When Anaphor Resolution Fails

24
When Anaphor Resolution Fails William H. Levine, Alexandria E. Guzma ´n, and Celia M. Klin State University of New York at Binghamton Implicit in the study of anaphors has been the assumption that anaphoric inferences must be resolved for a text to be coherent. In six experiments we examined this assumption. Using recognition and reading time measures, we found that antecedents (e.g., “tart”) were not accessed when an anaphor (e.g., “dessert”) was read if the antecedent was well backgrounded and if a highly salient same-category distractor (e.g., “cake”) intervened between the anaphor and its antecedent. We conclude that when an antecedent is difficult to retrieve, and when the failure to connect an anaphor to its antecedent does not create a coherence break, readers may simply read on rather than devoting additional time and attention to identifying the antecedent. © 2000 Academic Press Key Words: inference; anaphor; reading; discourse. Reading comprehension involves building a coherent mental representation of the informa- tion described by a text. A central aspect of this process involves making connections between currently read text and information from earlier in the passage. These connections are often sig- naled by anaphors, which are words or phrases that refer to some earlier-mentioned concept. The importance of anaphors in the comprehen- sion of text is reflected by the extensive litera- ture examining the process by which a reader makes a connection between an anaphor and its referent, or antecedent (e.g., Corbett & Chang, 1983; Dell, McKoon, & Ratcliff, 1983; Ehrlich & Rayner, 1983; Gernsbacher, 1989; Greene, McKoon, & Ratcliff, 1992; O’Brien, Duffy, & Myers, 1986; O’Brien, Raney, Albrecht, & Rayner, 1997; Sanford & Garrod, 1989). The process of connecting an anaphor to its referent usually feels effortless to the reader. Despite this perception, finding the antecedent for an anaphor is an inference. That is, it is up to the reader to determine that the noun phrase “the burglar,” for example, refers to the same person who was introduced earlier in a story as “the criminal” (McKoon & Ratcliff, 1980). Just as with other types of inferences, anaphoric inferences require readers to incorporate infor- mation into their text representation that is not explicit in the text. However, despite the simi- larities, the study of anaphors has tended to differ in important ways from the study of other classes of inferences. Most notably, research examining causal, forward, and instrumental in- ferences, for example, has focused on determin- ing if these inferences are drawn during reading, or more specifically, if they are drawn under an assortment of text conditions (e.g., Cotter, 1984; Klin, Guzma ´n, & Levine, 1999; Klin, Murray, Levine, & Guzma ´n, 1999; Klin & My- ers, 1993; Klin, 1995). In contrast, in much of the research examining anaphoric inferences, the starting assumption has been that they are drawn and drawn reliably. Thus, rather than asking if anaphoric inferences are drawn, the central focus has been on examining the pro- cesses involved in drawing these inferences. And rather than examining the influence of var- ious text conditions on the probability that ana- phoric inferences are drawn, research has fo- Portions of these data were reported at the Ninth Annual Meeting of the Society for Text and Discourse, Vancouver, Canada, 1999. The full set of experimental materials can be found at http://bingweb.binghamton.edu/;cklin/anaphor_ materials.html. We thank Rachel Altschuler, Sarah Burt, Stacey Greenblatt, Dawn Melzer, Oscar Molina, and Emily Scauri for their assistance with data collection and Kristin Weingartner for assistance with the article. We are also grateful to Jerome L. Myers for helpful comments on this research. Address correspondence and reprint requests to Celia M. Klin, Department of Psychology, State University of New York at Binghamton, Binghamton, NY 13902-6000. Fax: (607) 777-4890. E-mail: [email protected]. 594 0749-596X/00 $35.00 Copyright © 2000 by Academic Press All rights of reproduction in any form reserved. Journal of Memory and Language 43, 594 – 617 (2000) doi:10.1006/jmla.2000.2719, available online at http://www.idealibrary.com on

Transcript of When Anaphor Resolution Fails

Journal of Memory and Language43, 594–617 (2000)doi:10.1006/jmla.2000.2719, available online at http://www.idealibrary.com on

When Anaphor Resolution Fails

William H. Levine, Alexandria E. Guzma´n, and Celia M. Klin

State University of New York at Binghamton

Implicit in the study of anaphors has been the assumption that anaphoric inferences must beresolved for a text to be coherent. In six experiments we examined this assumption. Using recognitionand reading time measures, we found that antecedents (e.g., “tart”) were not accessed when ananaphor (e.g., “dessert”) was read if the antecedent was well backgrounded and if a highly salientsame-category distractor (e.g., “cake”) intervened between the anaphor and its antecedent. Weconclude that when an antecedent is difficult to retrieve, and when the failure to connect an anaphorto its antecedent does not create a coherence break, readers may simply read on rather than devotingadditional time and attention to identifying the antecedent.© 2000 Academic Press

Key Words:inference; anaphor; reading; discourse.

g amathiseelier

sse

epenra-dedng

hne

&

it

der.entup

raseme

astoricfor-noti-to

erarchl in-min-ing,

o na ter,1 ,Me oft ces,t ared ana thec pro-c ces.A ar-i na-p fo-

nuavern b

urtmilyistinalsthis

lia Mewax:

Reading comprehension involves buildincoherent mental representation of the infortion described by a text. A central aspect ofprocess involves making connections betwcurrently read text and information from earin the passage. These connections are oftennaled by anaphors, which are words or phrathat refer to some earlier-mentioned concThe importance of anaphors in the comprehsion of text is reflected by the extensive liteture examining the process by which a reamakes a connection between an anaphor anreferent, or antecedent (e.g., Corbett & Cha1983; Dell, McKoon, & Ratcliff, 1983; Ehrlic& Rayner, 1983; Gernsbacher, 1989; GreeMcKoon, & Ratcliff, 1992; O’Brien, Duffy, &Myers, 1986; O’Brien, Raney, Albrecht,Rayner, 1997; Sanford & Garrod, 1989).

The process of connecting an anaphor to

Portions of these data were reported at the Ninth AnMeeting of the Society for Text and Discourse, VancouCanada, 1999. The full set of experimental materials cafound at http://bingweb.binghamton.edu/;cklin/anaphor_materials.html. We thank Rachel Altschuler, Sarah BStacey Greenblatt, Dawn Melzer, Oscar Molina, and EScauri for their assistance with data collection and KrWeingartner for assistance with the article. We aregrateful to Jerome L. Myers for helpful comments onresearch.

Address correspondence and reprint requests to CeKlin, Department of Psychology, State University of NYork at Binghamton, Binghamton, NY 13902-6000. F(607) 777-4890. E-mail: [email protected].

5940749-596X/00 $35.00Copyright © 2000 by Academic PressAll rights of reproduction in any form reserved.

-

n

ig-s

t.-

rits,

,

s

referent usually feels effortless to the reaDespite this perception, finding the antecedfor an anaphor is an inference. That is, it isto the reader to determine that the noun ph“the burglar,” for example, refers to the saperson who was introduced earlier in a story“the criminal” (McKoon & Ratcliff, 1980). Jusas with other types of inferences, anaphinferences require readers to incorporate inmation into their text representation that isexplicit in the text. However, despite the simlarities, the study of anaphors has tendeddiffer in important ways from the study of othclasses of inferences. Most notably, reseexamining causal, forward, and instrumentaferences, for example, has focused on detering if these inferences are drawn during read

r more specifically,if they are drawn under assortment of text conditions (e.g., Cot984; Klin, Guzma´n, & Levine, 1999; Klinurray, Levine, & Guzma´n, 1999; Klin & My-rs, 1993; Klin, 1995). In contrast, in much

he research examining anaphoric inferenhe starting assumption has been that theyrawn and drawn reliably. Thus, rather thsking if anaphoric inferences are drawn,entral focus has been on examining theesses involved in drawing these inferennd rather than examining the influence of v

ous text conditions on the probability that ahoric inferences are drawn, research has

l,e

,

o

.

onsee

eliatntar a

re-

ingrtin

here.g,ofionor&

won

lveva

asderan

an,anin

tlena-

tainnd

tro-io &e inbebe-es-

andtoto

derssertitseen

nt of the

ay and

stop atfor the

Wnda

R

C

595WHEN RESOLUTION FAILS

cused on the influence of various text condition the timing of the process (for a review,Sanford & Garrod, 1989).

The belief that anaphoric inferences are rably drawn is based on the assumption threader’s goal is to build a coherent represetion of the text, and if the antecedent fopronoun, for example, is not identified (e.g.,Shewent to the store), this will result in a comphension failure (e.g., Who isshe?). Thus, ifreaders are not initially successful in identifyan antecedent, they should persevere, resoto alternate strategies for establishing coence, such as drawing bridging inferences (Haviland & Clark, 1974; Kintsch & van Dijk1978; van Dijk & Kintsch, 1983). In supportthis claim, there are a number of demonstratthat reading time increases when an anaphinference is difficult (e.g., O’Brien, Plewes,Albrecht, 1990).

In contrast with these assumptions, hoever, we hypothesize that there may be cditions under which readers fail to resoanaphors, even when an antecedent is a

TAB

Sample Tex

IntroductionWanda was throwing a surprise party for her best fr

company and some of his close friends wanted to

Distractor versionWanda even made him atart. She felt a little pres

she needed to prepare for that occasion as well. ShHer daughter loved their chocolatecakes.They were filfrosting on them. Wanda decided to have thecakedeco

No Distractor versionMeanwhile John was rushing around preparing

the store for decorations. Wanda was workintable and even made atart.

rap-upWanda hoped John’s party would be fun. The gues

Anaphor ending (Distractor and No Distractor versaid to save room for thedessert.

No Anaphor ending (Distractor version only)congratulated John on the promotion.

ecognition probeTART

omprehension questionWas John recently fired?

s

-a-

g-.,

sic

--

il-

able. Instead of viewing anaphor resolutionnecessary, there may be conditions unwhich readers’ “standard of coherence” (vden Broek, Risden, & Husebye-Hartm1995) is met even when the referent ofanaphor is not identified. For example,contrast with pronouns, which convey litinformation and are almost always used aphorically, noun phrase anaphors conmore information (e.g., Fletcher, 1984) acan either be used anaphorically or to induce a new concept (Fraurud, 1990; PoesVieira, 1998). Consider the sample passagTable 1. Although comprehension mightenhanced if readers make a connectiontween the categorical noun anaphor, “the dsert,” mentioned in the anaphor ending,“the tart,” described earlier, the failuremake this connection may not be disruptivethe comprehension process. That is, reamay be content to know that some desexists without specifically identifying whatis. Thus, if no single entity from the discourrepresentation is sufficiently activated wh

1

Experiment 1

d John. John had just been promoted to Vice Presidengratulate him.

ed because her daughter’s graduation was the next dtill had to get decorations and stop at the bakery for acake.with a very rich cream and had gooey dark chocolated with white flowers and a white border.

his daughter’s graduation the next day. He still had toard preparing for the evening. She had bought flowers

rrived right on time. As everyone sat down to eat, Was)

LE

t—

ienco

sure sledrate

forg h

ts asion

gaim

7)eyigma

rmthentrener-thatheen

flu-tiond faen

so-theheforievene rthehotivthet.tedhi

omthe

coeris tunntraytiodo

riber-f thor-orsnc

oftexton-ndssi-byphorn ofbe-

rs &tionm-a-ry,

thenateastual

asana-ed innter-dent.d aen-ack-l ofeenring

ity of79;r,aseuseoreer-y-

the

ed,useeralas-en athetures

596 LEVINE, GUZMAN, AND KLIN

the anaphor is encountered, rather than ening in additional processing, readers may sply read on.

