Journal of Memory and Languagedoi:10.1006/jmla.2001.2815, available online at http://www.academicpress.com on

Anchoring Comprehension in Linguistic Precedents

Dale J. Barr

Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana–Champaign

and

Boaz Keysar

The University of Chicago

Past research has shown that when speakers refer to the same referent multiple times, they tend to standardizetheir descriptions by establishing linguistic precedents. In three experiments, we show that listeners reduce un-

ents in a prece-adhere toce that

to ben-ers use

In ix fi-

MHBiddMar2. CfedeGreof thmou

ABecUnivews

certainty in comprehension by taking advantage of these precedents. We tracked listeners’ eye movemreferential communication task and found that listeners identified referents more quickly when specificdents existed than when there were none. Furthermore, we found that listeners expected speakers to precedents even in contexts where it would lead to referential overspecification. Finally, we provide evidenthe benefits of linguistic precedents are independent of mutual knowledge—listeners were not more likelyefit from precedents when they were mutually known than when they were not. We conclude that listenprecedents simply because they are available, not because they are mutually known. © 2001 Academic Press

Key Words:psycholinguistics; mutual knowledge; comprehension; precedents; reference.

Graham Greene’s novelThe End of the Af- of a public figure. After the meal, Bendr

r

en

hehn

e-

ndn-e

eks,

at,o

yes,r-e

fair, the narrator Bendrix relates the aftermaof his tryst with Sarah, the wife of his friendHenry. Sarah had ended her affair with Bendfor mysterious reasons, which caused himsuspect that she had started seeing someelse. Bendrix’s jealousy compelled him to scretly hire a private detective to spy on her. Oday, Bendrix invites Henry to lunch to tell himabout some reports that the detective had pvided. Bendrix avoids the topic over lunch, chating instead about Henry’s work in the BritisMinistry of Pensions. Recently, Henry had beinvolved in a Royal Commission that the Britispress reported on daily, making him somethi

Experiments 1 and 2 were supported by a PHS Grant

1

e-r-e,

pler-e

inar-

of-

49685 to the University of Chicago. We thank Megainger, Johnny Chueh, Sue Paik, Havoc Pennington,

ina Peterson for their assistance with Experiments 1 hristine Lee performed an invaluable service as the crate for Experiment 3. We thank Anne Henly a

gory Murphy for valuable comments on an earlier dris article. We also thank Susan Brennan and two anos reviewers for criticism and suggestions.

ddress correspondence and reprint requests to Dale Bkman Institute for Advanced Science and Technoloersity of Illinois at Urbana–Champaign, 405 N. Math Ave., Urbana IL 61801. E-mail: dalebarr@uiuc.edu.

0749-596X/01 $35.00

th

ixtoone-e

ro-t-

n

g

nally discloses to Henry that he had hired a dtective to follow Sarah and tells him, “I thinkyou ought to read the reports.” Just then, a frieof Bendrix’s approaches to say hello, and Bedrix introduces the newcomer to Henry. Thnewcomer tells Henry, “I’ve been following threports every day.” Henry is confused and as“What reports?” Bendrix’s friend replies, “TheRoyal Commission.” The narrator observes th“for once, Henry’s work had not come first tmind when that word was uttered.”

When the newcomer said “the reports,” whdid Henry think first of the reports on his wifinstead of the press reports? In linguistic termHenry’s behavior was egocentric—he intepreted “the reports” according to Bendrix’s usof the term, which caused him to think of the rports on his wife. He ignored the fact that refeential meaning depends on mutual knowledgor the set of beliefs and assumptions that peoshare and know that they share (Clark & Mashall, 1981). Henry had little reason to assumthat the newcomer would be using the termthe same way that Bendrix had because this elier linguistic experience was not, in the termsClark and Marshall, “co-present” between him

R29n

andandon-ndaftny-

arr,gy,-

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

i

i

e

i

a

r-geedualndforthateynndoes

onandthatitp.wl-nersenhar-imalely

arting

ngro-ar-ro-

t anteetn-elringst-outrate

r-as-ffi-e,ualce.r-

ti-

2 BARR AND

self and the newcomer. Instead, Henry shouhave realized that the only reports that weknown to the two of them were the press repoon the Royal Commission.

We propose that Henry’s misunderstandireflects a normal feature of language proceing: the egocentric anchoring of comprehensin linguistic precedents. When Bendrix uses tword “reports” to refer to the documents prpared by the detective agency, he establisheprecedent whereby the word “reports” desnates a particular set of documents. Over span of the conversation, the precedent servea linguistic index to a representation of the rerent in memory.

While precedents might occasionally caumisunderstandings like Henry’s, we propothat they generally enhance comprehension cause they reduce variability in the content athe meaning of speech. When specific predents exist, listeners can recognize words aidentify referents more swiftly than when theare absent. Although precedents are establisby specific speakers for specific purposes,suggest that, like Henry, listeners use precedebecause they are egocentrically available,because they are mutually known.

Three experiments provide support for theproposals. We used eyetracking techniques “communication game” in which listeners intepreted speakers’ referential expressions. Expment 1 establishes the basic benefit of predents to comprehension and examines interplay of precedents with preexisting, coventional names. Experiments 2 and 3 twhether the precedent must be mutually knoto obtain this benefit.

The concept of mutual knowledge, or common ground, has a long and controversial htory, from early debate regarding its plausibilias a psychological mechanism (Clark, 198Clark & Carlson, 1982; Clark & Marshall, 1981Sperber & Wilson, 1982) to its later widespreacceptance as an essential element in a gentheory of language use. According to the prevaing view, mutual knowledge is a kind of “modeof the interlocutor” that language users derithrough the application of a set of copresen

heuristics, which they use when they process

KEYSAR

ldrerts

ngss-onhee-s ag-thes asf-

sesebe-ndce-ndyhedwents

not

sen ar-eri-ce-then-estwn

-is-ty2;;deralil-l

vece

terances (Clark & Carlson, 1981; Clark & Mashall, 1981). From the standpoint of languaproduction, Clark and Murphy (1982) suggestthat speakers tailor utterances to their mutknowledge with specific addressees. Clark aCarlson (1981) presented a similar proposal language comprehension when they claimed listeners should restrict the information thconsider to mutual knowledge, or commoground: “ . . . when a listener tries to understawhat a speaker means, the process he gthrough can limit memory access to informatithat is common ground between the speaker his addressees” (p. 328). They suggested comprehension process will be optimal, “if limits its access to that common ground” (328). Both proposals characterize mutual knoedge as something that speakers and listewill spontaneously and routinely consider whthey process utterances. Furthermore, they cacterize language processing systems as optto the extent that these systems rely exclusivon mutually known information.

The Perspective Adjustment model of Keysand colleagues opposes a strongly contrasconception of optimality to the normative viewof Clark and colleagues (Keysar & Barr, ipress; Keysar, Barr, & Horton, 1998). Accordinto this conception, the design of a language pcessing system is “optimal” not because it guantees mutual understanding, but because it pvides adequate real-time understanding aminimal cognitive cost. Perspective Adjustmeposits that language processing systems mthis standard by employing a “egocentric achoring and adjustment heuristic.” The modassumes a rapid, automatic egocentric anchoprocess coupled with a slower, optional adjument process that computes information abperspective. These processes need not opestrictly in serial fashion, but can function in paallel or in cascade. Perspective Adjustmentsumes that the egocentric heuristic often is sucient for mutual understanding. Henclanguage users need not routinely check mutknowledge every time they process an utteranThis view predicts that the kinds of errors a peson will make in conversation will be systema

ut-cally biased toward that person’s own knowl-

O

ete

rd

th&i

we

t

s

udrr

6n

ho

le

aeuredc

e the

n

fi

arethekergedi-th

ac-ofhenthe

just Wevers Ine-entre-ves-ntsInf aeentg

re-ts”n-ly,

ailthethe

aveageect Wene-er-pe-reb-

lishes this basic benefit of linguistic precedents

ind

ANCHORING C

edge. By anchoring egocentrically and thusing mutual knowledge as needed to adjusthe interlocutor’s perspective, language ustrade-off failsafe understanding for cognitive eficiency (Keysar & Barr, in press; Keysar, Ba& Horton, 1998). In recent work, Keysar ancolleagues have found evidence supportingmodel for both language production (HortonKeysar, 1996) and language comprehens(Keysar, Barr, Balin, & Brauner, 2000; KeysaBarr, Balin, & Paek, 1998).

Our interest in this article is to test these tviews of the role of mutual knowledge with rspect to linguistic precedents. We ask, Whpeople understand referential expressions,they interpret them egocentrically or against background of mutual knowledge?

One good reason for listeners to use linguiprecedents is the fact that speakersuse them.When speakers refer to the same referent mple times in the same conversation, they tenstandardize their descriptions. Later desctions tend to be shorter and less lexically divethan early ones (Krauss & Weinheimer, 191966). When the members of a conversatiodyad take turns describing a referent, they coto describe the referent in the same way. It been proposed that descriptions can becstandardized through common ground (ClarkWilkes-Gibbs, 1986) or through language useattempts to establish consistency between guage production and comprehension syst(Garrod & Anderson, 1987).

In a recent proposal, Brennan and Cl(1996) argued that interlocutors coordinate rerence by forming partner-specific “conceptpacts,” or agreements concerning how a refeis to be conceptualized. These conceptual-lagreements affect how referents are encolinguistically. For instance, dyads formed pato refer to a particular shoe as the loaferwhen itwas presented in the context of another shosneaker). Later, speakers continued to call itloafer, even without the constraining contextthe sneaker (i.e., when the basic-level te“shoe” would have been sufficient). They fouthat when speakers switched to a new addresthey eventually stopped using the overspeci

term they had established with the old partn

MPREHENSION 3

ntors

f-r,

is

onr,

o-endo

he

tic

lti- toip-se4,al

measme &rs’an-ms

rkf-alentveledts

(a

They claimed that such conceptual pacts “partner specific” in the sense that they are end product of interaction between the speaand listener: “the references in our task emernot from solitary choices on the part of the drector, but from an interactive process by bodirector and matcher” (p. 1491).

For our current purposes, the issue of intertivity can be viewed as orthogonal to that precedent use. People can use precedents wspeaking or understanding either because precedents were interactively established or because prior use has made them available.seek a general theory of precedent use that coboth interactive and noninteractive situations.this article, we use the term “linguistic precdents” to broadly characterize the word-refermappings that listeners establish while comphending discourse. The precedents that we intigate are established interactively in Experime1 and 2, but noninteractively in Experiment 3. addition, although a speaker’s initial naming oreferent involves classification (Brown, 1958), wremain agnostic regarding whether the precedthat is the result of this initial naming is anythinother than lexical. For instance, if Bendrix had ferred to the reports using the word “documeninstead of “reports,” it seems that Henry’s misuderstanding would have been much less likeeven though “documents” and “reports” entquite similar conceptualizations. The issue of level at which precedents persist is beyond scope of our experiments.

Precedents such as conceptual pacts hbeen examined only as they pertain to languproduction. Our question is, Do listeners expspeakers to adhere to linguistic precedents?propose that precedents could potentially befit referential understanding by reducing unctainty in speakers’ intended meanings, escially when preexisting, conventional names anot readily available. The first experiment esta

ofrmdsee,ed

to comprehension.

EXPERIMENT 1

We used a referential communication task which one participant (the “director”) instructe

er.a second (the “addressee”) to rearrange objects

D

td

W

a

n

o

u

e

e

- in-t-

s-yn-e

rekeln

to-t,-

eoe-rtscn-

4 BARR AN

in a grid (see Fig. 1) to match a picture. Wtracked the position of addressees’ eyes as followed directors’ instructions (EberharSpivey-Knowlton, Sedivy, & Tanenhaus, 1995We compared the length of time it took themfind previously mentioned referents versus rerents that were previously unmentioned. expected that addressees would take advanof linguistic precedents to facilitate identifiction of referents.

As an example, consider an exchange abthe set of objects located in the grid in Fig.(For clarity of exposition, the figure contaifewer objects than a typical item in the expement.) The set includes a folded piece of pain the shape of an upside-down V (top right cner). When a speaker first refers to this objecthe “tent,” the addressee would need to estabthat the folded paper is the object that comost plausibly be conceptualized as a “tent.”accepting the label, then, the addressee eslishes a precedent by which this object is labe“tent.” In future instances when the speaker uthe label “tent,” the addressee could take advtage of the precedent to quickly identify the rerent.

