coordination of knowledge in communication: effects of ... · coordination of knowledge in...

Journal of Personality and Social Psychology1992, Vol.62, No. 3, 378-39!

Copyright 1992 by the American Psychological Association, Inc.0022-3514/92/S3.00

Coordination of Knowledge in Communication: Effects of Speakers'Assumptions About What Others Know

Susan R. Fussell and Robert M. KraussColumbia University

Two pairs of studies examined effects of perspective taking in communication, using a 2-stagemethodology that first obtained people's estimates of the recognizability to others of specificstimuli (public figures and everyday objects) and then examined the effects of these estimates onmessage formulation in a referential communication task. Ss were good at estimating stimulusidentifiability but were biased in the direction of their own knowledge. The amount of informationin a referring expression varied inversely with the perceived likelihood that addressees couldidentify the target stimulus. However, effects were less strong than anticipated. Although commun-icators do take others' knowledge into account, the extent to which they do so involves a trade-offwith other sorts of information in the communicative situation.

Communication is assumed to rest on a foundation of sharedunderstandings that are integral to both speaking and listening(e.g., Clark & Marshall, 1981; Krauss, 1987; Krauss & Fussell,1990,1991a, 1991b; Mead, 1934; Rommetveit, 1974). A centralcomponent in the construction of these mutual understand-ings is what Mead (1934) termed taking the perspective of theother—assessing the background knowledge, plans, attitudes,beliefs, outlooks, and so on, of one's fellow interlocutors. Here,we focus on one specific type of perspective taking: determin-ing the background knowledge of one's cocommunicators.Drawing on Clark and Murphy's (1982) terminology, we call thehypothesis that speakers create messages with their listeners'knowledge in mind the audience design hypothesis.

In the current studies, we examine audience design in onespecific aspect of language use: reference. As Brown (1958) hasnoted, things may be named in many ways—at different levelsof specificity as well as by different terms at the same level ofspecificity. An important consideration in choosing a referringexpression is the knowledge shared between speaker and lis-tener.1 For example, how one refers to one's occupation (e.g.,professor, psychologist, social psychologist, or social psycholo-gist with an interest in social cognition) depends on (among other

This article is based on a dissertation submitted by Susan R. Fussellin partial fulfillment of the PhD at Columbia University, under theguidance of Robert M. Krauss. Preparation of this article was aided byDepartment of Health and Human Services National Research Ser-vice Award MH09715-01A1 to Susan R. Fussell from the NationalInstitute of Mental Health and National Science Foundation GrantBNS-86-16131 to Robert M. Krauss.

We gratefully acknowledge the comments of the members of thedoctoral committee: Barbara Landau, Barbara Dosher, Robert Re-mez, and Clifford Hill. We also thank Nicholas Benimoff, EugeneBorgida, and five anonymous reviewers for their helpful comments onearlier versions of the article and Christina Colasante and Sue Scelzofor their assistance in conducting the experiments.

Correspondence concerning this article should be addressed to Rob-ert M. Krauss, Department of Psychology, Schermerhorn Hall, Co-lumbia University, New York, New York 10027.

things) one's perceptions of the addressee's familiarity with thefield.

Perspective Taking in Communication

Clark and his colleagues (Clark, 1985; Clark & Carlson, 1982;Clark & Marshall, 1981) have argued that successful communi-cation requires a more complex sort of perspective taking thanMead hypothesized. Specifically, they propose that communi-cators must identify their common ground or mutual knowl-edge—the information, beliefs, attitudes, and so on, that partici-pants share and know they share and know that all other partici-pants know they share, ad infinitum. However, it has yet to bedemonstrated that communicators can perform even the mostbasic perspective-taking task: assessing others' knowledge. If aspeaker wishes to determine whether a piece of knowledge he orshe possesses is part of the common ground shared with theaddressee, a sensible first step is to assess whether that ad-dressee knows it. Little is known about people's abilities tomake such judgments and their consequences for message pro-duction.

If speakers are to tailor their messages to their addressees,they must have some expectations about what those addresseesknow and hence what might be part of common ground. Theseexpectations about others' knowledge can come from severalsources. One important source is the interactive dynamics ofthe communicative situation. Krauss and Weinheimer (1966)and Clark and his colleagues (e.g., Clark & Wilkes-Gibbs, 1986;Isaacs & Clark, 1987) have demonstrated that over the course ofa conversation speakers and hearers accumulate a body ofshared knowledge that they draw on when formulating theirsubsequent messages. For instance, Isaacs and Clark (1987)found that speakers who were experts on New "York City quicklyadapted their references to New York City landmarks to their

' We use the term speaker to indicate the initiator of a message andlistener or hearer to indicate the intended recipient of a message, re-gardless of the modality of communication.

