

328 SCHNEIDER

The effects of verb-class and reading-spgroup observed in our experiments confirm pvious findings in the literature about the effeof lexical item-specific information in parsinand the variability of these effects across speers. However, these effects interact with omain findings in a very simple manner. They fect the decisions that are made at one cyclparsing, and thus they automatically affect commitments that the parser honors at sub

sonrn-

h

14 387 370

quent steps of parsing.

ACKNOWLEDGMENTS

We thank the audience at the 12th Annual CUNY Stence Processing Conference (March, 1999), Keith Rayand two anonymous reviewers for valuable commentsthis work. We thank Ted Eastwick, John Whalen, and JaLilley for help in running the experiments, and Susan Gasey for sharing the results of her sentence completion eximents with us. These studies were conducted when botthors were at the University of Delaware. The work wsupported in part by grants to the second author from
NSF (BCS-0196004) and the University of Delaware Rsearch Foundation.
2,

t

14 386 406 423 401

2,

APPENDIX A: RAW READING TIMES

TABLE 4

Raw, Untrimmed Reading Times (ms/word), Experimen

High attachment Low attachment

Ambig. Unambig. Ambig. Unambig.

1 422 422 428 4242 — — — 4283 448 434 444 4194 495 583 528 4395 420 491 431 4006 419 431 441 4037 432 471 444 4348 448 449 539 4479 463 435 567 421

10 389 382 458 38611 419 390 467 40312 443 425 470 42813 — — 447 416
14 409 423 434 413
ND PHILLIPS

anre-ts

gak-urf- of

hese-

en-ner, on

TABLE 5

Raw, Untrimmed Reading Times (ms/word), ExperimentNP-only Verbs

High attachment

Ambig. Unambig.

1 486 4603 470 4094 499 5095 404 4306 462 4117 451 4368 457 4979 421 424

10 361 37511 388 40312 393 42413 — —

per- au-as thee-

1

TABLE 6

Raw, Untrimmed Reading Times (ms/word), Experiment Strongly NP-Biased Verbs



1 465 468 457 4513 427 438 427 4724 515 544 473 4825 441 520 433 4656 450 428 459 4267 426 474 471 4378 470 497 511 4689 441 427 582 447

10 375 382 383 39511 382 394 414 39612 399 419 433 42113 — — 436 439

TABLE 7

Raw, Untrimmed Reading Times (ms/word), ExperimentWeakly NP-Biased Verbs



1 470 451 480 4603 434 437 421 4254 479 524 492 4965 425 425 445 403
6 447 458 449 428

C
GRAMMATICAL SEAR
TABLE 7—(Continued)



7 435 461 481 4608 480 468 562 4619 425 453 562 479

10 375 380 436 40111 365 391 413 39512 434 417 452 43913 — — 469 445

ass

a--

nte to

ck

of

14 395 401 421 423

TABLE 8

Raw, Untrimmed Reading Times (ms/word), ExperimentS-Biased Verbs



1 475 461 481 4823 437 470 459 4604 485 514 479 4875 399 554 437 4356 455 487 477 4567 460 476 484 4878 601 506 490 4869 558 458 490 454

10 400 399 390 38311 403 402 407 39812 400 423 407 42013 — — 452 455
c-
lf}ked

helf}f he

ewe

hop

zyutde

idtothe

elf}lly

zy

t

a

rih

io sc

es

APPENDIX B: EXPERIMENTALMATERIALS FOR EXPERIMENT 1

Each of the items in this list represents a full set of suli from Experiment 1. The complementizer that was usedas a disambiguator in the low unambiguous condition,a case-marked pronoun (him/her) was used as a disambiguator in the high unambiguous condition. The first flexive (himself/herself) listed in each example agrees wthe main clause subject and is the one used in the higtachment conditions; the second reflexive listed was uin the low-attachment conditions. The first completlisted was used in the high-attachment conditions; theond completion listed was used in the low-attachment ditions.

A full set of conditions from item number (1) can be sbelow, with regions that varied across conditions in italic

high ambiguous

The smart fellow who mentioned the senator’s wifegot a job for himselfwith the powerful lobbyist soon

14 387 384 393 412

thereafter.


high unambiguous

The smart fellow who mentioned her got a job forhimselfwith the powerful lobbyist soon thereafter.

low ambiguous

The smart fellow who mentioned the senator’s wifegot a job for herselfwith the powerful lobbyist gotsome hush money.

low unambiguous

The smart fellow who mentioned that the senator’swife got a job for herselfwith the powerful lobbyistgot some hush money.

The embedding verbs in the strongly NP-biased clwere discover, acknowledge, hear, appreciate, warn, andunderstand, and the weakly-biased verbs were know, doubt,mention, fear, and notice.