Yet this benefit might be due to factors oththan the establishment of a linguistic preced

n

fr-he-li--

t amte16

rid

gh

ment 1.

KEYSAR

ehey,).toef-

etage-

out1.sri-perr-

t aslishld

Bytab-ledsesan-f-

ernt.d-

ing previously mentioned objects simply because they learned the location of the objectsthe grid. To control for this, we also manipulated whether referents had or lacked preexising conventional names (“conventional” versu“unconventional” referents). Conventional referents were familiar objects that have widelknown conventional names, like the apple iFig. 1. We would expect near uniform agreement among speakers on how they would namthese referents. Unconventional referents weobjects that lacked these common names, lithe folded paper in Fig. 1. Unlike conventionareferents, one would expect wide variation ihow speakers would refer to them.

Objects such as the apple are less likely benefit from a specific naming precedent compared to unconventional referents. In contrasknowledge of a referent’s location should contribute equally to identification speed for bothkinds of referents. Therefore, if the benefit wfind for unconventional referents is due tknowledge of location, then we should find thsame effect for conventional referents. Any additional benefit for unconventional objects, oveand above the effect for conventional objecwould constitute a true benefit due to linguistiprecedents. Our first experiment seeks to cofirm the existence of this differential benefit.

Method

Participants. Twenty college students (12males and 8 females) from the University oChicago, who were all native speakers of Ameican English, participated as addressees in texperiment for payment. Nine additional participants did not provide useable data due to cabration problems or problems with the recording equipment.

Apparatus. The director and addressee sat asmall table, facing one another. Between thewe placed a vertical grid made of opaque whifiberglass. The grid was composed of a set of boxes arranged in a 4 3 4 pattern (see Fig. 1).The measurements for each square in the gwere 12.5 3 12.5 3 12.5 cm. The squares wereuncovered, so it was possible to see throu

each square to the other side. Thus, objects ini-

Specifically, addressees might be faster in fi

FIG. 1. Example of an Experimental Item in Exper

the grid were mutually visible to both people.

O

iide

mert

ao

ryda d

r t

e oi

tt te

di

eict

eyonts inuldbe not

toroaltor the.c-di-n-ati-

heedhewerethe-lly

ngar-e-chnotr-eded

thatevedtri-ridbex-or’sseets.t-di-ci- to

ANCHORING C

We used an Applied Science Laborator(ASL) Series 4000 head-mounted eyetracksystem to monitor the position of the adressee’s left eye. The addressee wore a hband on which an eye camera and a magnhead tracker were mounted. This setup perted addressees to move their heads about frwhile performing the task. The eye camera pvided data concerning the position of the paripant’s eye relative to the head, while the hetracker corrected for movements of the heTogether, these two values determined the ption of the participant’s gaze.

A free-standing video camera filmed the gfrom the addressee’s point of view. The etracker superimposed a crosshair on the viimage corresponding to the position of the dressee’s gaze at a temporal resolution of 30or approximately 1 sample every 33 ms. Adtionally, the real-valued coordinates of the adressee’s gaze relative to the grid were stodigitally by a PC running ASL E4000 softwaat a rate of 60 Hz. Two microphones placedthe table recorded the conversation onto videotape.

Procedure. To ensure that all addresseheard the same critical utterances during theperiment, the director was a trained (male) cfederate who produced scripted utterances seemingly spontaneous manner. To enhancerealism of the experimental situation, we usseveral credibility cues throughout the expement. The confederate arrived 5 min late to experiment and pretended never to have meexperimenter. Later, during a set of practice als (described below), the confederate madfew errors that a naïve subject might have ma

Once the confederate arrived, the expementer told the pair that the experiment wouinvolve two different roles, that of director anaddressee. We then staged a “random” assment of roles. The experimenter asked the pto choose from among two slips of paper thwere facing down that were supposedly markwith the two different roles. In reality, both wermarked “Addressee.” In this way, the real partipant always drew the role of addressee, andconfederate simply announced that he had

ceived the role of director.

MPREHENSION 5

esng-ad-

eticit-elyo-ic-add.si-

ide-eod-Hz,i-d-redeonhe

sex-n-

n a theedri-hetheri- a

de.ri-ld

gn-airated

-here-

The experimenter then told the pair that thwould be playing a series of communicatigames where the goal was to rearrange objeca grid to match a goal state. The director wobe given a photo showing what needed to done to reach this state. The addressee wouldbe allowed to view the photo and the direccould not move the objects. To reach the gstate, they had to work together, with the direcinstructing the addressee to move objects andaddressee following the director’s instructions

We designed the next portion of the instrutions to lead the participant to believe that therector would be producing the instructions spotaneously. The experimenter explained thatthe beginning of each trial, he would give the drector a card containing a photograph of tgrid. The cover story was that the card containa photograph of the grid in its initial state, witarrows indicating how objects were to bmoved. Therefore, the director would only knowhich objects needed to be moved and whthey should go, but would have to come up wihis own way of describing the objects. Unbknownst to the addressee, the director actuahad scripted instructions printed on the cards.

We pretrained the confederate to avoid usieye gaze to communicate the position of the tget. The confederate looked back and forth btween the card and the grid while preparing eautterance, making sure that the target wasthe last object that he looked at. While deliveing the critical utterance, he kept his eyes fixaton the card. After the utterance, he was allowto look up at the addressee.

We used three practice trials to make sure the addressee understood the task and belithe cover story. The director’s cards for these als contained actual photographs of the gwith arrows showing how objects were to moved. At the end of the first two trials, the eperimenter showed the addressee the directcard with the arrows so that he or she could that there were no written labels for the objecAfter completing two practice trials, the parners switched roles. Playing the role of the rector made it extremely clear to the partipant/addressee that the director would have

come up with labels on his own.

6 BARR AND KEYSAR

yove dthe

nm

t thehthut ri

sttt

bet des

terentes-)

-nd 22 ob- to to_.”setmeson-tedentsent.we

usedthede-the weificmes

are

The experimenter then introduced the etracking equipment and explained its operatiTo avoid drawing undue attention to eye moments, the pair was told that the eye camrecorded features of the eye “such as pupilameter.” Finally, the experimenter mounted eyetracking equipment on the addressee’s hand performed the calibration procedure.

Next, the experimenter gave the participatwo final instructions. First, he instructed thenot to talk about or touch the objects, excepthe task required. This was mainly to preventaddressee from spontaneously naming obj(e.g., “look at this cute toy monkey!”) before tdirector could refer to them. Second, he told confederate to say “Ready” before each instrtion. This was a cue for addressees to look acenter position on the grid, so that for every ttheir gaze would start in the center.

After completing all of the experimental trialthe experimenter gave both participants a wriquestionnaire with general questions about experiment. It ended with a question that prowhether the addressee suspected that the dirwas a confederate. The question stated thasome of the pairs in the experiment we usetrained confederate and offered financial inctive for the participants to accurately gue

whether their partners were confederates. Aft

e-n.e-rai-ead

ts

ase

ctseec-theal

,enheedctorfor an-s

completing the experiment, the experimenfully debriefed participants about the experimand allowed the participants to ask any qutions. Of the 20 participants, only 3 (15%guessed that the director was a confederate.

Materials and design. We performed an informal norming study to select objects with awithout conventional names. We presentednative English speakers with 50 objects, oneject at a time. Each participant was askedname the objects as they would if they hadcommand someone to “Pick up the ______We drew 12 conventional referents from the of objects that generated agreement on nafor at least 70% of participants, and 12 uncventional referents from those that generaagreement for less than 30%. These referwere used as target objects in the experimFor conventional targets, the names that chose were those that respondents had most frequently. For unconventional objects,respondents produced highly idiosyncratic scriptions. Therefore, instead of choosing most frequent name, we chose names thatthought were not easily predictable but specenough to uniquely identify the object froamong the set of alternatives in the grid. Namand a short description for all 24 targets

erlisted in Table 1.

ber

s

e.

TABLE 1

Conventional and Unconventional Referents, Experiment 1

Conventional objects Unconventional objects

Basket: A small wicker basket. Link: A metal link open at one end.Crayon: A Crayola crayon. Belt: A pink velcro strap or belt.Straw: A drinking straw. Foam: An odd-shaped piece of green foam.Eraser: A white pencil eraser. Probes: A pair of multimeter probes, with wires coiled and rub

banded.Candle: A small green candle. Adapter: A “Y” adapter which allows two sets of headphones to

be plugged into one jack.Pen: An ordinary ball-point pen. Wires: A hanger for collectible plates, composed of two spring

and two bent pieces of wire.Pencil: An everyday pencil. Hook: A metal coat hook typically found mounted on the back

of doors.Soap: A bar of soap, still in its Razor: A small retractable knife.

wrapper.Chapstick: A tube of Chapstick. Mustard: A small packet of mustard.Sunglasses: A pair of men’s Wrench: A piece of a door latch that looks “wrench-like”.

sunglasses.Matches: A bundle of wooden Nail: A small eye-screw.

matches.Battery: An AA battery. Tent: A piece of paper tri-folded to make an inverted “V” shap

ANCHORING COMPREHENSION 7

o

cees

o

ie

ooocoe

t

-

eo

eb

i innce

o

n-e-

-

achtsastialg.,s

thera-e’s

h-d-ig-thetify fix-

d-or

ce,hisouldn-te

head-ook to re-

d-se-

thechhisheions

tifyansnder

words, the existence of a precedent shortened

1 The criterion for a “fixation” was that the eye must re-main within a particular square of the grid for 100 consecu-

We embedded these target objects amfiller objects in the grid to create items for thexperiment. Each experimental item consisteda preparatory instruction and a critical instrution. The preparatory instruction always prceded the critical instruction, and eithgrounded the target object by mentioning it alandmark or left it ungrounded. In Fig. 1, the taget object is the folded paper (“tent”) in the tright-hand corner. For this grid, the preparatoinstruction in the grounded condition might b“the toy truck goes below the tent.” To attamaximal consistency across conditions, whthe preparatory instruction left the target objeungrounded, it required the addressee to perfthe same action as in the grounded conditithough without mentioning the target object. FFig. 1, in the ungrounded condition this instrution would be “the toy truck goes above the btle.” Thus, in both conditions, the instruction rquires the addressee to move a filler object (ethe toy truck) to the same square, though the get object is grounded in only one condition.

After the preparatory instruction came thcritical instruction, occasionally with a filler instruction intervening. The critical instruction foeach item was always the same. For the gridFig. 1, the critical utterance was, “the tent gobelow the apple.” We compared how long it toaddressees to identify the tent in the groundand ungrounded conditions.

Between trials, the experimenter would rplace objects from the previous trial with ojects for the next trial. To minimize setup timbetween trials, each grid contained two expemental items. The two items within a single gralways appeared in opposite conditions. Forstance, if one item was unconventional agrounded, the other item would be conventioand ungrounded. The preparatory and critiinstructions for the two items were interleavin each grid, and some grids had addition“filler” instructions to obscure the purpose the experiment.

The design of the experiment was a (Conventionality: Conventional vs Unconvetional) 3 2 (Mention: Precedent vs No Precdent) within-participant design.

Coding and analysis. An undergraduate re

search assistant, who was blind to the hypoth

nge of--r ar-pryenn

ctrmn,r-t--.g.,ar-

e

r insked

--eri-dn-dalaldalf

2

ses of the study, coded the videotapes. For ecritical utterance, the coder located two pointhat defined response time. The first point wthe onset of the target word, defined as the inisyllable of the name of the target object (e.the first syllable of “tent”). The last point wathe “decision point” or “final fixation,” definedas the point at which the addressee identifiedtarget object as the referent. This was opetionalized as the beginning of the addresselast fixation1 on the target object before toucing it. We also extracted the latency of adressees’ first fixation on the target from the dital data. In some cases, upon locating referent addressees would immediately idenand move the target. In these cases, the sameation was both first and final.

Additionally, the coder noted whether the adressee asked the director for clarification confirmation about the referent. For instanthe addressee might ask, “is this the tent?” Twas noted because such overt exchanges wpossibly inflate the response time in the ugrounded condition and thereby overcontributo the effect.

Results and Discussion

Response times for four trials (0.8% of tdata) could not be computed because the dressees used peripheral vision and did not lat the target square or the crosshair failedenter the square due to poor calibration. Wemoved these data points from the analysis. Aditionally, to reduce any inflation in respontime due to clarification or confirmation requests, we truncated response times, usingtop of the distribution of items where no suexchanges occurred as the cutoff point. Ttruncation procedure affected only 2.3% of tdata. Complete means and standard deviatfor the measures are provided in Table 2.

