Perspective-free pragmatics: Broken precedents and the recovery-from-preemption hypothesis

Journal of

www.elsevier.com/locate/jml

Journal of Memory and Language 56 (2007) 436–455

Memory andLanguage

Perspective-free pragmatics: Broken precedentsand the recovery-from-preemption hypothesis q

Edmundo Kronmuller *, Dale J. Barr

Department of Psychology, University of California, Riverside, CA 92521, USA

Received 9 February 2006; revision received 11 May 2006Available online 14 July 2006

Abstract

When speakers refer to the same referent multiple times in a conversation, they tend to follow established patterns ofusage, known as conversational precedents. Research has found that listeners expect speakers to follow precedents, andthat this expectation guides their search for referents (Barr, D. J., & Keysar, B. (2002). Anchoring comprehension inlinguistic precedents. Journal of Memory and Language, 46, 391–418). Recently, Metzing and Brennan (2003) (Metzing,C., & Brennan, S. E. (2003). When conceptual pacts are broken: partner-specific effects on the comprehension of refer-ring expressions. Journal of Memory and Language, 49, 201–213) reported a speaker-specific effect for broken prece-dents that suggests early use of speaker information when precedents are broken. Results from two eyetrackingexperiments show that this speaker effect results from the late use of speaker information to recover from an early, part-ner-independent preemption effect. When a new description is heard, existing precedents preempt the mapping of thenew description to an old referent. Later, listeners use speaker-information to inhibit precedents that are not known tothe current speaker. Time-course data, as well as the results of a cognitive load manipulation, suggest that the preemp-tion and speaker effects are supported by distinct processing systems. Our findings indicate that certain pragmatic effectsin language comprehension are based on general expectations about language use, rather than assumptions about thebeliefs and goals of particular speakers.� 2006 Elsevier Inc. All rights reserved.

Keywords: Language comprehension; Pragmatics; Common ground; Referential communication; Eye-tracking

Introduction

An important topic in psycholinguistic research ishow comprehenders integrate linguistic and contextualinformation to interpret conversational references.

0749-596X/$ - see front matter � 2006 Elsevier Inc. All rights reserv

doi:10.1016/j.jml.2006.05.002

q We thank the TalkLab undergraduate research assistantswho assisted in collecting the data for these experiments: LillianFarjeat, Karla Gonzalez, Alejandro Manzo, Wajeeha Munir,Vanessa Olguin, Jeff Olney, Marvin Sesuca, and Erica Severan.

* Corresponding author.E-mail address: [email protected] (E. Kronmuller).

Speakers and listeners face uncertainty in processingreferring expressions because any given object can becategorized in multiple ways, each corresponding to adifferent linguistic encoding; for example, a given bookcan be referred to as the book, the thing on the table,the large object, and so on. In dialogue, where referentsare often referred to multiple times, interlocutors canreduce some of this uncertainty by establishing tempo-rary conventions regarding how particular referents areto be described (Barr & Keysar, 2002; Brennan & Clark,1996; Clark & Wilkes-Gibbs, 1986; Garrod & Anderson,1987). These conventions, known as conceptual pacts

ed.

mailto:[email protected]

E. Kronmuller, D.J. Barr / Journal of Memory and Language 56 (2007) 436–455 437

(Brennan & Clark, 1996) or as precedents (Barr & Key-sar, 2002), impact how people produce and comprehendreferring expressions.

Precedents are important for theories of language usebecause research indicates that consistency in namingcan overrule other factors in determining how a referentis to be described (Brennan & Clark, 1996). Speakers donot simply choose a description for a referent based sole-ly on the referent’s similarity to a given category, buttend to favor descriptions that they have used in thepast. Evidence for this can be seen in the phenomenonof entrainment (Garrod & Anderson, 1987). Throughrepeated use, an expression comes to designate a partic-ular referent more and more rigidly, progressively takingon more name-like properties (Carroll, 1980). Entrain-ment has been shown to influence speakers’ lexicalchoices: although speakers normally refer to objects ata ‘generic’ or basic-level (Cruse, 1977; Rosch, Mervis,Gray, Johnson, & Boyes-Braem, 1976), they sometimeswill overspecify a referent to maintain consistency withan established precedent (Brennan & Clark, 1996). Thus,a speaker who has habitually referred to a shoe using thespecific term loafer is likely to continue using the specificterm loafer rather than shoe, even when no other shoe ispresent in the set of contextually given objects.

Precedents have important implications for languagecomprehension as well, because listeners expect consis-tency in speakers’ descriptions, and these expectationsguide how listeners search for referents (Barr & Keysar,2002; Metzing & Brennan, 2003). Past research onprecedent use in language comprehension has observedtwo key effects. First, precedents benefit comprehensionby reducing uncertainty and making processing moreefficient. When a speaker follows a precedent, listenerscan access the appropriate referent from memory,thereby circumventing an effortful search of context(Barr & Keysar, 2002). Second, when a precedent isbroken—that is, when an old referent is described ina new way—comprehension is impaired (Metzing &Brennan, 2003).

In this article, our general aim is to investigate thecognitive basis of precedent effects in language compre-hension. Does precedent use involve assessments ofshared knowledge, or can it be explained on the basisof simpler kinds of processing that do not involve suchassessments? More specifically, do listeners interpretspeech against the background of precedents establishedwith the current conversational partner, or do theyaccess any precedents that are available, regardless ofwho established them?

Precedent use in comprehension: Partner-specific or

partner-independent?

One possible explanation for how precedents benefitcomprehension assumes that listeners access a

specialized body of knowledge known as common ground

(Clark & Carlson, 1981; Clark & Marshall, 1981), the setof information that interlocutors share and know thatthey share. Precedents form part of the common groundbetween interlocutors by virtue of their linguistic co-pres-

ence; that is, because the speaker and listener were bothpresent when the precedent was established. Commonground theory assumes that listeners constrain the setof information they consider during comprehension totheir common ground with the speaker (Clark & Carl-son, 1981). Under this explanation, the benefit of a par-ticular precedent should be partner-specific: it should beobserved when listeners interpret speech from the speak-er who established the precedent, but not when theyinterpret speech from another speaker who lacks knowl-edge of the precedent.

Barr and Keysar (2002) tested this prediction in a ser-ies of eyetracking experiments. In their Experiment 2,listeners played a referential communication game inwhich they rearranged objects in a grid based on spokeninstructions from two speakers. These objects were unfa-miliar and lacked conventional names. Listeners’ eyemovements were monitored as they interpreted certaintest instructions from a speaker in which a target objectwas mentioned. Barr and Keysar manipulated two vari-ables: (1) whether or not a precedent had been previous-ly established for the target object; and (2) whether theprecedent was in common ground with the currentspeaker. To examine the benefit of the precedent, Barrand Keysar measured the latency of listeners’ first fixa-tion to the target object from the onset of the referringexpression. They found that listeners were faster to iden-tify referents when precedents existed than when theydid not, but this benefit was just as large when the prec-edent was in common ground as when it was not. Thissupported the idea that precedent use is partner-inde-pendent: it is based on the availability of the precedent,and not on common ground. Barr and Keysar (2002)suggested that a listener’s use of precedents that arenot in common ground could be a systematic source ofmisunderstanding in conversation, since there is noguarantee that two people who use the same term areusing it to refer to the same referent.

Metzing and Brennan (2003) suggested that animportant additional test case can be found in the com-prehension of broken precedents, wherein a new descrip-tion is used to refer to an old referent. Like Barr andKeysar (2002), Metzing and Brennan conducted an eye-tracking experiment in which participants interpretedreferences from two different speakers. They manipulat-ed whether an old precedent was broken or maintainedin referring to a target object, as well as the identity ofthe speaker referring to the target. In one condition,the speaker who referred to the target was the samespeaker who had originally established a precedent forthat object. In another condition, the speaker was

438 E. Kronmuller, D.J. Barr / Journal of Memory and Language 56 (2007) 436–455

different from the one who established the precedent.When the old precedent was maintained, no speakereffect was found, replicating Experiment 2 of Barr andKeysar (2002). However, when the precedent was bro-ken, there was a significant effect of speaker, with listen-ers taking longer to initially fixate the target whenspeakers broke their own precedents than when newspeakers broke previous speakers’ precedents.

Preemption and recovery from preemption in interpreting

new expressions

At the heart of the speaker effect obtained by Met-zing and Brennan is the phenomenon of preemption,according to which an existing association between agiven linguistic form and a given meaning preemptsthe use of a new linguistic form to denote that samemeaning (Clark & Clark, 1979; Clark, 1991). Preemp-tion is a principle with broad scope in language use.Clark and Clark (1979) originally discussed its rolein the creation of denominal verbs, in which a novelverb is created from a noun (e.g., Jana gurneyed the

patient, where to gurney refers to the act of puttinga patient in a gurney). They suggested that preemptionwill influence which denominalized verbs languageusers will find acceptable and which they will not.For instance, because of preemption, language userswould not accept denominalized use of the noun hos-

pital (e.g., to hospital a patient) to refer to the act ofputting a patient into a hospital: the existence of theconventional verb hospitalize preempts the creationof a new form to refer to this same act. Preemptionalso plays a role in language acquisition (Clark,1990): children tend to avoid mapping a novel wordto an object with an established name (Markman &Wachtel, 1988).