Weakly NP-Biased Conditions

1. The smart fellow who mentioned (that) {the sentor’s wife/her} got a job for {himself/herself} with the powerful lobbyist {soon thereafter/got some hush money}.

2. The quiet woman who mentioned (that) {the violeman/him} caused big trouble for {herself/himself} with thlocal police {by discussing the incident at all/was thoughtbe an informant}.

3. The busy woman who mentioned (that) {her sison/him} picked up the prescription {herself/himself} fromthe neighborhood pharmacy {after work/was accusedneglect}.

4. The nervous man who mentioned (that) {the acused woman/her} cursed very loudly at {himself/hersebefore the important hearing {about the robbery/was asabout the robbery}.

5. The attentive waitress who mentioned (that) {tsick man/him} opened the bathroom door {herself/himsein a big hurry {so he could get in/sent someone to see ineeded help}.

6. The experienced monk who doubts (that) {the nnun/her} wrote in the ledger {himself/herself} about thmembership increase {of the last five years/told the bisabout the potential problem}.

7. The angry woman who doubts (that) {the laman/him} called the state police {herself/himself} at abothree o’clock {to report her suspicions/thinks his wife mathe call}.

8. The pig-headed girl who doubts (that) {the stupboy/him} reported the transgressions {herself/himself} the principal {after school/thinks the class snitch was culprit}.

9. The distrustful woman who doubts (that) {thstrange man/him} locked the front door {herself/himseafter the big party {for the neighbors/asked who realocked the door}.

10. The famous knight who doubts (that) {the la

im-

nd-e-

th-at-

sednec-

on-

en.

2,

girl/her} fed the hungry horses {himself/herself} before the

A

o

oe t

icest

hy

louy

fes

she

nesh

vheit

s aig

tft

fim

gh

igmy

{tev

dats

usteed

syeo-

sedutved

dlf}of-

he}uld

inge- with

eer-r/

helf}n’s

thepee

user-as

hyr-an-

n-m-he

at)er-fter

330 SCHNEIDER

evening meal {of potatoes and roast beef/looked for fprints in the dirt}.

11. The handsome man who noticed (that) {the famactress/her} called the fashion magazine {himself/hersafter the gala fundraiser {for the charity/was surprisedwoman’s assistant didn’t call}.

12. The nice man who noticed (that) {the swoman/him} opened the bathroom door {himself/hersdespite the strong protestations {of the staff/ of the tried to reassure her}.

13. The observant woman who noticed (that) {the mtreated man/him} informed the personnel department {self/himself} about the frightening incident {a few daafter it happened/was pleased with the response}.

14. The young mother who noticed (that) {the disabman/him} warned the hospital staff {herself/himself} abthe broken wheelchair {as soon as she could/felt sorrhim}.

15. The middle-aged man who knows (that) {the mous woman/her} invited the political press {himself/hself} to the news conference {about the big scandal/wapleased by her action}.

16. The smart girl who knows (that) {the fooliboy/him} told the long story {herself/himself} during thstory hour {at the library/gained a new respect for him}.

17. The creative woman who knows (that) {the funman/him} wrote some comedy sketches {herself/himsabout the amusing escapades {she had seen/thinks he publish them}.

18. The bilingual man who knows (that) {the well-traeled woman/her} translated the travel books {himself/self} without any extra help {from a dictionary/was quimpressed with the result}.

19. The nosy guy who knows (that) {the depreswoman/her} told some hilarious jokes {himself/herself}the doctor’s office {in order to cheer her up/thinks she mbe faking her depression}.

20. The paranoid man who fears (that) {the decepwoman/her} locks the barroom doors {himself/herself} athe last call {every night/thinks he will be locked in}.

21. The grumpy woman who fears (that) {the scruman/him} called the building management {herself/hself} after the late-night disturbance {in the stairwell/triedappease him}.

22. The anxious man who fears (that) {the deranwoman/her} wrote the accusatory statement {himself/self} at the police station {to get her arrested/expects thelice to visit}.

23. The sleepless woman who fears (that) {the nwatchman/him} checks the window blinds {herself/hiself} during the late movie {every night/is planning to budog}.

24. The inexperienced nanny who fears (that) cranky handyman/him} fixed the squeaky door {hershimself} with some cooking oil {so the man wouldn’t ha
to come in the house/knows the door will need to be fixagain soon}.
ND PHILLIPS

ot-

uslf}he

klf}aff

is-er-s

edtfor

a-r-not

ylf}ould

-r-

e

edtht

iveer

fy-

to

eder-po-

ht-

a

helf/eed

Strongly NP-Biased Conditions

25. The quiet man who heard (that) {the terrifiewoman/her} called the county police {himself/herself} around four o’clock {in the morning/in the morning washocked at the lack of any response}.