In general, addressees took longer to idenreferents when no precedent existed (the mewere 3018 ms in the No Precedent condition a1710 ms in the Precedent condition). In oth

e-tive ms, or three video frames.

Wa

e

foh

l

-

e

a5

h

e

euenem

f-

erisl.t?d-lly

rn-

-

x-l

stu-

rs

offrly

e

theiseo-b-

ndld

e

nt

4)Unconventional 1701 (1117) 1246 (837) 3950 (2921) 1695 (1093)

referential search by an average of 1308 ms.submitted the data to an analysis of variance report subject and item analyses as F1 and F2,respectively. As predicted, we found a main fect of mention,F1(1,19) 5 45.10,p , .001,MSE 5 779123; F2(1,11) 5 55.17, MSE 5368153,p , .001.

More importantly, this benefit due to an exising precedent was much more pronounced unconventional (2255 ms difference) than fconventional objects (376 ms difference). Tinteraction was significant,F1(1,19) 5 59.30,p , .001, MSE 5 303140; F2(1,22) 5 17.97,MSE5 587876,p , .01, as were the individuaprecedent effects for unconventional referent(19) 5 7.88,p , .001, and conventional referents,t(19) 5 2.39,p , .05.

Analysis of the first fixations reveals the sambasic pattern. We lost the digital data for thrparticipants, so they were excluded from thanalysis. Addressees took longer to first fixon referents when there was no precedent (1versus 1289 ms), yielding a benefit of 288 mThe main effect of mention was significanF1(1,16) 5 5.26, p , .05; MSE 5 852;F2(1,22) 5 14.42,MSE5 304,p , .01. The in-teraction, however, was only marginal but in tpredicted direction,F1(1,16) 5 2.70,p , .13,MSE5 413; F2(1,22) 5 2.49,MSE5 758,p ,.13.

The finding that precedents provide a largbenefit for unconventional referents excludthe possibility that the effect was due to knowedge of location of the referent. The larger befit for unconventional referents was mainly dto a long search time when they were unmtioned (3950 ms on average). Unconventioand conventional referents that had been mtioned were statistically indistinguishable fro

one another [mean RTs 1695 and 1725 ms,

end

f-

t-orre

ts,

ee

iste77s.t,

e

ers

l-n-en-aln-

spectively,t(19) 5 0.29,ns]. In effect, the lin-guistic precedent made the unconventional reerents look conventional.

In sum, linguistic precedents produce greatbenefits for comprehension when a referent unconventional than when it is conventionaWhat mechanism is responsible for this benefiOne possibility, consistent with the standartheory, is that linguistic precedents benefit comprehension because they are part of a mutuaknown background of information. Anothepossibility is that they benefit comprehensioindependently of mutual knowledge, simply because they are available to the addressee.

The collaborative model of Clark and colleagues (Clark & Wilkes-Gibbs, 1986; Clark &Brennan, 1991) implies that listeners should ehibit more certainty when they interpret a labefor which a mutually known precedent existthan when the precedent exists but is not mually known. This implication stems from twosources.

First, it is claimed that speakers and listenecollaboratively “ground” the meanings of lin-guistic expressions. One of the consequencesgrounding is the creation of the mutual beliethat some term or utterance has been propeunderstood (Clark & Wilkes-Gibbs, 1986; Clark& Brennan, 1991). It follows that listeners aronly truly justified in applying a linguisticprecedent to understand subsequent uses ofterm when this mutual belief exists. When thbelief is lacking, such as when they interpret thutterances of a new speaker who was not cpresent when the original precedent was estalished, they should seek to once again grouthe term because the meaning of the term woube less certain.

The second way in which the collaborativ

8 BARR AND KEYSAR

TABLE 2

Mean Fixation Latencies (and Standard Deviations) by Mention and Conventionality, Experiment 1

First fixation Final fixation

No Precedent Precedent No Precedent Precede

Conventional 1453 (940) 1331 (800) 2101 (1655) 1725 (107

re-model implies that grounding is most effective

O

u

oyoh

e

h

e-

tollyheectityhes-

hisc-alonna,It

ex-is,ses

ANCHORING C

in relation to mutual knowledge originates claims concerning the nature of the comprehsion system itself. Clark and Carlson (198suggest that when addressees comprehend ances, they should restrict the information thconsider to mutual knowledge, or commground. It follows that the comprehension stem should ignore any precedents that exist side of this shared body of knowledge. Twould prevent it from making otherwise sytematic errors, such as when Henry interpre“I’ve been reading the reports” as referring the reports about his wife.

Unlike the common ground model, Persptive Adjustment predicts that listeners use predents simply because prior use has made tegocentrically available. If this is the case, thlisteners should be no less likely to use a predent when it is not part of common ground th

when it is. Experiments 2 and 3 use multip

e

t

fitu-as

es-beas

. In-in-ameer

easteenleor,

gnt.

ver-uc-e

ad-ker.italle

speakers to test these two models.

EXPERIMENT 2

In this experiment, we used the same taskin Experiment 1, except that addressees listeto instructions from not one, but two directorslive or “present” confederate director as in Eperiment 1 and a “recorded” director, whose structions the addressee heard through a hphone. The fact that the present director conot hear the recorded director’s instructions sured that the addressee would have to estabseparate precedents with each director. In experiment, a description for a referent wmentioned first by one of these two directoand then referred to in a target instruction by present director. In other words, while there walways a precedent for the second mention,precedent had been established by either present or recorded director. The question wWould comprehension benefit more from precedent when it was used by the same direwho had established it than when it was useda director who had not? Alternatively, would thbenefit be independent of the identity of the rector who established the precedent?

We designed the experiment to distinguiamong three specific hypotheses: (1) strong

partner specificity, (2) weak partner specificity,

MPREHENSION 9

inen-1)tter-eyns-ut-is

s-tedto

c-ce-em

ence-anle

asned: ax-in-ead-uldn-lishtheasrs

theashattheas:a

ctor bye

di-

sh

and (3) partner independence. The strong ver-sion of the partner-specificity hypothesis prdicts that the existence of a precedent will onlyproduce a benefit when the director that refersa test object is the same one who originamentioned it. Because in our experiment tpresent director always refers to the test objthe second time, the strong partner-specifichypothesis predicts a benefit only when tpresent director was the one who originally etablished the precedent. A weak version of thypothesis predicts that the identity of the diretor will matter somewhat. The idea that mutuknowledge can be a probabilistic constraint comprehension was advanced by HanTrueswell, Tanenhaus, and Novick (1997). predicts the largest benefit when a precedentists and is part of mutual knowledge; that when the present director is the one who uthe label first. There should still be a benewhen a precedent exists, but is not part of mtual knowledge; that is, when the referent hbeen mentioned previously, but not by the prent speaker. However, this benefit should smaller than in the case where the target wmentioned both times by the same speakercontrast to these two hypotheses, the partnerdependence hypothesis predicts the exact sbenefit irrespective of the identity of speakwho established the precedent.

Method

Participants. Thirty-six college students fromthe University of Chicago, who were all nativspeakers of American English, participated addressees in exchange for payment. Sevenadditional participants did not provide useabdata due to poor calibration, experimenter error problems with the audio equipment.

Apparatus. We used the same eyetrackinsetup as described in the previous experimeIn addition, addressees wore a headphone coing one ear, where they heard spoken instrtions from a “secret” prerecorded director. Thinstructions were recorded by a male undergruate student who was a native English speaThese instructions had been stored as digaudio files on the hard drive of a PC compatib

computer equipped with a sound card. The pres-

D

u

o

es

to setu

rs

thrwepvreia eofts

na. As not

ac-chirec-en-tor

omssee

ex-tal

ri-ici-ate.eri-riti-

alles

bylerex- in-idl-in-entndas

-o

ntitysenturat

s.ch

therd

efix-fix-

ation from the digital data.

2 Specifically, we replaced the “nail” (see Table 1) with alarge blue soap shaped like Batman.

3 Unfortunately, the distribution of conditions within twoversions of the items was not perfectly balanced. This wasdue to a programming error in the software that generatedthe four versions of the items. To correct for possible arti-facts, in addition to our overall analysis we ran statisticalanalyses on the two versions of the items that were properlybalanced. All effects that were significant in the overallanalysis were also significant in the analysis that excluded

10 BARR AN

ent, confederate director activated the sofiles using a computer keyboard.

Procedure. In many ways, the procedure fthis experiment was similar to that used in Eperiment 1. Therefore, in this section we focmainly on changes to the procedure that wintroduced to accommodate the use of the ond (recorded) speaker.

As in the previous experiment the direcwas a confederate, a female undergraduatedent from the University of Chicago who wanative speaker of American English. We chosmale speaker as the recorded director to conwith the female confederate. To motivate the of both a present director and a prerecordedrector we informed participants that we weresearching how people respond to live verrecorded speech.

As in the previous experiment, we used cover story that the director would receive cacontaining photographs of objects, with arroindicating how they were to be moved. The perimenter explained that the arrows on the ture were marked to indicate who would delieach instruction, the present or recorded ditor. Some of the arrows would be marked wletters (e.g., A, B, AA, and AB) to indicate ththe corresponding instruction was a secretstruction to be delivered by the recorded dirtor. The director entered the letters on a cputer keyboard to activate the playback osound file that could be heard only through addressee’s headphone. When an arrow walabeled, this meant that the instruction was togiven by the present (confederate) director hself. In actuality, the director would be readiscripted utterances printed on the cards and vating recorded instructions when necessaryin Experiment 1, the confederate was trainedto indicate the target using her eyes.

As in the first experiment, we used three prtice trials, including a third role reversal whimade it clear to the addressee that (1) the dtor had a picture with arrows and needed to gerate labels for objects and (2) the direcwould be unable to hear the instructions frthe recorded director because the addrewould hear them through an earphone.

After the practice trials, the eyetrackin

KEYSAR

nd

rx-usre

ec-

rstu- a a

rastsedi-e-us

edssx-ic-erc-

thtin-c-m- ahe notbeer-gcti-

equipment was mounted, calibrated, and the periment began. Before the first experimengrid, the partners completed a warm-up grid.

As in the previous experiment, a postexpement questionnaire examined whether partpants thought that the director was a confeder

Materials and design. The target objects werthe 12 unconventional referents from Expement 1, except that we replaced one of the ccal objects because it turned out to be too smand therefore somewhat difficult for addresseto locate.2 We created 12 experimental items embedding these referents in a grid with filobjects. As in the previous experiment, each perimental item consisted of one preparatorystruction and one critical instruction. Each grcontained two experimental items, which aways appeared in opposite conditions. For stance, if one item was mentioned by the presdirector, the other item was unmentioned athe preparatory instruction for that item wgiven by the recorded director.

The experiment had two, two-level withinparticipant factors: Mention (Precedent vs NPrecedent) and Precedent Speaker, the ideof the speaker who set the precedent (PreDirector or Recorded Director). We created fodifferent versions of the experiment, so theach item appeared in all of four condition3

The order of the grids was randomized for eaparticipant.

Coding and analysis. As in Experiment 1, acoder who was blind to the hypotheses of study located the beginning of the critical woon the videotape for each experimental trial. Walso used the same criteria to locate the final ation on the target object and extract the first

gthe two unbalanced versions.

e toeo

inrer

td

tyll

a-vs4;

lyntlly

44thetheintn-

o-i-en

ANCHORING CO

Results and Discussion

Data from 25 trials (5.48%) could not bused because there was no final fixation ontarget due either to peripheral vision or pocalibration. We excluded 12 trials from thanalysis where addressees moved the wrobject (2.63% of the data) and three triawhere there was experimenter error in settup the grid (less than 1%). In addition, thewere a few very long search times in this expiment (longer than 10 s), so we truncated sponse times at the 98th percentile (appromately 10 s).

As in Experiment 1, unmentioned referentook much longer to identify than mentionereferents (the means are 4374 and 2902 ms,spectively).4 This main effect of mention wassignificant,F1(1,35)5 83.72,MSE5 910645,

p , .01; F2(1,11)5 16.45,MSE 5 1532168, dge

sel-

thet ofan

al

FIG. 2. Fixation Latency 3 Preceden

4It may be noted that the response times in Experimenappear to be slightly longer than those observed in Expement 1. This is mainly attributable to the greater number objects in the grids in Experiment 2, where each grid had average of 15 objects, versus 11 in Experiment 1. Therefoaddressees in Experiment 2 had more objects to seaamong, resulting in longer search times.