Recent findings from O’Brien et al. (199provide initial support for this hypothesis. Thused a rapid serial visual presentation paradat a rate of 250 ms/word. After the anaphor wpresented, participants were asked to perfonaming task on the adjective modifier oftarget antecedent. When there was substadistance between the anaphor and its refeno facilitation was found. In the current expiments we go a step further and suggesteven when readers are allowed to read atown pace they may not identify the antecedfor an anaphor. Further, we explore the inence of a variety of text factors on the resoluprocess. We hypothesize that readers shoulto resolve anaphors (1) when the antecedare difficult to retrieve and (2) when this relution failure does not create a break incoherence of the passage. When both of tconditions are met, it should be inefficientreaders to invest attentional resources to retrthe antecedent. In contrast, when antecedare easily accessible, readers should resolvgardless of the need to do so. Similarly, iffailure to locate an antecedent for an anapcreates a break in the coherence of the narrareaders should be more likely to devoteresources needed to retrieve the anteceden

In the current set of experiments we testhese hypotheses by creating passages in wantecedents should be difficult to retrieve frmemory and in which the failure to resolveanaphors should not create a break in theherence of the text. In a series of six expments, participants read passages, such aone in Table 1, that either ended with a nophrase anaphor (e.g., dessert) or had a coending. An antecedent (e.g., tart) was alwpresent in the passage. The primary queswas if conditions exist under which readersnot resolve the anaphor. In addition to contuting to our understanding of anaphoric infences, these experiments provided a test ofactors that influence the accessibility of infmation in long-term memory and of the factthat influence readers’ assessment of cohere

g--

,sa

ialt,

tirt

ilts

se

etse-

re,

ch

--he

olsn

-

e

e.

We start by considering the accessibilityantecedents. Given a bottom-up theory ofcomprehension, such as Kintsch’s (1988) Cstruction-Integration model or Myers aO’Brien’s (1998) resonance model, the accebility of an antecedent should be influencedfactors such as the distance between the anaand the antecedent, the amount of elaboratiothe antecedent, and the degree of matchtween the anaphor and its antecedent (MyeO’Brien, 1998). This is based on the assumpthat anaphors act like any other input to meory, automatically activating related informtion in memory. All related concepts in memoresonate in parallel (cf. Gillund & Shiffrin1984; Hintzman, 1986; Ratcliff, 1978) anddegree to which potential antecedents resoshould depend on their level of activationwell as the number of semantic and concepfeatures that overlap with the anaphor.

The first text factor that we included wsubstantial referential distance between thephor and its antecedent. In the passages usthe current experiments, several sentences ivened between an anaphor and its anteceMoreover, the intervening material constituteshift in topic. Thus, when the anaphor wascountered, the antecedent was always bgrounded, or out of focus. Given any modememory, increasing the amount of time betwencountering the antecedent and encountethe anaphor should decrease the accessibilthe antecedent (e.g., Clark & Sengul, 19Duffy & Rayner, 1990; Ehrlich & Rayne1983; O’Brien, 1987). Distance may decrethe activation of the antecedent either becaits strength decays over time or because mrecently mentioned concepts are likely to ovlap with the contents of working memory (Mers & O’Brien, 1998).

The second text factor we explored waspresence of other potential antecedents, ordis-tractors. As Greene et al. (1992) suggestfailure to locate a referent may occur becaeither no entity matches well or because seventities match about equally well. In the psages used in the current experiments, whsame-category distractor was included inpassage, two entities shared semantic fea

wabeth

n,764;

lsondoripaa

atelyotseth

bymi-lu

lowill

sad inofo

obaucerereanunedbod aenithe

di

thtorke

ore.grayuldngth

valeasn toas

ofseead-horsicalhen

lityuse

Ac-ed&o

horhorro-rk-

em-byy-

edhens itas-n been-beestro-

s aly

ike-be

ent.horouldnts,weree.g.,ountandcancane a

597WHEN RESOLUTION FAILS

with the anaphor. Thus, when the anaphorread, its signal should have been dividedtween the referent and distractor, reducingactivation of the referent (Myers & O’Brie1998) by creating a fan effect (Anderson, 19Myers, O’Brien, Balota, & Toyofuku, 198Reder & Ross, 1983).

In support of this, Corbett (1984; see aCorbett & Chang, 1983; Mason, 1997) fouthat reading times were longer on an anaphphrase (e.g., frozen vegetables) when thesage contained both a referent (e.g., frozenparagus) and a distractor from the same cgory (e.g., fresh corn) than when it oncontained a referent. Corbett argued that bpotential antecedents were initially acceswhen the anaphoric phrase was read, andcorrect referent was eventually identifiedusing the information provided by the prenonal adjective (i.e., frozen). Whereas the incsion of a distractor has been shown to scomprehension, it is not clear whether it walso eliminate resolution.

The third factor we examined was thelience of the distractor. This was manipulatetwo ways. First, the amount of elaborationthe distractor was varied. A large amountelaboration should serve to increase the prbility that the distractor is encoded and proda richer, more highly integrated memory repsentation. When the anaphor is encounteactivation should be drawn away from thetecedent. Support for this hypothesis was foin O’Brien et al. (1990); an elaborated antecent was retrieved more quickly than an unelarated antecedent even though the elaboratetecedent was more distant. In the currexperiments an elaborated distractor may ecause readers to erroneously reinstate thetractor or to reinstate neither entity.

Salience was also manipulated by varyingtypicality of the distractor. When the distracis a more typical category exemplar (e.g., cathan the referent (e.g., tart), it will share msemantic features with the category label (edessert) than the referent (Rosch, Mervis, GJohnsen, & Boyes-Braem, 1976). This shocause the distractor to resonate more strobecause of its greater featural overlap with

s-e

;

cs-s--

hde

-

-

f-

-d,-d--n-trs-

e

)

.,,

lye

anaphor, which, in turn, should make retrieof the antecedent more difficult. Again, wherthe degree of match has not been showinfluence the probability of resolution, it hbeen shown to influence the time coursecomprehension. Duffy and Rayner (1990;also Garrod & Sanford, 1977) found that reers took longer to resolve categorical anap(e.g., bird) when the antecedent was an atypmember of the category (e.g., goose) than wit was a highly typical member (e.g., robin).

In addition to manipulating the accessibiof the antecedents, we also attempted toanaphors that did not demand resolution.cording to the bottom-up framework introducearlier (Kintsch, 1988; McKoon, Gerrig,Greene, 1996; Myers & O’Brien, 1998), if nentity is sufficiently activated when an anapis read, or if several entities match the anapabout equally well, readers may continue pcessing by refocusing on the contents of woing memory and sending new signals to mory until an antecedent is retrieved orengaging in some sort of problem solving (Mers & O’Brien, 1998). Further, this continuprocessing should be most likely to occur wthe failure to retrieve the antecedent makedifficult to integrate the anaphor into the psage. In contrast, when a noun phrase caeasily integrated into the ongoing text represtation without access to its referent, it maymore efficient to continue reading than to invadditional time and attention continuing to pcess the anaphor.

Without a full understanding of what maketext coherent, it is difficult to define precisethe set of factors that should influence the llihood that an anaphoric noun phrase cancomprehended without retrieving its antecedHowever, both the explicitness of the anapand the degree of contextual constraint shhave an influence. In the current experimenoun phrase anaphors (e.g., the dessert)used rather than pronominal anaphors (she). The greater semantic content of nphrases should allow them to more easily son their own; whereas written pronounsonly be used anaphorically, noun phraseseither be used anaphorically or to introduc

usetexoucitaslkeretheno

tedde

veageced sulf thg aceckal,waa

dewenctratheesri-

hausld iastioen

der

nylves aownail-tente

na-

e Ta-er-thatniteto annedarere-in

thectordis-t. Inoneare

tion,thethem-

thebothlityon-the

hor-wasntialan-con-agetheer’s

aged foran-to

obeersthe.or,ionna-thetheto

598 LEVINE, GUZMAN, AND KLIN

new concept. And second, anaphors werethat were constrained by the passage conFor example, because “dessert” is part ofdinner party schema, it should be in “implifocus” (Sanford & Garrod, 1981) when the psage is read (Garrod & Sanford, 1983; Wa& Yekovich, 1987). Therefore, even if thewere no antecedent explicitly included inpassage, it could be used to refer to somespecific dessert.

In summary, in the six experiments reporhere we asked if there are any conditions unwhich noun anaphors fail to be resolved, ewhen an antecedent is available in the passPassages were constructed so that the antents should have been difficult to retrieve anthat the failure to resolve the anaphors shonot have created a break in the coherence opassages. In Experiment 1, after processinanaphor (e.g., dessert), there was no evidenresolution when the antecedent was bagrounded by a description of a highly typicsame-category distractor (e.g., cake). Thisbased on the failure to find facilitation onrecognition probe representing the antececoncept (e.g., tart). In Experiments 2 and 3investigated the influence of referential distaand of the salience of the same-category distor, respectively. Experiment 4 replicatedfindings from Experiment 1 with reading timinstead of recognition. The results of Expement 5 ruled out the possibility that readersactually resolved the anaphor by erroneoselecting the distractor as the referent, anExperiment 6 an instructional manipulation wused to demonstrate that despite the resolufailures in Experiments 1 and 4, the antecedwas accessible and could be identified if readevoted additional time.

EXPERIMENT 1

In Experiment 1, we asked if there are aconditions under which readers fail to resocategorical anaphors when the antecedent icessible and readers are free to read at theirpace. In contrast to the assumption that aphoric inferences are reliably drawn, if the faure to resolve an anaphor does not constitucoherence break, and the retrieval of the a

dt.r

-r

n-

rne.d-

odenof-

s

nt

ec-

dyn

nts

c-n-

a-

cedent is difficult, readers may leave the aphor unresolved.

Passages appeared in three versions (seble 1). The first was a Distractor–Anaphor vsion, in which passages ended with a linecontained a categorical anaphor in a definoun phrase (e.g., the dessert) that referredantecedent (e.g., tart) that had been mentioearlier in the passage. To determine if thereconditions under which anaphors are notsolved, three text factors were manipulatedan attempt to decrease the accessibility ofantecedent. First, a same-category distrawas included in the passage. Second, thetractor was more salient than the antecedenthe sample passage in Table 1, there is onlyshort sentence mentioning the tart, but therefour sentences related to the cake. In addithe distractor was a more typical member ofcategory instantiated by the anaphor thanantecedent (Battig & Montague, 1969). Assuing that the distractor is activated whenanaphor is processed (e.g., Corbett, 1984),the extensive elaboration and the high typicashould increase its level of activation and csequently decrease the level of activation ofantecedent. Third, the antecedent was toughly backgrounded when the anaphorencountered. Not only was there a substaamount of material intervening between thetecedent and the anaphor, but this materialstituted a shift in topic. In the sample passthe intervening material shifted the topic ofnarrative from the dinner party to the daughtgraduation.