26. The quiet girl who heard (that) {the mischievoboy/him} opened the front door {herself/himself} despimany strong warnings {that the boy shouldn’t be allowin/hoped he wouldn’t get in trouble}.

27. The intelligent man who heard (that) {the clumwoman/her} turned on the light {himself/herself} on thfront porch {so she wouldn’t fall/decided to install an autmatic light}.

28. The caring woman who heard (that) {the depresman/her} called the mental hospital {herself/himself} abothe serious side-effects {he was experiencing/ was relieat the news}.

29. The kind man who discovered (that) {the injurewoman/her} called the emergency room {himself/hersefrom a nearby payphone {after the late-night accident/fered to drive her to the hospital}.

30. The surprised woman who discovered (that) {tdrunk man/him} locked the front door {herself/himselfwith the spare key {to keep him away/was amazed he coeven walk}.

31. The hungry man who discovered (that) {the starvwoman/her} stole the big roast {himself/herself} from the dserted kitchen {of the restaurant/asked her to share somehim}.

32. The angry woman who discovered (that) {thbadly-beaten man/him} protested to the authorities {hself/himself} about the rampant violence {for over an houwas proud of his courage}.

33. The fickle actress who understood (that) {tstrange man/him} wrote the bizarre stories {herself/himsein just three days {despite being sick/wondered at the mabrilliance}.

34. The generous nobleman who understood (that) {persuasive woman/her} bought the powerful telesco{himself/herself} for the small college {in order to promotscience/promised to fund a new observatory}.

35. The pious man who understood (that) {the religiowoman/her} financed the church restoration {himself/hself} with the large inheritance {from a distant cousin/wimpressed with her devotion}.

36. The unhappy man who understood (that) {the swoman/her} reported the continuing problems {himself/heself} in an angry phone message {to the apartment mager/visited the woman to settled the dispute}.

37. The kind man who acknowledged (that) {the cosiderate woman/her} typed the generous contribution {hiself/herself} into the charity database {after receiving tcheck/showed appreciation for the help}.

38. The magnanimous prince who acknowledged (th{the quiet woman/she} accepted all the blame {himself/hself} for the many errors {of his subjects/pardoned her athe investigation}.

39. The grateful woman who acknowledged (that) {themaintenance man/him} answered all the questions {her-

C

ry

t

n

rt

hea

einth

leths

h

n

lto h

l

h-

als,n of

h

P-

-

GRAMMATICAL SEAR

self/himself} about the serious incident {at the refinethanked him profusely}.

40. The beloved duchess who acknowledged (that) {rich man/him} saved the well-known charity {herself/himself} despite the financial problems {following the scadal/praised him at the banquet}.

41. The big policeman who warned (that) {the angsaleswoman/her} ran down the street {himself/herself} afthe getaway car {following the robbery/hoped the othcops wouldn’t shoot her}.

42. The neighbor woman who warned (that) {the macboy/him} shot the rabid dog {herself/himself} with a largshotgun {before going to work/thought she should call his pents}.

43. The concerned priest who warned (that) {the mdling woman/her} opened all the mail {himself/herself} the church office {so she wouldn’t read the letters/told worried parishioner to contact him by phone}.

44. The young widow who warned (that) {the littboy/him} opened the locked cabinet {herself/himself} wia small screwdriver {while he was outside/told the babyter to watch him closely}.

45. The health-conscious man who appreciates (t{the beautiful woman/her} takes good care of {himself/heself} throughout the entire year {by eating regularly awalking everywhere/wants to ask her on a date}.

46. The ambitious boy who appreciates (that) {hsmart sister/her} does lots of homework {himself/herseafter swim team practice {because he wants to get ingood college/understands that she doesn’t have time tohim}.

47. The good-natured landlady who appreciates (th{the nice man/him} cleans the big house {herself/himsebefore the day’s end {so the house will be clean for himthinking about lowering his rent}.

48. The thoughtful woman who appreciates (that) {tyoung man/him} pays all the utilities {herself/himself} be
fore the due date {to keep the man happy/wants to keep
gn l adrarrioesly

weearese

thatr’slf}sh

usyhey

tten-ed

the bathroom door {herself/himself} in a big hurry {so he

man as a roommate}.