MPREHENSION 11

her

nglsger-e-xi-

s

re-

p , .01. There was no main effect of the identiof the director nor interaction with mention (aFsø 1).

Figure 2 presents the results for the final fixtion on the target (means, No Precedent Precedent: Present Director, 4428 vs 298Recorded Director, 4319 vs 2819). It clearshows that the benefit of a linguistic precedewas the same regardless of who had originamentioned the referent, with a benefit of 14ms when the referent was mentioned by same (present) director and of 1500 ms for different (recorded) director. There was no hof an interaction between mention and the idetity of the first speaker (F1 and F2 , 1); indeed,the 56 ms difference in benefit was in the oppsite direction of what would be predicted by ether of the partner-specificity hypotheses. Whaddressees could not assume mutual knowleof the precedent, they were no less likely to uit than when they could assume mutual knowedge. These results unambiguously support partner-independence hypothesis: The benefithe precedent is due to its availability rather ththe fact that it is mutually known.

The first fixation measure also failed to reve

t 2ri-ofanre,rch

any effect of the identity of the first speaker on

t and Precedent Speaker in Experiment 2.

oe

4

ae

me

oa

o

h

f

rrdfi

id

nsce-

ffer- aon-ra-he of

ltec-eshatat-ld

the thesa-ert, inr-rs.n-irst,theial.entip-rs

inri-the

hatofortsend-

eesoavent

ven-thevi-

orstoe-

12 BARR AND

initially locating the target object. Latencies fthe first fixation were truncated at the 98th pcentile (5833 ms). When the precedent speawas the present director, the mean was 15292098 when there was no precedent). Whenwas the recorded director, the mean was 1(vs 2013, no precedent). While the differencebenefit was only 5 ms, the difference of 80 m(1529 vs 1449) between the two Precedent cditions was in the opposite direction of whpartner specificity would predict. Thereformutual knowledge of linguistic precedents donot seem to impact the early moments of coprehension.

It is possible that we found no evidence fpartner specificity because some participantsalized that the present director was a confedate. Perhaps these participants treated bothrectors as if they were drawing upon the samutual knowledge. The postexperiment qutionnaire revealed that 47% of participants the experiment guessed that the director waconfederate. Most of these participants (65%though, reported that this possibility did ncross their mind during the experiment. To evuate whether guessing the identity of the cofederate made a difference, we compared data of participants who guessed with thoseparticipants who did not guess. If partner speficity depends on participants’ belief that thpresent director is a true participant and noconfederate, then one would expect to find mevidence for partner specificity with participanwho did not guess than with participants wdid guess that the director was a confederate

However, no such pattern emerged. In fawe saw an opposite trend: The benefit nonguessers was 150 ms larger on average wthe recorded director first mentioned the tethan when the present director did. The factowhether addressees correctly guessed faileexplain any variation in the differential bene(r2 5 0.004,ns). These findings demonstratesolid pattern of partner independence.

The fact that we found no effect of speakidentity on the length of time needed to identa mentioned referent is surprising, in that it incates that the comprehension system ignore

potentially valuable source of information

KEYSAR

rr-

ker(vs it49ins

on-t,

esm-

orre-er- di-

es-ins a),tl-n-the ofci-et are

tso

.ct,orhenm of tota

erfyi-s a

Though the meanings of referential expressioare less certain when a mutually accepted predent does not exist, addressees seemed indient to the mutuality of the precedent. Fromtheoretical standpoint, it is possible that reasing about common ground is a largely delibetive process that is too slow to constrain trapid and relatively nonreflective processesreferential search (Keysar et al., 2000).

A possible objection is that it was too difficufor addressees to keep track of the two persptives because of the intermixing of utterancfrom two different speakers, a circumstance tsome might regard as unnatural. However, a nural analogy to the experimental situation woube when someone talks to one person overphone and another person, who is present inroom, occasionally interjects a side convertion. We think that it is implausible that undsuch circumstances an addressee could noprinciple, easily distinguish between the pespectives of their two conversational partneStill, we went to great lengths to clearly accetuate the differences between the speakers. Fwe required addressees to play the role of present director in a role-reversal practice trThis permitted them to observe that the presdirector would not hear the recorded descrtions. Second, the voices of the two directowere very different because of the differencegender. Third, at the beginning of the expement it was emphasized to participants that recordings were secret instructions that theyshould not reveal to the present director. Twe found no evidence for the mutuality precedents despite these procedures suppthe idea that these processes operate indepently of common ground.

A similar objection is that perhaps addressinferred mutual knowledge between the twspeakers to make sense of what may hseemed an improbable event: two differespeakers repeatedly using the same uncontional names to describe targets. If this is case, then listeners would have stronger edence for mutual knowledge between directin later trials than in earlier trials. According this explanation, we should find a partner-sp

.cific effect only in the beginning of the experi-

o

r

e

,

s4

i

iu

a

ye-

esi--ofn-mt-

ter-d ofd-kereri-adersd.en-eoble

n-m-e ofnd-r

fe-ate-b-arkll aedre

ai-

ANCHORING CO

ment, before the addressee had evidence thadirectors would use the same descriptions.test this, we considered only the first trial feach subject in each of the four conditions. Tmeans for No Precedent vs Precedent were 4vs 2876 in the Recorded Director condition, obenefit of 1820 ms; in the Present Director codition, the means were 4696 vs 2978, or a befit of 1832 ms. While there was a significant main effect of precedent,F(1,35) 5 51.63,MSE5 2652393,p , .001, there was no tracof a partner-specific effect (F , 1).

A final concern might be that addressees hto remember the locations of referents froamong a large array, leading to some losearch times that may have masked any earlyfect of mutual knowledge. To address this cocern, along with the others mentioned aboveExperiment 3 we replicate our findings usingsimpler task.

Experiment 3 explores the additional issueoverspecification. The more frequently speaers refer to a referent, the more they tendstandardize their description (Clark & WilkeGibbs, 1986; Krauss & Weinheimer, 1961966). This can cause them to overspecify rerents, such as when they call the only shoe display of objects a “loafer” (Brennan & Clark1996). When speakers overspecify referenthey violate Grice’s Maxim of Quantity (Grice1975) because they convey more informatthan necessary. This raises the interesting qtion of how listeners respond to this overspefication. If listeners expect speakers to uprecedents, then comprehension should becilitated. If they expect speakers to adhestrictly to the Maxim of Quantity, then overspecification vis-à-vis precedents should imp

sx- totnerly

theyt is

le ex-

After establishing precedents, at the end ofthe experiment listeners completed a set of

5 We also tracked movements of the computer mouse andfound results that were closely time-locked to the eye-move-ment data. Mouse tracking presents an easy-to-use, low-budget alternative to eyetracking. Information about themouse tracking data and the technique of mouse tracking

comprehension.

EXPERIMENT 3

The motivation for this experiment watwofold: (1) to investigate whether listeners epect speakers to overspecify referents dueprecedent use and (2) to test the issue of parspecificity, i.e., whether listeners are more liketo expect a speaker to use a precedent whenhave a basis for inferring that the preceden

shared. As in the previous experiments, this e

MPREHENSION 13

t theTor

he696 an-ne--

admng ef-n-in

a

ofk-to-,

ef-n a,ts,,ones-

ci-sefa-re-ir

periment explored these questions using etracking.5

In addition, this experiment improved on thdesign of Experiment 2, offering greater sentivity to small, early effects of mutual knowledge and a more compelling separation speakers. There were only two images for listeers to select from, and listeners could view thesimultaneously, reducing memory demands. Uterances from the two speakers were not inleaved but blocked, following Brennan anClark (1996), who investigated speakers’ useprecedents by blocking trials with different adressees. Thus, the identity of the speachanged at most once during the entire expment, unlike Experiment 2, where listeners hto switch back and forth between two speakwith whom they had different common grounFurthermore, speakers used preexisting convtional names to identify referents, avoiding thpotential problem in Experiment 2 where twdifferent speakers used the same improbanames.

In the experiment, pairs of pictures of coventional, everyday objects appeared on a coputer screen. Listeners heard a speaker namone of the pictures and selected the correspoing picture by clicking on it with the computemouse.

During the experiment, listeners heard a male speaker (Speaker A) establish subordinlevel precedents for pictures of everyday ojects, as speakers did in Brennan and Cl(1996). For instance, listeners heard her capicture of a car a “sportscar” when it appearin the context of a station wagon, and a pictuof a flower a “carnation” in the context of daisy. Speaker A “entrained” on the subordnate-level precedents by using them multiptimes in these contexts over the course of theperiment.

x-can be obtained by contacting the first author.

D

r te

c ixithn

sah trenaio

etn t lia wncenr

tu

b

rda

s rene

heithhear in-hatofheatill

thend-

fer-

esre-

d bede-testre-

ingrs’ in-ts.hattheeo-dis-rers

k,en-veceson,

thatker Inlvestheysee in-dureratethem-tionly

par-

14 BARR AN

“posttest” trials where target pictures appeain contexts where basic-level names wouldappropriate. These trials probed whether lisers’ experience with the precedents led themexpect to hear subordinate-level overspecifition or conventional, basic-level names. Forstance, the carnation appeared in the contethe sportscar. Thus, listeners could expect ethe basic-level names “flower” and “car” or tsubordinate-level precedents “carnation” a“sportscar.”

In these trials, the speaker used the balevel name for the target referent, e.g., “cNote that “car” overlaps phonologically with tonset of the subordinate-level precedent forother picture, “carnation.” Therefore, if listeneare expecting the subordinate-level preced“sportscar” and “carnation” instead of “car” a“flower,” then when they hear “car,” the “carntion” precedent should interfere with selectof the car as the referent.

This experiment tested the strong and wpartner-specificity hypotheses against parindependence. All listeners first entrained onprecedents with Speaker A. Later, half of theteners completed the posttest trials with SpeA, while the other half completed them withnew, male speaker, Speaker B. Listeners continued with Speaker A (Same Speaker cotion) could infer mutual knowledge of the predents through linguistic and social coprese(Clark & Marshall, 1981). However, listenewho switched to Speaker B had no evidencetheir copresence and could not infer muknowledge.

In order to ascertain the interference dueprecedents, the posttest performance of groups was compared to their performancethese same items during an earlier “pretest” wSpeaker A. The pretest was conducted at theginning of the experiment, before the subonate-level precedents were established,therefore provides an important baseline. Athe posttest, listeners saw the car and flowethe screen and heard “car.” Strong partner spficity, where listeners fully constrain comprehesion to mutual knowledge, predicts that listen

who hear Speaker B say “car” in the postteshould be no more likely to look at the carnatio

o

KEYSAR

edben- toa-

n-t ofhered

ic-r.”ehesntsd-n

aker

hes-kerahodi--

cesforal

toothonithbe-i-ndin

onci--rsst

than they were when they heard “car” in tpretest. Likewise, if listeners who continue wSpeaker A use precedents, then when they Speaker A say “car,” they should experienceterference from the precedent “carnation.” Tis, they should exhibit a higher probability looking at the carnation than in the pretest. Tweak version of partner-specificity predicts thlisteners in the Different Speaker condition wstill show interference, but less than those inSame Speaker condition. The partner-indepeence hypothesis predicts equally strong interence in both conditions.

Unlike Experiment 2, the spoken utterancused as stimuli in this experiment were all precorded. This is because real speakers woulsomewhat unlikely to produce the precise scriptions with overlapping onsets needed to the hypotheses. Moreover, the use of precorded materials permitted stringent matchof the timing and duration of the two speakeposttest utterances to reduce variability andcrease the likelihood of detecting small effec

We were concerned about the possibility tprerecorded materials would compromise generality of our findings. This is because thries of language use make a fundamental tinction between addressees and overhea(Clark & Carlson, 1982; Schober & Clar1989). Listeners who know that they are listing to prerecorded materials would beliethemselves to be “overhearers” of utteranthat were not designed for them. For this reaswe strove to convince one set of participants they were actually listening to a real speawho spoke to them, live, from another room.other words, these listeners believed themseto be nonparticipating addressees because could not interact with the speaker (Addrescondition). From a research standpoint, thistroduced undesirable aspects into the procebecause it requires deception and elabotechniques to incorporate credibility cues. In interest of simplifying future research, we copared the Addressee condition to the condiwhere we told listeners that they would simpoverhear prerecorded utterances from real ticipants (Overhearer condition). If we find n

ndifference, this suggests that future research on

O

r

ethd

ek

ig

eAeo

m

3ten

ne

t

twd

tewt

rra u ep

it-edofonfix- ton.rd

ande.

m-een

he inntftels,end

dnt,”esn-of of

heof

tri-ac-eri-ichreere theets

rs toereat

on, of

etheing

ANCHORING C

precedent use may not require such elaboprocedures.