Immediately after the last line of the passa probe word, the antecedent, was presenterecognition. If the anaphor is resolved, thetecedent should be highly active, leadingfacilitation when it is presented as a test pr(e.g., Dell et al., 1983). In contrast, if readfail to resolve the anaphor, responding toantecedent probes should not be facilitated

The second version, Distractor-No Anaphwas identical to the Distractor-Anaphor versexcept that the last line did not contain an aphor. Instead, it continued with the theme ofpassage (“Wanda congratulated John onpromotion”) and in no way made reference

astheci-anheortw

r-icaorthiblth

soraceen

.g.heverntreeanreenn t

astena-edg

cetors.

atm-r aom5%erd

as-s torA is-t b

nden-as

wedro-ametyp-&

–7thethet re-

Dis-na-ssgth

on,ingin-andaredr, inin

ge of

fol-theagese athe

e 18bexi-nti-60%theas-

-wasas-ully.hor.tal

theor,wohirdeachone-

599WHEN RESOLUTION FAILS

the antecedent concept. This version servedcontrol. If readers resolve the anaphor inDistractor-Anaphor version, recognition desions should be faster and more accurate ththe Distractor-No Anaphor version. On tother hand, if they fail to resolve the anaphthere should be no difference between theversions.

Finally, in the third version, No DistractoAnaphor, the last line contained the categoranaphor, as it did in the Distractor-Anaphversion, but the passages were written sothe antecedent should be easily accessThere was less intervening material betweenanaphor and the antecedent and there wasame-category distractor. The lengthy elabtion about the cake, for example, was replawith a shorter section of neutral material. Givthe findings from several prior studies (eDell et al., 1983; O’Brien et al., 1997), tanaphor should be easily resolved in thission; the anaphor is unambiguous and the acedent is only backgrounded by two to thlines of text. This condition was included asadditional baseline. If the anaphor is notsolved in the Distractor-Anaphor version, latcies and error rates should be the same as iDistractor-No Anaphor version, leading tonull result. In contrast, latencies should be faand error rates lower in the No Distractor-Aphor version. This condition was also includto control for semantic priming; any priminfrom the anaphor (e.g., dessert) to the anteent (e.g., tart) should be equal in the DistracAnaphor and No Distractor-Anaphor version

Method

Participants. Forty-eight undergraduatesthe State University of New York at Binghaton participated as part of the requirement fointroductory psychology course. The data frfour participants who made more than 2probe or comprehension question errors wreplaced. Therefore data analyses are base44 participants.

Materials. There were 18 experimental pages, with three versions of each: Distracnaphor, Distractor-No Anaphor, and No D

ractor-Anaphor. A sample passage can

a

in

,o

l

ate.eno-d

,

-e-

--he

r

d--

n

eon

-

e

found in Table 1. The Distractor-Anaphor athe Distractor-No Anaphor versions were idtical until the last line: The antecedent wmentioned once early in the passage, folloby a topic shift where the distractor was intduced. The distractor was a member of the scategory as the antecedent but was a moreical member than the antecedent (BattigMontague, 1969). In addition, there were 4lines of elaboration about the distractor. InDistractor-Anaphor version, the last line ofpassage included a categorical anaphor thaferred to the antecedent. In contrast, in thetractor-No Anaphor version the categorical aphor was not included in the last line. Acroversions the last lines were equated for lenby number of characters. In the third versiNo Distractor-Anaphor, the section describthe distractor was replaced by neutral fillerformation that was unrelated to the anaphorwas shorter. In addition, antecedents appean average of only 3 lines before the anaphocomparison to the two distractor versionswhich the antecedents appeared an avera10 lines before the anaphor.

All three versions of the passage werelowed by a recognition probe, which wasantecedent. There were also 25 filler passthat varied in length. For each filler passagrecognition probe was generated. None offiller passages ended with anaphors. For thfiller passages that were followed by prowords that required “no” responses, appromately 40% of the probe words were semacally related to the passages and the otherwere unrelated words. Over the course ofexperiment participants read a total of 43 psages, 25 of which required ayes probe response. A yes/no comprehension questiongenerated for each experimental and filler psage to encourage participants to read carefThese questions never referred to the anap

Design. For each participant, experimenpassages were quasirandomly assigned toDistractor-Anaphor, Distractor-No Anaphand No Distractor-Anaphor versions with tconstraints: (a) Each participant saw one-tof the passages in each version and (b)version of each passage was presented to

talerpe

id-roxonici-siotheXTantinerretegek

fora

reara

thebox

se

ed-intoE-sfol-

ches-e-red

sjectinbil-x-

-eseiersal

r Rates

r rate

600 LEVINE, GUZMAN, AND KLIN

third of the participants. The order of presention was the same for all participants. Filpassages were interspersed among the exmental passages.

Procedure.Participants were tested indivually in an experimental session lasting appimately 45 min. The texts were presentedline at a time on a computer monitor. Partpants were instructed to read for comprehenat their own pace. Each trial began withpresentation of the phrase READY FOR NESTORY at the center of the screen. Participcontrolled the presentation of the text with a ladvance key. Each key-press caused the culine to be erased and the next to be presenImmediately after the final line of the passahad been presented, pressing the advancecaused the string “XXX” to be presented500 ms. Participants did not know they werethe end of the passage until this cue appeaThis was replaced by a recognition probe. Pticipants had been instructed to indicate,quickly and as accurately as possible, whethe probe had been in the preceding textpressing a YES or NO key on a response bAfter responding, the probe word was era

TAB

Experiments 1, 2, 3, and 6: Last Line Reading Time

Version

Experiment 1: Highly salient distractorNo Distractor-AnaphorDistractor-AnaphorDistractor-No Anaphor

Experiment 2: No distractorAnaphorNo Anaphor

Difference

Experiment 3: Less salient distractorNo Distractor-AnaphorDistractor-AnaphorDistractor-No Anaphor

Experiment 6: Highly salient distractor and instructionsAnaphorNo Anaphor

Difference

* p , .05.

-

ri-

-e

n

s

ntd.

ey

td.-sr

y.

d

from the screen and participants received feback about their response time. On trialswhich participants took longer than 1500 msrespond, the phrase PLEASE TRY TO RSPOND TO PROBES MORE QUICKLY wapresented for 1500 ms. Each probe waslowed by the string “???” for 500 ms, whiwas followed by a Yes/No comprehension qution. For trials in which the participants rsponded incorrectly, the word “Error” appeain the center of the screen for 1500 ms.

Results and Discussion

In all experiments,t 1 andF 1 refer to analysein which the error term is based on subvariability and t 2 and F 2 refer to analyseswhich the error term is based on item variaity. An a level of .05 was used across all eperiments.

Probe recognition times.Mean correct recognition times are presented in Table 2. Thvalues were computed after discarding outl(Tukey, 1977),1 3.8% of the data. Most centr

1 According to Tukey’s (1977) criterion, a score,X, isdefined as an outlier ifX . H 1 (1.5) 3 (H 2 H ) or

2

nd Probe Response Times (in Milliseconds) and Erro

eading time Probe response times Probe erro

1746 763 4.9%1779 804 14.0%1775 804 15.5%

1713 806 8.2%1736 874 8.2%

23 268* 0.0%

1699 833 8.3%1720 824 9.2%1658 876 11.0%

2155 865 13.6%2027 907 14.4%128* 242* 0.8%

LE

s a

R

U U L

foNopitorievth

ngrt”ften,tedThongt.cenode

renrceo

2)erthin

ionster-ver tnoNoby

2 ntc is-t tot usw eer on,t reco

rst tesw hora and1 eN ni-t alsol redw2

er-

aect

sterdentre-r, aibil-

dentn inis-dentnotnotof

ncerd-we

lvedsedforNocemore

thena-na-andor

sen-n-iva-

ortved

keards pe

601WHEN RESOLUTION FAILS

to the hypotheses, probe recognition timesthe Distractor-Anaphor and the Distractor-Anaphor versions were identical. Thus, desprevious findings demonstrating that anaphinferences are reliably drawn, there was noidence that readers resolved the anaphor inDistractor-Anaphor version. After readi“save room for the dessert,” the concept “tawas no more accessible in memory than areading “congratulated John on the promotiodespite the fact that it had been explicitly stathat Wanda was making a tart for dessert.signal from the anaphor was clearly not strenough to reactivate the intended antecedenaddition, because the inference was not nesary for comprehension (i.e., readers didneed to know that the dessert was a tart), reaseem to have been willing to leave the refeunresolved rather than investing the resouneeded to complete the resolution process. Csistent with McKoon and Ratcliff’s (199Minimalist hypothesis, it appears that readdid not draw an anaphoric inference becauseantecedent was not easily available and theference was not necessary for comprehens

In contrast, recognition latencies were fafollowing the target line in the No DistractoAnaphor version. That is, after reading “saroom for the dessert,” participants were fasterespond to the probe “tart” when there wasdistractor in the passage. A contrast of theDistractor-Anaphor version with the averagethe two distractor versions was significantsubjects,t 1(43) 5 3.52, SEM 5 11.46,p 5.001, and byitems, t 2(17) 5 2.16, SEM 5

0.14, p 5 .04. This provides an importaomparison for the null effect for the two Dractor versions; clearly the task is sensitivehe activation level of the probed concept. The can conclude that if the anaphor had b

esolved in the Distractor-Anaphor versihen it would have been reflected in fastergnition times.

X , H L 1 (1.5) 3 (HU 2 H L). HU and H L refer to thescores that cut off the upper and lower 25% of the ranscores, respectively. Further, note after data were discthere were slight inequalities in numbers of observationcondition.

r

ec-e

r”

e

Ins-trsts

n-

se-.r

o

of

,n

-

Error rates.The pattern of error rates mirrohe pattern of recognition latencies. Error raere almost identical in the Distractor-Anapnd Distractor-No Anaphor versions, 14.05.5%, respectively,ps . .5. In contrast, in tho Distractor-Anaphor version where recog

ion latencies were faster, the error rate wasower (4.9%). This was reliable when compaith the Distractor-Anaphor version,t 1(17) 5.83, SEM 5 .47, p 5 .012, andt 2(43) 5

3.15, SEM 5 .174, p 5 .003 and whencompared with the Distractor-No Anaphor vsion, t 1(17) 5 3.24, SEM 5 .48, p 5 .005,

nd t 2(43) 5 5.21, SEM 5 .12, p , .001.Again, this provides a contrast for the null effin the two distractor versions.

Although we have suggested that the falatencies and lower error rates on the anteceprobe in the No Distractor-Anaphor versionsulted from readers resolving the anaphosecond possibility exists. The greater accessity may have been the result of the anteceappearing more recently in the passage thathe other two versions. The reduction in dtance alone might have made the anteceprobe more accessible, even if readers hadresolved the anaphor. Although the data doallow us to distinguish between the influenceanaphor resolution and the influence of distain the No Distractor-Anaphor version, regaless of the reason for the greater accessibilitycan conclude that the anaphor was not resoin the Distractor-Anaphor version. This is baon the null effect in latency and error ratesthe Distractor-Anaphor and the Distractor-Anaphor versions. This lack of a differendemonstrates that the antecedent was noactive after the anaphor was encountered inDistractor-Anaphor version than when no aphor was encountered in the Distractor-No Aphor version. Further, the shorter latencieslower error rates in the No Distractor-Anaphversion provide evidence that the task wassitive to the activation level of the probed cocept, whatever the reason for the higher acttion in this condition. And finally, inExperiment 4 we provide converging suppfor the claim that the anaphor was not resolin the Distractor-Anaphor version.

dedr

eslatey’sheer

wckis-deenry.ty i

thvinin

tedricat aertextionasucaseegoth

tedlawe

ter-re.of

thyd ipreor

cus, &fi-n

artckan

diero-

nd itser,

udyor-tedrent

roxi-

phortheor

dis-eu-

hetillthe

e aForle 3)r’s

p a

ngna-e

atm-ata5%ere

ed on

as-andd inpas-ac-x-

wasthehectione inpic.d ofake

602 LEVINE, GUZMAN, AND KLIN

Reading times.Last line mean reading timare presented in Table 2. These were calcuafter discarding outliers as defined by Tuke(1977) criterion; this eliminated 5.2% of tdata. Reading times did not differ across vsions, bothFs , 1. Although we cannot drastrong conclusions from this null result, the laof a slowdown on the anaphor line in the Dtractor-Anaphor version suggests that readid not attempt to retrieve the antecedthrough a slow, effortful search of memoInstead, we conclude that because no entimemory resonated above a threshold, andpassage was coherent even without resolthe anaphor, readers had little incentive tovest additional resources.