Preemption might also operate at the level of conver-sational precedents, leading listeners to associate newdescriptions with new referents. One possible explana-tion for the speaker effect obtained by Metzing andBrennan—that is, for the finding that listeners compre-hended a broken precedent faster when the speakerbreaking the precedent was new rather than old—is thatpreemption is modulated by common ground, a hypoth-esis that we call partner-specific-preemption. Thishypothesis assumes that only those precedents that arein common ground with the current speaker can give riseto preemption. Under this hypothesis, if a speaker firstreferred to a given referent as the shiny thing and thenlater called it the metal object, comprehension of the met-

al object should be impaired because of preemption bythe precedent shiny thing. In contrast, comprehensionshould not be impaired if the speaker calling the referentthe metal object is not the same as the one who previous-ly called it the shiny thing, because the precedent is not incommon ground.

However, there is an alternative explanation for thespeaker effect for broken precedents, which we call therecovery-from-preemption hypothesis. This hypothesisassumes that listeners initially experience preemptionfrom any available precedents, not just those in commonground, and localizes the speaker effect to a later ‘recov-ery’ process. The initial preemption from availableprecedents will lead the listener to attempt to map thenew description to any new (i.e., previously unmen-tioned) object. At some point in the search, listenersmay realize that the new description fits an old referentbetter than any new referent, causing them to search fora reason why the precedent was broken. In effect, listen-ers must answer the question: why did this speakerdescribe an old referent in a new way? Listeners wouldbe able to answer this question more quickly in the caseof a new speaker than an old speaker, because in theformer case, the old precedent is not in common groundwith the current speaker; therefore, there is no reasonwhy the speaker should follow the precedent. In the caseof an old speaker, listeners would experience difficultyreconciling the speaker’s breaking of his or her ownprecedent with the assumption of cooperativeness. Asa result, listeners would be slower to recover from pre-emption when speakers break their own precedents thanwhen speakers break the precedents of other previousspeakers. We emphasize that recovery-from-preemptionis not intended as a special-purpose mechanism, butrather is assumed to reflect the operation a more generalmonitoring process that uses common ground (Keysar,Barr, Balin, & Brauner, 2000).

That Metzing and Brennan (2003) obtained thespeaker effect as early as the first look to the targetobject might seem to rule out the possibility that itwas due to a late-onset recovery process. However, thisassumption is not warranted: the latency of fixation tothe target object is too coarse of a measure to supportstrong inferences about the time course of processing.Metzing and Brennan reported a mean latency in theold partner-new expression condition of 1253 ms, witha standard deviation of 465 ms. Assuming a normal dis-tribution, 75% of the distribution would fall between theinterval from 718 to 1787 ms. Clearly, such an intervalwould contain many fixations that are relatively ‘late.’The observed differences between the two conditionscould be driven entirely by the ‘late’ segment of this dis-tribution, which might tap into recovery effects, whilethe ‘early’ segment, if considered alone, would showno such effects.

Clearly, distinguishing between these two hypothesesrequires considering detailed time-course information.To this end, we conducted two eyetracking experiments.In these experiments, listeners’ eye movements weremonitored as they interpreted descriptions of objectsfrom either of two speakers (a male and a female) thatlacked common ground with one another. At certain


points, listeners heard a test description of a previouslymentioned target object (e.g., the metal object). Wemanipulated the factor of Precedent, corresponding towhether the precedent that had been established forthe target object was consistent (Maintain condition)or inconsistent (Break condition) with the test descrip-tion. For instance, a target object called the metal object

as the test description would previously have been calledthe metal object in the Maintain condition and the shiny

thing in the Break condition. We also manipulated com-mon ground via the factor of Speaker, corresponding towhether the speaker who established the precedent wasthe same (Old Speaker condition) or different from(New Speaker condition) the speaker who provided thelater test description. Experiment 2 also included a cog-nitive load manipulation, which was intended to test thehypothesis that preemption effects and recovery effectsare supported by distinct processing systems. In bothexperiments, we analyzed fixations to the target objectover a series of 300 ms time windows, starting at theonset of the test description.

The specific predictions of the two hypotheses withrespect to the time course data are as follows. First,the effect of preemption would be revealed as a lowerproportion of fixations to the target object in the Breakcondition, in which a new description is provided forthe target object, than in the Maintain condition, wherean established precedent is followed. According to thepartner-specific-preemption hypothesis, listeners wouldonly experience preemption in the Old Speaker condi-tion, in which the violated precedent is in commonground. In the New Speaker condition, there shouldbe no such preemption. Thus, in the Break conditionthere should be a speaker effect present from the earli-est moments of comprehension, consisting of higherfixations to the target in the New Speaker conditionthan in the Old Speaker condition. This speaker effectshould not be present in the Maintain condition, repli-cating Barr and Keysar (2002) and Metzing and Bren-nan (2003). Thus, this hypothesis specifically predictsan interaction between Precedent and Speaker thatshould be present from the earliest moments ofcomprehension.

In contrast, the recovery-from-preemption hypothe-sis predicts that listeners will experience preemptionfrom precedents any time a new expression is used, inde-pendently of who established the precedent that is beingviolated. Thus, it predicts an early main effect of Prece-dent, with lower fixations to the target in the Break con-dition than in the Maintain condition. Critically, thismain effect should initially appear in the absence ofany Speaker by Precedent interaction. The hypothesisalso predicts that a Speaker by Precedent interactionshould emerge in the data, but only once the recoveryprocess begins. As with the partner-specific-preemption

hypothesis, this interaction should be driven by a speak-er effect localized to the Break condition, with no sucheffect in the Maintain condition.

In sum, both hypotheses predict a Speaker by Prece-dent interaction driven by the Break condition, in whichfixations to the target would be higher when the speakeris new than when the speaker is old. However, thehypotheses differ with respect to the predicted timecourse of this interaction. If the first appearance of aSpeaker by Precedent interaction is later than a maineffect of Precedent, then the recovery-from-preemptionhypothesis would be supported. In contrast, if its firstappearance is prior to, concurrently with, or in theabsence of a main effect of Precedent, then the part-ner-specific-preemption hypothesis would be supported.

Experiment 1

In this experiment, participants’ eye movements weretracked as they followed instructions to move pictures ofobjects around a grid displayed on a computer screen(see Fig. 1). All of the objects were unfamiliar andlacked conventional names.

Although intended as a replication of Metzing andBrennan (2003), the experiment incorporated severalmodifications to the design. First, as in Experiment 3of Barr and Keysar (2002), we used pre-recorded expres-sions rather than live speakers, so that all listeners wouldhear the same utterance token for each test instruction.Listeners were led to believe that they were hearingrecorded speech from two previous speakers who playedthe communication game on two different occasionsusing the same materials, with different partners playingthe role of listener.

The use of pre-recorded materials made it possible toreduce variability and therefore gave us greater power todetect small effects. A further reason for the use of pre-recorded materials was the concern that live confeder-ates might introduce confounds. It did not seem possibleto conduct the experiment in a manner such that theconfederates would be blind to the experimental condi-tion, which would be necessary to ensure that theywould produce the target instruction consistently acrossconditions and not, for example, unwittingly hesitate forlonger when breaking their own precedents.

Second, unlike Metzing and Brennan (2003), we keptthe test instruction constant across conditions andmanipulated the precedent that was established priorto the test instruction. Likewise, we also kept the speak-er of the test instruction constant and manipulated theidentity of the speaker who established it (see Fig. 1).This means that participants heard the same utterancetoken as the test instruction in every condition. There-fore, any differences between the conditions could not

Fig. 1. Examples of test-trial displays from Experiment 1 (left) and Experiment 2 (right). The target object appears in the lower-leftcorner of the displays. The text in normal typeface represents speech by the male speaker; the text in italics represents speech by thefemale speaker.


be attributed to differences in the stimuli itself, but onlyto the effects of the experimental variables.

Finally, in Metzing and Brennan (2003) all objects inthe grid had been previously mentioned prior to the crit-ical test instruction. Thus, when the speaker who hadnamed all of the objects suddenly used a new expression,the only conclusion that listeners could have drawn wasthat a precedent was being broken. In contrast, prior tothe test instruction in our experiment, we introduced anobject that had not yet been mentioned (the unmentioned

object). This object was introduced into the grid severalinstructions before the test instruction so that it wouldbe familiar at test, although it remained unmentionedby the speaker. Importantly, this object did not matchthe description of the target object.

Method

Participants

Fifty-two undergraduate students from the Uni-versity of California, Riverside participated in this

experiment in exchange for credit in an introducto-ry psychology course. Six additional participantswere excluded from the analysis, five due to poorcalibration and one due to a failure to followinstructions.

Design

The design included two two-level factors: Speaker(Old, New) and Precedent (Break, Maintain). These fac-tors were administered within-participant.

Materials

Eight arrays of pictures of objects were used (seeFig. 1 for an example). The pictures were of unusualobjects that were collected from various web sites onthe Internet. The picture files were converted to170 · 170 pixel black and white bitmap files. Each gridcontained eight pictures. The computer image for theentire screen had a resolution of 1024 · 768 pixels. AnLCD projector was used to project the image onto a wallin the laboratory.


The spoken stimuli were recorded in two sessions of aproduction version of the experiment. The spokeninstructions involved moving a target object from itscurrent location to a destination location on the grid.The destination was usually specified relative to the loca-tion of another object in the grid, e.g., ‘‘put the shinything below the watch-like object.’’

The two speakers were recorded in separate sessions.In these sessions the speaker (a naıve participant) and alistener (a research assistant) performed a referentialcommunication task. We gave the speakers a diagramof the grid with all the movements they were supposedto ask the listener to perform. They were allowed to referto the objects using any name they chose, except the sin-gle target object in each grid, for which we asked them touse a specific name (which we chose from a set of namesspontaneously provided in an informal survey of 15undergraduate students). We asked the speakers to tryto produce the names given to them as though theyhad spontaneously generated them. To get two differentexpressions for each target, we instructed the malespeaker to use one expression and the female speakerto use another expression. Each instruction was record-ed into a separate sound file.