APPENDIX C: EXPERIMENTAL MATERIALS FOR EXPERIMENT 2

The stimuli in Experiment 2 followed the same patternthe stimuli in Experiment 1 except for the followinchanges: Four verb classes were used (NP-only, StroNP-biased, Weakly NP-biased, S-biased); both high andunambiguous conditions used a case-marked pronoundisambiguator; all stimulus sentences were embedwithin a clause whose subject provided an antecedent fodisambiguating pronoun. The antecedent for the disbiguating pronoun in the unambiguous conditions cosponded to the ambiguous NP in the ambiguous condit(e.g.,the senator’s wifein example (1) below). Many of thstimuli with strongly or weakly NP-biased verbs were baon stimuli used in Experiment 1. In items with NP-on
verbs there were no low conditions, since these would be

/

he--

yerer

o

r-

d-

e

it-

at)r-d

isf} aelp

at)f}/is

e

the

as

glyows aed

them-e-ns

ed

grammatical. Items 1–40 were used in one block of triitems 41–80 in a second block. The order of presentatiothe two blocks was counterbalanced across subjects.

A full set of stimuli from item (1) is shown below, witregions that varied across conditions in italics.

high ambiguous

The congressional stafferswere not surprised that thesmart fellow who mentioned the senator’s wifeleakedthe important news himself to the powerful lobbyistsoon thereafter.

high unambiguous

The senator’s wifewas not surprised that the smartfellow who mentioned her leaked the important newshimselfto the powerful lobbyist soon thereafter.

low ambiguous

The congressional stafferswere not surprised that thesmart fellow who mentioned the senator’s wifeleakedthe important news herself to the powerful lobbyistgot some hush money.

low unambiguous

The senator’s wifewas not surprised that the smartfellow who mentioned sheleaked the important newsherselfto the powerful lobbyist got some hush money.

The verb classes contained the following: 6 strongly Nbiased verbs,hear, acknowledge, appreciate, discover,warn, understand; 4 weakly NP-biased verbs,know, men-tion, doubt, notice; 3 S-complement biased verbs,claim, be-lieve, suspect, and 19 verbs which only allow NP-complements, abuse, admire, adore, annoy, defy, despise,disappoint, feed, fire, harass, humiliate, love, overcharge,pity, prosecute, rescue, supervise, support, treat. Somespeakers judge that despise, adore, and loveallow sententialcomplements under certain circumstances. However,have found no informants that allow these verbs to appwith that-less sentential complements, and therefore thverbs are unambiguous in our experimental materials.

Weakly NP-Biased Conditions (Block A)

1. The congressional staffers were not surprised the smart fellow who mentioned {the senatowife/her/she} leaked the important news {himself/herseto the powerful lobbyist {soon thereafter/got some humoney}.

2. The neighborhood gossips reported that the bwoman who mentioned {the sick boy/him/he} picked up tprescription {herself/himself} from the local pharmac{after work/was accused of neglect}.

3. The restaurant managers were glad that the ative waitress who mentioned {the sick man/him/he} open

un-could get in/sent someone to see if he needed help}.

A

ap

te

e

oehe

ac

r

ael

n

nne.

h

o

dau

ati

w

m

itythe

efulhe}eri-

anthely

hoaildby

edthee-

tivethegting

girlat-ngout

manage--

be-intse she

anntson

be-rti-e

}.tedhe-

it}.ndtheoink

332 SCHNEIDER

4. The family friends stated that the angry womwho doubts {the apathetic man/him/he} called the state lice {herself/himself} at about three o’clock {to report hesuspicions/thinks the man’s son made the call}.

5. The curious children observed that the anxiowoman who doubted {the strange man/him/he} locked front door {herself/himself} after the big party {for thneighbors/asked who really locked the door}.

6. The editors proved that the handsome man whoticed {the famous actress/her/she} called the fashion mazine {himself/herself} after the gala fundraiser {for thcharity/tried to interrupt the call}.

7. The witnesses discovered that the conscientiwoman who noticed {the mistreated man/him/he} informthe personnel department {herself/himself} about the frigening incident {a few days after it happened/had fired seral employees}.

8. The rowdy boys remarked that the thoughtful mwho noticed {the young girl/her/she} closed the firepladoors {himself/herself} before the birthday party {at thmansion/congratulated her on her forethought}.

9. The proud parents boasted that the smart girl wknows {the childish boy/him/he} told the long story {heself/himself} during the story hour {at the library/gainednew respect for him}.

10. The book publisher heard that the bilingual mwho knows {the well-traveled woman/her/she} translatthe travel books {himself/herself} without any extra he{from a dictionary/was quite impressed with the result}.

Strongly NP-Biased Conditions (Block A)

11. The neighbors were surprised that the quiet mwho heard {the terrified woman/her/she} called the coupolice {himself/herself} at around four o’clock {in themorning/in the morning didn’t say anything to anybody}.