EXPERIMENT 2

The results of Experiment 1 demonstrathat readers do not always resolve categoanaphors, even when they are free to read arate and an antecedent is available. In Expments 2 and 3 we examined some of thefactors that were responsible for the resolufailure. Three features were included in the psages in Experiment 1 in an attempt to redthe accessibility of the antecedent: there wsubstantial amount of backgrounding betwthe anaphor and its antecedent, a same-catedistractor was included in the passage, anddistractor was highly salient; it was elaboraand it was a more typical category exempthan the referent. In Experiments 2 and 3examined these factors individually to demine their contribution to the resolution failu

In Experiment 2 we examined the rolereferential distance. Simply having a lengepisode intervening between the anaphor anantecedent might have been sufficient tovent resolution. Although noun phrase anaphdo not require their antecedents to be in foas do pronominal anaphors (Gordon, GroszGilliom, 1993), even a semantically rich denite noun phrase, such as “the dessert,” maybe sufficient to reactivate an entity (e.g., tthat was mentioned only once and was bagrounded by approximately 100 words andentire episode. In support of this, several stuhave found a slowdown in the resolution p

d

-

rst

neg-

lnyi-t

-eanrye

r

ts-s,

ot)-

s

cess as the distance between the anaphor aantecedent increased (e.g., Duffy & Rayn1990; Mason, 1997). Further, in a recent st(O’Brien et al., 1997) the referent for a categical anaphor was not immediately activawhen the anaphor (e.g., clothes) and its refe(e.g., baby clothes) were separated by appmately 60 words.

Passages appeared in two versions, Anaand No Anaphor. These were similar toDistractor-Anaphor and Distractor-No Anaphversions from Experiment 1 except that thetractor elaboration section was replaced by ntral filler material, equated for length with tdistractor elaboration. This neutral material sbackgrounded the antecedent by changingfocus of the passage, but it did not includdescription of a same-category distractor.example, in the sample passage (see TabWanda must still prepare for her daughtegraduation but instead of needing to pick ucake she must buy anecklace.The primaryquestion is if the inclusion of backgroundimaterial that is completely unrelated to the aphor will still reduce the accessibility of thantecedent and eliminate resolution.

Method

Participants. Forty-three undergraduatesthe State University of New York at Binghaton participated for course credit or $5. The dfrom one participant who made more than 2probe and comprehension question errors wreplaced. Therefore data analyses were bas42 participants.

Materials. There were 18 experimental psages, with two versions of each: AnaphorNo Anaphor. A sample passage can be founTable 3. The passages were similar to thesages in the Distractor-Anaphor and Distrtor-No Anaphor versions in Experiment 1 ecept that the distractor elaboration sectionreplaced by material that was unrelated toanaphoric concept. This “neutral filler” was tsame length as the distractor elaboration sefrom Experiment 1, was as similar as possiblits content, and constituted a change in toFor example, in the sample passage, insteadescribing Wanda’s quest for a chocolate c

ow

e igeav

hoo

phoon)mth

recthgeri-

-t 1ex

ctledin3%ult

eretheec-b-

.1 -

ersck-phor

nedcuswas

t totheno

e lastw oten-t besr t areu sert)a horv on-s aphora ec-o

nt of the

eded to

ily

nda

603WHEN RESOLUTION FAILS

for her daughter’s graduation party, she is nin search of a pearl necklace.

Passage endings were identical to thosExperiment 1; in the Anaphor version, passaended with the categorical anaphor (e.g., sroom for the dessert) and in the No Anapversion the last line continued with the themethe passage and did not include the ana(e.g., congratulated John on the promotiThe last lines were equated for length by nuber of characters. Further, in both versions,last line of the passage was followed by aognition probe that was the antecedent inAnaphor version (e.g., tart). The filler passafrom Experiment 1 were used in this expement.

Design and procedure.The design and procedure were identical to those in Experimenexcept that there were two versions of theperimental passages instead of three.

Results and Discussion

Probe recognition times.The mean correprobe response times are presented in TabThese values were computed after discaroutliers (Tukey, 1977), which consisted of 3.of the data. In contrast to the pattern of res

TAB

Sample Tex

Introduction and antecedentWanda was throwing a surprise party for her best fr

company and some of his close friends wanted to

Neutral fillerThe day had been busy because Wanda’s daughter

prepare for that as well. She had first stopped atdaughter’s gift. Her daughter had asked for a peatradition for quite some time to give the girls a pe

Wrap-upWanda hoped John’s party would be fun. The gues

Anaphor endingsaid to save room for thedessert.

No Anaphor endingcongratulated John on the promotion.

ProbeTART

Comprehension questionWas John recently fired?

nserfr

.-e-es

,-

2.g

s

from Experiment 1, probe response times w68 ms faster in the Anaphor version than inNo Anaphor version, 806 and 874 ms, resptively. This difference was significant by sujects, t 1(41) 5 6.33, SEM 5 10.63, p ,001, and byitems, t 2(17) 5 5.17, SEM 54.29,p , .001. Thefaster recognition laten

cies in the Anaphor version show that readresolved the anaphor, reinstating the bagrounded antecedent after reading the analine (e.g., save room for the dessert).2 Impres-sively, although the antecedent was mentioonly once and backgrounded by a shift in foand approximately 100 words, the anaphorresolved.

The finding that distance was not sufficieneliminate anaphor resolution is in contrast toresults of O’Brien et al. (1997), who found

2 Because the category name (e.g., dessert) is thord of the passage in the Anaphor version, there is p

ially a concern about priming. However, (1) the proepresent low typical category exemplars (e.g., tart) thanlikely to be primed by the category names (e.g., desnd (2) the null effect in Experiment 1 (Distractor-Anaps Distractor-No Anaphor version) empirically demtrates that any semantic association between the annd the probe is insufficient to produce facilitation in rgnition latencies.

3

Experiment 2

d John. John had just been promoted to Vice Presidengratulate him. Wanda even made him atart.

s graduating from high school tomorrow. She had nedry cleaner’s and then at the jeweler’s to pick up herecklace for her graduation present. It had been a famnecklace when they graduated from high school.

rrived right on time. As everyone sat down to eat, Wa

LE

t—

ienco

watherl narl

ts a

esb

haceis

e-urt a

aves,he.asnis-1.

litytheesses

nt te ocried

in

ble2 rror fact phov rror inE rror ofE thd pha f tha

reeeneno

oren

ry, th

ond

fer-

2as

d innot

ifor

w an-t or ift thef nceb esss icalc eri-m ndc is ifr ast oft

M

heS onp Thed hom sionq hod al-y

as-s plep rac-t er-s thatt lesst nedo is-t inw edb thw tedt is-t orc lastl idn

604 LEVINE, GUZMAN, AND KLIN

facilitation on an antecedent (e.g., baby clothwhen it was separated from the anaphorapproximately 60 words, substantially less tthe distance between the anaphor and anteent in the current experiment. Although itdifficult to pinpoint the critical differences btween the two studies, in contrast with the crent study, O’Brien et al. presented texts arate of 250 ms/word. Thus, resolution may hbeen delayed rather than eliminated. That ireaders had been given additional time, tmay have completed the resolution process

The results of Experiment 2 are in contrwith the lack of a difference in recognitiolatencies for the Distractor-Anaphor and Dtractor-No Anaphor versions in ExperimentAlthough this does not rule out the possibithat referential distance interacted withsame-category distractor to reduce the accbility of the antecedent in Experiment 1, it dosuggest that distance alone was not sufficieeliminate anaphor resolution. The presencthe same-category distractor was thereforeical in reducing the accessibility of the antecent. This hypothesis is investigated furtherExperiment 3.

Error rates.Error rates are presented in Ta. In contrast with the response latencies, eates did not differ between versions. Thehat shorter response latencies in the Anaersion were not accompanied by lower eates is in contrast to the pattern of resultsxperiment 1. This suggests that the lower e

ates in the No Distractor-Anaphor versionxperiment 1 may have been the result ofecrease in the distance between the anand the antecedent rather than the result onaphor being resolved.Reading times.Last line reading times a

presented in Table 2 and did not differ betwversions. Although the last lines were not idtical, they were equated for length. The lacka slowdown in the Anaphor version therefsuggests that the anaphoric inferences didinvolve a slow, strategic search of memoInstead, because there was no distractorintended referent should have resonated mstrongly than any other entity in memory a

)ynd-

-

ify

t

i-

of

t--

rtr

r

r

eore

-f

ot.e

re

should have been quickly selected as the reent of the anaphor.

EXPERIMENT 3

The major conclusion from Experimentwas that the inclusion of a salient distractor wnecessary for the resolution failure observeExperiment 1; backgrounding alone wassufficient. In the current experiment we asksimply includingany same-category distract

ould have reduced the accessibility of theecedent and eliminated anaphor resolutionhe salience of the distractor was a factor inailure. We examined the influence of saliey replacing the salient distractors with lalient distractors. These were both less typategory exemplars than those used in Expent 1 (e.g., “eclair” instead of “cake”) a

ontained less elaboration. The questioneaders will still fail to resolve the anaphor,hey did in Experiment 1, or if the saliencehe distractor was critical.

ethod

Participants.Forty-five undergraduates at ttate University of New York at Binghamtarticipated in exchange for course credit.ata were replaced from four participants wade more than 25% probe or comprehenuestion errors and from two participants wid not follow instructions. Therefore the anses were based on 39 participants.Materials. There were 18 experimental p

ages, with three versions of each. A samassage can be found in Table 4. The Dist

or-Anaphor and Distractor-No Anaphor vions from Experiment 1 were used, excepthe salient distractor was replaced with aypical category exemplar that was mentionly once. The third version was the No D

ractor-Anaphor version from Experiment 2,hich the distractor information was replacy filler information that was equated for lengith the distractor elaboration but was unrela

o the anaphor. The last line of both the Dractor-Anaphor and No Distractor-Anaphontained the categorical anaphor and theine of the Distractor-No Anaphor version dot.

c-typderorkwodistheexwlof

jecseer

ors

-t 1

ve

,

Fctle 2.dingostes

horer-as

S r-

inginghist was-

ced-eri-

nt of the

eded to

esent

nda

605WHEN RESOLUTION FAILS

Rating experiment.To ensure that the distrators selected for this experiment were lessical than those used in Experiment 1, 27 ungraduates at the State University of New Yat Binghamton were asked which of the texemplars was more typical. One was thetractor from Experiment 1 and the other wasdistractor selected for this experiment. Forample, the question for the sample passage“Which of the following is the more typicadessert? A) Eclair B) Cake.” The orderchoices was counterbalanced across subThe distractors used in Experiment 1 werelected as more typical 86% of the time. Furthfor every item, the Experiment 1 distractwere rated as more typical.