To create the four cells in the design, it was necessaryto record additional speech. To this end, we made a sec-ond set of recordings for each speaker, for which theywere asked to attempt to imitate the other speaker’sinstructions. By combining the spontaneous instructionsand the imitations, we were able to create the four exper-imental conditions in which the instructions for the firstthree rounds were varied across conditions, while thetest instruction in the last round was held constant (seeFig. 1). During the recording sessions, each speaker lis-tened to each recording from the other speaker severaltimes, and was asked to produce their own imitationof the speech. We asked them to adopt the name usedby the speaker, but discard all the other words and tryto express it in their own way, as well as to attempt toconvey the same level of certainty or uncertainty con-veyed in the original expression. The use of an imitationversus an original expression as the test instruction wasnearly balanced, with five of the eight items using an imi-tation and the other three using an original expression.

We created four stimulus lists by rotating each itemthrough its four conditions across lists. Thus, each par-ticipant saw each item only once, and saw two items ineach of the four conditions created by factorially com-bining the two-level variables of Speaker and Precedent.

Apparatus and data collection

We used a ISCAN ETL-400 remote eyetracker thatwas placed on a table in front of the participant. Theeyetracker provides information about the participant’spoint of gaze at a sampling rate of 60 Hz. The experi-ment software time-stamped the eye data relative to

the onset of the test instruction, and stored the data ina database.

Procedure

Participants were tested individually in a quiet room,seated facing the wall-projected image. They sat in achair with an adjustable headrest to partially stabilizethe head.

Participants read a description of the experimentalprocedure that appeared on the computer screen. Thedescription stated that the purpose of the experimentwas to examine the comprehension of instructions fromdifferent speakers. Their task would be to move picturesof objects around a grid, following the instructions fromtwo different speakers, a male and a female, who wereprevious participants in the experiment. The idea thatthe speakers had no connection to each other wasemphasized by telling the participants that the speakersparticipated in the experiment on separate occasions.Next, the eyetracking apparatus was calibrated.

The task in each trial was to move pictures around a4 · 4 virtual grid generated by a computer. Eight differ-ent pictures of objects appeared in the different slots ofthe grid. We used a free-viewing paradigm in which lis-teners were allowed to freely view the objects in the dis-play, instead of starting each trial from a central fixationpoint. The participant heard the speakers’ instructionsthrough a loudspeaker. Participants used a control padto move a cursor and select objects.

If the participants selected the wrong object or thewrong location, a pre-recorded voice (different fromthose of the speakers) announced that they had madean error, and asked them to listen once more to theinstructions and try again. Then the same instructionwas played. If they made a mistake the second time, thenthe computer moved the correct object to the correctlocation and played the next move. In total, there weresixteen instructions for each set of objects. These sixteenmoves were divided in four rounds of four moves each,similar to the scheme used in Metzing and Brennan(2003). At the beginning of each round, the positionsof the objects in the grid were shuffled.

The schema of four rounds and four moves for eachround allowed the use of the same expression for refer-ring to the same object across rounds (i.e., the establish-ment of the precedent). If there was a speaker switch, itoccurred during the first move of the last round. Thecritical test instruction was always delivered as the sec-ond or the third move of the last round.

Analysis

In the analysis below, we report one off-line measure,the rate of non-target selection, as well as one on-linemeasure, an analysis of fixations to the target object.The rate of non-target selection is defined as the rateat which listeners selected some object other than the


target as the referent for the test description. We did notexclude such cases from our analyses because they arepart and parcel of the phenomenon of preemption thatwe are exploring. A listener might select an object otherthan the target to avoid assuming that the speaker brokea precedent (and was therefore behaving uncooperative-ly). Thus, it would be misleading to consider such cases‘errors.’ Unfortunately, due to programming limita-tions, we were unable to obtain more detailed informa-tion about which non-target object was eventuallyselected (e.g., whether it was the unmentioned object,or some other already-mentioned object). Hence, weonly report the non-target selection rate.

Fixations on various regions were determined byquerying the experiment database, relating the timingof the onset of the referring expression within eachsound file to the timing of the eye movements, and relat-ing the position of the point of gaze to the location ofthe target picture on the screen. A human coder wasinvolved only in identifying the onset of the referringexpression within each sound file for each test instruc-tion—which, as indicated above, was the same for allconditions. Thus, it is not possible that any of our anal-yses would be subject to the biases that a human codermight introduce.

To gain insight into the time course of processing, weexamined fixations to the target object. As speechunfolds, listeners will become progressively more certainabout the identity of the target object. Correspondingly,they will increasingly spend a greater proportion of timelooking at the target object and looking away from non-target objects. Thus, one way of evaluating the timecourse of processing is to compare the proportion of fix-ations to the target object during a given time windowto the proportion of fixations to any other object inthe grid. To this end, for each test trial we computeda ‘‘target advantage’’ score for each time window thatwe analyzed. The target advantage score is defined asthe proportion of time spent fixating the target over agiven time window minus the average proportion oftime spent fixating any non-target object. Similar mea-sures have been used in other published studies thatinclude a time-course analysis (e.g. Nadig & Sedivy,2002).

The target advantage score is a more informativemeasure of processing than just the proportion of fixa-tions on the target, because it scales the amount of timespent looking at the target relative to any other object.In the current study, listeners were required to controlthe movement of a cursor, and thus would spend someamount of time looking at the cursor itself or at the des-tination of the movement rather than at any given objectin the grid. Therefore, it was desirable to compute ameasure that would reflect preferential looking at thetarget, controlling for the likelihood of fixating any giv-en object in the display. The target advantage score

takes on values between �1 and +1 with a score of 0meaning that the participant was equally likely to lookat the target as any other object; a positive score reflectsa greater likelihood of fixating the target; and a negativescore reflects a greater likelihood of fixating non-targetobjects. Note that the non-target objects in each gridconsisted of six mentioned objects and one unmentionedobject. To make the proportion for non-target objectscomparable to the target, for each time window and tri-al, the proportions for each of the seven objects wereaveraged and the resulting value subtracted from theproportion of fixations to the target for that time win-dow and trial.

Because of our interest in both early and late compo-nents of processing, we considered fixations over a largetime span: a series of seven 300 ms time windows start-ing at the onset of the referring expression and endingat 2100 ms. It is assumed that the planning of an eyemovement requires approximately 180 ms (Matin, Shao,& Boff, 1993). Therefore, the first time window spanned180–300 ms after the onset of the expression instead of0–300 ms. There was considerable variability across tri-als in the timing of the selection of the target, raisingthe question of how to handle cases of an ‘‘earlyresponse,’’ wherein a response was entered before theend of the last window (this occurred in 27 cases outof 416, or 6%). In the experiment, once the listenerentered a response, the screen was cleared, recordingof eyetracking data was halted, and the trial ended; thus,a trial with an early response would not contribute datato the later time windows. So that these cases would notbe excluded, we made the target advantage score ‘‘cumu-lative to target’’: from the point at which the listenerpressed the response button and ended the trial, theremaining part of the window in which the responseoccurred, as well as all subsequent windows, were treat-ed as though the listener remained fixated on the target(except in cases where an object other than the targetwas selected).

Results and discussion

For all analyses, we conducted a series of 2 · 2 with-in-factor ANOVAs treating participants and items asrandom factors (F1 and F2, respectively).

Rate of non-target selection

Overall, participants selected a non-target object asthe referent instead of the target in 40 out of 416 trials(9.6%). Preemption predicts that when a precedent isbroken, listeners should be less likely to select the targetobject because another name has already been estab-lished for that object. Thus, the non-target selection rateshould be higher in the Break than in the Maintain case,even though listeners heard the exact same test descrip-tion in both conditions. Furthermore, if common


ground interacts with preemption (either by directlymodulating the effect or enabling recovery), then thereshould be an effect of Speaker in the Break condition.Specifically, listeners should be more likely to select anon-target object in the Old Speaker-Break conditionthan in the New Speaker-Break condition. In contrast,there should be no effect of Speaker in the Maintain con-dition, which would be consistent with the finding thatcommon ground does not modulate the processing ofprecedents when they are maintained (Barr & Keysar,2002; Metzing & Brennan, 2003).

The observed rates of non-target selection were asfollows: 20% in the Old Speaker-Break condition, 14%in the New Speaker-Break condition, 3% in the OldSpeaker-Maintain condition, and 1% in the New Speak-er-Maintain condition. In support of preemption, listen-ers selected the target less often when a precedent wasbroken, as evidenced by a main effect of Precedent,F1(1,51) = 31.086, MSe = .040, p < .001; F2(1,7) =7.344, MSe = .026, p < .05; minF 0(1,11) = 5.94, p < .05.The main effect of Speaker was not significant,F1(1,51) = 1.805, MSe = .043, p = .19; F2(1,7) = .848,MSe = .014, p = .39; minF 0(1,15) = .577, p = .46.Although the pattern of means suggests a 6% higher rateof non-target selection in the Old Speaker-Break condi-tion than in the New Speaker-Break condition, thehypothesized Speaker by Precedent interaction was notreliable, F1(1,51) = .567, MSe = .034, p = .455;F2(1,7) = .219, MSe = .014, p = .65; minF 0(1,13) = .158,p = .70.

In sum, listeners occasionally selected a non-targetobject at a high rate in the Break condition, even though

Fig. 2. Mean Target Advantage Score By Condition, Experiment 1. Ththe two Speaker conditions for each level of Precedent (Break, Ment

they heard the same test description as listeners in theMaintain condition. However, although the pattern ofmeans was in the direction of partner-specificity, the pre-dicted interaction was not observed, suggesting thatcommon ground had only a weak effect on listeners finaljudgments.