378

COORDINATION OF KNOWLEDGE IN COMMUNICATION 379

addressee's knowledge or lack thereof. Clark and Wilkes-Gibbs(1986) have developed a collaborative model of communicationin which speaker and listener work jointly to establish thatsuccessful reference has occurred.

However, although interactional feedback is one importantsource of information about others' perspectives, it is neithernecessary nor sufficient for audience design. Even in commu-nicative contexts where feedback is available, speakers mustcreate an initial message before they can make use of this feed-back. On their first trial, speakers in Isaacs and Clark's (1987)study had to choose some referring expression—name, descrip-tion, or a combination of the two—before they could use feed-back to assess the listener's New \brk City expertise. In ordi-nary conversation, the things to which one refers are typicallymore diverse than the landmarks in Isaacs and Clark's study,and the addressee's background knowledge of a particular refer-ent often will not be directly predictable from the success orfailure of preceding acts of reference. Although speakers in suchsituations may use interactional feedback to determine whetherthey should expand or replace a particular referring expression,they must rely on their prior assumptions about others' knowl-edge to construct that referring expression in the first place.

Moreover, feedback signals are open to multiple interpreta-tions. A simple back-channel response such as "uhuh" can be asign of attention, understanding, and so on, and speakers mayrely on prior expectations about the listener's knowledge tointerpret these signals. Subtle judgments about the social distri-bution of knowledge are also important in projecting what lis-teners are likely to know from their vocal responses. Identify-ing the Empire State Building from its name will create weakerexpectations about an addressee's ability to identify other New\brk City landmarks than identifying Grand Central Station orthe Soldiers' and Sailors' Monument. Thus, an understandingof speakers' implicit theories of others' knowledge is importantto an understanding of how they utilize feedback in messageconstruction.

Finally, there are many communicative situations in whichfeedback is delayed or nonexistent: exchanging letters or elec-tronic mail, teaching large lectures, and so on. In such situa-tions, communicators must rely on their a priori beliefs aboutthe social distribution of knowledge to create messages thatfollow the principle of audience design.

Several studies of communication in noninteractive contextshave shown that speakers' prior beliefs about others' knowledgeaffect both the form and communicativeness of their messages(Danks, 1970; Fleming & Darley, 1991; Fleming, Darley, Hil-ton, & Kojetin, 1990; Fussell & Krauss, 1989a, 1989b; Innes,1976; Kaplan, 1952, cited in Werner & Kaplan, 1963; Krauss,Vivekananthan, & Weinheimer, 1968). For example, in onestudy (Fussell & Krauss, 1989a), we found that, on average,referring expressions for nonsense figures were more thantwice as long when intended for another student than whenintended for oneself. In a second study (Fussell & Krauss,1989b), we found that messages directed to a specific friendcommunicated better to that friend than to other students.These studies suggest that prior knowledge or suppositionsabout anothers' characteristics are incorporated into commun-icators' hypotheses about information that is (mutually) known

and information that is not (and, hence, must be stated explic-itly). However, several issues remain unresolved.

First, the strongest effects are found when messages for theself and for another person are compared. Because these differ-ences could result from sources other than audience design (e.g.,a closer relationship between messages to the self and the lan-guage of thought), such comparisons offer only weak supportfor the audience design hypothesis.

Second, the audience design hypothesis should be evidencedin speech to a single individual—that is, referring expressionsintended for a particular addressee should be more or less spe-cific or detailed depending on that addressee's knowledge.Within-addressee effects would indicate that audience design isa basic "on-line" component of the message production system.However, these effects have not been demonstrated empiri-cally; rather, previous studies have used paradigms in which thespeaker could make a general categorization of the addressee atthe outset and then tailor all messages according to this initialcategorization.

Third, the effects of prior beliefs on message constructionhave generally been demonstrated in noninteractive settings inwhich such beliefs are the communicators' sole source of infor-mation about their partners' perspectives. It has yet to be dem-onstrated that prior beliefs affect communication in situationsin which feedback is also available. An exception is a study bySchober and Clark (1989) in which speakers drew upon knowl-edge shared with addressees but not overhearers to make mes-sages that were communicative only to the former. However, itis not clear precisely when and how speakers in this study uti-lized prior beliefs about their addressees' knowledge, becausethese beliefs were not ascertained independent of their mes-sages.

Fourth, in most previous studies, speakers had no informa-tion about the message recipient beyond the fact that he or shewas "another student." In everyday life, even when interactingwith strangers, communicators often can identify each other'scategory memberships and use them to draw inferences aboutwhat is known (e.g, Schegloff, 1972). The Isaacs and Clark(1987) study provided indirect evidence that an addressee's cate-gory membership can influence a speaker's referring expres-sions; however, Isaacs and Clark did not establish that speakersactually classified their addressees on this basis, nor did theyspecify what speakers assumed about the addressee's knowledgeon the basis of this classification.