Design and procedure.The design and procedure were identical to those in Experimen

Results and Discussion

Error rates and reading times.As in Exper-iment 2, there were no differences across

TAB

Sample Tex

Introduction and antecedentWanda was throwing a surprise party for her best fr

company and some of his close friends wanted to

Neutral backgroundingThe day had been busy because Wanda’s daughter

prepare for that as well. She had first stopped atdaughter’s gift. Her daughter had asked for a pea

Distractor versionIt had been a family tradition for some time. Sh

daughter’s graduation dinner.No Distractor version

It had been a family tradition for quite some timfrom high school.

Wrap-upWanda hoped John’s party would be fun. The gues

Anaphor ending (Distractor and No Distractor versaid to save room for thedessert.

No Anaphor ending (Distractor version only)congratulated John on the promotion.

ProbeTART

Comprehension QuestionWas John recently fired?

--

-

-as

ts.-,

.

r-

sions in error rates (Fs , 1) or reading timesF 1(2,76) 5 1.23, MSe 5 31,443,p 5 .30;

2(2,34) 5 .83, MSe 5 26,730,p 5 .45.Probe recognition times.The mean corre

probe response times are presented in TabThese values were computed after discaroutliers (Tukey, 1977), 5.3% of the data. Mcentral to our question, probe reaction timwere 52 ms shorter in the Distractor-Anapversion than in the Distractor-No Anaphor vsion: 824 ms vs 876 ms. This difference wsignificant by subjects,t 1(37) 5 3.87,SEM513.36,p , .001, and byitems,t 2(17) 5 2.09,

EM5 30.88,p 5 .05. In contrast to Expeiment 1, readers were faster to respondyes tothe probe word “tart,” for example, after read“save room for the dessert” than after read“congratulated John on the promotion.” Toccurred despite the fact that the antecedensubstantially backgroundedand a same-category distractor intervened between the anteent and the anaphor. However, unlike Exp

4

Experiment 3

d John. John had just been promoted to Vice Presidengratulate him. Wanda even made him atart.

s graduating from high school tomorrow. She had nedry cleaner’s and then at the jeweler’s to pick up herecklace for her graduation present.

lso still had to stop at the bakery foreclairs for her

give the girls a pearl necklace as their graduation pr

rrived right on time. As everyone sat down to eat, Was)

LE

t—

ienco

watherl n

e a

e to

ts asion

nory.ratx-

vedthe.ct1,eentiontorac-nodisre

s esss rov inge trt thea redt et( hea nol thea thec

s oe or,w fl cals prev allo utiom t ot ons eet inei thea lae alsn plyd uyu uat

oesn at

theultsto

dersis-thatthelveNott 1gorydent

oricEx-imesEx-e of

ci-na-ilityhe

sageafteruse it

is, thet int 1essentn astionthe

ana-or

iffi-ofad-ntlyk ofandingsdoner,

606 LEVINE, GUZMAN, AND KLIN

ment 1, the distractor was not elaboratedwas it a highly typical member of the categoAlthough these data do not allow us to sepathe effects of elaboration and typicality, in Eperiment 1 their combined effect clearly serto divert activation from the antecedent todistractor when the anaphor was processed

Not only does the 52-ms facilitation effecontrast with the null effect in Experimentbut the facilitation of the probe following thAnaphor line was not significantly larger whthere was no distractor present: The facilitaon the antecedent probe in the No DistracAnaphor version, compared with the Distrtor-No Anaphor version, was 43 ms. Thus,only can we conclude that the less salienttractor providedlessinterference than the moalient distractor, but the inclusion of the lalient distractor did not appear to have pided any interference at all. That is, havncountered a sentence describing Wanda’s

o the bakery to buy eclairs did not hinderccessibility of “tart” when readers encounte

he anaphor, “the dessert.” Although Corb1984) did not include the same conditions,lso found an effect of typicality: There was

onger a slowdown on the anaphor whenntecedent was a highly typical example ofategory and the distractor was atypical.Because we did not separate the effect

laboration from the typicality of the distracte can only conclude that the presence o

engthy episode describing a highly typiame-category distractor was sufficient toent resolution. However, it is possible thatf these features were not necessary; resolay have also been prevented if a subse

hese factors were included. For example, cider an elaborated episode intervening betwhe antecedent and the anaphor that contanformation from the same semantic field asntecedent, but did not mention a particuxemplar. That is, the anaphor may haveot been resolved if the backgrounding simescribed Wanda’s trip to the bakery to bnspecified sweets for her daughter’s grad

ion party.In summary, although this experiment d

ot allow us to specify the full set of factors th

r

e

-

t-

-

ip

t

f

a

-

nf-nd

ro

-

were responsible for the failure to resolveanaphor in Experiment 1, the combined resof Experiments 2 and 3 do allow us to beginestablish some boundary conditions unwhich anaphoric inferences are drawn. Content with the many studies demonstratinganaphoric inferences are drawn reliably,conditions under which readers fail to resocategorical anaphors appear to be limited.only was the resolution failure in Experimendependent on the inclusion of a same-catedistractor, but it seems to have been depenon the distractor being highly salient.

EXPERIMENT 4

In contrast to the assumption that anaphinferences are reliably drawn, the results ofperiment 1 established that readers sometleave categorical anaphors unresolved. Inperiments 4 and 5 we investigate the naturthe resolution failure.

In the current experiment, we ask if partipants would have ultimately resolved the aphor in Experiment 1, despite the inaccessibof the antecedent, if they had more time. Tanaphor was always the last word of the pasand the probe was presented only 500 msreaders pressed the line advance key. Becais not clear if readers wait until all processingcomplete before pressing the advance keyreduction in accessibility of the antecedenthe Distractor-Anaphor version in Experimenmay have simply slowed the resolution procrather than eliminated it. Thus, the antecedmay have begun to accrue activation as soothe anaphor was encountered but the resoluprocess may not have been complete whenprobe was presented. This may be becausephor resolution is not complete until a wordtwo after the anaphor when resolution is dcult (Ehrlich & Rayner, 1983) or becausetask demands associated with line-by-line reing and probe tasks. Concerns have recebeen raised about the interpretation of a lacfacilitation of antecedent probes; Gordonhis colleagues have argued that these findare the result of task-specific strategies (Gor& Chan, 1995; Gordon, Hendrick, & Fost2000).

ntd agn

othleteates a

isas

, inyis-2inb

ederith

ase

noen

na-n-ion.d totatedontheedit,

y theit isthe

edncequalthethenalad-ne;ceptost-

nt of the

re for that

them.

d neededup herfamily

W

A

T

P

C

607WHEN RESOLUTION FAILS

To address these concerns, in Experimereading time on a postanaphor line was usethe dependent measure instead of the recotion latencies used in Experiment 1. This bprovides additional time for readers to compthe processing of the anaphor and eliminconcerns about task demands. Passagepeared in two versions (see Table 5): The Dtractor-Anaphor version from Experiment 1 wused (now labeled the Distractor version)which the antecedent was backgrounded bhighly salient distractor, as well as the No Dtractor-Anaphor version from Experiment(now labeled the No Distractor version),which the distractor elaboration was replacedneutral filler material. Both versions includthe anaphor line. Given the results of Expments 2, the anaphor should be resolved inNo Distractor version. This was used as a bline.

In contrast to Experiment 1, there wasprobe task and two lines were added to the

TAB

Sample Tex

Introduction and antecedentWanda was throwing a surprise party for her best fr

company and some of his close friends wanted to

Distractor versionShe felt a little pressured because her daughte

occasion as well. She also still had to get detheir chocolatecakes.They were filled with a vWanda decided to have thecakedecorated with

No Distractor versionThe day had been busy because Wanda’s dau

to prepare for that as well. She had first stopdaughter’s gift. Her daughter had asked for atradition for quite some time to give the girls

rap-upWanda hoped John’s party would be fun. The gues

naphor lineWanda said to save room for thedessert.

arget lineThe tart looked perfect.

osttarget lineEveryone was so happy for John.

omprehension questionWas John recently fired?

4si-

sp-

-

a

y

-e-

d

of the passages. A target line followed the aphor line in both versions and explicitly metioned the antecedent in the subject positFor example, after reading that Wanda saisave room for dessert, the target sentence sthat “The tart looked perfect.” Reading timethis line was the dependent measure. Ifinclusion of the salient distractor simply slowthe resolution process, but did not eliminatereaders should have resolved the anaphor btime they processed the target line; althoughpossible that resolution is not completed onanaphor itself, it is unlikely that it is delaymore than a word or two beyond the senteboundary. Thus, reading times should be eon the target line in the two versions. Onother hand, if the anaphor is not resolved inDistractor version, even given the additiotime and the elimination of the probe task, reing times should be longer on the target litime should be needed to reinstate the con“tart.” Finally, each passage ended with a p

5

Experiment 4

d John. John had just been promoted to Vice Presidengratulate him. Wanda even made him atart.

raduation was the next day and she needed to prepations and stop at the bakery for acake.Her daughter lovedrich cream and had gooey dark chocolate frosting onite flowers and a white border.

r was graduating from high school tomorrow. She haat the dry cleaner’s and then at the jeweler’s to pick

arl necklace for her graduation present. It had been aearl necklace when they graduated from high school.

rrived right on time. As everyone sat down to eat,

LE

t—

ienco

r’s gcoraerywh

ghtepedpe

a p

ts a

the

heon

anom5%cepa

as-s anN uni torv s ae er,t phT thp nut ntA ithr pt ves n ot at at lerp ep

t in

s int robt

le 6.ers,nona-

rgetomvedas-99

thes,

by

neadctorwasthead

beenion,er-

toory.hortheersi-oftionideri-

line

608 LEVINE, GUZMAN, AND KLIN

target line that was neutral with respect toanaphor.

Method

Participants.Thirty-six undergraduates at tState University of New York at Binghamtparticipated as part of the requirement forintroductory psychology course. The data frthree participants who made more than 2comprehension question errors were replaTherefore data analyses were based on 33ticipants.

Materials. There were 18 experimental pages, with two versions of each: Distractoro Distractor. A sample passage can be fo

n Table 5. As in Experiment 1, the Distracersion contained an antecedent as well alaborated, highly typical distractor. Furth

he passages contained the categorical anawo sentences were added to the end ofassage. The target line, which was the pe

imate line, explicitly mentioned the antecedend the final, posttarget line was neutral w

espect to the anaphor and was included simo complete the passage. The No Distractorion was identical except that the descriptiohe distractor was replaced by neutral filler merial, which was the same number of lineshe distractor elaboration. Twenty-five filassages were interspersed throughout theerimental passages.Design.The design was the same as tha

Experiment 2.Procedure.The procedure was the same a

he previous experiments except that the pask was omitted.