Target advantage score

Although the analysis of non-target selection rate didnot offer evidence for the use of common ground withbroken precedents, it is possible that a more sensitivemeasure, such as the target advantage score, mightreveal such effects.

The partner-specific-preemption and recovery-from-preemption hypotheses both predict an interactionbetween Speaker and Precedent, in which the targetadvantage score should be higher in the New Speaker-Break condition than in the Old Speaker-Break condi-tion. However, whereas partner-specific-preemption pre-dicts that the interaction should appear from the earliestmoments of comprehension—i.e., simultaneously withthe earliest visible effects of Precedent—recovery-from-preemption predicts that the speaker effect shouldemerge later than any main effect of Precedent.

The mean target advantage scores by condition andtime window are presented in Fig. 2. Before presentingthe full statistical analysis, we first describe the differenc-es that are apparent on the graph. Note that up to the1200–1500 ms window, the only apparent differencesare between the Break and Maintain conditions. Thatthe effect of Precedent is not modulated by an effect ofspeaker supports the hypothesis that early preemption

e error bars represent the standard error of a difference betweenion), as computed from the analysis by Participant.


is independent of common ground. The hypothesizedSpeaker by Precedent interaction does eventuallyappear, but only at the 1500–1800 and 1800–2100 mswindows, where, as predicted, fixations to the targetare higher in the New Speaker-Break condition than inthe Old Speaker-Break condition. This pattern of resultsseems to support the recovery-from-preemptionhypothesis.

These observations were largely corroborated by thestatistical analyses. We submitted the set of mean targetadvantage scores to a series of ANOVAs by participantand by item. Because of the large number of analysesperformed, we report the complete set of analyses inTable 1, and in the main text we focus on those analysesthat resulted in significant or marginally significanteffects by participants or by items.

Table 1Mean square error and F test on the target advantage score for Expe

Time window Participants

Source of variance MSe F1

180–300 ms

Speaker .029 .01Precedent .029 .61Speaker · Precedent .029 .04

300–600 ms

Speaker .033 .03Precedent .035 2.32Speaker · Precedent .033 .28

600–900 ms

Speaker .049 1.06Precedent .041 8.86**

Speaker · Precedent .057 .95

900–1200 ms

Speaker .080 .15Precedent .127 6.01*


1200–1500 ms

Speaker .117 .31Precedent .088 3.46�


1500–1800 ms

Speaker .096 3.94*

Precedent .080 .00Speaker · Precedent .141 2.14

1800–2100 ms

Speaker .102 .20Precedent .093 1.45Speaker · Precedent .122 4.06*

Note. Degrees of freedom (df) for all effects = 1; Error df for all Particfor MinF0 in the table.

� p 6 .10.* p 6 .05.

** p 6 .01.

The main effect of Precedent was significant by par-ticipants and items in the 600–900 and 900–1200 ms win-dows. The Speaker by Precedent interaction wasmarginally significant by item, but not by participant,in the 1500–1800 ms window, and was significant byboth participant and item in the last time windowobserved, 1800–2100 ms. The early effect of Precedentin the absence of any interaction with Speaker confirmsthe prediction of the recovery-from-preemption hypoth-esis that preemption is experienced even from precedentsthat are not part of the common ground with the currentspeaker. The relatively late onset of the Speaker by Prec-edent interaction suggests that listeners consult commonground in the course of recovering from the preemptioneffect. Given that longest test description used in theexperiment lasted 1324 ms (the mean length for all 8 test

riment 1

Items MinF0

MSe F2 Df MinF0

.003 .07 1,58 0.01

.003 1.04 1,37 0.38

.002 .09 1,44 0.03

.000 .05 1,36 0.02

.012 3.50� 1,34 1.40

.001 .34 1,29 0.15

.006 1.30 1,29 0.58

.011 5.26* 1,17 3.30�

.005 1.74 1,38 0.61

.007 .26 1,37 0.10

.012 9.56* 1,35 3.69�

.005 .89 1,53 0.20

.013 .41 1,30 0.18

.037 1.26 1,13 0.92

.015 .00 1,7 0.00

.020 2.95 1,20 1.69

.030 .00 1,51 0.00

.011 4.40� 1,41 1.44

.021 .15 1,20 0.09

.028 .75 1,16 0.49

.014 5.37* 1,30 2.31

ipant analyses = 51; Error df for all item analyses = 7; Error df


descriptions was 880 ms), the speaker effect does notappear until after the entire test instruction has been per-ceived, which is quite late indeed.

Planned comparisons were conducted comparing theSpeaker effect for the Break and Maintain conditions forthe two windows in which a significant interaction or amarginally significant interaction were found. In the1500–1800 ms window, listeners in the New Speaker-Break condition (M = .37, SD = .39) were more likelyto look at the target than listeners in the Old Speaker-Break condition (M = .21, SD = .33). This differenceexceeded the 95% confidence interval for the differencebetween means for the participant analysis(CI = ± .15), but was marginal in the item analysis(CI = ± .18). In contrast, during this same window therewas no significant effect of Speaker for the two Maintainconditions; the means were .29 (SD = .30) in the NewSpeaker-Maintain condition and .28 (SD = .33) in theOld-Speaker Maintain condition, a difference that wascontained within the 95% confidence intervals for a dif-ference between means by participant (CI = ± .12) andby item (CI = ± .09).

The pattern of means for the 1800–2100 window werenumerically in the predicted direction, with higher fixa-tions to the target in the New Speaker-Break than inthe Old Speaker-Break condition. The means were .32(SD = .41) in the New Speaker-Break condition com-pared to .20 (SD = .31) in the Old-Speaker Break condi-tion. However, this difference was not reliable (95%CI = ± .14 by participants and 95% CI = ± .15 byitems). As expected, the speaker contrast for the Main-tain condition was not significant; the means were .27(SD = .31) in the New Speaker-Maintain condition ver-sus .35 (SD = .32) in the Old Speaker-Maintain condi-tion, a difference that was contained within the 95%confidence intervals by participants (CI = ± .12) andby items (CI = ± .16).

In sum, the current study replicated two main find-ings from previous studies: first, as in Barr and Keysar(2002), we found a benefit for maintained precedentsthat was independent of the identity of the speaker; sec-ond, like Metzing and Brennan (2003) we found a speak-er effect for broken precedents. The current experimentindicates that this speaker effect has a later onset thanthe effect of preemption, and therefore supports therecovery-from-preemption hypothesis. This can be seenmost clearly in Fig. 2, in the later divergence of the tar-get advantage scores for the two Old Speaker-Break andNew Speaker-Break conditions relative to the earlydivergence of the Break versus Maintain conditions, col-lapsed over Speaker. Thus, we suggest that precedentsare used independently of common ground, even whenthey are broken. Listeners experience preemption when-ever they hear a new description for an old referent, evenfrom precedents that are not in their common groundwith the current speaker. The speaker effect for broken

precedents reflects the operation of a late-onset monitor-ing process that uses speaker-specific information torecover from preemption, and which comes to comple-tion faster in the case of a new speaker.

Even though the hypothesized Speaker by Precedentinteraction was obtained in this experiment, and the pat-tern of means was numerically in the predicted direction,the effects were not fully robust statistically. The speakereffect may have been present during the last two timewindows, but was swamped by variability arising fromthe complex nature of the task. Recall that listenershad to search for the target from a set of eight possibleobjects, and follow instructions of the form ‘‘put thething that looks like a stereo below the flying-saucer-likeobject.’’ Because the task was so time-consuming, wewere only able to fit eight test items within a one-hoursession, which resulted in limited statistical power. Inaddition, there would be substantial difficulty in locatingthe target object given the large number of objects toconsider, especially when the description was new.Finally, further variability would be added to the targetmeasure due to the need to locate a reference object(e.g., ‘‘the flying-saucer-like object’’), relative to whichthe target object was to be placed. Listeners would pre-sumably begin to search for this reference object afterthey had identified the target, which would result in adrop in fixations to the target. This drop could occurearlier in those conditions where the target descriptionwas comprehended more easily (i.e., the two Maintainconditions). Support for this possibility is suggested bya drop in the Maintain conditions at around 1500–1800 ms.

Although the experimental task was designed to mir-ror previous studies on precedent use in comprehension,some reduction of task complexity would be desirable togain a clearer picture of processing. To this end, weconducted Experiment 2, which used a similar experi-mental design, but with a task that was drasticallysimplified.

A more central goal of Experiment 2 was to explorethe nature of the precedent and speaker effects obtainedin the current experiment. Based on the different time-courses observed for these effects, we hypothesized thatthey reflect the operation of functionally distinct pro-cessing systems. Specifically, the precedent effect can beviewed as the result of a memory-based, relatively auto-matic associative system that is influenced by the relativesalience or cognitive availability of information regard-less of its common ground status. In contrast, the latespeaker effect is due to the operation of a more con-trolled, inferential system that can take common ground(or any other kind of information, for that matter) intoaccount to resolve ambiguity. In this sense, our proposalis similar to the perspective adjustment view (Keysar,Barr, Balin, & Paek, 1998; Keysar et al., 2000), whichproposes a fast egocentric system coupled with a slower


monitoring and adjustment system that uses commonground.

Our proposal predicts that controlled, inferentialaspects of processing should be impaired when listenersare placed under cognitive load. Experiment 2 tested thisprediction by having participants perform a secondarymemory task while listening to test instructions. Previ-ous studies have found that cognitive load manipula-tions inhibit perspective-taking processes while sparingother aspects of language processing (Horton & Keysar,1996; Rossnagel, 2000). Our hypothesis was that thisload task should disproportionately disrupt the abilityto use common ground, making it difficult for listenersto recover from preemption, while leaving early part-ner-independent effects of precedents relatively intact.