A recurring problem in these studies is a circularity in identi-fying the particular shared knowledge the speaker is taking intoaccount during message formulation. For example, the fact thatspeakers who are New Yotk City experts create shorter mes-sages when speaking to other experts cannot be used to showboth that speakers classified their addressees as New York Cityexperts and that the effect of this classification is a briefer mes-sage. To examine the effects of prior beliefs about others' knowl-edge on message formulation, the nature of these beliefs mustbe established independent of the messages themselves.

Perspective-Taking Process

Suppositions about what others know are derived from a vari-ety of sources varying along a continuum of directness of knowl-

380 SUSAN R. FUSSELL AND ROBERT M. KRAUSS

edge (Clark & Marshall, 1981; Krauss & Fussell, 1990,1991a,1991b). Here, we focus on one particular information source—aperson's social category memberships. Every individual is amember of a number of social categories, and knowing that aperson belongs to a particular category can allow one to predictsome of the things that person is likely to know. So, for example,it is reasonable to assume that a doctor will know anatomicalterms, that a taxi driver will be familiar with a city's mainthoroughfares, and that a baseball fan will know the details ofthe infield fly rule. In specific cases, such assumptions may bewrong; nevertheless, they provide some basis for assuming thatshared knowledge exists. However, it is not clear how peopleestablish the boundaries of a person's category-related knowl-edge. Although it seems reasonably certain that a New \brkerwill know the location of the Empire State Building and St.Patrick's Cathedral, it is more difficult to say whether he or shewill be familiar with the Soldier's and Sailor's Monument or theMuseum of Colored Glass and Light.

One possibility is that judges rely on a variety of inferenceheuristics and simplifying knowledge structures, such as sche-mata and stereotypes (cf. Cantor, Mischel, & Schwartz, 1982;Fiske & Taylor, 1984; Hastie, 1981; Markus & Zajonc, 1985;Taylor & Crocker, 1981), when reasoning about the social dis-tribution of knowledge. Such structures and heuristics facilitatethe task of drawing inferences, but they may also induce system-atic errors or biases (e.g., Kahneman, Slovic, & Tversky, 1982;Nisbett & Ross, 1980) that lead to miscalculations of what ismutually known. One form of bias particularly relevant to theperspective-taking process is the false consensus effect, whichresults when subjects erroneously believe that others are moresimilar to themselves than they actually are (Ross, Greene, &House, 1977).

Bias in estimates of others' knowledge was studied by Nick-erson, Baddeley, and Freeman (1987), who examined the rela-tionship between a person's ability to answer a variety of gen-eral knowledge questions and his or her estimate of how manyother people could answer these questions. Nickerson et al.found that subjects who could answer an item tended to overes-timate the proportion in the population who knew the answer.An equally important issue is the overall accuracy of such judg-ments—how good are subjects at estimating the likelihood thatan item will be known? Nickerson et al. do not report overallcorrelations between estimated and actual values, but an exami-nation of one of their figures, in which these two values areplotted against each other, suggests that accuracy was quitehigh for subjects who could answer an item and may have beenbetter than chance for those who could not. In the current stud-ies, we examine both accuracy and bias in estimates of others'knowledge of public figures and everyday objects.

Present Paradigm

To support the audience design hypothesis, one must use aparadigm that compares messages for different addressees orestablishes in some other way what assumptions the speaker ismaking about the addressee, independent of the message(Krauss & Fussell, 1988). The referential communication task isone paradigm that permits this. In such tasks, communicatorsname or describe a set of stimuli so that they themselves, or

another person, can select the correct referent from the fullarray. By varying characterizations of the intended message re-cipient, one can evaluate the audience design hypothesis. In thecurrent studies, a conversational version of the task was used,in which speakers and listeners were allowed to communicatefreely. Using this paradigm, we tested whether prior beliefs af-fected message construction even when feedback from the lis-tener was also available.

In the present experiments, we used a two-step research strat-egy to demonstrate that messages are shaped by the principle ofaudience design. First, people's assumptions about the socialdistribution of knowledge were assessed by asking subjects toestimate others' knowledge of specific items. Second, thesegroup judgments were then used to infer the assumptions indi-vidual speakers had as they created messages in a referentialcommunication task. For this strategy to be successful, the so-cial categories for which knowledge judgments are obtainedmust be psychologically real and salient to communicators, andpeople's perceptions of the social distribution of knowledgemust be socially shared. In the studies presented below, thesocial categories used—undergraduates at a particular institu-tion and men versus women—fulfilled these criteria.