Finally, as in Experiment 2, we used a postperiment manipulation check to examine success of the cover story. We also addequestion and a recall test to the DiffereSpeaker condition to investigate whether listers treated the perspectives of the two speadifferently.

Method

Design. The experiment used a mixed desconsisting of two two-level between-particpants factors, Speaker (Same or Differspeaker in the posttest) and Listener Status (dressee or Overhearer). There was one two-lwithin-participant factor, Test (Pretest Posttest).

Participants. Sixty-four undergraduates frothe University of Illinois at Urbana–Champaigparticipated for course credit. There were males and 32 females, equally distribuamong the four cells that result from combinithe factors of Speaker and Listener Status.

Apparatus. A computer controlled the presetation of stimuli and playback of the prercorded instructions, synchronizing these evewith the recording of eyetracking data. We usan Eyelink eyetracking system instead of ASL eyetracker used in the previous expements. The main difference between the systems is that the Eyelink system acquires at a sample rate of 250 instead of 60 Hz.

General procedure. In this section, we firsprovide a general overview of the experimfor all conditions, except where noted. Then,provide details about the two Listener Staconditions.

Participants arrived at the laboratory and ceived instructions, and the experimenter intduced the cover story (details below). The pticipant sat at a comfortable distance in fronta computer screen and operated a compmouse. After operation of the eyetracker wasplained to the participant, the experimenter it on the participant’s head and completedbrief calibration procedure.

In each trial, two images appeared on a co

puter screen, spaced at an equal distance fr

MPREHENSION 15

ate

x-e a

ntn-ers

ni-ntd-velr

n2dg

--

ntsedheri-o

ata

nte

us

e-o-r-

ofterx-ut a

m-

the center of the screen. Listeners were permted to freely view the images for 1 s. We decidto use this free-viewing technique instead forcing participants to stare at a central fixatipoint because a pilot study revealed that the ation point technique encouraged participantsselect targets using only their peripheral visioAfter the 1-s inspection period, listeners heathe speaker name one of the two images,they clicked on the referent with the mousThey did not receive feedback on their perforance. After a picture was selected, the scrwas cleared and the next trial began.

For readers’ convenience, a flowchart of tkey aspects of the experiment is presentedFig. 3. The temporal order in which the differeitems appeared follows the flowchart from leto right. For simplicity, the figure omits threpractice trials at the beginning, some filler triaand the recall test and questionnaire at the of the experiment.

It is important for the reader to keep in minthat our division of the experiment into differephases (“pretest,” “entrainment,” “posttestetc.) is for expository purposes only and donot correspond to different phases of the listeers’ experience. Listeners were not aware these divisions, experiencing only a sequencetrials that was interrupted once, just prior to tpretest, to accommodate the introduction Speaker B in the Different Speaker condition.

The experiment began with three practice als (not shown in Fig. 3). The images in the prtice trials were not seen again during the expment. Then came eight pretest items, whprovided a comprehension baseline befoprecedents were established. The items wpresented in a random order. Listeners heardspeaker use basic-level terms to name targ(e.g., “car” and “butterfly”). Because listenewould see these displays again, we wantedmake sure that the two pictures of each pair wreferred to with roughly equal frequency, so ththere were no “favored” targets. For this reasimmediately following the pretest came a setfiller items (not shown in Fig. 3) where the samdisplays appeared, but the listener heard eight opposite pictures referred to, also us

ombasic-level terms (e.g., “flower” and “knife”).

nd excludes practice trials, certain filler items, the reen-

16B

AR

R A

ND

KE

YS

AR

FIG. 3. Design of Experiment 3. “A” and “B” refer to Speaker A and Speaker B. The figure is schematic atrainment phase with Speaker A prior to the posttest, and the postexperiment recall test.

O

got

vcmh

cse einw

a n” r

c

isre

str” bmhttn

o.

trc

tt

the

a-eeirak

ls; 8ly

ntentin

,eis

ofn.ond

.tri-foreetestiszekerny theithhadd,”

erce-

top-if-

ithkere

sic-nesst)

and “carnation” six times (four entrainment, one

ANCHORING C

After the pretest, listeners completed eiblocks of items where they entrained on subdinate-level precedents. Every block presendifferent pictures and different subordinate-leterms to entrain upon. The order of the blowas randomized. Before every two blocks cathe same three filler items (shown in Fig. 3), twe later used in the recall test.

In each block, the two images from eapretest item appeared as targets four timecontexts that now required subordinate- instof basic-level names. See the figure for anample of four entrainment displays. For stance, the flower appeared with another floand was called “carnation”; the car, which apeared with another car, was called “sportscListeners also heard the speaker refer to theflower (“daisy”) and new car (“station wagonfour times each. The 16 trials for each blockthe entrainment phase were presented in adom order.

Each block ended with a retest item to prowhether listeners would expect to hear predents in a basic-level context. Figure 3 shothe retest item for one block, in which the ltener sees the car and flower and hears “cation.” For this item, the question was whethlisteners would expect “flower”/“car” or “carnation”/“sportscar.” Right after the probe item, liteners saw a similar display and heard the opicture, the car referred to (as “sportscaAgain, this was so that both pictures wouldreferred to equally often. Listeners’ perforance on retest items is not crucial to our potheses and is redundant with the positems; therefore, in our results we focus maion posttest items.

In the next block, listeners would entrain precedents for a different set of pictures (e“monarch,” “butter knife,” “black butterfly,” and“steak knife”) and then complete two retest als for that set. Because each block introdufour new precedents, for a total of 32, there wa concern that listeners might have forgottenprecedents from early blocks by the time of posttest. To remind them, after the last blockentrainment and just prior to the interruptioparticipants underwent “reentrainment” on

precedents. The eight images that would be u

MPREHENSION 17

htr-edelkse

at

h inadx--er

p-r.”ew

)ofan-

bee-

ws-na-r

--her).e-y-estly

ng.,

i-edasheheofn,

as subordinate-level competitors (e.g., “carntion” and “butter knife”) in the posttest phasappeared one time each as targets with thsame category images (e.g., “daisy” and “steknife”). These appeared in a random order.

To summarize, listeners completed 183 triawith Speaker A. There were 3 practice trialstrials in the pretest, plus 8 fillers immediatefollowing; 12 filler trials (3 before every twoblocks); 144 entrainment trials (16 entrainmeand 2 retest for each block); and 8 reentrainmtrials. Listeners required approximately 20 mto complete Speaker A trials.

Following completion of Speaker A trialsthere was a 3-min “interruption” in both thSame and Different Speaker conditions. Thwas to accommodate the introduction Speaker B in the Different Speaker conditioThe details of this procedure differ for the twSpeaker and two Listener Status conditions aare provided in the respective sections below

The posttest was preceded by three filler als, the same displays that had appeared beevery two blocks with Speaker A. Half of thlisteners completed these fillers and the postitems in the Different Speaker condition. In thcondition, we used the filler trials to emphasito listeners that Speaker B did not know SpeaA’s precedents before listeners completed aposttest items. To this end, Speaker B namedfiller targets in a way that was inconsistent wSpeaker A’s precedents. Whereas Speaker A called these targets “skates,” “flashlight,” an“couch,” Speaker B called them “rollerblades“penlight,” and “sofa.” In the Same Speakcondition, Speaker A continued using her predents for the fillers.

Next, listeners in both conditions went on complete the eight posttest trials, which apeared in a random order. Listeners in the Dferent Speaker condition completed them wSpeaker B, and listeners in the Same Speacondition, with Speaker A. Note that by the timof the posttest, listeners had heard the balevel terms “car” and “flower” one time each (ithe pretest), the precedent “sportscar” five tim(four during entrainment and once in the rete

sedretest, and one reentrainment). Therefore, ex-

D

em

ronr th

hec

eate

ut

t

dgo

noafae

f .”

na

(am

unt

de-.” ofp-entndn.” a

e-emered-

c-theenreen-n-

asnd

oneeri-there

. Atd-est

ter-

ofnteri- toct

be-oc-nt.ely

asx-ac-ithingese

interactions occurred upon arrival to the testing

18 BARR AN

pectation of the subordinate-level precedshould be strong, especially in the SaSpeaker condition.

After completing the posttest, participantsthe Different Speaker condition completed a call test, which was intended to examine hwell they remembered the different namiprecedents the two speakers established fothree filler targets. They were shown picturesthe three filler objects and asked, “What did first speaker call these objects?” Then, “Wdid the second speaker call these objects?”counted the number of pictures for which tlistener correctly recalled both terms. (This twas not administered in the Same Speaker dition.)

Next, an oral questionnaire was administerwhich is described under “Materials.” This wto determine whether listeners believed cover story and the degree to which they kthe speakers’ perspectives separate.

Finally, participants were fully debriefed. Wnow turn to differences in this general procedthat were introduced in order to manipulate conversational status of listeners.

Addressee procedure. Participants arrived athe laboratory where they met the experimenand the female confederate who had recorthe stimuli used in the experiment. We thouthat meeting the confederate in person befhand would make it compelling to listeners thshe was actually there when they later listeto her utterances (though in reality, she was nThe female confederate pretended to be a nparticipant who was running the experiment course credit. She arrived 5 min after the pointed time and pretended not to know the perimenter.

The experimenter introduced the idea o“third” male participant who was “missingThis was to motivate the later change of speain the Different Speaker condition. Participain the Different Speaker condition would hethis person “arrive” later in the experiment the beginning of the posttest). In the SaSpeaker condition, it simply would appear listeners that the third person never showed

As in Experiment 2, before the experime

began there was a staged “drawing” of instru

KEYSAR

nte

ine-wgtheof

heatWeeston-

d,shept

ere

he

teredhtre-atedt).ive

orp-x-

a

kertsrte

top.

tion sheets that participants believed would termine the roles of “director” and “matcherEach person then silently read his or her copythe instruction sheet, which we include as Apendix A. The stated purpose of the experimwas to investigate “the role of feedback amultiple speakers in language comprehensioThis justified the separation of the listener intodifferent room from the director. Listeners blieved that the speaker communicated with ththrough a one-way microphone so that thwould be no way for them to provide any feeback.

After the pair finished reading the instrutions, the experimenter wondered aloud at whereabouts of the third participant: “Hmm, ththird guy still hasn’t shown up. Let’s just go owith the experiment anyway. Technically, theare supposed to be two directors, and if he evtually shows up, he can just take over [the cofederate’s name]’s place.”

Next, listeners observed the confederateshe sat down in front of a computer screen aput on a set of headphones and a microphthat was mounted on a headset. Then the expmenter left the confederate behind and led listener down the hall to the testing room whehe or she was seated in front of a computerthis point, unbeknownst to listeners, the confeerate departed and listeners would, for the rof the experiment, hear her prerecorded utances.

The computer was connected to a pair audio speakers through which the participaheard the confederate’s utterances. The expmenter wore a headset with a microphonemake it appear that he could verbally interawith the confederate. Staged interactions tween the experimenter and confederate curred at various points during the experimeThese were introduced to make it extremplausible to listeners that the confederate wreally there, in the other room, talking to the eperimenter and to them. During these intertions, the experimenter pretended to talk wthe confederate. In reality, he was synchronizhis utterances with prerecorded dialogue. Th

c-room, before and after the practice trials, and in

lerhgi

o

sa

bti

t

e

e

s

ht

o

r

totedthe

rsvering

theyntertheow bevi-

in-e-kernt

gleshowre

that

ardtheyap-yn-hore-es

r theneon. for the theter

he

em-y ofrs,ize

were shown on a computer monitor set at a reso-

ANCHORING CO

the interruption phase just prior to the posttewhere participants in the Different Speaker codition heard a second director take the confedate’s place.