Experiment 2

The design of Experiment 2 was similar to Experi-ment 1, except that in a subset of trials, participantshad to interpret target instructions while simultaneouslyperforming a secondary cognitive load task, whichrequired them to maintain a string of digits in workingmemory. The primary experimental task was also muchsimplified. Unlike Experiment 1, there were only threeobjects in a given display (see Fig. 1). For test trials,these objects were the (previously mentioned) target,another previously mentioned object, and an unmen-tioned object. Additionally, the listener’s task was sim-ply to select the target with a cursor controlled by agame controller, and did not involve moving the targetto another location. These simplifications enabled usto increase the number of experimental blocks to 32.Because the method was otherwise quite similar toExperiment 1, in the following section we only focuson those aspects of the experiment that differed.

Method

Participants

Fifty-six undergraduate students from the Universityof California, Riverside participated in the study inexchange for credit in an introductory psychologycourse. All were native English speakers. Two additionalsubjects were excluded from the analysis because of poorcalibration.

Design

To the Speaker and Precedent factors used in Exper-iment 1, we added the two-level factor of Load (Load,No Load), which was administered within-participant.

Materials

There were 32 items, each of which consisted of ablock of displays. Each display contained three pictures

that were drawn from a larger set of five differentobjects. The five pictures for each block were uniqueto that block, and not used in any other block. The threeobjects in each display were presented in a triangulararrangement.

To produce the recorded instructions, we used a pro-cedure similar to Experiment 1, except that we did notask the naıve speakers to name target objects in a pre-de-termined way, but allowed them to produce their ownspontaneous referring expressions. The mean lengthfor the 32 referring expressions was 647.81 ms.

The stimuli for the secondary task consisted of 256random strings of 6 digits from 0 to 9. 64 of these werestrings to be remembered and the remaining 192 wereused as distractors in the recognition test that followedeach test trial.

Eight stimulus lists were created by rotating the 32experimental items. Each participant saw each item inone of the eight conditions, and saw exactly four itemsin each of the eight cells created by factorially combiningSpeaker, Precedent, and Load.

Procedure

Each block consisted of a sequence of 10 instructionsto click on one of three objects presented on the screen.Over the 10 instructions, the target as well as one fillerobject were each mentioned three times, two other fillerobjects were mentioned twice, and a fifth object was nev-er mentioned. Up to and including the seventh instruc-tion, only one of the two speakers gave instructionsand established precedents. The next three instructionswere delivered either by same speaker who establishedthe precedents (the Old Speaker) or a different speaker(the New Speaker). For half of the blocks, the testinstruction took place as the first of these three instruc-tions. For the other half, it was the second instruction ofthe three.

For trials in which there was cognitive load, a stringof six digits was presented for 2500 ms on a screen thatwas otherwise blank. The test display containing thethree items then appeared, and immediately afterward,playback of the sound file began.

As soon as participants selected the object, theirmemories for the digit string were tested in a recognitiontask, in which the target digit string was presentedamong three other six digit distractor strings. Therewas an additional ‘filler’ memory load task during oneof the first seven instructions. This filler task helpedobscure the identity of the test trial.

Analysis

In addition to the analyses pursued in Experiment 1,we include two additional analyses, with the goal ofdetecting any performance tradeoffs between the twotasks. Specifically, we analyzed error rate on the second-ary memory task as well as response time on the primary

Table 2Mean non-target selection rate for Experiment 2 (Standarddeviation in parentheses)

Precedent No Load Load Collapsed

Old Speaker

Break .13 (.33) .10 (.30) .11 (.32)Maintain .01 (.09) .01 (.09) .01 (.09)Collapsed .07 (.25) .05 (.23) .06 (.23)

New Speaker

Break .05 (.22) .08 (.27) .06 (.25)Maintain .01 (.09) .01 (.09) .01 (.09)Collapsed .03 (.17) .04 (.21) .04 (.19)

Marginal .05 (.22) .05 (.21)


task. Response time was defined as the latency betweenthe onset of the test description and the moment atwhich the target was selected.

Results and discussion

We removed three observations from the analysiswhere participants entered their responses before theonset of the referring expression (2 cases) or so early thatit was implausible that they had heard a sufficientamount of the referring expression (336 ms; 1 case). Asfor the previous experiment, we conducted a series of2 · 2 · 2 within-factor ANOVAs treating participantsand items as random factors (F1 and F2, respectively).

In this experiment we sought to replicate the findingfrom Experiment 1 of a late onset for the speaker effectfor broken precedents, reflected in a late Speaker byPrecedent interaction. Another goal of the currentexperiment was to test whether this speaker effect wasthe result of a cognitively effortful monitoring process.If so, then it should be strongly reduced, delayed, oreliminated by the secondary cognitive load task. Thus,we predicted a three-way interaction between Speaker,Precedent, and Load, which would be driven by a largerspeaker effect for broken precedents in the No Loadcondition than in the Load condition.

Rate of non-target selection

Listeners selected a non-target object on only 89 (5%)of 1792 test trials. In the previous experiment, there wasthe suggestion of a speaker effect for broken precedentsin the rate of non-target selection. Specifically, listenerswere numerically more likely to select a non-targetobject in the Old Speaker-Break condition than in theNew Speaker-Break condition, although the differencewas not significant. With the greater statistical powerprovided by this experiment, we expected that the differ-ence would be significant, at least in the No Load subsetof the design. We also predicted that the load manipula-tion would interfere with listeners’ ability to use com-mon ground to recover from preemption, leading themto a higher rate of non-target selection.

The data are presented in Table 2. The prediction ofa speaker effect for broken precedents was confirmed,with an overall Speaker by Precedent interaction,F1(1,55) = 4.89, MSe = .010, p < .05; F2(1,31) = 4.50,MSe = .006, p < .05; minF 0(1,77) = 2.34, p = .13. AsTable 2 shows, when a precedent was broken by theold speaker, listeners were more likely overall to selecta non-target object. Collapsing over Load, the differencebetween the Old Speaker-Break and New Speaker-Breakconditions exceeded the 95% confidence interval for thecomparison (CI = ± .04 by participants and ±.05 byitems). This effect demonstrates that listeners use com-mon ground at some point in the processing of brokenprecedents. However, the prediction that the Load

condition would affect the ability to use common groundwas not supported, given the absence of a higher-orderinteraction between Speaker, Precedent, and Load,F1(1,55) = 2.59, MSe = .001, p = .11; F2(1,30) = 1.59,MSe = .006, p = .22; minF 0(1,65) = .99, p = .32. It isworth noting, however, that the means were in thedirection of a smaller speaker effect for broken precedentsin the Load condition. Other significant effects werethe main effect of Speaker: F1(1,55) = 6.45, MSe = .014,p < .05, F2(1,31) = 5.13, MSe = .007, p < .05, minF 0(1,74) = 2.86, p = .10; and Precedent: F1(1,55) = 67.84,MSe = .010, p < .001, F2(1,31) = 11.32, MSe = .036,p < .01, minF 0(1,42) = 9.70, p = .003. There was no maineffect of Load, nor did Load interact with any otherfactor.

Clearly, then, listeners are using speaker-specificknowledge in making decisions about the identity ofthe target. Although we predicted that the ability touse this information would be impaired under load, thisprediction was not supported. It is possible that the abil-ity to recover from preemption was not eliminated dueto a task trade-off. Listeners could have allocated lesseffort to the secondary task when they experienced diffi-culty on the primary task. Additionally, listeners mayhave engaged in a speed/accuracy trade-off, in whichthey took longer to respond to the primary task whenunder load than when not, giving them more time toengage in inferential processing. The following twoanalyses address these possibilities.

Error rate for the secondary task

If listeners allocated less effort to the secondary taskin those cases where they experienced difficulty on theprimary task, then this should by reflected in a Speakerby Precedent interaction in the error rate for thesecondary task.

Overall, listeners were quite accurate at the secondarymemory task. The mean error rate in the memory taskwas .14, with a standard deviation of .09 and a rangefrom .00 to .31. The means by condition were as follows:


.18 (SD = .18) for Old Speaker-Break, .10 (SD = .15)for Old Speaker-Maintain, .14 (SD = .16) for NewSpeaker-Break, and .15 (SD = .16) for New Speaker-Maintain. The pattern of means is consistent with theidea of a trade-off, with poorest performance in theOld Speaker-Break condition, where recovery from pre-emption would be most difficult, and best performancein the Old Speaker-Maintain condition. An analysis ofthese values partly supported the prediction of a tasktrade-off, with an interaction between Speaker andPrecedent that was significant by participants but notby items, F1(1,55) = 4.23, MSe = .003, p < .05; F2(1,31) = 1.27, MSe = .004, p = .27; minF 0(1,50) = .98,p = .33. No other effects were significant.

Response time

It is possible as well that the non-target selection rateremained relatively constant across the load conditionsbecause listeners who were under load simply allowedthemselves a longer time to rehearse the digit string aswell as to complete any additional linguistic or inferen-tial processing. Such a speed/accuracy trade-off wouldbe reflected in a main effect of Load on RT.