12. The family members believe that the intelligeman who heard {the handicapped woman/her/she} turon the light {himself/herself} on the front porch {so thwoman wouldn’t fall/decided to install an automatic light}

13. The hospital staff was surprised that the kind mwho discovered {the injured woman/her/she} called temergency room {himself/herself} from a nearby paypho{after the late-night accident/offered to drive her to the hpital}.

14. The vagrants stated that the hungry man who covered {the starving woman/her/she} stole the big ro{himself/herself} from the deserted kitchen {of the restarant/asked her to share some with him}.

15. The movie studio was surprised that the fickle tress who understood {the strange man/him/he} wrote bizarre stories {herself/himself} in just three days {despbeing sick/wanted to put on the stories}.

16. The church leaders thought that the pious man understood {the religious woman/her/she} financed the churestoration {himself/herself} with the large inheritance {froa distant cousin/was impressed with her devotion}.

17. The institute employees believe that the kind m
who acknowledged {the considerate woman/her/she} typ
ND PHILLIPS

no-r

ushe

no-ga-

usdt-v-

ne

e

ho-a

nd

p

anty

ted

ane

nes-

is-st-

c-hete

horch

an

the generous contribution {himself/herself} into the chardatabase {after receiving the check/mentioned her to board members}.

18. The corporate office was pleased that the gratwoman who acknowledged {the maintenance man/him/answered all the questions {herself/himself} about the sous incident {at the refinery/thanked him profusely}.

19. The store managers thought that the big policemwho warned {the angry saleswoman/her/she} ran down street {himself/herself} after the getaway car {immediateafter the robbery/told the other cops not to shoot}.

20. The nuns thought that the concerned priest wwarned {the meddling woman/her/she} opened all the m{himself/herself} in the church office {so she wouldn’t reathe letters/told the worried parishioner to contact him phone}.

S-Biased Conditions (Block A)

21. The truancy office knows that the concernwoman who suspected {the short boy/him/he} called youth center {herself/himself} after the recent fight {to rport what she knew/intends to talk to the boy’s parents}.

22. The executive committee thinks that the percepwoman who suspects {the friendly man/him/he} started malicious rumor {herself/himself} during the civic meetin{because she wanted to keep the man from getelected/didn’t tell anyone of her suspicions}.

23. Some classmates reported that the scheming who suspects {the grumpy boy/him/he} ate the incrimining evidence {herself/himself} before the school meeti{so the boy couldn’t get her in trouble/told the teacher abher suspicions}.

24. The loan company learned that the depressed who suspects {the devious woman/her/she} sold the engment ring {himself/herself} at the pawn shop {for five-hundred dollars/went to the pawn shop to look for the ring}.

25. The coaches are amused that the young boy wholieves {the neighbor girl/her/she} buys some peppermgum {himself/herself} before every baseball game {becauthe gum is supposed to bring the team good luck/hopeswill give him some gum}.

26. The dealership knows that the dark-haired womwho believes {the knowledgeable salesman/him/he} wathe blue car {herself/himself} despite the peeling paint {the car/the car is a good deal}.

27. The police stated that the charming woman who lieves {the dishonest man/him/he} bought the valuable afacts {herself/himself} from the disreputable store {at thman’s insistence/thinks the man deserves to be cheated

28. The agency staff was surprised that the exciwoman who claimed {the small boy/him/he} opened tfront door {herself/himself} without any help {despite holding the boy with both hands/made such a big deal about

29. The sales clerks couldn’t believe that the kiwoman who claimed {the errant boy/him/he} purchased expensive clothing {herself/himself} with a credit card {sthat the store would not prosecute him for theft/doesn’t th

edthe boy’s father will believe her}.

C

p

r

l

o

b

i

r

ar

a

oo

bg

or

p

no

ousyta-

anomof

edr-

she

dle-tedce

andydes

hoouso

irlnt

to

anntalhe

sedhem

an toce

er-her/hend

anon-s-ettle

artc-

er-.ved

GRAMMATICAL SEAR

30. The dispatch office stated that the dismayed mowho claimed {the runaway boy/him/he} called the state lice {herself/himself} from a pay phone {because the phoat home was broken/thinks the police should have gonthe pay phone to pick the boy up}.

NP-Only Conditions (Block A)

31. The homeless people think that the brilliant womwho supported {the impoverished man/him} called the lacharity herself after the fund drive to see if he had appfor aid.

32. The office supervisors know that the nice man wfired {the incompetent woman/her} wrote the critical evaation himself during the long flight back to the home offic

33. The smuggling ring fears that the stern woman wprosecuted {the violent man/him} called the immigratioffice herself after the deportation hearing to make surehe couldn’t get back in.

34. The alert siblings remarked that the vengeful who defied {the young woman/her} opened the large paage himself on the kitchen table.

35. The concerned relatives were glad that the valwoman who rescued {the injured man/him} battled the cfire herself for over fifteen minutes until the firefightearrived.