The studies examined two sets of hypotheses. The first setconcerned the accuracy and bias of people's inferences aboutothers' knowledge. We hypothesized that subjects would accu-rately estimate the likelihood that others could identify particu-lar public figures (Experiment 1) and the likelihood that theywould know the names of everyday objects (Experiment 3), butwe also anticipated that these estimates would be biased in thedirection of the subjects' own knowledge.

The second set of hypotheses concerned the use of such attrib-uted knowledge in communication. We expected that theamount of informative detail speakers provided in their mes-sages would be a function of the perceived recognizability ofthat stimulus to the addressee: The less recognizable the stimu-lus, the more descriptive information would be provided. It wasalso hypothesized that the presence or absence of such descrip-tive information would be related to the listener's ability toidentify the referent.

Experiment 1

By definition, public figures are recognizable to at least somesubset of the population, but some public figures are more rec-ognizable than others. For instance, almost everyone can iden-tify ex-President Ronald Reagan, but only those with a specialinterest in politics may be able to recognize low-rankingmembers of his cabinet. In Experiment 1, students viewed pic-tures of public figures and rated how identifiable each was bothfor themselves and for other undergraduates. To estimate theactual proportion of people in the population who could nameeach target, subjects also indicated the names of those theyknew.

Method

Materials. Asstimuli we used pictures of 15 men (listed in Table 1)prominent in business, politics, or entertainment. The approximatelyV* x 1 in. (1.905 X 2.54 cm) pictures were mounted on 3 X 5 in. (7.62 X


Table 1Percentage Correct and Mean Identifiability Ratings for Public Figures

Public figure

Woody AllenClint EastwoodPaul NewmanGary HartLeelacoccaGeorge BushIvan BoeskyAlexander HaigHoward BakerJudd NelsonDonald ReganTed TurnerKevin KlineCarl IcahnT. B. Pickens

% correct

9380737373604747404040271370

All observations

Others

6.46.26.35.85.75.24.14.94.14.03.93.34.02.52.0

Self

6.56.46.15.95.74.84.14.83.54.14.33.13.51.61.6

Mean identifiability rating

Named

Others

6.86.66.85.86.46.45.65.65.35.74.35.35.0

—

Self

6.96.86.96.57.06.46.76.66.06.36.27.05.5

—

Unnamed

Others :

4.75.05.83.83.32.84.43.32.93.62.63.92.42.0

Self

4.73.84.52.32.31.93.31.92.72.91.73.21.61.6

Note. Identifiability ratings were made on a scale of 1 (none) to 7 (high). Named indicates observationson which the target was correctly named; unnamed indicates observations on which the target was notnamed. Dashes indicate that there were no judgments of a particular type made for that public figure.

12.7 cm) notecards; the notecard's color indicated the stimulus person'sprimary occupational category (business, entertainment, or politics).

Subjects rated how identifiable each stimulus person was on a 7-point scale (ranging from not identifiable to very identifiable), both forthemselves and for other Columbia University undergraduates. Theywere also asked to indicate the stimulus person's name if they knew it.Demographic information was collected in a postexperimental ques-tionnaire.

Subjects and procedure. Fifteen undergraduates (9 men and 6women) participated alone or in small groups. All subjects here and insubsequent studies were native English speakers of college age (17-25years). They received written instructions (including explanation ofthe notecard color coding scheme), followed by oral instructions inwhich identifiability was defined as the likelihood of knowing the stim-ulus person's name. They received the cards in random order andworked at their own pace.

Results and Discussion

Correct identifications of the 15 public figures ranged from0% to 93% (M = 48%, SD = 28%; see Table 1). Individual subjectscorrectly identified between 13% and 87% of the targets, with amedian of 53%.

Mean ratings of identifiability to others (others) ranged from2.0 to 6.4 (M= 4.6, SD =2.1); ratings for identifiability to self(self) had a similar range (M = 4.4, SD = 2.6; see Table 1).Group estimates for others and self, formed by averaging alljudgments for a particular target, were highly correlated withactual proportions correct (for others, r = .95; for self, r = .97;both/?s<.001).

Although these high correlations are consistent with the prop-osition that subjects are sensitive to others' knowledge, theycould also result from a strong false consensus bias: If all sub-jects who identified a given picture assumed that everyone elsecould identify it, and all who could not identify a picture as-

sumed that no one could identify it, the correlation between themean identifiability rating for each picture and the proportionin the population who could identify the person in the picturewould be perfect. However, the correlation would derive from aprimitive assumption of similarity rather than an awareness ofthe way knowledge is distributed socially.