Before submitting the mean RT scores to analysis, wefirst trimmed the top 2.5% of the distribution (5171 msand above) to remove the influence of extreme values.Overall, listeners took an average of 2182 ms(SD = 935) to select the target. The means by conditionare provided in Table 3. As predicted, there was a maineffect of Load, F1(1,55) = 13.22, MSe = 209924, p < .01;F2(1,31) = 18.46, MSe = 85867, p = .01; minF 0(1,85) =7.70, p = .007. There was also a main effect of Speaker,F1(1,55) = 9.72, MSe = 207859, p < .01, F2(1,31) =8.76, MSe = 133374, p < .01, minF 0(1,77) = 4.61, p =.04; and Precedent, F1(1,55) = 231.90, MSe = 216200,p < .01, F2(1,31) = 42.35, MSe = 674826, p < .01, minF 0

(1,43) = 35.81, p = .00. None of the interactions weresignificant. Thus, listeners were able to maintain arelatively constant rate of non-target selection across

Table 3Mean RT for target selection for Experiment 2 (Standarddeviation in parentheses)

Precedent No Load Load Collapsed

Old Speaker

Break 2573 (1175) 2837 (1751) 2705 (1495)Maintain 1771 (670) 2020 (732) 1896 (712)Collapsed 2174 (1037) 2429 (1402)

New Speaker

Break 2391 (1185) 2610 (1463) 2501 (1335)Maintain 1761 (592) 1843 (610) 1802 (601)Collapsed 2075 (987) 2227 (1184)

Marginal 2124 (1013) 2328 (1300)

conditions by taking longer to respond on the primarytask when they were under cognitive load.

Target advantage score

Although the cognitive load manipulation seems tohave not impaired listeners’ ultimate interpretations ofthe test descriptions, it may have disrupted thetime-course of inferential processing. To test this, we com-puted target advantage scores for each of seven 300 mswindows in the manner described for Experiment 1. Themean target advantage scores by condition and timewindow are plotted in Fig. 3.

First, the No Load condition (top panel) closely mir-rors the general pattern of findings in Experiment 1: anearly, partner-independent effect of Precedent is fol-lowed by an effect of Speaker in the broken precedentcondition. The main difference from Experiment 1 is inthe more rapid time course of these effects, due to thesimplification of the task. Critically, there was no changein the order in which these effects appeared: just as inExperiment 1, the new speaker advantage for brokenprecedents was preceded by a partner-independent maineffect of Precedent. In Experiment 1, the precedent effectfirst surfaced at around 600 ms, while the speaker effectfirst appeared at 1500 ms. In contrast, in the currentexperiment the precedent effect first appeared at300 ms, followed by a speaker effect that first appearedat 600 ms. The fact that the speaker effect only appearsafter a strong, partner-independent main effect ofPrecedent contradicts the hypothesis that preemptionis partner specific, and supports the recovery frompreemption hypothesis.

Further evidence that the speaker effect is due to late-onset inferential processing rather than to the partner-specific use of precedents can be found by comparingthe No Load condition to the Load condition (Fig. 3,bottom panel). Whereas the No Load condition suggeststhat a speaker effect for broken precedent is presentbetween 900 and 1500 ms, there is no trace of such aneffect in the Load condition. Nonetheless, during thissame period there is a clear partner-independent effectof precedents.

These observations were largely corroborated by sta-tistical analysis (see Table 4). As for Experiment 1, wediscuss in the main text only those effects that are rele-vant to our hypotheses and that were significant (ormarginally significant) by participants or items. In the300–600 and 600–900 ms windows, there were significantmain effects of Precedent and Load, as well as a Prece-dent by Load interaction. These effects were alsoobserved in the 900–1200 ms time window, with theaddition of a three way Speaker by Precedent by Loadinteraction that was marginal in the participantanalysis and significant in the item analysis. Separate2 · 2 ANOVAs for each of the two load conditions inthis window revealed that the three way interaction

Fig. 3. Mean Target Advantage Score By Condition, Experiment 2. The error bars represent the standard error of a difference betweenthe two Speaker conditions for each level of Precedent (Break, Mention), as computed from the analysis by Participant.


was driven largely by a Speaker by Precedent interactionin the No Load condition, F1(1,55) = 4.25, MSe = .100,p < .05; F2(1,31) = 7.96, MSe = .030, p < .01; minF 0(1,86) = 2.77, p = .10. For this same window, the Speakerby Precedent interaction in the Load condition did notreach significance, F1(1,55) = .316, MSe = .117, p =.58; F2(1,31) = .48, MSe = .044, p = .49; minF 0(1,86) =.19, p = .66.

We followed up the significant two-way interaction inthe No Load condition for the 900–1200 window with apair of planned comparisons of the two Speaker condi-tions for each level of Precedent. As predicted, these

comparisons confirmed the existence of a speaker effectfor broken precedents, but not for maintained prece-dents. The mean of the New Speaker-Break-No Loadwas .49 (SD = .37) compared to.35 (SD = .42) in theOld Speaker-Break-No Load condition, a difference thatwas on the boundary of the 95% confidence interval forthe participant analysis (CI = ± .14), and that exceededthe interval for the item analysis (CI = ± .12). For theMaintain condition, there was a mean of .80(SD = .22) in the New Speaker-Maintain-No Loadcondition, compared to a mean of .83 (SD = .17) inthe Old Speaker-Maintain-No Load condition, a

Table 4Mean square error and F test on the target advantage score for Experiment 2

Time window Participants Items MinF0

Source of variance MSe F1 MSe F2 Df MinF0

180–300 ms window

Speaker .063 1.83 .062 1.01 1,64 0.65Precedent .061 .14 .080 .07 1,61 0.05Load .066 .47 .074 .24 1,62 0.16Speaker · Precedent .097 .45 .061 .40 1,77 0.21Speaker · Load .075 .00 .086 .00 1,55 0.00Precedent · Load .071 1.49 .084 .70 1,60 0.48Speaker · Precedent · Load .084 .10 .068 .07 1,70 0.04

300–600 ms window

Speaker .062 .32 .053 .21 1,68 0.13Precedent .050 53.63** .091 16.77** 1,51 12.78**

Load .087 10.31** .041 12.62** 1,83 5.67*

Speaker · Precedent .087 .01 .042 .01 1,79 0.01Speaker · Load .069 .73 .070 .40 1,64 0.26Precedent · Load .081 8.83** .072 5.74* 1,68 3.48�

Speaker · Precedent · Load .091 .23 .033 .37 1,86 0.14

600–900 ms window

Speaker .121 1.64 .050 2.34 1,85 0.96Precedent .073 109.84** .105 42.97** 1,55 30.89**

Load .117 26.18** .045 39.57** 1,85 15.76**

Speaker · Precedent .086 .192 .038 .31 1,86 0.12Speaker · Load .080 .26 .053 .25 1,78 0.13Precedent · Load .098 4.28* .055 4.23* 1,79 2.13Speaker · Precedent · Load .095 2.02 .047 2.47 1,83 1.11

900–1200 ms window

Speaker .113 2.43 .049 3.20� 1,84 1.38Precedent .079 169.01** .165 46.43** 1,48 36.42**

Load .071 14.89** .033 18.38** 1,83 8.23**

Speaker · Precedent .108 .97 .031 1.96 1,86 0.65Speaker · Load .055 .03 .046 .02 1,69 0.01Precedent · Load .090 3.53� .044 4.17* 1,83 1.91Speaker · Precedent · Load .108 3.28� .044 4.64* 1,85 1.92


Speaker .083 3.81� .036 4.97* 1,84 2.16Precedent .086 149.07** .148 49.48** 1,52 37.15**

Load .066 15.21** .037 15.28** 1,79 7.62**

Speaker · Precedent .069 .23 .028 .31 1,84 0.13Speaker · Load .060 .01 .032 .01 1,79 0.01Precedent · Load .071 .64 .041 .66 1,80 0.32Speaker · Precedent · Load .071 7.23** .040 7.30** 1,80 3.63�


Speaker .057 3.13� .036 2.71 1,76 1.45Precedent .062 165.50** .118 49.99** 1,50 38.39**

Load .086 8.19** .034 11.75** 1,85 4.83*

Speaker · Precedent .062 1.10 .044 .84 1,73 0.48Speaker · Load .053 .58 .024 .78 1,84 0.33Precedent · Load .070 .00 .034 .00 1,55 0.00Speaker · Precedent · Load .054 .37 .032 .32 1,76 0.17


Speaker .056 6.46* .032 6.46* 1,79 3.23�

Precedent .051 149.37** .096 45.43** 1,50 34.84**

Load .055 2.07 .035 1.81 1,76 0.97

(continued on next page)


Table 4 (continued)

Time window Participants Items MinF0

Source of variance MSe F1 MSe F2 Df MinF0

Speaker · Precedent .045 .55 .021 .64 1,82 0.30Speaker · Load .050 .03 .014 .05 1,86 0.02Precedent · Load .046 .66 .033 .54 1,74 0.30Speaker · Precedent · Load .039 .56 .020 .60 1,81 0.29

Note. Degrees of freedom (df) for all effects = 1; Error df for all Participant analyses = 55; Error df for all item analyses = 31; Error df

for MinF0 in the Table.� p 6 .10.* p 6 .05.

** p 6 .01.


difference that was not statistically reliable (95%CI = ± .07 by participants and ± .08 by items).

For the 1200–1500 ms time window, there was also asignificant Speaker by Precedent by Load interactionthat was driven by the No Load condition. The 2 · 2ANOVA for this window for the No Load conditionalone revealed a significant Speaker by Precedentinteraction, F1(1,55) = 5.90, MSe = .060, p < .05; F2(1,31) = 8.97, MSe = .022, p < .01; minF 0(1,86) = 3.56,p = .06. In contrast, the interaction for the Load condi-tion was not significant, F1(1,55) = 2.18, MSe = .079,p = .15; F2(1,31) = 2.16, MSe = .046, p = .15; minF 0(1,79) = 1.08, p = .30. Planned comparisons for the NoLoad condition found, once again, a speaker effect forbroken precedents. In the Break condition, the NewSpeaker-Break-No Load condition had a mean of .60(SD = .30), compared to a mean of .47 (SD = .32)in the Old Speaker-Break-No Load condition (95%CI = ± .11 by participants and items). For the Maintaincondition, there was a mean of .88 (SD = .18) in theNew Speaker-Maintain-No Load condition, comparedto a mean of .90 (SD = .13) in the Old Speaker-Main-tain-No Load condition, a difference that did not exceedthe 95% confidence intervals, neither by participants(CI = ± .06) nor by items (CI = ± .07).