36. The office staff said that the compassionate mwho treated {the depressed woman/her} filed the insuraforms himself with the insurance company the woman h

37. The many bystanders thought that the concewoman who adores {the teenage boy/him} called the ressquad herself on a cell phone in her purse.

38. The delivery staff said that the well-known mwho annoyed {the quick-tempered woman/her} sent thecusatory letter himself with a polite note taped to the top

39. The police officers claimed that the cruel man wabused {the penniless woman/her} locked the thick dhimself with the monstrous key so that she couldn’t get

40. The children’s teachers thought that the sweetwho pitied {the poor boy/him} bought the new jacket hersas a birthday gift for the boy.

Weakly NP-Biased Conditions (Block B)

41. The private detectives think that the quiet womwho mentioned {the violent man/him/he} caused big troufor {herself/himself} with the local police {by discussinthe incident at all/is a big gossip}.

42. The court officials were amazed that the nervman who mentioned {the accused woman/her/she} cuvery loudly at {himself/herself} before the important heaing {about the robbery/was asked about the robbery}.

43. The concerned parishioners heard that the exenced monk who doubts {the new nun/her/she} wrote inledger {himself/herself} about the attendance increase {oflast five years/told the bishop about the potential problem}

44. The nosy classmates assumed that the pig-hegirl who doubts {the stupid boy/him/he} reported the tragressions {herself/himself} to the principal {after scho
thinks the class snitch was the culprit}.

thero-nee to

angelied

hou-e.honthat

oyck-

antars

annced.

nedcue

nac-.hoorut.girlelf

anle

ussedr-

eri-thethe.adeds-l/

45. The talkative servants suspected that the famknight who doubts {the lazy girl/him/he} fed the hungrhorses {himself/herself} before the evening meal {of potoes and roast beef/looked for footprints in the dirt}.

46. The anxious family appreciated that the nice mwho noticed {the sick woman/her/she} opened the bathrodoor {himself/herself} despite the strong protestations {the staff/of the staff tried to reassure her}.

47. The visitors said that the young mother who notic{the disabled man/him/he} warned the hospital staff {heself/himself} about the broken wheelchair {as soon as could/put it in the corner}.

48. The persistent reporters remarked that the midaged man who knows {the famous woman/her/she} invithe political press {himself/herself} to the news conferen{about the big scandal/was not pleased by her action}.

49. The talent agency thinks that the creative womwho knows {the funny man/him/he} wrote some comesketches {herself/himself} about the amusing escapa{she had seen/wants to publish them}.

50. The office staff learned that the nosy guy wknows {the depressed woman/her/she} told some hilarijokes {himself/herself} at the doctor’s office {in order tcheer her up/thinks she might be hiding her depression}.

Strongly NP-Biased Conditions (Block B)

51. The anxious friends were glad that the quiet gwho heard {the mischievous boy/him/he} opened the frodoor {herself/himself} despite many strong warnings {leave the door locked/doesn’t plan to tell anyone}.

52. The co-workers reported that the caring womwho heard {the depressed man/him/he} called the mehospital {herself/himself} about the serious side-effects {was experiencing/was relieved to hear the news}.

53. The freezing friends were upset that the surpriwoman who discovered {the drunk man/him/he} locked tfront door {herself/himself} with the spare key {to keep hiaway/took the key from him}.

54. The concerned relatives heard that the angry womwho discovered {the badly-beaten man/him/he} protestedthe authorities {herself/himself} about the rampant violen{for over an hour/was proud of the man’s courage}.

55. The faculty members were grateful that the genous benefactor who understood {the persuasive woman/she} bought the powerful telescope {himself/herself} for tsmall college {in order to promote science/promised to fua new observatory}.

56. The irritated neighbors noted that the unhappy mwho understood {the shy woman/her/she} reported the ctinuing problems {himself/herself} in an angry phone mesage {to the apartment manager/visited the woman to sthe dispute}.

57. The anxious family was delighted that the smprince who acknowledged {the quiet woman/her/she} acepted all the blame {himself/herself} for the judgement rors {of his subjects/pardoned her after the investigation}

58. The board members were gratified that the belo
duchess who acknowledged {the rich man/him/he} saved the

A

m

o

be}

aht

wzro

eos

itthg

eher

m

u

e

nteo

ecew

}

anm-

nteing. fa-n-

hoor to

the}m-

anor-uld

yewuch

anne

ibleheon

edt-

s-

-

334 SCHNEIDER

popular charity {herself/himself} despite the financial prolems {following the scandal/praised him at the banquet}.