TAB

Experiments 4 and 5: R

Version

Experiment 4

Anaphor line Target line

Distractor 1697 1468No Distractor 1718 1369

Difference 221 99*

* p , .05.

d.r-

dd

n

or.el-.

lyr-f-s

x-

e

Results and Discussion

Mean reading times are presented in TabThese were calculated after discarding outli6.7% of the data (Tukey, 1977). There werereading time differences on the pretarget, aphor line in the two versions;p 5 .7. In con-trast, the pattern of reading times on the taline provides support for the conclusions frExperiment 1; the anaphor was not resolwhen there was a salient distractor in the psage. Reading times on the target line werems longer in the Distractor version than inNo Distractor version, 1468 ms vs 1369 mrespectively. This difference was reliablesubjects,t 1(32) 5 4.11, SEM 5 24.11,p ,.001, and byitems t 2(17) 5 2.08, SEM 542.38, p 5 .03. Following the anaphor, aexplicit mention of the antecedent was rmore slowly when a same-category distrawas included in the passage than when itnot. If the anaphor had been resolved inDistractor version by the time participants rethe target line, the antecedent should haveas active as it was in the No Distractor versleading to similar reading times in the two vsions. However, time was clearly neededretrieve the antecedent concept from memGiven that readers did not resolve the anapby the time they reached the second word offollowing sentence, it is unlikely that they evwould have. In addition to ruling out the posbility that the reduction in the accessibilitythe antecedent simply delayed the resoluprocess, the results of this experiment provconverging support for the finding of Expe

6

ing Times (in Milliseconds)

Experiment 5

Anaphor line Target line Posttarget

1817 1900 19951788 1511 177029 389* 225*

LE

ead

ureriou00er-

wnreolu&

iersx-thencesce

thethebisnting

de-theontothmath

ili-tortheineed

enthes inedthe

hetheeatateus

abletoral ofas-

pas-lab-

forck-byac-g.,at is,nd

o n)a ialw erd e thea thet on-s derse is-t ualo n-t thed thet it isi

M

att m-t dit.T han2 re-p d on3

as-s tora undi iefi .g.,t hisw ofe or( torv ly

609WHEN RESOLUTION FAILS

ment 1 with a different dependent measWhereas probe tasks are susceptible to vastrategic processes (e.g., Gordon et al., 2Potts, Keenan, & Golding, 1988), in this expiment there were no secondary tasks.

In summary, previous studies have shothat text factors such as distance and the pence of distractors may make anaphor restion more difficult (e.g., Corbett, 1984; DuffyRayner, 1990) or perhaps even eliminatewhen presentation rate is not under readcontrol (O’Brien et al., 1997). The current eperiments add to this by demonstrating thatresolution of anaphors may be eliminatedtirely when the antecedent is not easily acsible and resolution is not critical for coheren

EXPERIMENT 5

In the current experiment we considerpossibility that readers actually did resolveanaphor in Experiments 1 and 4, but did soerroneously selecting the distractor. Thatreaders may have selected “cake” as the acedent of dessert rather than “tart.” Assumthat the accessibility of entities in memorypends on their level of activation as well asnumber of features that overlap with the ccepts in working memory, the salient distracshould have resonated more strongly withanaphor than the referent. Thus, readershave erroneously selected the distractor asreferent. This would explain the lack of factation of the antecedent probe in the DistracAnaphor version of Experiment 1 as well asslowdown in Experiment 4 on the target lthat explicitly mentioned the intended antecent.

To test the hypothesis that readers did idtify a unique referent, but did so by selectingdistractor, we used a similar manipulation aExperiment 4. A target line was again addafter the anaphor line, but it now instantiateddistractor. For example, after reading “. . . saveroom for the dessert,” the target line read, “Tcake looked perfect.” If readers resolvedanaphor incorrectly, the target line should rsmoothly. In contrast, if readers did not reinsthe distractor, this line should be anomaloleading to a slowdown in reading.

.s

0;

s--

t’

e--.

y,e-

-reye

-

-

-

d

,

Passages appeared in two versions (see T7), Distractor and No Distractor. The Distracversion was the same as that used in severthe previous experiments. In the sample psage, the tart was mentioned early in thesage and was followed by several lines of eoration about the cake that Wanda neededher daughter’s graduation party. This was bagrounded by a couple of lines and followedthe anaphor line. In contrast, in the No Distrtor version, the high typical distractor (e.cake) was now used as the antecedent. ThWanda was making acakefor John’s promotioparty instead of atart. “Cake” was mentione

nce (parallel to “tart” in the Distractor versiond was followed by neutral filler materhich detailed Wanda’s efforts to buy haughter a necklace for graduation. Becausntecedent was “cake” rather than “tart,”

arget line (“The cake looked perfect”) was cistent with the passage. Therefore, if rearroneously reinstated the distractor in the D

ractor version, reading times should be eqn the target line in the two versions. In co

rast, if readers did not mistakenly reinstateistractor, reading times should be longer on

arget line in the Distractor version becausenconsistent with the passage.

ethod

Participants. Thirty-three undergraduateshe State University of New York at Binghaon participated in exchange for course crehe data of one participant who made more t5% comprehension question errors werelaced. Therefore data analyses were base2 participants.Materials. There were 18 experimental p

ages, with two versions of each: No Distracnd Distractor. A sample passage can be fo

n Table 7. In the Distractor version, the brntroduction was followed by the referent (eart), which was mentioned only one time. Tas followed by an average of five lineslaboration of the high typicality distracte.g., cake). In contrast, in the No Distracersion, after a brief introduction, the high

troonhishin

edethaoryhed ais-uciontyth

-t 4

Thenot,ms

Nopec-ts,

. nc vs1 le,tt

ro-hatdis-ugh

nt of the

igh’s andrwhen

ay andbakeryy

Wnda

A

T

N

C

610 LEVINE, GUZMAN, AND KLIN

typical category member (e.g., cake) was induced as the referent. It was mentionedtime. In place of the distractor elaboration, twas followed by neutral filler material whicwas equated with the distractor elaborationnumber of lines.

Both versions included a line that containthe categorical anaphor (“. . . save room for thdessert”) and a penultimate, target linemade reference to the high typicality categexemplar (“The cake looked perfect”). In tDistractor version, this was the distractor ansuch constituted a contradiction; in the No Dtractor version this was the referent and as swas consistent with the passage. Both versended with a neutral posttarget line. Twenfive filler passages were interspersed amongexperimental passages.

Design and procedure.The design and procedure were identical to those in Experimen

TAB

Sample Tex

IntroductionWanda was throwing a surprise party for her best fr

company and some of his close friends wanted to

No Distractor versionWanda even made him acake.The day had been

school tomorrow. She had needed to preparethen at the jeweler’s to pick up her daughter’graduation present. It had been a family tradthey graduated from high school.

Distractor versionWanda even made him atart. She felt a little pres

she needed to prepare for that occasion as wfor a cake.Her daughter loved their chocolatecadark chocolate frosting on them. Wanda deciborder.

rap-upWanda hoped John’s party would be fun. The gues

naphor linesaid to save room for thedessert.

arget lineThe cakelooked perfect.

eutral posttarget lineEveryone was so happy for John.

omprehension questionWas John recently fired?

-e

t

s

hs

-e

.

Results and Discussion

Reading times are presented in Table 6.first thing to note is that reading times diddiffer on the anaphor line,p . .5. In contrastreading times on the target line were 389longer in the Distractor version than in theDistractor version, 1900 ms vs 1511 ms, restively. This difference was reliable by subject 1(31) 5 6.84,SEM5 56.74,p , .001, andby items,t 2(17) 5 4.17, SEM 5 82.08,p ,001. Theslowdown in the Distractor versioontinued onto the posttarget line, 1995 ms770 ms. This difference was also reliab

1(31) 5 4.25, SEM 5 53.06, p , .001;2(17) 5 3.80,SEM5 75.03,p 5 .001. The

large slowdown in the Distractor version pvides strong support for the conclusion treaders did not incorrectly instantiate thetractor when they read the anaphor. Altho

7

Experiment 5

d John. John had just been promoted to Vice Presidengratulate him.

sy because Wanda’s daughter was graduating from hr that as well. She had first stopped at the dry cleanerft. Her daughter had asked for a pearl necklace for hefor quite some time to give the girls a pearl necklace

ed because her daughter’s graduation was the next dShe also still had to get decorations and stop at the.They were filled with a very rich cream and had gooeto have thecakedecorated with white flowers and a white

rrived right on time. As everyone sat down to eat, Wa

LE

t—

ienco

bufo

s giition

surell.kesded

ts a

ionn tenthectos ot ihath

aviont isectookedt o, inmenenowintagri-

nolinromtheconrennono

oun.gssm

areatemoriona

ice-re

theon-y ofced-hoseum-gesec-To

la-ten-thatas-of

theas-therma-line,n-

86tely. Inthe

ions.par-lly,to

thetionro-ershosetingwasnot

washadfor

been

atm-dit.oreer-

611WHEN RESOLUTION FAILS

“cake” must have received some activatwhen readers processed the anaphor, givefinding in Experiments 1 and 4 that a salidistractor interfered with the accessibility ofreferent, readers clearly rejected the distraas the referent. Given the slow reading timethe target line, they were clearly aware thawas inconsistent with the passage. If theyselected the distractor as the referent ofanaphor, reading time on this line should hbeen the same as in the No Distractor vers

Finally, comparing Experiments 4 and 5, iinteresting to note the large difference in effsizes. In the current experiment, readersalmost 400 ms longer to read “The cake looperfect” when the cake was not the referenthe anaphor as when it was. In contrastExperiment 4 readers took only about 100longer to read “The tart looked delicious” whthe anaphor had not been resolved than whhad. In the current experiment the large sldown suggests that readers were unable tograte the target line with the rest of the passIn contrast, the smaller effect size in Expement 4 suggests that although “tart” wasactive when readers encountered the targetthey were able to retrieve the antecedent fmemory and integrate the target line intopassage. Thus, although the contextualstraints reduced the accessibility of the refeso that the signal from the anaphor wasstrong enough to retrieve it, the referent wascompletely inaccessible; when readers enctered an explicit mention of the antecedent (ethe tart), with an additional 100 ms of proceing time they were able to retrieve it fromemory.

EXPERIMENT 6

The results of the first five experimentsinconsistent with the view that anaphors initia search of memory that is carried out to copletion. Instead, it appears that when anaphinferences are not needed for comprehensreaders may not choose to invest additioresources to complete the inference.

In the final experiment, we examine a predtion implicit in this account: if readers had dvoted additional resources, they would have

het

rntdee.

tk

f

s

it-e-e.

te,

-ttt-

.,-

-icn,l

-

-

solved the anaphor. That is, as suggested inconclusions of Experiment 5, although the ctextual constraints reduced the accessibilitthe antecedent, they did not make the anteent impossible to retrieve. Rather, readers cnot to devote the additional resources. Presably this resulted from the fact that passawere coherent even without making the conntion between the anaphor and its referent.test this prediction, an instructional manipution was used to increase the amount of attion readers paid to the anaphor. The linecontained the anaphor was highlighted withterisks and participants were told that somethe comprehension questions that followedpassages would involve information in theterisked line. Across the experiment, 50% ofcomprehension questions addressed infotion that was presented on an asteriskedmaking it an efficient strategy to carefully ecode that information. However, only 9 ofcomprehension questions, or approxima10%, specifically asked about an anaphoraddition, there were numerous anaphors inpassages that were not queried in the questThus, although the instructions encouragedticipants to encode the anaphor line carefuthey did not go as far as asking readersresolve the anaphors.

If readers resolve the anaphors, despitesame set of text factors that led to a resolufailure in Experiments 1 and 4, this would pvide support for the conclusion that readtreated these inferences as optional and cnot to devote additional resources to reinstathe antecedents when their accessibilityreduced. In contrast, if the anaphors are stillresolved, this will suggest that the anaphoran inadequate retrieval cue; even if readersinvestigated additional resources to searchan appropriate antecedent, they might haveunsuccessful.