After the eyetracker was mounted, listenecompleted the practice trials. In a staged intertion that occurred right after the practice trialisteners heard the experimenter ask if thwere any questions, and heard the confederespond, “Does [participant’s name] see texact same two pictures that I see?” We thouthat when listeners heard Speaker A menttheir names, they would feel compelled ththere was really someone there, talking to them.Then the experiment began.

After listeners completed all eight blocks entrainment trials, and just prior to the posttethe experiment was interrupted by a stagevent. In the Different Speaker condition, liteners heard a staged interaction where SpeB was heard to “arrive” at the room down thhall where Speaker A gave instructions. Listeers heard someone knocking on the door of room where Speaker A was believed to Speaker A said, “I think there’s someone at door.” The experimenter replied, “I wonder that’s the third guy. I’ll be right there,” and theleft the room. While the experimenter wagone, listeners heard a prerecorded interacbetween experimenter, confederate, aSpeaker B. This established four important bliefs on the part of the listener, namely (1) thSpeaker B had never previously met the expmenter nor the first speaker; (2) that he hbeen absent during the first part of the expment and thus could not possibly have knowedge of A’s descriptions; (3) that, like the litener, he was also a naive participant; and that B was replacing A, who they could hear dpart. This part of the procedure required aproximately 3 min.

To make the interruption comparable in tSame Speaker condition, this staged interacwas replaced by a 3-min episode where the cfederate complained that her computer hcrashed. The experimenter left the testing roand listeners heard the experimenter enter room, chat informally with A, and restart he

computer.

MPREHENSION 19

st,n-er-

rsac-s,reateeht

onat

fst,ed-keren-thee.

hefnsionnde-atri-adri-l--(4)e-p-

eionon-adm

A’s

In both Speaker conditions, upon returningthe testing room, the experimenter recalibrathe eyetracking equipment and began posttest trials.

After completing the experiment, listenewere asked to rate the credibility of the costory on a scale of 1 to 10, with 1 correspondto didn’t believe it at alland 10 to fully believedit.

Overhearer procedure. Participants arrived athe laboratory and met the experimenter. Tdid not meet the confederate. The experimetold them that the experiment investigated role of feedback and multiple speakers in hpeople understand words and that they wouldlistening to prerecorded utterances from preous participants. They were told to read thestructions (Appendix A) that the participants bfore them were given. In the Same Speacondition, the experimenter told the participathat they would hear descriptions from a sinspeaker because the second speaker did not up. In the Different Speaker condition, they wetold that the second speaker arrived late andthey would not hear him until the end.

They completed the practice trials and hethe same staged interactions, but of course,believed that these were all things that had hpened before. Instead of the experimenter schronizing his utterances with someone wwas not there, the listener simply heard his precorded voice in the playback mix. The namthat listeners heard the confederate use aftepractice trials were not their own, but that of oof the participants in the Addressee conditiThe staged events in the interruption phasethe Different/Same Speaker conditions weresame as in the Addressee condition. UnlikeAddressee condition, however, the experimenremained in the room with the participant for tentire duration.

Materials. Forty-two bitmap images werused in the construction of stimuli. These iages depicted everyday objects from a varietcategories, including animals, vehicles, flowekitchen utensils, furniture, and so on. The sof the bitmaps was 278 3 278 pixels, and they

lution of 1024 3 768 pixels.

20 BARR AND KEYSAR

ennre

enci

oinmem or

”

tt

tua

w

Tre

h

l

t r

s-(“m

it iss-

ion,na-

shr B.ak-ondrst

rmslly ofv-’sfor

red

ldedorget.ure

toforam-s atachm-an-

cu-hatti-ityng

Each display consisted of two images, ctered vertically on the computer monitor aspaced equidistantly from the vertical midlion the horizontal axis. Each display was paiwith a spoken stimulus to constitute an expmental item.

There were eight sets of items in the expment, created from combining four differebitmap images from two categories (e.g.,and flower). Each set contained two critical dplays that were used in the pre- and posttewhere the subordinate-level precedent for picture (e.g., “carnation”) shared an overlappphonological onset with the basic-level nafor the other (e.g., “car”). The two other imag(e.g., “station wagon” and “daisy”) were frothe same category as the basic-level targetsubordinate-level competitor. The basic/subdinate level pairings for all eight sets we“car–carnation,” “pitcher–pitbull,”“butterfly–butterknife,” “tape–tablespoon,”“coat–cobra,” “clock–clawhammer,“plant–plank,” “tent–tennis shoe.”

Target images appeared on either the lefright. There was no systematic placement of gets within or across item sets.

We modeled the auditory stimuli after narally produced names. The female confederwho later served as Speaker A, participatedspeaker in a mock version of the experimewith the experimenter as addressee. This before she knew the purpose of the experimor had ever seen the images. We recordedvoice as she named the target images. recordings were transcribed to generate a scSome of the subordinate-level descriptions wchanged in the script so that they shared same onset with the basic-level term. All of tnames that were later used in the experimwere rerecorded from the script, and the earrecordings were discarded.

For the entrainment phase, four different kens of each name were recorded. To makespeech sound realistic, the linguistic charactetics of each token depended on the speaker’miliarity with the referent. Initial referring expressions included such features as repairs car . . . I mean, the sports car”), hesitations (“u. . the . . . stained tile of wood”), and a rising fin

intonation. For example, when referring to a pa he

n-deedri-

ri-tars-sts,neges

andr-e

orar-

-te,

asnt,as

entherheipt.re

thee

entier

o-theis- fa-

the .

alr-

ticular dog for the first time in the context of a pbull, the speaker said, “oh, what kind of dogthat? um . . . golden retriever?”. Later expresions from the speaker had a shorter duratfewer hesitations, and a more confident intotion (e.g., “the golden retriever”).

A male native speaker of American Englirecorded the posttest utterances for SpeakeTo make the acoustic features of the two speers’ utterances as similar as possible, the secspeaker modeled his own after those of the fispeaker. For each posttest item, the wavefoof the two speakers’ utterances were visuacompared to ensure that timings of the pointdisambiguation were roughly matched. The aerage point of disambiguation for Speaker Atest utterances was 235 ms, and 230 ms Speaker B.

The oral questionnaire that was administeat the end of the experiment was as follows:

How strongly would you endorse thefollowing statement(s), on a scale of 1 to10, with 10 being strongest?

1. I believe that the second speaker hadno knowledge of the first speaker’s descrip-tions. (Different Speaker condition only).

2. I believe that I was listening to (a) livespeaker(s).

Analysis. Interference in the posttest coumanifest itself in two ways: (1) as an increaslikelihood of gazing at the competitor image (2) as a delay in eye movements to the tarFor statistical analysis we used a single measthat combines these two possibilities, the targetpreference score. We grouped the samples in24 ms “bins,” computing a preference score each bin. Each 24 ms corresponded to six sples of eye data (the EyeLink system samplea rate of 250 Hz, or 1 sample every 4 ms). Ebin was labeled with the median sample nuber. Thus, bin 12 corresponded to samples spning 0 and 23 ms.

The preference score for each bin was callated by subtracting the number of samples tvisual activity was directed toward the competor from the number of samples that the activwas directed toward the target, and then dividithis value by the total number of samples in t

ANCHORING COMPREHENSION 21

m

ae

eir

&Teeenfoa

-thetw

dh

fhe .mife

ta

natrstndi-

ce

ts,

ec-ble

ak-ge,ct

ead-see

is-

r-vser ast,

testionn-heionns,

tedts in

he

onref-ashis

was identical to the corresponding values in the

6 Test 3 Listener Status,F1(1,60) 5 .4, MSE 5 .7076,F2(1,7) 5 2.20, MSE 5 .0664,p , .19; Test 3 ListenerStatus 3 Bin, F1(35,2100) 5 .42, MSE 5 .0349,F2(35,245) 5 .4, MSE5 .0184; Test 3 Listener Status 3Speaker,F1(1,60) 5 .24,MSE5 .7076,F2(1,7) 5 .42,MSE5 .1933; Test 3 Listener Status 3 Speaker 3 Bin,F1(35,2100) 5 .90, MSE 5 .0349, F2(35,245) 5 .61,MSE5 .0254.

time span. A positive value for a given spa(maximum of 1) means that the listener spemore time looking at the target than at the copetitor. Correspondingly, a negative value (minmum of 21) means that the listener lookedthe competitor more. A value of zero means thlooked at them equally often (or not at all).

We analyzed a temporal window spannininclusively, bins 226 and 1068. The formnumber represents the earliest point at whone would expect to find eye movements pgrammed on the basis of the linguistic sign(after approximately 50 ms of input), given 18ms for “saccadic overhead” (Matin, Shao,Boff, 1993). Bin 1068 represents median Rafter which point data is lacking for half of thtrials. In those trials where the participant rsponded before the median RT, the empty cup to 1068 were filled with ones (representipreference for the target). In other words,this analysis the final fixation on the target cbe regarded as “cumulative.”

Results and Discussion

Manipulation check. We first report our manipulation check measures to show that cover story was effective and that listeners wable to separate the perspectives of the speakers.

First, results from the postexperiment quetionnaire show that the cover story was etremely effective. The participants in the Adressee condition gave the credibility of tcover story an average rating of 9.8, with a moof 10. The minimum rating for the credibility othe cover story was 9. Furthermore, all of tparticipants in this condition indicated that thstrongly believed that they were listening tolive speaker, providing a mean rating of 9(SD5 .712), with a mode of 10 and a minimurating of 8 (4 participants). This was reliably dferent from the Overhearer case, where the mwas 1.6 [SD5 1.23, mode 5 1, with a maximumrating of 6 from one participant and missing dafrom one participant; t(61) 5 32.15,p , .0001].

Listeners in the Different Speaker conditioindicated that they very strongly believed ththe second speaker did not know the fispeaker’s precedents. In the Addressee cotion, the mean rating was 9.8 (SD 5 .712,

nnt-

i-ty

g,rcho-al0

,

-llsgrn

ereo

s-x--ede

eya6

-an

mode 5 10) compared to 9.2 (SD5 1.23, mode 510) in the Overhearer condition. This differenwas marginal,t(28) 5 1.68,p , .10. The lowestrating obtained was a 7, given by two participanone in each of the two cover story conditions.

It was clear that listeners kept the persptives separate, because they were typically ato recall the different names that the two speers used for the three filler items. On averathey got 2.52 of the three recall items corre(SD5 .58, mode 5 3; 4 of 32 data points wermissing because the experimenter forgot to minister the test). The mean for the Addrescondition was 2.58 (SD5 .51, mode 5 3), and2.46 for the Overhearer condition (SD 5 .64,mode 5 3). There was no difference across Ltener Status,t(61) 5 .39. p 5 .70.

Main results. We submitted the target prefeence scores to an ANOVA with Test (Pre Post), Speaker (Different vs Same), ListenStatus (Addressee vs Overhearer), and Binfactors. We found a main effect of TesF1(1,60) 5 36.57,MSE5 0.7076,p , .0001;F2(1,7) 5 35.18,MSE 5 0.3690,p , .0001.Mean target preference was .62 in the precompared to .47 in the posttest. The interactof Bin with Test was significant (after Greehouse–Geiser correction), indicating that ttemporal characteristics of the comprehensprocess was different in the two test conditioF1(35,2100) 5 13.89,MSE5 0.035,p , .0001;F2(35,245) 5 3.74, MSE 5 0.0651,p , .05.This finding suggests that listeners expecspeakers to use subordinate-level precedenthe basic-level context of the posttest.

Listener Status did not interact any of tother factors (all Fs ø 1).6 We therefore col-lapsed over Listener Status in order to focusthe Speaker manipulation. The mean target perence for the Different Speaker condition w.62 in the pretest and .47 in the posttest. T

22 BARR AND KEYSAR

rc

e

aa5

a

es

a

ernti-ticofrt

,

nedardentsxtnde264er-chedsts

pri-kerseenhatis-

he ince-cehate-

adell32t ofith in-ds,theesic-

-

kertse).hes).ss)

Same Speaker condition. An ANOVA confirmethat there was no difference in interference btween these conditions [Test 3 Speaker interac-tion, F1(1,60) 5 .42,MSE5 .7076; F2(1,7) 5.06,MSE5 .3716; Test 3 Speaker 3 Bin inter-action, F1(35,2100) 5 .84, MSE 5 .0349,F2(35,245) 5 .59,MSE5 .0250]. These resultssupport the partner-independence hypotheListeners attempted to apply linguistic precdents in the posttest simply because they wavailable, causing interference in selecting ttarget. Even though listeners in the DiffereSpeaker condition possessed strong evideagainst mutual knowledge of the precedenthey were no less likely to use them to interpspeech than listeners in the Same Speaker dition.