59. The teenage friends feared that the neighbor wowho warned {the aggressive boy/him/he} shot the wild d{herself/himself} with the powerful rifle {before going twork/would call the boy’s parents}.

60. The visiting children thought that the young widowho warned {the little boy/him/he} opened the locked canet {herself/himself} with a small screwdriver {while thboy was outside/told the babysitter to watch him closely

S-Biased Conditions (Block B)

61. The newspaper said that the scared man whopects {the irate woman/she/her} notified the police depment {himself/herself} after the recent burglary {at tneighborhood store where the woman works/hopes thawoman will be arrested}.

62. The supervisors realized that the astute man suspected {the brilliant woman/she/her} made the amadiscovery {himself/herself} in the government lab {befothe woman announced her results/thinks that she shhave taken the credit}.

63. The police think that the ailing woman who supects {the ill-tempered boy/him/he} took the cash box {hself/himself} from the unlocked cabinet {so that the bcouldn’t steal it/told the department manager what thought had happened}.

64. The foundation employees heard that the destman who believes {the kind woman/her/she} visited soup kitchen {himself/herself} after the severe storm {to a good meal/asked about her health}.

65. The court jesters joked that the beautiful princwho believes {the dim-witted knight/him/he} contacts twar committee {herself/himself} before every committmeeting {to try to get the knight promoted/should womore about someone else}.

66. The neighbors mentioned that the desperate who believes {the cheerful woman/her/she} bought the ncar {himself/herself} at the car dealership {that {thwoman/she} recommended/thinks she made a foolish dsion}.

67. The office staff wasn’t surprised that the insecman who believes {the reassuring woman/her/she} opethe mysterious package {himself/herself} before the loflight {because she told him to look inside/called to seshe liked the gift}.

68. The security company thought that the concerfather who claimed {the screaming girl/her/she} broke small window {himself/herself} with a small rock {so hcould open the door to rescue {his daughter/her}/said sstay out of trouble}.

69. The family said that the aging uncle who claim{the teenage girl/her/she} purchased the new {himself/herself} at the car dealership {so {the girl/shcould have his old car/denied that he had anything to do the purchase}.

70. The security officers stated that the scheming m
who claimed {the little girl/her/she} removed the name ta
ND PHILLIPS

b-

anog

wi-

.

sus-rt-e the

hoingeuld

s-r-yhe

uteeet

sseery

anew

{himself/herself} with a quick jerk {so no one would knowthat {the girl/she} was not his/tried to take {the girl/herfrom the hospital}.

NP-Only Conditions (Block B)

71. The office workers remarked that the arrogant mwho humiliated {the sad woman/her} paid the large fee hiself with a personal check to show how rich he was.

72. The family members heard that the intelligewoman who loves {the good-looking man/him} called thfancy restaurant herself for dinner reservations that even

73. The school classmates guessed that the smilingther who loves {the precocious girl/her} bought the expesive doll himself as a birthday present for the girl.

74. The hotel staff noticed that the thoughtful man wadmires {the ambitious woman/her} opened the front dohimself at the sea-side resort just in time for the womanenter.

75. The sympathetic lawyers were disappointed that devious woman who deceived {the stupid man/himclaimed the large reward herself from the insurance copany after the trial.

76. The staff supervisors remarked that the cruel mwho despises {the stupid woman/her} presented the imptant report himself at the afternoon meeting so he coclaim credit for it.

77. The family members noticed that the grumpwoman who harassed {the lazy man/him} stained the ncouch herself with a greasy hamburger just after the cowas delivered.

78. The nursing staff said that the conscientious mwho fed {the sick woman/her} answered the ringing phohimself in the hospital room since she was so sick.

79. The company president was glad that the senswoman who supervises {the industrious man/him} wrote tglowing recommendation herself for the prestigious positithat the man had applied for.

80. The accounting office noticed that the underhand

eeci-

renedng if

edhe

to

dar}ith

an

man who overcharged {the likable woman/her} wrote the criical invoice himself on the company letterhead late at night.

REFERENCES

Abney, S. P. (1987). Licensing and parsing. In Proceedingsof NELS 17. Amherst, MA: GLSA.

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

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

Behagel, O. (1932). Deutsche Syntax: Eine GeschichtlicheDarstellung, vol. IV, Wortstellung, Periodenbau.Hei-delberg: Carl Winters.

Crocker, M. (1996). Computational psycholinguistics. Dor-drecht: Kluwer.

Ferreira, F., & Clifton, C. (1986). The independence of syntactic processing. Journal of Memory and Language,

g 25,348–368.

C

n

-

n

n

s

.ee

M

n

n

.

n

sr

n

a

9y

u

al

-tic

-es

ltic

f

the

-

-

b-

rid

fore,

ing.