To examine this possibility, correlations between mean iden-tifiability ratings and actual percentages correct were calcu-lated separately for observations in which subjects knew thestimulus person's name (named) and for those in which they didnot (unnamed). (Estimates of subjects who incorrectly identi-fied the stimulus person were excluded, as were mean estimatesbased on a single subject's estimate.) Mean estimates fromnamed and unnamed observations are shown in Figure 1, inwhich the solid lines represent the best fit to each set of esti-mates. Subjects were clearly sensitive to the relative identifiabi-lity of different targets, regardless of whether they themselvesknew the person's name (rs = .82 for the named and .70 for theunnamed estimates, both ps < .005; for the comparison, z =.66). However, although the regression lines for the two distri-butions have virtually identical slopes, their intercepts differsubstantially (4.35 for named estimates vs. 2.32 for unnamedestimates). Subjects who could identify a stimulus personjudged him to be more identifiable to others than those whocould not.

Individual subjects' correlations between estimates and per-centages correct across the 15 stimuli showed the same patternas the group values: The mean individual correlation was .67,and no subject's correlation was less than .46. Even with thesmall number of observations, 12 of the 15 individual correla-tions exceeded the minimum required (.51, df= 13) for signifi-cance at the .05 level, and ten exceeded the minimum (.64) atthe .01 level. When the analysis was limited to observationsabout which the subject was correct, correlations were still rea-


MeanIdentifiability

Rating

10 20 30 40 50 60 70

Actual Percent Correct

80 90 100

Figure 1. Mean identifiability ratings of public figures versus percentage of correct identifications.(Solid regression lines represent the best fit to each set of estimates.)

sonably high: Of 13 subjects with at least two correct observa-tions, 8 had correlations above .40. Subjects may have relied ontheir own familiarity with the stimulus person in estimatingtheir identifiability to others. Individual correlations betweenestimates for self and others were high, with 14 of the 15 valuesexceeding the minimum required (.51) for significance at the.05 level.

Experiment 1 showed that people are reasonably sensitive towhat others know, especially when they themselves are knowl-edgeable about the stimulus. Considering the small number ofobservations, the results were impressively strong. Further-more, the fact that most subjects' correlations were significantsuggests that theories or intuitions about the distribution ofknowledge are shared, at least within the Columbia Universitystudent population. Hence, it is reasonable to use these judg-ments to estimate what a speaker would assume about an un-known addressee's likelihood of knowing a particular stimulusperson's name. The next experiment examined the effects ofthese assumptions on message formulation.

Experiment 2

In Experiment 2, subjects communicated about the publicfigures used in the preceding study in a referential communica-tion task. According to the audience design hypothesis, sub-jects using the name of a target should provide more identifyinginformation (e.g., description of features and occupational cate-gory membership) with that name as the perceived likelihoodof the listener's recognizing the target declines.

We also evaluated the success or failure of speakers' messagesby examining the type of response given by the matcher. Whendescriptive content is omitted from lesser known targets, longeroverall speaking exchanges should be required for speaker andhearer to agree that successful reference has occurred.

Method

Materials. Stimuli consisted of the pictures of the 15 public figuresused in Experiment 1; pictures were mounted on notecards indicating

their occupational categories. Postexperimental questionnaires wereadministered to collect personal data and reactions to the task.

Subjects and procedure. Twenty-eight undergraduates served assubjects. They were randomly assigned in pairs to either the role ofdirector, (the person who described each of the pictures) or matcher (theone who tried to select the correct stimulus from the full array). In 6 ofthe pairs, both director and matcher were women, in 4 both were men,and in 4 there were mixed gender pairs. Their discussions were taperecorded.

Subjects were seated at desks separated by a small screen. Directorsreceived the notecards in a prearranged order of three rows by fivecolumns; matchers received the notecards in random order and wereinstructed to lay them out on the desk so that they were clearly visible.Directors were instructed to convey the pattern of cards to theirpartners one by one in the prescribed order, using any form of refer-ence they liked (e.g., name, description, or occupational category).Speed and accuracy were stressed.2 Matchers were permitted to makebrief comments or to ask questions as needed.

Subjects repeated the task three times with each set of stimuli. Adifferent order was used for each trial, but the order of the cards withina trial was constant for all subjects.3 Errors made by the matcher wererecorded by the experimenter, although very few mistakes were found.Afterward, subjects completed the follow-up questionnaire.

Preparation and coding of transcripts. All utterances (includingspeech disfluencies and pauses) were transcribed and checked by Su-

2 Speed was stressed because we found in pilot studies that speakerswould not only refer to the target person but discuss everything theyknew about him when there were no time restrictions. From casualobservation, this restriction did not appear to alter the naturalness ofthe dialogues.