Method

Participants. Forty-eight undergraduatesthe State University of New York at Binghaton participated in exchange for course creThe data from 7 participants who made mthan 25% probe or comprehension question

yse

as-s anN d iT thep Dist 1,e lal aw agw en1 onr Twc pass ntsi re-h ero tfi add ; int ret amp des theA on,w th

n,”as

as-sionster-

n-twoage.

peri-hethe

500pill-sk.therote

ndor

ffero

eslated

nt of the

r that

nda

Wnda

612 LEVINE, GUZMAN, AND KLIN

rors were replaced. Therefore data analwere based on 41 participants.

Materials. There were 18 experimental pages, with two versions of each: Anaphoro Anaphor. A sample passage can be founable 8. The materials were identical toassages used in Distractor-Anaphor and

ractor-No Anaphor versions of Experimentxcept that asterisks were placed around the

ine of the passages. The probe word wasays the antecedent concept. The filler passere also identical to those used in Experimexcept that asterisks were placed around

andomly selected line in each passage.omprehension questions followed eachage. In contrast to the first five experime

nstead of having exclusively Yes–No compension questions, many of the questions wpen-ended. In the experimental passagesrst question was the target question andressed information from the asterisked line

he Anaphor version, it required readers torieve the antecedent. For example, in the sle passage, the target question was “Whatert did Wanda prepare for John?” innaphor version. In the No Anaphor versihere the last line of the passage stated

TAB

Sample Tex

Introduction and antecedentWanda was throwing a surprise party for her best fr

company and some of his close friends wanted to

Distractor elaborationShe felt a little pressured because her daughter’s g

occasion as well. She also still had to get decorachocolatecakes.They were filled with a very rich cdecided to have thecakedecorated with white flowe

rap-upWanda hoped John’s party would be fun. The gues

Anaphor ending**** said to save room for thedessert.****

No Anaphor ending**** congratulated John on the promotion. ****

ProbeTART

Comprehension questionWas John recently fired?

s

dn

-

stl-esteo-,

ehe-

--

s-

at

Wanda “congratulated John on the promotiothe target question was “What did Wanda doeveryone sat down to eat?” For the filler psages, the second of the two comprehenquestions addressed information from the aisked line.

Design and procedure.The design was idetical to Experiment 1 except that there werecomprehension questions after each passThe procedure was also the same as in Exment 1 except that the cue “XXX” was on tscreen for 750 ms before being replaced byprobe word. This was increased from thems used previously to avoid processing sover from the asterisked line to the probe taFinally, instead of pressing a key to answercomprehension questions, participants wtheir answers on a sheet of paper.

Results and Discussion

Error rates. Probe error rates were 13.6 a14.4% in the Anaphor and the No Anaphversions, respectively. These did not diacross versions,t’s , 1, and were similar tthose in Experiment 1.

Reading times.Last line mean reading timare presented in Table 2. These were calcu

8

Experiment 6

d John. John had just been promoted to Vice Presidengratulate him. Wanda even made him atart.

ation was the next day and she needed to prepare fos and stop at the bakery for acake.Her daughter loved theirm and had gooey dark chocolate frosting on them. Waand a white border.

rrived right on time. As everyone sat down to eat, Wa

LE

t—

ienco

radutionrears

ts a

%ing

hanadte

ionrge

ants.

A od imed ml iont cewSr

inuaerslineinul

eheierthepe. I1,na65

fer-

intheth

thahethhehi

es

thel, or

s totheer-

theavere-

avever,derson.

esssarywn.wendi-notble.of

rs:toand

, iftan-theced-otet thises inted

ierents

-upas

ono-s actmes ofthebe-

ngtheo-in

tage

613WHEN RESOLUTION FAILS

after eliminating outliers (Tukey, 1977), 9.6of the data. The first thing to note is that readtimes were approximately 300 ms slower tin Experiment 1, providing evidence that reers were spending additional time on the asisked lines, as they were instructed. In additreading times were quite variable on the taline; almost 10% of the data were classifiedoutliers. This was due to some participaspending as much as 4 or 5 s on thetarget line

lthough this variability makes it difficult traw strong conclusions from the reading tata, it is notable that readers spent 128

onger on the target line in the Anaphor vershan in the No Anaphor version. This differenas reliable by subjects,t 1(40) 5 2.48,EM 5 51.57, p 5 .017, but did notquite

each significance by items,t 2(17) 5 1.78,SEM 5 70.96,p , .09. This difference is incontrast with the results from Experiment 1which reading times were approximately eqin the two versions. Thus, not only did readspend additional time processing the targetas instructed, but this time was spent resolvthe anaphor; otherwise reading times shohave been about equal in the two versions.

Probe recognition times.Mean correct probresponse times are presented in Table 2. Tvalues were computed after discarding outl(Tukey, 1977), 3.8% of the data. Althoughpassages were exactly the same as in Exment 1, a different pattern of results emergedcontrast to the null effect in Experimentprobe recognition times were faster in the Aphor version than the No Anaphor version, 8ms vs 907 ms, respectively. This 42-ms difence was significant by subjects,t 1(40) 52.70,SEM5 15.53,p 5 .01, andmarginallysignificant by items,t 2(17) 5 1.75, SEM 523.32, p 5 .10.

These data, as well as the last line readtime data, show that the failure to resolveanaphor in Experiments 1 and 4 was due toantecedent being difficult to access ratherto it being completely inaccessible. With tidentical materials, participants reinstatedantecedent and resolved the anaphor wasked to pay attention to the anaphor line. Tprovides support for the conclusion that th

-r-,ts

s

l

,gd

ses

ri-n

-

g

en

ense

inferences were not needed for maintainingcoherence of the passage, but were optionaelaborative. Assuming that a reader’s goal ibuild a coherent mental representation ofinformation described by the text, if the cohence of the passage depended on identifyingreferent for an anaphor, readers should hdevoted additional resources. And given thesults of the current experiment, this would hbeen sufficient to resolve the anaphor. Howeinstead of devoting additional resources, reasimply left the anaphor unresolved and read

GENERAL DISCUSSION

Implicit in the study of anaphoric inferenchas been the assumption that they are neceinferences and are therefore reliably draHowever, in a series of six experimentsexamined the hypothesis that there are cotions under which anaphoric inferences aredrawn even when an antecedent is availaFurther, we hypothesized that the probabilityresolution should be a function of two factothe extent to which resolution is necessarycreate a coherent discourse representationthe accessibility of the antecedent. That isthere are conditions under which readers’ sdard of coherence is met without makingconnection between an anaphor and its anteent, it may be inefficient for readers to devresources reinstating the antecedent. To teshypothesis, we attempted to create passagwhich noun phrases could either be interprereferentially or without referring to an earlmentioned concept and in which antecedwere difficult to retrieve.

Placing these questions within a bottomtheory of discourse processing, suchKintsch’s (1988) Construction-Integratimodel or Myers and O’Brien’s (1998) resnance hypothesis, we assume that anaphorlike any other input to memory. Thus, the saset of factors that influence the effectivenesany memory retrieval cue should determineaccessibility of the antecedent: the matchtween the cue and target, the memory streof the cue, and so on. Further, given this thretical framework, after related conceptsmemory are activated, in a subsequent s

ateti-eror

ry.obeen

dedends

hot 2bethenen3,

lienanlie

er-aseeivayy

s anoofntl

e entecheou

is-thef th88ing2ioningghmeee

k-ad-de-

s tobe-hatn-ncereasient4)ory

esson.m-

eri-ullytyp-, asof

rowionin.os-na-tor.re-ingers

or.tionssertm-ion;avevenionent

tionld,as a

asethet is aders.

toit is

614 LEVINE, GUZMAN, AND KLIN

some of the reactivated concepts are integrwith the current text input. Although the acvation process is initiated automatically, readcan continue processing, resampling memwith the current contents of working memoWe assume that continued processing is mlikely to occur when a noun phrase cannoteasily integrated into the ongoing text represtation.

The six experiments reported here provisupport for our hypotheses. When readerscountered the anaphors in Experiments 1 anthe antecedent was not reactivated when alient distractor intervened between the anapand its antecedent. In contrast, in Experimenfacilitation was found for the antecedent proafter the anaphor was read despite the factthere was an extensive amount of text interving between the anaphor and the antecedFacilitation was also found in Experimentdespite the fact that there was a less sadistractor intervening between the anaphorthe antecedent. Thus, the presence of the sadistractor in Experiments 1 and 4 clearly intfered with resolution; when the anaphor wread, its signal may have been divided betwthe referent and distractor, reducing the acttion of the referent (Myers & O’Brien, 1998) bcreating a fan effect (e.g., Anderson, 1976; Mers et al., 1984; Reder & Ross, 1983). Aresult, the activation of the antecedent didreach a critical threshold. Or, the activationthe antecedent may not have been sufficiehigher than that of the distractor. As Greenal. (1992) suggested, failure to locate an acedent may occur because no entity matwell or because several entities match abequally well.

Although we might expect referential dtance to play a role in the accessibility ofantecedent, either because the strength omemory traces decay with time (Kintsch, 19or because of interference from the interventext (Myers & O’Brien, 1998), in Experimentdistance was not sufficient to prevent resolutHowever, with a greater amount of interventext or a stronger topic shift, the results mihave been different. In addition, the saamount of referential distance might have b

d

sy

st

-

-4,a-r,

at-t.

tdnt

n-

-

t

yt-st

e)

.

t

n

sufficient to disrupt resolution if the bacgrounding material were more complex, if reers were attending to some other importanttail of the passages, and so on. This pointwhat is undoubtedly a complex interactiontween text factors, making it unsurprising tO’Brien et al. (1997) found evidence for noresolution when there was substantial distabetween anaphors and their referents, whewe found that distance alone was not sufficto prevent resolution. Similarly, Corbett (198concluded that the inclusion of a same-categdistractor simply slowed the resolution procwhere we found that it prevented resolutiAlthough we examined the influence of a nuber of text factors in the current set of expments, additional research is needed to funderstand the influence of factors such asicality, referential distance, and elaborationwell as their interaction, on the reactivationbackgrounded information.

Although these experiments began to nardown the possibilities of what the resolutfailure entailed, several possibilities remaThe results of Experiment 5 ruled out the psibility that readers actually resolved the aphors by erroneously selecting the distracThus, although the distractor must haveceived at least temporary activation, interferwith the accessibility of the antecedent, readdid not ultimately integrate it with the anaphThis may have been the result of an integrastage, where “cake” was rejected as the deof the dinner party. Alternately, the cue to meory may have consisted of an entire propositrather than the cue being “dessert,” it may hbeen “the dessert at the dinner party.” Githis, although “cake” received some activatdue to the partial match, this was not sufficifor full retrieval.

One possibility is that because the activaof the referent did not reach a critical threshoreaders treated the anaphoric noun phrasenew entity. That is, although the noun phrwas used anaphorically by the character inpassage (i.e., Wanda knows that the dessertart), it was not treated as an anaphor by reaAlthough definite noun phrases often referentities that have already been established,

r too &ontheav

theific

asenrlapaveaoldan-ve

nteadwa

itrew

ndahisrasen

at-as-di

heo

labanac

l fowetheonrt oeedeeet. Ifoofavin

orythe

u-as-t wethe

ofrs inm-

ow-notrcesing

rentthatadus.

idee asts 1ershus,ese

xper-is

ves-oniga-ofe.

ht.

msnd

nts

dis-l

big-

ces:

ti-

615WHEN RESOLUTION FAILS

not uncommon for them to be used to refeentities that are new (Fraurud, 1990; PoesiVieira, 1998). Thus, given the lack of activatiof “tart,” readers may not have processed “dessert” as an anaphor at all. They may hunderstood “the dessert” to be attributive, rathan referential, referring to some non-specdessert.