-e-

le-

.m-

g ofin

ral

GRAMMATICAL SEAR

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

Fodor, J., & Ferreira, F. (Eds.) (1998). Reanalysis in sentence processing. Dordrecht: Kluwer.

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

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

Frazier, L. (1987). Sentence processing: A tutorial reviewM. Coltheart (Ed.),Attention and Performance XI(pp.559–587). Hillsdale, NJ: Erlbaum.

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

Frazier, L., & Clifton, C., Jr. (1998). Sentence reanalyand visibility. In Fodor & Ferreira (Eds.),Reanalysis insentence processing(pp.143–176). Dordrecht: Kluwer

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

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

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

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

Gibson, E. A. F., Pearlmutter, N., Canseco-González, EHickok, G. (1996). Recency preference in the humsentence processing mechanism. Cognition,59,23–59.

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

Hawkins, J. (1994).A performance theory of order and costituency. Cambridge, UK: Cambridge Univ. Press.

Juliano, C., & Tanenhaus, M. K. (1994). A constraint-balexicalist account of the subject/object attachment perence. Journal of Psycholinguistic Research, 23,459–471.

Jurafsky, D. (1996). A probabilistic model of lexical asyntactic access and disambiguation. Cognitive Sci-ence, 20,137–194.

Just, M., Carpenter, P., & Woolley, J. (1982). Paradigms processes in reading comprehension. Journal of Exper-imental Psychology: General, 111,228–238.

Kennedy, A., Murray, W. S., Jennings, F., & Reid, C. (198Parsing complements: Comments on the generalitthe principle of minimal attachment. Language andCognitive Processes, 4, 21–49.

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

a-e-

d

ce

tic,

. InI

is,

r-ntses.

.il-us

s-,

f

, &an

-

edef-

d

nd

). of

re

Marcus, M., Hindle, D., & Fleck, M. (1983). D-theory:Talking about talking about trees. Proceedings of the21st Meeting of the Association for ComputationLinguistics.

MacDonald, M., Just, M. A., & Carpenter, P. (1992). Working memory constraints on the processing of syntacambiguity. Cognitive Psychology, 24,56–98.

Mitchell, D. C., & Holmes, V. M. (1985). The role of specific information about the verb in parsing sentencwith local ambiguity. Journal of Memory and Lan-guage, 24,542–559.

Pearlmutter, N., & MacDonald, M. C. (1995). Individuadifferences and probabilistic constraints in syntacambiguity resolution. Journal of Memory and Lan-guage, 34,521–542.

Phillips, C., & Gibson, E. A. F. (1997). On the strength othe local attachment preference. Journal of Psycholin-guistic Research, 26,323–346.

Pritchett, B. L. (1988). Garden path phenomena and grammatical basis of language processing. Language,64,539–576.

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

Rayner, K., & Frazier, L. (1987). Parsing temporarily ambiguous complements. Quarterly Journal of Experi-mental Psychology, 39A, 657–673.

Schneider, D. (1999). Parsing and incrementality. Unpulished Ph.D. dissertation, University of Delaware.

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

Stevenson, S. (1994b). A competitive attachment model resolving syntactic ambiguities in natural languagparsing. Ph.D. dissertation, University of MarylandCollege Park, MD.

Stevenson, S. (1998). Parsing as incremental restructurIn J. Fodor & F. Ferreira (Eds.),Reanalysis in sentenceprocessing(pp.327–363). Dordrecht: Kluwer.

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

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

Sturt, P., Pickering, M., Scheepers, C., & Crocker, M(2001). The preservation of structure in language coprehension: Is reanalysis the last resort? Journal ofMemory and Language.

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

Trueswell, J., & Tanenhaus, M. K. (1991). Tense, tempocontext and syntactic ambiguity resolution. Language
and Cognitive Processes, 6, 303–338.

336 SCHNEIDER AND PHILLIPS

Wanner, E. (1980). The ATN and the sausage machine:Which one is baloney? Cognition, 8, 209–225.

cet.

Trueswell, J., Tanenhaus, M. K., & Kello, C. (1993). Verb-specific constraints in sentence processing: Separating
lexical effects from garden paths. Journal of Experi-mental Psychology: Learning, Memory and Cognitio,19,528–553.
Vosse, T., & Kempen, G. (2000). Syntactic structure assbly in human parsing: A computational model basedcompetitive inhibition and a lexicalist grammar. Cogni-tion, 75,105–143.

Weinberg, A. (1993). Parameters in the theory of sentenprocessing: Minimal commitment theory goes easJournal of Psycholinguistic Research, 22,339–364.

(Received May 22, 2000)(Revision received November 7, 2000)Published Online June 21, 2001

n

em- on

grammatical search and reanalysis

Documents