3 To obtain a large corpus of utterances generated under as similarcircumstances as possible (i.e., with the same alternatives in mind), wedecided to present the stimuli in the same order for all subjects. Thisstrategy allows one to estimate the effects of both order and attributedknowledge, as long as they are not confounded, and it should helpreduce extraneous variance that makes it difficult to uncover the ef-fects of attributed knowledge. It is also consistent with other sorts ofcommunication tasks, such those that require the speaker to relate aprespecified sequence of events or propositions (e.g., Kraut, Lewis, &Swezey, 1982).


san Fussell and trained assistants. The transcribed texts were punc-tuated in conventional fashion, with greater than 90% agreement.

The texts of directors' first turns were divided into idea units (e.g.,Butterworth, 1975) that were denned in terms of their content matter:Each major attribute of a target—including clothing, body position,and facial expression—counted as a separate unit regardless ofwhether it was produced in the same phrase as other attributes. Agree-ment on the number of idea units was greater than 92%. Six types ofidea units were distinguished: occupational category, proper names(correct and incorrect), description (the physical features of the personin the picture), personal information (e.g., specifics of the person's occu-pation, such as movies in which he had appeared), other informationthat might less directly aid identification (e.g., statements about famil-iarity), and noninformation (e.g., repetitions, repairs, or extraneouscomments). Units were coded by Susan Fussell. All Trial 1 messages(624 units) were also independently coded by an assistant, with excel-lent agreement (95%).

The majority of units belonged to one of the first 3 categories—name, occupational category, and description. These three categoriesplus personal information and other information were considered con-tent units. We tallied the number of content units and the number ofwords of description for directors' first-turn messages. To determinethe time it took pairs to establish reference, we tabulated the number ofspeaking turns by both partners required after the director's first mes-sage until the pair moved on to the next target. Because the distribu-tions for all dependent measures were positively skewed, values weretruncated to three standard deviations above the mean.


Because the overall pattern of results across the three trials isof limited importance for testing the current hypotheses, it isdescribed here only briefly (full details are provided in Fussell,1991). Directors provided significantly less information, bothin terms of content units and words of description, on latertrials, and fewer speaking turns were required to establish refer-ence. This pattern held for both messages in which the targetwas named correctly (named) and those in which the target wasnot named (fmnamed), although the unnamed messages tendedto be substantially longer than the named ones (mean contentwords for Trials 1-3 were 3.18,1.95, and 1.44, respectively, fornamed observations; and 12.19,10.45, and 9.27, respectively,for unnamed observations). Similarly, fewer speaking turnswere required to establish reference when names were used(mean turns for Trials 1-3 were 2.95, 2.35, and 2.14, respec-tively, for named observations; and 3.57,2.67, and 2.37, respec-tively, for unnamed observations). These results are in accor-dance with previous studies of this type (eg., Clark & Wilkes-Gibbs, 1986; Krauss & Weinheimer, 1966).

To examine the effects of communicators' prior theoriesabout the social distribution of knowledge on message construc-tion, one must look at messages constructed on the first trial,before the director has received any feedback from the matcher.According to our hypotheses, as the addressee's perceived famil-iarity with the target declines, the director will accompany thetarget's name with more information.

The descriptive content of all Trial 1 messages containingnames (n = 80) was examined as a function of perceived stimu-lus identifiability (identifiability, here defined as the proportionof correct identifications in Experiment I).4 As we anticipated,

the amount of information provided with the name of a stimu-lus increased as perceived identifiability declined (for contentunits, r= -.38, p< .01; for description words, r= -.28, p< .01).These findings were examined by regressing the number ofcontent units per message onto presentation order, identifiabi-lity, and a set of dummy-coded subject variables. Significanteffects in the expected direction were found for identifiability,/(64) = 3.99, p < .001, B = -2.19; order, f(64) = 2.92, p <. .005,B = -.08; and subject, ^(13, 64) = 2.54, p < .01.5 In addition,there was a significant Order X Identifiability interaction, F(l,63) = 5.38, p < .05: The lengthening effects of lower recogniza-bility decrease as the speaker gets to the end of the sequence ofstimuli. When number of words of description is similarly ana-lyzed, significant effects are found only for identifiability,/(64) = 2.36, p < .05; B = -6.25; and the Order X Identifiabilityinteraction, F(l, 63) = 3.98, p < .05.

To examine the adequacy of directors' first messages as afunction of the amount of information provided with a target'sname, we counted the number of speaking turns required toestablish reference (as indicated by the pair moving on to thenext stimulus). Correlations between both measures of messagecontent (content units and words of description) and the num-ber of turns required were low and nonsignificant. Multipleregressions confirmed that the amount of information speakersprovided with the targets' names did not have a substantialimpact on the hearer's ability to identify the target, at least bythe measure used here.