Once listeners observed in the posttest tthe speaker was using basic-level terms instof the precedents, they prepared for this eveIn both Speaker conditions there was a decrein interference from the first to the second hof the posttest trials, as manifest by an increin the target preference score from .42 to .F1(1,60) 5 5.81,p , .05; F2(1,7) 5 10.07,p ,.05. However, this factor did not interact witthe Speaker variable (all Fs # 1); the meanswere .41 vs .53 for the Same Speaker conditi.43 vs .52 for the Different Speaker condition.

To more closely determine the time courseprecedent use, we undertook a bin-by-bin anasis of the preference scores using a multistBonferroni procedure, with an alpha level of .0for each set of bin-by-bin comparisons. We pformed both a participant and an item analyto determine at what point preference for the tget became statistically reliable. We considthe average of the two analyses to representmost probable real value. We first report resucollapsing across the Speaker conditions athen individually compare the two conditions.

In the pretest, where listeners heard baslevel descriptions, they became more likely look at the target than the competitor at 312 mAllowing 180 ms for programming the eymovement to the target (or for deciding not tolook away from the target), listeners identifiethe target after only 132 ms of speech inpThis is well before the earliest disambiguatiopoint (185 ms) and about 104 ms before the

de-

sis.e-erehentncets,eton-

hatadnt.aselfse2,

h

on,

ofly-ge5r-is

ar-er theltsnd

ic-tos.

e

dut.nv-

erage disambiguation point (236 ms). In othwords, even without precedents, listeners idefied referents on the basis of minimal linguisinput. This is consistent with the predictions word recognition models such as Coho(Marslen-Wilson, 1987; Marslen-Wilson &Welsh, 1978) or TRACE (McClelland & Elman1986).

In the retest, after the speaker had entraion subordinate-level precedents, listeners hethe speaker continue to use these preced(e.g., “carnation”), but in a basic-level conte(e.g., while viewing the pictures of the car aflower). They were reliably more likely to fixatthe target by 444 ms, on the basis of about ms of input. This was just 26 ms after the avage disambiguation point of 238 ms (whiranged from 185 to 367. That listeners identifithe carnation as the target so quickly suggethat even when basic-level terms are approate, listeners are not surprised when speacontinue using precedents. If listeners had bsurprised by the precedents, it seems likely ta delay much larger than 26 ms, relative to dambiguation, would have been observed.

Finally, in the posttest, listeners heard tspeaker return to the basic-level name (“car”)a basic-level context. If they expect the predent (“carnation”), then they should experieninterference in selecting the target. This is wwe found. Preference for the target did not bcome significant until 528 ms, after listeners hheard about 348 ms of the input. This was wafter the average disambiguation point of 2ms and even the latest disambiguation poin282 ms. Direct comparison of the posttest wthe pretest revealed reliably greater posttestterference from 420 to 744 ms. In other worparticipants expected speakers to violate Maxim of Quantity (Grice, 1975) and continuto use linguistic precedents, even though balevel terms would have been sufficiently informative in that context.

We now focus our attention on the Speamanipulation. Figure 4 shows probability plofor these two conditions (Different and SamEach chart displays fixation probabilities for ttarget (squares) and the competitor (circleAdditionally, each chart shows probabilitiefrom the pretest (filled in squares and circle

ANCHORING COMPREHENSION 23

h

hlhusth

t

ti

hs

ec

t

v

ut

einn e

r

l ea

htni

ral,eomac-ightent

n beundryw-Ex-but

is thattivevelndersxis-ote,cts,uselearwasr to ofce,

pec-ly

eers

ce-eci-ibleee-x- aer-ces isalven

rac-

edi-ter

and the posttest (open squares and circles).x axis is labeled with bin number. At bin 12 tprobabilities do not start at zero because offree-viewing paradigm.

It can be seen very clearly from the plots texpectation of the precedent on the part of teners interfered with their interpretation of tbasic-level term in both Speaker conditions, ring out the strong partner-specificity hypotheIn the Different Speaker pretest, fixations to target (“car”) became reliably more likely thafixations to the competitor (“carnation”) at 28ms. In the posttest, this did not happen until 5ms, amounting to a delay of 276 ms relativethe pretest. In the Same Speaker condition,get fixations were reliably more likely by 37ms in the pretest and 576 ms in the posttesother words, there was a delay of 204 ms, whis numerically smaller than the delay of 276the Different Speaker condition; in other wordthis is in the opposite direction from what tstrong and weak partner-specificity hypothewould predict.

To summarize, our results indicate that listers, like speakers, entrain on linguistic predents, and expect speakers to overspecify reents. When they heard basic-level terms in posttest instead of the precedents, they expenced interference in selecting the target, ethough the basic-level term would have besufficiently informative.

Listeners used linguistic precedents becathese precedents were available to them, nocause they were mutually known. Subordinalevel precedents such as “carnation” did not rably produce more interference in recognizthe word “car” when it was mutually knowthan when it was simply available. There wasevidence for either weak or strong partnspecificity.

To conclude, we have shown that listeneexpectations that speakers will use linguisprecedents is strong; in fact, so strong that teners expect to hear them even when theyoverinformative—and surprisingly, even whthey have no reason to believe that they shared with the speaker.

Clearly, this experiment presents tstrongest case against the claims of the parspecificity of linguistic precedents. However,

Theethe

atis-el-

is.e

n864toar-2. Inchins,ees

n-e-fer-heeri-en

en

sebe-te-li-g

nor-

s’ticis-arenre

eer-

ts

task might strike some readers as unnatuwhich could possibly limit the generality of thresults. One way that the task diverges frface-to-face conversation is the lack of intertion between speaker and listener. How mthe lack of interaction cause listeners’ preceduse to differ from interactive situations?

First, perhaps the only way a precedent camutually accepted and added to common grois through live interaction, as grounding theosuggests (Clark & Brennan, 1991). Note, hoever, that under this assumption, listeners in periment 3 should never use precedents—clearly, we show that they do. Therefore, itpossible that certain effects of precedent usehave been claimed to support the collaboramodel may instead be attributed to low-leavailability effects. For instance, Brennan aClark (1996) present the finding that speakoverspecify referents as evidence for the etence of conceptual pacts. However, as they nthere are factors other than conceptual pasuch as frequency or recency, that might caspeakers to continue to use a pact. It is not cwhether the overspecification they observed due to the existence of conceptual pacts othese other factors because the interactivitythe situation was not manipulated. For instanit is possible that speakers would have oversified referents even without interaction simpbecause of the availability of the precedents.

A more troubling criticism would be that thabsence of true interaction prevented listenfrom sufficiently associating Speaker A’s predents with that speaker to yield partner-spficity effects. We acknowledge this as a posslimitation of this experiment, but it should bkept in mind that we did not find partner-spcific effects in the face-to-face situation of Eperiment 2, where listeners interacted withconfederate. The common view that partnspecific effects in processing single utteranare most likely to be found in live interactionsimply an a priori assumption with no empiricsupport. In fact, we suggest that there are ebetter a priori reasons to expect that live intetion is where one is least likely to find partner-specific effects because the presence of immate, multimodal feedback permits greaegocentrism (Barr, 1999).

24 BARR AND KEYSAR

FIG. 4. Fixation Probability Plots in Experiment 3. Error bars represent the standard error of the mean for each meas-ure.

n

x-tsut aakatail-

3ofd-

thet,ntsfens

at

ANCHORING CO

Another reason the task of Experiment 3 mstrike some as unnatural is that the naming tis not embedded in any broader conversatiotask. This too could fail to generalize in mothan one way. On the one hand, it may heighattention to the task itself and allow the develoment of processing strategies that would arise in a more conversational task. On the othand, in a conversational task, attention is dtributed over many more processes, perhmaking it even more difficult for listeners tkeep track of who knows what. In other wordthe task of Experiment 3 could have potentiaoverestimated partner-specificity effects because there listeners could concentrate moretention who used what name, given that thwere not distracted by other things. Finally, nothat we failed to find partner-specific precedeeffects in Experiment 2, where the naming wembedded in the conversational task of re

a

a

e

tdoed-- theeyisin,nttb-inn-reforereatetoiledrsdt

ranging objects.

GENERAL DISCUSSION

Our experiments demonstrate that the stdardization that occurs in language productihas a counterpart in comprehension. Linguisprecedents enable listeners to quickly accesserential meanings and can override the assumtion that speakers will be optimally informativin their referential descriptions. Listeners expespeakers to follow precedents, whether tprecedent has been used only once or mtimes. Of course, the strength of a precedent wdepend on its frequency and recency, thoughexperiments do not examine these factors.

In the first experiment, the benefit was mularger for unconventional than conventional rerents, clearly implying that it was linguistic: itdepended on an established relation betweenlabel and the object. The second experimfound benefits at second mention, while the thexperiment found benefits due to entrainmentthe precedent. Both Experiments 2 and 3 sported partner independence over the partnspecificity hypotheses: listeners were just likely to use precedents with a new speakerthey were with the one who established the

This led to an equal comprehension benefit in E

MPREHENSION 25

ayasknalretenp-ot

heris-

apsos,lly- at-eytentasar-

n-onticref-

p-ecthenyill

our

chf-

theentird onup-er-as asm.

periment 2 in the two Speaker conditions. In Eperiment 3, listeners’ expectation of precedenimpaired comprehension of basic-level terms, blisteners were no less likely to expect them withnew speaker than with an old one. Not even wepartner specificity was observed, suggesting thlisteners use precedents because they are avable, not because they are mutually known.

As a point for future research, Experiment suggests interesting possibilities for models word recognition. Consistent with Cohort anTRACE models, we found that in the pretest listeners could access referential meanings on basis of limited phonological input. In the reteslisteners accessed subordinate-level precedeextremely quickly, even though in the context othe display, the basic-level name would havbeen appropriate. Models of word recognitiomight easily incorporate linguistic precedents aa boost in the baseline activation of words thcorrespond to precedents.

One criticism of our studies might be thathey require people to represent what others notknow, something that has been characterizas computationally intractable (Polichak & Gerrig, 1998). This claim is not very compelling because people can and very often do representfact that others do not know something that ththemselves know; in fact, this state of affairs what often stimulates people to communicate the first place. Even if we grant this criticismour studies do not require a listener to represeeverythingthat a speaker does not know, busimply to ignore the precedents that were estalished by the previous speaker. For instance,Experiment 3, during the one second that listeers were provided to scan the display befohearing the name, they could have prepared the upcoming basic-level name from the speakby inhibiting the precedent and activating thbasic-level term. In fact, listeners showed ththey could inhibit precedents, as reported abovin the decrease in interference from the first the second half of the posttest. Yet even whlisteners in the Different Speaker condition hagreater justification than Same Speaker listenefor both inhibiting Speaker A’s precedents ananticipating basic-level terms, they were no

x-more likely to do so.

u

ge

out

o

t

y

l

a

ts

s

o

s

e

re

sentx-n-ak-hat

o-ge

on-ageadek,-ongalerseyity, arerkef-ary in-er”owd ins isd-stab-edur-useex-ingith. In-spe-andat

er”ueed.en-sarners

use mutual knowledge only to diagnose and cor-

26 BARR AND

Besides, this criticism misses the point ththe design we use in our studies is essentiademonstrating that a process distinguishes tween what is and what is not mutually knowStudies that offer evidence for the use of mutknowledge typically demonstrate that languausers rely on information that is mutual (e.Clark, Schreuder, & Buttrick, 1983; GreenGerrig, Mckoon, & Ratcliff, 1994). But asKeysar (1997) demonstrated, these studies cfound mutuality with information available tthe self. Every piece of information that is mtually known to A and B is also simply known A. Therefore, to support the argument that coprehension relies on mutual knowledge, it is nsufficient to show that mutually known information was used—one must also show that infmation that was not mutual was not used. Thiswould demonstrate that the process in quesdistinguishes between mutual and available formation.