Finally, the last two time windows (1500–1800 and1800–2100) showed only main effects but no interac-tions. In the 1500–1800 ms window, there were maineffects of Precedent and Load. In the 1800–2100 mswindow there was a main effect of Precedent and a maineffect of Speaker. No other main effects or interactionswere significant.

Lastly, we address a remaining puzzle in our data: thesecondary task seemed to affect the time course of infer-ential processing, but did not ultimately impact listeners’final judgments. Listeners were somehow able to main-tain a fairly constant rate of non-target selection acrossload conditions, which suggested that they were able torecover from preemption in the Load condition as well.However, we failed to find a Speaker by Precedent inter-action in the Load condition for the windows weobserved, which is the signature of such a recovery

process. It seems possible that this effect was simplypostponed beyond the final window analyzed (1800–2100 ms). To assess this possibility, we conductedadditional analyses on two subsequent windows in theLoad condition alone (2100–2400 and 2400–2700 ms).We found a significant Speaker by Precedent interactiononly in the latter of these windows, though it was mar-ginal by items, F1(1,55) = 4.25, MSe = 1.031, p < .05;F2(1,31) = 3.61, MSe = .013, p = .07; minF 0(1,75) =1.95, p = .17. There was a significant speaker effect inthe Break condition: the mean target advantage scorein the New Speaker-Break condition was .83(SD = .12) compared to .77 (SD = .13) in the OldSpeaker-Break condition (95% CI = ± .02 by partici-pants and 95% CI = ± .04 by items). In contrast, therewas no such speaker effect in the Maintain condition:the mean target advantage score in the New Speaker-Maintain condition was .98 (SD = .05) compared to.97 (SD = .06) in the Old Speaker-Maintain condition(95% CI = ± .02 by participants and by items). This lateSpeaker by Precedent interaction indicates that listenersin the Load condition were eventually able to consultcommon ground to aid in the recovery process, althoughthey did so much later than the listeners in the No Loadcondition. Thus, when listeners find themselves undercognitive load, they are able to skilfully allocatetheir attention and effort to maintain high rates ofperformance on language processing.

In sum, the current experiment provides strongsupport for the recovery-from-preemption hypothesis.First, the speaker effect for broken precedents emergedlater than the partner-independent effects of precedent.Second, the speaker effect for broken precedents wasdrastically delayed when listeners were placed undercognitive load, while the partner-independent effect ofpreemption was only partly diminished.

General discussion

Our primary goal in the current article was toexamine the nature of the speaker effect for broken


precedents. We explored two different hypotheses forthis effect. The partner-specific-preemption hypothesisassumes that an old precedent can preempt a newdescription for a given referent only when it is in com-mon ground with the current speaker. In contrast, therecovery-from-preemption hypothesis assumes that thepreemption effect is partner-independent: old prece-dents preempt new descriptions because they are avail-able, independently of whether they are in commonground with the current speaker. Under recovery-from-preemption, the speaker effect is explainedthrough a recovery process in which preemption effectsare cancelled faster in the case of a new speaker com-pared to the case of an old speaker. It was hypothe-sized that the recovery effects reflect the operation ofa more general inferential system that uses commonground, similar to the monitoring process proposedby Keysar et al. (2000).

These hypotheses were tested via two eyetrackingexperiments. Both experiments replicated the speakereffect for broken precedents that was originally observedby Metzing and Brennan (2003): overall, listeners lookedat the target faster when the current speaker was not thespeaker who originally established the precedents. Wealso found a speaker effect in the rate of non-target selec-tion, in that listeners who heard a precedent broken wereless likely to select the target and more likely to selectsome other object when the current speaker was old.Two sources of evidence supported a recovery-from-pre-emption explanation for this effect. First, an analysis ofthe time-course of the speaker effect revealed that itsonset was later than the partner-independent effect ofprecedents. Fixations to a previously mentioned targetwere equivalently low for both old and new speakersin the early moments of processing of a test expressionthat broke an established precedent. Thus, listenersexperienced preemption from established precedentseven when they were not in common ground. A higherlikelihood of fixating the target in the case of a newspeaker emerged only after a partner-independent differ-ence between old and new expressions. The secondsource of evidence was that the cognitive load manipula-tion in Experiment 2 greatly postponed the recoveryprocess, making it difficult for listeners to use commonground to recover from preemption.

In research on language use, it is always challengingto strike a balance between methodological rigor andecological validity. Previous experiments on precedentuse have all placed a slightly greater emphasis on ecolog-ical validity, using live naıve or confederate speakers. Incontrast, we opted to use pre-recorded materials, whichmade it possible to use the exact same utterance tokenacross all conditions. Admittedly, this is a relativelyunnatural situation, but we note two considerations infavor of the ecological validity of our results. First, pre-vious research suggests that—at least with respect to the

use of precedents—there is no difference between believ-ing one is listening to live versus pre-recorded speech(Barr & Keysar, 2002): effects of precedents are just asstrong in either case. The second consideration is oursuccess in replicating previous findings that were basedon studies with live confederates. Clearly, knowing oneis listening to recorded speech does not stop listenersfrom expecting speakers to follow precedents, nor fromusing common ground when interpreting broken prece-dents. It would be difficult to explain how we managedto obtain these effects if listeners had suspended thepragmatic principles they follow during interactivedialogue.

Moreover, it is important to consider not onlywhether our effects might fail to generalize, but alsohow they might fail to generalize. Is language processingin interactive dialogue more or less partner-specific thanin the laboratory? It might be argued that partner effectsmight be stronger in interactive dialogue, since interac-tion could strengthen the association between theexpression-referent mapping and the speaker. However,it must also be noted that the current results might over-

estimate the effects of partner-specificity that actuallytake place during real conversation. Experiment 2 showsthat partner-specific effects are especially disrupted bycognitive load. Unlike in the current experiments, whereparticipants’ only task was to identify the object thespeaker was describing, in real conversations listenersare likely to be cognitively much busier and thereforemight be less able to pay attention to partner informa-tion. Furthermore, listeners in conversation are able totake advantage of the possibility for interaction toreduce their computational load (Barr & Keysar, 2005).

Ultimately, such concerns are of little consequencegiven that the body of findings on precedent use, despitedifferences in methodology, presents a picture that isnearly uniform. Despite four experiments that have pro-vided opportunities to observe a speaker effect for main-tained precedents—Experiment 2 in Barr and Keysar(2002), the experiment in Metzing and Brennan (2003),and the two experiments in the current article—nonehas obtained such an effect. Thus, the null finding thatprecedents provide a partner-independent benefit tocomprehension would seem to reflect the absence ofany existing effect of common ground, rather than toan experimental artifact or low statistical power.

For the case of broken precedents, we, as well asMetzing and Brennan, have found an effect of speaker.However, Experiment 3 of Barr and Keysar failed tofind any speaker effect for the case of broken precedents.Apart from this exception, the findings suggest an asym-metry in common ground use: listeners use commonground when precedents are broken, but not when theyare maintained. The remaining challenge is to provide acoherent explanation for this asymmetry. In what fol-lows, we suggest that current memory-based accounts


(Horton & Gerrig, 2005; Metzing & Brennan, 2003), aswell as constraint-based accounts that assume thatspeaker information modulates processing from the ear-liest moments of comprehension (Hanna, Tanenhaus, &Trueswell, 2003; Nadig & Sedivy, 2002) cannot parsimo-niously explain these results. In contrast, the view thatcommon ground effects are mediated by a functionallydistinct monitoring system (Keysar et al., 2000)provides a comprehensive explanation that can evenaccommodate the apparent anomaly of Barr andKeysar’s Experiment 3.

The memory-based approach (Gerrig & McKoon,1998; Horton & Gerrig, 2005) is based on the idea that‘‘conversational partners act as contextual cues for theretrieval of associated information’’ (Horton & Gerrig,2005, p. 130). Under this view, partner information ismade available to language processes through a passivecuing of memory. Metzing and Brennan (2003) soughtto explain their speaker effect for broken precedents interms of the memory-based approach. They suggestedthat the effect of speaker may have emerged becausethe new speaker did not cue the old precedents as strong-ly: ‘‘because the original speaker was absent when thenew speaker was present, it is possible that the originalreferent-expression mapping may have decayed some-what, making it less likely to interfere with a new map-ping involving the same referent’’ (p. 211). However,such an explanation is inadequate on two counts. First,it is difficult to reconcile with the absence of a speakereffect for maintained precedents. In that condition, sucha decay following the partner switch would surely takeplace as well, leading to greater facilitation in processingold expressions for the original than for the new speaker.It could be argued that a speaker effect is absent becausethe expression itself, through the entrainment process,has become such a strong cue that it overwhelms thespeaker cue. However, this seems unlikely, since speakereffects are absent even when there has been only a singlepresentation of the old expression (Barr & Keysar, 2002,Experiment 2). Second, the assumption that both prece-dent and speaker effects are supported by the same mem-ory system would have difficulty explaining why the loadmanipulation delayed the speaker effect about five timesas much as it delayed the precedent effect (about1500 ms, as compared to about 300 ms).