On the whole, these findings provide some support for thehypothesis that speakers use their prior beliefs about the ad-dressee's knowledge in constructing their messages, even incommunicative contexts where feedback is available. However,the effects of these prior beliefs were weaker than we had antici-pated. In many cases, directors added little descriptive infor-mation to the target's name regardless of his recognizability.Just over half of the messages mentioned the occupational cate-gory of the target, and only 21% of these messages containedother descriptive information. In contrast, when names werenot used, 91% of the messages indicated the occupational cate-gory of the target, and 68% of them contained descriptive infor-mation. On the whole, speakers seemed to feel that the namealone was sufficient for identification.

It is possible that we would have found stronger effects ofattributed knowledge with more observations. The small set of

4 Mean estimates from correct observations in Experiment 1 werealso used as a measure of perceived stimulus recognizability, with vir-tually identical results. However, because these mean estimates werebased in some cases on very few estimates, and because the distribu-tion of actual proportions correct was more diverse, we used this valueinstead.

5 Here and in all regression analyses reported in this article, signifi-cance tests were performed on individual factors only when the overallregression equations were found to account for a significant propor-tion of the variance. The effects of interactions were examined after themain effects had been entered simultaneously into the equation (cf.Cohen & Cohen, 1983). In the interest of conserving space, statisticsfor the overall regressions and for interactions failing to attain signifi-cance at the .05 level are not discussed in the text.


observations was a consequence of unanticipated changes inthe recognizability of some of the stimulus persons. During asummer school break in which the experiment was temporarilyhalted, some of the targets changed dramatically in their identi-fiability. The next two studies used a different sort of stimuli,everyday objects, that were less likely to vary in recognizabilityover time, permitting better tests of the hypotheses.

Experiment 3

In Experiment 1, subjects' judgments about others' expertisewere correlated with their own familiarity with the stimulusperson. Because the target population was Columbia Universityundergraduates, this strategy made good sense—the subject'sown knowledge should be a reasonable guide to what othermembers of the same category will know (cf. Dawes, 1989).However, members of different social categories often have par-ticular domains of expertise. To create successful referring ex-pressions, a speaker must go beyond his or her own familiaritywith the referent and take such group differences into account.

To examine the effects of social category membership onmessage production, one must find social categories that aresocially shared, that elicit differential amounts of attributedknowledge, and that can be made relevant to communication.One such category is gender: Gender is or can be easily madesalient to the interactants (e.g., Taylor, Fiske, Etcoff, & Ruder-man, 1978), it elicits shared expectations about implications forothers' characteristics (e.g., Deaux, 1976), and it can be maderelevant to performance in referential communication tasks byusing stimuli that are of differential familiarity to men andwomen. One source of such differential knowledge is house-hold activities. Evidence indicates that household tasks such ascooking and maintenance are allocated differentially to the twosexes (Nyquist, Slivken, Spence, & Helmreich, 1985; Robinson,1988), and these differences in experience should lead to sexdifferences in domain-related expertise. If people are aware ofthese discrepancies in knowledge, there should be differencesin their expectations about what men and women know.

In Experiment 3, subjects estimated separately the percent-ages of male and female undergraduates who could correctlyidentify pictures of everyday things. Some of the objects werehypothesized to be more familiar to men, some more familiarto women, and some equally familiar to both. As in Experi-ment 1, population knowledge was estimated by asking all sub-jects to provide the name of the item (if they knew it) prior tomaking their judgments. It was anticipated that subjects wouldbe sensitive both to overall proportions of men and women ableto identify each of the objects and to the differences in propor-tions of men and women able to do so.

In addition, to assess bias more accurately, we obtained judg-ments on a rating scale that was comparable to the scale onwhich performance was measured. By having subjects estimatethe proportions of students able to identify each stimulus, esti-mates and actual values could be compared directly, permittinga quantitative assessment of accuracy and bias.

As noted above, one heuristic judges might use, especiallywhen they do not know the name of the stimulus, is their sub-jective feeling of familiarity. In Experiment 1, subjects' esti-mates of identifiability to self were significantly correlated with

their estimates of identifiability to others; however, it is notclear exactly what these identifiability-to-self ratings measure.An intuitively more meaningful index is one's feeling of know-ing—the subjective feeling that one could recognize the nameor retrieve it at a later point (Brown & McNeill, 1966; Tulving &Pearlstone, 1966). Nickerson et al. (1987) found that subjects'ratings of feeling of knowing were significantly correlated withestimates of others' knowledge. To see if similar effects wouldbe found with these stimuli, subjects were asked to rate theirfeeling of knowing if they did not know the name of the object.

Method

Materials. The stimuli consisted of black and white line drawings of10 objects in each of eight categories: sports equipment, householdappliances, car parts and tools, kitchen utensils, tools, electronicequipment, vegetables, and musical instruments. Categories were cho-sen that were expected to be more familiar to women (e.g., kitchenutensils), more familiar to men (e.g., tools), or equally familiar to both(e.g., musical instruments). Each set of 10 pictures (with each picturelabeled with an identifying letter) was arranged in arbitrary order onone sheet of paper; these eight pages of pictures were then randomlyarranged in 12 orders.