Another possibility is that the referent wpartially activated when the anaphor wascountered, as a result of the conceptual ovebetween the anaphor and the referent. Thistial retrieval may have caused readers to ha“feeling of knowing” (e.g., Hart, 1965) orsense of familiarity. Although this subthreshactivation was not sufficient to retrieve thetecedent from long-term memory, it may habeen sufficient to inform readers that an acedent was present in the passage. Thus, remay have been aware that “the dessert”referring to a specific entity. But becausewould have required additional resources totrieve this entity, and because the passagecoherent without knowing what dessert Wawas serving, readers simply moved on. Tsuggests that readers may process noun phas anaphoric, or referential, even without idtifying a particular referent.

Finally, in this series of experiments wetempted to maximize the probability that psages would read smoothly even if readersnot connect the targeted anaphors with tintended antecedents. This was based onhypothesis that when an antecedent is avaiin memory readers should only fail to resolveanaphor when the antecedent is difficult tocess and when its retrieval is not essentiacomprehension. We selected anaphors thatconstrained by the context and that could eibe used referentially or to introduce a new ccept. For example, because “dessert” is paour dinner party schema, even if there had bno antecedent included in the passage, theinite noun phrase “the dessert” could have bused to refer to some nonspecific dessercontrast, if Wanda had said to leave room“the frozen drinks,” an entity that is not partour dinner party schema, readers might hdevoted additional resources to draw a bridg

er

-pr-a

-erss

-as

es-

dirurle

-rrer-fnf-nnr

eg

inference or spent more time searching memin an attempt to integrate this concept intopassage.

Although we did not systematically maniplate factors that should influence readers’sessment of coherence in the same way thamanipulated factors that should influenceaccessibility of the antecedent, the patternresults suggests that the unresolved anaphoExperiments 1 and 4 did not constitute a coprehension failure. Based on the lack of a sldown in reading on the anaphor line, it doesappear that readers devoted additional resouin an attempt to resolve the anaphors. Assumthat readers’ goals are to construct a coherepresentation of the text, it seems unlikelythey would have simply read on if they hfound the narratives incoherent or anomaloThe results from Experiment 6 also provsupport for this conclusion; antecedents werdifficult to access as they were in Experimenand 4, but with additional processing readsuccessfully reinstated the antecedents. Twe conclude that readers did not devote thresources when reading the passages in Eiments 1 and 4. However, this discussionspeculative, and in the same way that we intigated the influence of a variety of factorsthe accessibility of the antecedent, an investtion is needed of the influence of a varietyfactors on readers’ assessment of coherenc

REFERENCES

Anderson, J. R. (1976).Language, memory, and thougHillsdale, NJ: Erlbaum.

Battig, W. F., & Montague, W. E. (1969). Category norof verbal items in 56 categories: A replication aextension of the Connecticut category norms.Journalof Experimental Psychology,80, 1–46.

Clark, H. H., & Sengul, C. J. (1979). In search of referefor nouns and pronouns.Memory & Cognition, 7,35–41.

Corbett, A. T. (1984). Pronominal adjectives and theambiguation of anaphoric nouns.Journal of VerbaLearning and Verbal Behavior,23, 683–695.

Corbett, A. T., & Chang, F. R. (1983). Pronoun disamuation: Accessing potential antecedents.Memory &Cognition,11, 283–294.

Cotter, C. A. (1984). Inferring indirect objects in sentenSome implications for the semantics of verbs.Lan-guage and Speech,27, 24–45.

Dell, G. S., McKoon, G., & Ratcliff, R. (1983). The ac

singl

ndand

andentl

nou

ty ind

oricional

t efcaiser-

ref

r

G , an-

G ).n i

G an-

and

G un-ni-

H ing

H w?en

or,

H lti-

K rsedel.

K of

-

K sall

K om

nd

K .ces.

Kng-

M the-t.

M ouns of-

M ionana-r-

M d-

M is--

M .role

O d they:

O a--on,

O 0).l

O , K.licit

P iga--

P 8).ces:

y

616 LEVINE, GUZMAN, AND KLIN

vation of antecedent information during the procesof anaphoric reference in reading.Journal of VerbaLearning and Verbal Behavior,22, 121–132.

Duffy, S. A., & Rayner, K. (1990). Eye movements aanaphor resolution: Effects of antecedent typicalitydistance.Language and Speech,33, 103–119.

Ehrlich, K., & Rayner, K. (1983). Pronoun assignmentsemantic integration during reading: Eye movemand immediacy of processing.Journal of VerbaLearning and Verbal Behavior,22, 75–87.

Fraurud, K. (1990). Definiteness and the processing ofphrases in natural discourse.Journal of Semantics,7,395–433.

Fletcher, C. R. (1984). Markedness and topic continuidiscourse processing.Journal of Verbal Learning anVerbal Behavior,23, 487–493.

Garrod, S., & Sanford, A. (1977). Interpreting anaphrelations: The integration of semantic informatwhile reading.Journal of Verbal Learning and VerbBehavior,16, 77–90.

Garrod, S., & Sanford, A. J. (1983). Topic-dependenfects in language processing. In G. B. Flores d’Ar& R. Jarvella (Eds.),The process of language undstanding.Chichester: Wiley.

Gernsbacher, M. A. (1989). Mechanisms that improveerential access.Cognition,32, 99–156.

Gillund, G., & Shiffrin, R. M. (1984). A retrieval model foboth recognition and recall.Psychological Review,91,1–67.

ordon, P. C., & Chan, D. (1995). Pronouns, passivesdiscourse coherence.Journal of Memory and Language,34, 216–231.

ordon, P. C., Grosz, B. J., & Gilliom, L. A. (1993Pronouns, names, and the centering of attentiodiscourse.Cognitive Science,17, 311–347.

ordon, P. C., Hendrick, R., & Foster, K. C. (2000). Lguage comprehension and probe list memory.Journalof Experimental Psychology: Learning, Memory,Cognition,26, 766–775.

reene, S. B., McKoon, G., & Ratcliff, R. (1992). Pronoresolution and discourse models.Journal of Experimental Psychology: Learning, Memory, and Cogtion, 18, 266–283.

art, J. T. (1965). Memory and the feeling-of-knowexperience.Journal of Educational Psychology,56,208–216.

aviland, S. E., & Clark, H. H. (1974). What’s neAcquiring new information as a process in comprehsion.Journal of Verbal Learning and Verbal Behavi13, 512–521.

intzman, D. L. (1986). “Schema abstraction” in a muple-trace memory model.Psychological Review,93,411–428.

intsch, W. (1988). The role of knowledge in discoucomprehension: A construction-integration moPsychological Review,95, 163–182.

intsch, W., & van Dijk, T. A. (1978). Toward a model

s

n

-

-

d

n

-

text comprehension and production.Psychological Review,85, 363–394.

lin, C. M., & Myers, J. L. (1993). Reinstatement of cauinformation during reading.Journal of ExperimentaPsychology: Learning, Memory and Cognition,19,554–560.

lin, C. M. (1995). Causal inferences in reading: frimmediate activation to long-term memory.Journal ofExperimental Psychology: Learning, Memory aCognition,21, 1483–1494.

lin, C. M., Guzman, A. E., & Levine, W. H. (1999)Prevalence and persistence of predictive inferenJournal of Memory and Language,40, 593–604.

lin, C. M., Murray, J. D., Levine, W. H., & Guzma´n, A. E.(1999). Forward inferences: From activation to loterm memory.Discourse Processes,27, 241–260.

ason, R. A. (1997).The role of multiple antecedents intime course of anaphor resolution.Unpublished Master’s thesis, University of Massachusetts, Amhers

cKoon, G., Gerrig, R. J., & Greene, S. B. (1996). Pronresolution without pronouns: Some consequencememory-based text processing.Journal of Experimental Psychology: Learning, Memory, & Cognition,22,919–932.

cKoon, G., & Ratcliff, R. (1980). The comprehensprocesses and memory structures involved inphoric reference.Journal of Verbal Learning and Vebal Behavior,19, 668–682.

cKoon, G., & Ratcliff, R. (1992). Inference during reaing. Psychological Review,99, 440–466.

yers, J. L., & O’Brien, E. J. (1998). Accessing the dcourse representation during reading.Discourse Processes26, 131–157.

yers, J. L., O’Brien, E. J., Balota, D. A., & Toyofuku, M(1984). Memory search without interference: Theof integration.Cognitive Psychology,16, 217–242.

’Brien, E. J. (1987). Antecedent search processes anstructure of text.Journal of Experimental PsychologLearning, Memory, and Cognition,13, 278–290.

’Brien, E. J., Duffy, S. A., & Myers, J. L. (1986). Anphoric inference during reading.Journal of Experimental Psychology: Learning, Memory, & Cogniti12, 346–352.

’Brien, E. J., Plewes, P. S., & Albrecht, J. E. (199Antecedent retrieval process.Journal of ExperimentaPsychology: Learning, Memory and Cognition,16,241–249.

’Brien, E. J., Raney, G. E., Albrecht, J. E., & Rayner(1997). Processes involved in the resolution of expanaphors.Discourse Processes,23, 1–24.

oesio, M., & Vieira, R. (1998). A corpus-based investtion of definite description use.Computational Linguistics,24, 183–216.

otts, G. R., Keenan, J. M., & Golding, J. M. (198Assessing the occurrence of elaborative inferenLexical decision versus naming.Journal of Memorand Language,27, 399–415.

-

dgefanrn-

, &cat

S

S andnce

T ,

v , E.ce in. F.ce

v -

W of

((

617WHEN RESOLUTION FAILS

Ratcliff, R. (1978). A theory of memory retrieval.Psychological Review,85, 59–108.

Reder, L. M., & Ross, B. H. (1983). Integrated knowlein different tasks: The role of retrieval strategy oneffects.Journal of Experimental Psychology: Leaing, Memory, and Cognition,9, 55–72.

Rosch, E., Mervis, C. B., Gray, W. D., Johnsen, D. M.Boyes-Braem, P. (1976). Basic objects in naturalegories.Cognitive Psychology,9, 382–440.

anford, A. J., & Garrod, S. C. (1981).Understandingwritten language.New York: Wiley.

anford, A. J., & Garrod, S. C. (1989). What, when,how? Questions of immediacy in anaphoric refereresolution.Language and Cognitive Processes,4, 235–262.

-

ukey, J. W. (1977).Exploratory data analysis.ReadingMA: Addison–Wesley.

an den Broek, P. W., Risden, K., & Husebye-Hartman(1995). The role of readers’ standards for coherenthe generation of inferences during reading. In RLorch, Jr. & E. J. O’Brien (Eds.),Sources of coherenin reading (pp. 353–373). Mahwah, NJ: Erlbaum.

an Dijk, T. A., & Kintsch, W. (1983).Strategies of discourse comprehension.New York: Academic Press.

alker, F., & Yekovich, C. (1987). Activation and usescript-based antecedents in anaphoric reference.Jour-nal of Memory and Language,26, 673–691.

Received July 13, 1999)Revision received February 18, 2000)