The asymmetry also poses a challenge to constraint-based accounts, which assume that common ground is aprobabilistic cue that the comprehension system usesfrom the earliest point in processing (Hanna et al.,2003; Nadig & Sedivy, 2002). These accounts mightattempt to explain the difference between broken andmaintained precedents by assuming that commonground is a more reliable cue when precedents are bro-ken, but not when they are maintained. However, thisis problematic because it renders the use of commonground contingent upon the listener’s recognition of

how an old or new expression is being used (e.g., tointroduce a new referent, to break a precedent, or tore-refer to an old referent). It seems implausible that lis-teners could discern that a new expression is being usedto break a precedent rather than introduce a newreferent without having already processed a significantportion of that expression.

In contrast, our account assumes that the inferentialprocesses that use speaker information are optional innature, and that the pragmatic effects of entrainmentand preemption are the result of low-level memory pro-cesses that do not use speaker information. Incomingspeech is not only processed for its semantic content,but is also matched against available precedents. Whenthe speech matches an established precedent, this cuesthe old referent from memory and attention is shiftedto that object (the entrainment effect). In contrast, whenthe speech is perceived as new, attention is directed tonew referents in the context and away from old ones(the preemption effect). Occasionally, successful compre-hension will be achieved on this basis, with no furtherinferencing required. However, if for any reason theselow-level processes do not yield an adequate solution,then the listener will use inferential processing to makesense of the discrepancy. For instance, the listener mightnote that a new description matches an old object betterthan any new object, and then search for a reason whythe current speaker would break the precedent for thatobject (e.g., because the established precedent is not incommon ground).

Our account provides a straightforward explanationfor the asymmetry in common ground use. Wheneverpreemption or entrainment leads to a coherent,unproblematic interpretation, inferential processing willnot be required. Thus, when a description follows anestablished precedent, listeners are unlikely to sponta-neously notice the potential for misunderstanding. Inexperiments that have examined the case in which anew speaker follows an old speaker’s precedent, therewas never any other object in the display that provideda good or better match to the description of the targetthan the target itself. Thus, following the entrainedprecedent would always lead to the appropriateresponse.

For the case of broken precedents, it is easy to seethat in the current experiments, as well as that of Met-zing and Brennan (2003), the preemption heuristicwould systematically lead to an unviable result, trigger-ing further processing. In Metzing and Brennan’s exper-iment, all of the objects in the display at the time of testhad been named multiple times. Therefore, there wasnever an unmentioned object onto which a new expres-sion could be mapped; the only option was to assumethat the speaker was violating an established precedent.In our experiments, there was an unmentioned object,but it was always one that was highly dissimilar to the


mentioned description, such that listeners would haveeasily noticed that it was a very poor match. In bothof these situations, listeners would need to take extrasteps to try to make sense of the speaker’s violation ofan established precedent.

An important characteristic of our account is that itdoes not assume that speaker information is only usedwhen precedents are broken. Instead, it assumes that lis-teners will consult this information whenever the defaultentrainment and preemption processes fail to yield anadequate solution. Studies that have been conducted todate have all involved situations in which the entrain-ment heuristic would tend to succeed and preemptionwould tend to fail. Our account therefore predicts thatit should be possible to find cases in which commonground is used when a precedent is maintained, andnot used when a precedent is broken. As to the formerpossibility, we would expect to find effect of commonground if listeners are made aware that two speakersmight use the same conventional term to refer to differ-ent objects. This awareness should cause listeners to becautious in following established precedents when thecurrent speaker is new. We hope to test this predictionin a future study.

In contrast, Experiment 3 of Barr and Keysar (2002)provides an example of a case in which there is no effectof common ground when a precedent is broken. In thatexperiment, a speaker established a precedent in which asubordinate-level name was used to refer to pictures oftwo common objects; e.g., a car was referred to as the

sportscar, and a flower was referred to as the carnation.In the test phase, pictures of the two objects were pre-sented and a speaker referred to one of the pictures usinga basic-level term (e.g., the car), thereby breaking theestablished precedent. Metzing and Brennan suggestthat the failure to find an effect of speaker knowledgein this experiment may have been due to strong primingfrom the subordinate-level precedent, which was repeat-ed five times prior to the test description, and which, as aconsequence, may have overwhelmed any smaller effectsof common ground. In contrast, our explanation for theabsent speaker effect is that the recovery-from-preemp-tion process did not require differential access to com-mon ground. When a speaker breaks a precedent byswitching from one unconventional expression to anoth-er (e.g., from the metal object to the shiny thing) therecovery process can be difficult. In contrast, when aspeaker breaks a precedent by switching from one con-ventional term to another (e.g., from sportscar to car),listeners could account for this switch by noticing thatthe subordinate-level term is too specific for this objectin the current context (e.g., because there was only onecar in the display). Because such an inference wouldnot require access to common ground, the recovery pro-cess would be equally fast regardless of the identity ofthe speaker.

Traditionally, research on language use has assumedthat pragmatic effects are tied to intentional inferencesabout speaker’s beliefs and shared knowledge. Ourfindings represent a significant theoretical advance bychallenging this well-received view. Specifically, we findevidence that entrainment and preemption effects mightreflect the operation of general expectations aboutlanguage use that are not tied to a listener’s beliefs abouta particular speaker. That is, the comprehension systemtends to associate old expressions with old information(entrainment), and new expressions with new informa-tion (preemption), without regard to whether the expres-sion is ‘old’ or ‘new’ for the particular speaker that iscurrently speaking to them. Our most compellingevidence for this is that the early preemption effect wasequally strong in the case of a new speaker as it was inthe case of an old one, a clear violation of commonground. It remains an open question how widespreadsuch ‘perspective-free’ pragmatic heuristics are, andhow successful they might be in explaining how speakersand listeners achieve shared understanding in the face ofambiguity.

References

Barr, D. J., & Keysar, B. (2002). Anchoring comprehension inlinguistic precedents. Journal of Memory and Language, 46,391–418.

Barr, D. J., & Keysar, B. (2005). Making sense of how we makesense: The paradox of egocentrism in language use. In H. L.Colston & A. N. Katz (Eds.), Figurative language compre-

hension: Social and cultural influences (pp. 21–41). Mahwaw,NJ: Erlbaum.

Brennan, S. E., & Clark, H. H. (1996). Conceptual pacts andlexical choice in conversation. Journal of Experimental

Psychology: Learning, Memory, and Cognition, 22,1482–1493.

Carroll, J. M. (1980). Naming and describing in socialcommunication. Language and Speech, 23, 309–322.

Clark, E. V. (1990). On the pragmatics of contrast. Journal of

Child Language, 17, 417–431.Clark, E. V., & Clark, H. H. (1979). When nouns surface as

verbs. Language, 55, 767–811.Clark, H. H. (1991). Words, the world, and their possibilities.

In G. R. Lockhead & J. R. Pomerantz (Eds.), The perception

of structure. Washington, DC: American PsychologicalAssociation.

Clark, H. H., & Carlson, T. B. (1981). Context for compre-hension. In J. Long & A. Baddeley (Eds.), Attention and

performance IX (pp. 313–330). Hillsdale, NJ, USA: Law-rence Erlbaum Associates.

Clark, H. H., & Marshall, C. R. (1981). Definite reference andmutual knowledge. In A. K. Joshe, B. L. Webber, & I. A.Sag (Eds.), Elements of discourse understanding (pp. 10–63).Cambridge: Cambridge University Press.

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


Cruse, D. A. (1977). The pragmatics of lexical specificity.Journal of Linguistics, 13, 153–164.

Garrod, S., & Anderson, A. (1987). Saying what you mean indialogue: a study in conceptual and semantic co-ordination.Cognition, 27, 181–218.

Gerrig, R. J., & McKoon, G. (1998). The readiness is all: thefunctionality of memory-based text processing. Discourse

Processes, 26, 67–86.Hanna, J. E., Tanenhaus, M. K., & Trueswell, J. C. (2003). The

effects of common ground and perspective on domains ofreferential interpretation. Journal of Memory and Language,

49, 43–61.Horton, W. S., & Gerrig, R. J. (2005). The impact of memory

demands on audience design during language production.Cognition, 96(2), 127–142.

Horton, W. S., & Keysar, B. (1996). When do speakers takeinto account common ground? Cognition, 59(1), 91–117.

Keysar, B., Barr, D. J., Balin, J. A., & Brauner, J. S. (2000).Taking perspective in conversation: The role of mutualknowledge in comprehension. Psychological Science, 11(1),32–38.

Keysar, B., Barr, D. J., Balin, J. A., & Paek, T. S. (1998).Definite reference and mutual knowledge: process models of

common ground in comprehension. Journal of Memory and

Language, 39(1), 1–20.Markman, E., & Wachtel, G. (1988). Children’s use of mutual

exclusivity to constrain the meanings of words. Cognitive

Psychology, 20, 121–157.Matin, E., Shao, K. C., & Boff, K. R. (1993). Saccadic

overhead: information-processing time with and with-out saccades. Perception & Psychophysics., 53(4),372–380.

Metzing, C., & Brennan, S. E. (2003). When conceptual pactsare broken: partner-specific effects on the comprehension ofreferring expressions. Journal of Memory and Language, 49,201–213.

Nadig, A. S., & Sedivy, J. C. (2002). Evidence of perspectivetaking constraints in children’s on-line reference resolution.Psychological Science, 13, 329–336.

Rosch, E., Mervis, C. B., Gray, W., Johnson, D., & Boyes-Braem, P. (1976). Basic objects in natural categories.Cognitive Psychology, 8, 382–439.

Rossnagel, C. (2000). Cognitive load and perspective taking:applying the automatic-controlled distinction to verbalcommunication. European Journal of Social Psychology,

30, 429–445.

Perspective-free pragmatics: Broken precedents and the recovery-from-preemption hypothesis

Documents

Transcript of Perspective-free pragmatics: Broken precedents and the recovery-from-preemption hypothesis