It is interesting that the comprehension stem is not designed to effectively use mutuknowledge, which is a potentially valuabsource of information. Clark and Carlson (198proposed that mutual knowledge can potentialimit the scope of contextual information ththe comprehension system must consider. Twould prevent listeners from making systemamistakes. It may be, however, that this benefiinformation reduction and increased succesoutweighed by the difficulty of using metaknowledge during real-time language proceing. In our view, listeners’ use of precedents thare outside mutual knowledge is symptomatica language processing system that is designeuse available information to settle matters referential ambiguity quickly and efficiently sthat it can keep up with the rapid influx of linguistic information. In other words, it is a sytem that is designed in response to the compmises of performance in the real world, not the exacting standards of pragmatic theories.

To preempt a possible misunderstandingour position, we point out that although listenein the current experiments did not use the idtity of the speaker to guide the identification the referent, this does not imply that they we

not keeping track of this information. It is obvi

KEYSAR

atl tobe-n.al

ge.,,

on-

-om-ot-r-

ionin-

s-ale1)llythistic in is-s-at ofd toof

--ro-to

ofrsn-

of

ous that language users can and do reprewho said what. For instance, participants in Eperiment 3 were typically able to recall the coceptual precedents that the two different speers had established. The important point is tthey simply did not usethis knowledge to iden-tify intended referents. There may be other dmains of language use where this knowledcould be effectively used.

It may be noted that our studies seem to cflict with research on precedent use in languproduction, where some researchers have mclaims for partner specificity (Brennan & Clar1996; Clark & Wilkes-Gibbs, 1986). For instance, Brennan and Clark (1996) made strclaims for the partner specificity of conceptupacts: “In the historical models so far, speakchoose their wording regardless of whom thlast spoke to. But according to partner specificthey do so for the specific addressees theynow talking to” (p. 1484). But Brennan and Cladid not find this. Instead, they state that, “the rerences in our task emerged not from solitchoices on the part of the director, but from anteractive process by both director and match(p. 1491). In other words, they appeared to shthat “dyads negotiate” references over prolongeinteractions, not that “speakers choose” themproducing single utterances. What this meanthat when directors started talking to new adressees, they used the precedents they had elished with the old matcher. They then receivfeedback from their addressee, who were sprised by the overly specific terms, and becaof this feedback they changed their referring pressions. Such pattern of results concernspeakers’ behavior is perfectly consistent wour results concerning addressees’ behaviorneither case need one assume that a partnercific pact was established between matcher director. Instead, it is sufficient to assume theach of them entrained on precedents (“loafand “carnation”) and later attempted to continusing them with new partners, but were rebuffSuch an account is consistent with the more geral anchoring and adjustment model of Keyand Barr (in press), where speakers and liste

-rect coordination problems.

7

ont

og

;ee

s

ee rhho awatyg

n

rdo

h ares theet”

u.thelec-ter

rsrec-ul.tures.. Be-willt”.tor onDi-ad-

re,nglly

aourt toun-okentus

di-er the

ANCHORING CO

Evidence for the partner-specific use precedents in language production is lackiOther researchers in the area have found certain accommodations that speakers appeamake for addressees are actually due to llevel effects such as availability or primin(Bard, Anderson, Sotillo, Aylett, Doherty-Sneddon, & Newlands, 2000; Brown & Dell, 1987Ferreira & Dell, 2000). More studies arneeded, especially like the above-mentionones, which overcome the confound discusby Keysar (1997).

The results from our three experiments hput Henry’s egocentrism in context. Henry rtrieves the reports from the detective agencythe referent of “the reports” because he anchohis understanding in the linguistic precedent tBendrix had established immediately prior to texchange. This led him to the incongruous cclusion that the newcomer had knowledge ofextremely private and sensitive matter. He knothis is implausible and attempts to adjust awfrom the anchor, but his wife’s possible infidelilies heavily on his mind and is not easily inored. Because of this, he asks the spea“What reports?” even though there is only oreferent that is uniquely defined by their mutuknowledge. In short, the nature of the comphension system makes us all prone to misunstandings such as Henry’s—even when

Please feel free to ask the experimenter any ques-

spouses are not cheating on us.

APPENDIX

kn.p-

-l--

r-

and

tots.

Clark, H. H. (1982). The relevance of common ground:

Instruction Sheet for Experiment 3

INSTRUCTIONS FOR: MATCHER

This experiment investigates the role of feedbacand multiple speakers in language comprehensioWe are interested in how people understand descritions from different people, under conditions wherefeedback is either present or absent.

You are in the FEEDBACK ABSENT condition.You have drawn the role of the MATCHER . The

other two participants will take turns being the DI-RECTOR. To be completely sure that directors areunable to get any verbal or nonverbal feedback fromyou, we will put you (the matcher) in a separate roomdown the hall (Room 1428).

You will be performing the task on a computer. Your

computer (which is located in 1428) is connected to the

MPREHENSION 2

fg.hatr tow-

-

ded

lp-asedaten-nsy

-kereale-er-ur

directors’ computer (the one here, in 1410) througlocal Intranet. The task is very simple. Two pictuwill appear simultaneously on your screen and ondirectors’ screen. One of these pictures is the “targpicture. A director will name that picture for yoBased on that person’s description, you will select picture that you think is the target. Indicate your setion by clicking on the picture using the compumouse. Because you are in the FEEDBACK ABSENTcondition, you will not be able to give the directofeedback or ask them for help. Furthermore, the ditors will not know if their descriptions are successfThus, if you are unable to determine the target picbased on the description, just make your best gues

Each director wears a private set of headphonesfore the pictures appear, one of the two directors hear a voice in the headphone say “right” or “lefThis tells that director that: (1) it is her/his turn name the target; and (2) that the target will appeathe corresponding (right or left) side of the screen. rectors are NOT given names for targets in the hephones—they only know their locations. Therefothey must come up with their own way of namithem. We will be recording their descriptions digitaonto the computer’s hard disk.

As you perform the task, you will be wearingheadband equipped with two cameras that film yeyes. Later, we will analyze this data with respecthe descriptions that you heard. This allows us to derstand how the eyes change in response to splanguage. The experimenter will explain the apparain more detail once you arrive in Room 1428.

Before the experiment begins, you and the tworectors will participate in three practice trials togethto make sure that everyone properly understandstask.

tions at this time.

REFERENCES

Bard, E. G., Anderson, A. H., Sotillo, C., Aylett, M., Doherty-Sneddon, G., & Newlands, A. (2000). Controling the intelligibility of referring expressions in dialogue. Journal of Memory and Language, 42, 1–22.

Barr, D. J. (1999). A theory of dynamic coordination foconversational interaction. Ph.D. thesis, The University of Chicago.

Brennan, S. E., & Clark. H. H. (1996). Conceptual pacts lexical choice in conversation. Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 22,1482–1493.

Brown, R. (1958). How shall a thing be called? Psychologi-cal Review, 65, 14–21.

Brown, P. M., & Dell, G. S. (1987). Adapting production comprehension: The explicit mention of instrumenCognitive Psychology, 19, 441–472.

Comments on Sperber and Wilson’s paper. In N. Smith

28 BARR AND KEYSAR

(Ed.),Mutual knowledge(pp. 124–127). London: Aca-demic Press.

mDns

re

:

e

c

nsic

of

o

&da

nn

e-

Rt e

&c

icrne

Horton, W. S., & Keysar, B. (1996). When do speakers takeinto account common ground? Cognition, 59, 91–117.

n-ey

s

).al

).ss

n-ing

ce in-

k,al-

n

-on

r-t

l

ndnd

d-

dith

Clark, H. H., & Brennan, S. E. (1991). Grounding in comunication. In L. B. Resnick, J. M. Levine, & S. Teasley (Eds.),Perspectives on socially shared cogtion (pp. 127–149). Washington, DC: American Pchological Association.

Clark, H. H., & Carlson, T. B. (1981). Context for comphension. In J. Long & A. Baddeley (Eds.),Attentionand Performance(Vol. IX, pp. 313–330). Hillsdale, NJErlbaum.

Clark, H. H., & Carlson, T. B. (1982). Speech acts and hers beliefs. In N. V. Smith (Ed.),Mutual knowledge(pp. 1–37). London: Academic Press.

Clark, H. H., & Marshall, C. R. (1981). Definite referenand mutual knowledge. In A. K. Joshe, B. Webber,I. Sag (Eds.),Elements of discourse understandi(pp. 10–63). Cambridge, UK: Cambridge Univ. Pres

Clark, H. H., & Murphy, G. L. (1982). Audience design meaning and reference. In J. F. Le Ny & W. Kints(Eds.), Language and comprehension(pp. 287–299).Amsterdam: North-Holland.

Clark, H. H., Schreuder, R., & Buttrick, S. (1983). Commground and the understanding of demonstrative reence. Journal of Verbal Learning and Verbal Behavio,22, 245–258.

Clark, H. H., & Wilkes-Gibbs, D. (1986). Referring as a claborative process. Cognition, 22, 1–39.

Eberhard, K. M., Spivey-Knowlton, M. J., Sedivy, J. C.,Tanenhaus, M. K. (1995). Eye movements as a wininto real time spoken language comprehension in nral contexts. Journal of Psycholinguistic Research, 24,121–135.

Ferreira, V. S., & Dell, G. S. (2000). Effect of ambiguity alexical availability on syntactic and lexical productioCognitive Psychology, 40, 296–340.

Garrod, S., & Anderson, A. (1987). Saying what you min dialogue: A study in conceptual and semantic codination. Cognition, 27, 181–218.

Greene, S. B., Gerrig, R. J., McKoon, G., & Ratcliff,(1994). Unheralded pronouns and managemencommon ground. Journal of Memory and Languag,33, 511–526.

Grice, H. P. (1975). Logic and conversation. In P. Cole Morgan (Eds.),Syntax and semantics 3: Speech a(pp. 41–58). New York: Academic Press.

Hanna, J. E., Trueswell, J. C., Tanenhaus, M. K., & NovJ. M. (1997) Consulting common ground during refeential interpretation. Poster presented at the 38th Anual Meeting of the Psychonomic Society, Philadphia, PA.

-.i-y-

-

ar-

e&g.

nh

ner-r

l-

owtu-

d.

anor-

.by

J.ts

k,--l-

Keysar, B. (1997). Unconfounding common ground. Dis-course Processes, 24, 253–270.

Keysar, B., & Barr, D. J. (in press). Self anchoring in coversation: Why language users don’t do what th“should.” In T. Gilovich, D. W. Griffin, & D. Kahne-man (Eds.),The psychology of judgment: Heuristicand biases. Cambridge, UK: Cambridge Univ. Press.

Keysar, B., Barr, D. J., Balin, J. A., & Brauner, J. S. (2000Taking perspective in conversation: The use of mutuknowledge for error correction. Psychological Science,11,32–38.

Keysar, B., Barr, D. J., Balin, J. A., & Paek, T. S. (1998Definite reference and mutual knowledge: Procemodels of common ground in comprehension. Journalof Memory and Language, 39, 1–20.

Keysar, B., Barr, D. J., & Horton, W. S. (1998). The egocetric basis of language use: Insights from a processapproach. Current Directions in PsychologicalScience, 7, 46–50.

Krauss, R. M., & Weinheimer (1964). Changes in referenphrases as a function of frequency of usage in socialteraction. Psychonomic Science, 77, 622–626.

Krauss, R. M., & Weinheimer (1966). Concurrent feedbacconfirmation, and the encoding of referents in verbcommunication. Journal of Personality and Social Psychology, 4, 343–346.

Marslen-Wilson, W. D. (1987). Functional parallelism ispoken word recognition. Cognition, 25, 71–102.

Marslen-Wilson, W. D., & Welsh, A. (1978). Processing interactions and lexical access during word recognitiin continuous speech. Cognitive Psychology, 10,29–63.

Matin, E., Shao, K. C., & Boff, K. R. (1993). Saccadic ovehead: Information processing time with and withousaccades. Perception & Psychophysics, 53, 372–380.

McClelland, J. L., & Elman, J. L. (1986). The TRACE modeof speech perception. Cognitive Psychology, 18, 1–86.

Polichak, J. W., & Gerrig, R. J. (1998). Common ground aeveryday language use: Comments on Horton aKeysar (1996). Cognition, 66, 183–189.

Schober, M. F., & Clark, H. H. (1989). Understanding by adressees and overhearers. Cognitive Psychology, 21,211–232.

Sperber, D., & Wilson, D. (1982). Mutual knowledge anrelevance in theories of comprehension. In N. Sm(Ed.), Mutual knowledge(pp. 61–87). London: Aca-demic Press.

(Received December 20, 2000)(Revision received January 4, 2001)