Answer sheets contained a response slot, in which subjects wrote thename of the stimulus if they knew it, and three ratings scales for eachpicture: an 8-point scale for rating feeling of knowing and two scalesranging from 0% to 100%, marked at 10% intervals, for estimating thepercentages of men and women who would know the name of the item.A postexperimental questionnaire collected pertinent background in-formation.

Subjects and procedure. Fifty subjects (28 women and 22 men) par-ticipated in groups of up to eight. They were instructed to go throughthe 80 pictures one at a time. They first determined whether they knewthe name of the pictured item and, if they did, wrote it in the spaceprovided. If they did not know the name, they rated their feeling ofknowing—the likelihood that they could retrieve the name at a laterpoint or would recognize the name if they saw it written. Then, allsubjects estimated the percentages of undergraduate men and womenwho would know that item's name. They worked at their own pace;when they were finished, they completed the follow-up questionnaire.


Proportions of subjects able to name the stimuli ranged from0 to 1.00 (M= .74, SD = .32 for male subjects; M= .67, SD = .35for female subjects). Differences between proportions ofcorrect identifications ranged across items from -.23(women > men) to .44 (men > women), with a mean of .06 (SD =.14). Because the categories were of a somewhat ad hoc natureand the variability within categories was great, we treated thestimuli as 80 individual items for the purposes of our analyses.

As in Experiment 1, there was good correspondence betweenmean estimates of identifiability and actual proportions ofcorrect identifications for both male and female targets. Resultsfrom regressions of estimates onto actual proportions correctare shown in Table 2. Although the overall correlation wassomewhat lower for male targets than for female targets, thetwo values were not significantly different (z = -1.23). Theserelationships are not simply the result of averaging over subjectswith different intuitions about the social distribution of knowl-edge: Individual subjects' correlations across the 80 stimuli


Table 2Summary of Regressions of Estimated Proportions CorrectOnto Actual Proportions Correct

Observations regressed

All targets (n = 80)MenWomenMen - women

Named (n = 78)MenWomenMen — women

Unnamed (n = 55)MenWomenMen — women

All targets (n = 42)MenWomenMen - women

Named (n = 41)MenWomenMen - women

Unnamed (n = 25)MenWomenMen - women

Intercept

Experiment 3

.54**

.49**

.00

.55**

.55**-.01

.47**

.43**-.02

Experiment 4

.50**

.48**

.01

.42**

.40**

.02

.40**

.42**

.00

B

.33

.40

.46

.34

.35

.45

.15

.17

.68

.40

.41

.64

.52

.51

.67

.17

.07

.73

r

.73**

.81**

.59**

.76**

.83**

.61**

.40*

.39*

.57**

.79**

.79**

.81**

.93**

.91**

.81**

.35

.16

.79**

R2

.53

.65

.34

.58

.68

.37

.16

.15

.32

.63

.63

.66

.87

.82

.66

.12

.03

.62

Note. Men - women = Estimated difference in percentage of correctidentifications of objects by men and women versus actual differencein percentage of correct identifications by men and women.*p<.005. **p<.001.

ranged from .31 to .71 for the male targets (M = .51, SD = .10)and .47 to .77 (M = .62, SD = .07) for the female targets—allgreater than the minimum value required (.29, df= 78) forsignificance at the .01 level.

To ensure that these high correlations did not result from afalse consensus bias or simple assumptions of similarity, esti-mates over correct and incorrect observations were examinedseparately. In Figure 2, the actual percentages of men andwomen who knew each item's name are plotted against theestimated percentages of men and women who would know it,separately for observations on which the target was correctlynamed (named) and those on which it was not named (un-named). The solid lines represent the best fit to the two sets ofestimates. Estimates on named observations show substantialsensitivity to level of knowledge in the target populations (seeTable 2), and sensitivity did not vary by target gender (z =— 1.14). Estimates on unnamed trials were significantly corre-lated with actual values but were also significantly less accuratethan those of the named group (zs = 3.03 and 4.11 for compari-sons of correlations for male and female targets, respectively,both ps < .005). This pattern of results not only holds for thepopulation means, but for individual subjects as well. Meanwithin-subject correlations were, for male targets, .02 and .53

for unnamed and named observations, respectively; the corre-sponding values for female targets were .23 and .58.

Despite their sensitivity to differences in degree of knowl-edge in the student population, subjects who know an item'sname might be biased in the direction of their own knowledge(cf. Nickerson et al, 1987). If estimates are compared to thebroken lines in Figure 2, which represent the unit lines, it isclear that subjects who knew the name of less recognizable ob-jects substantially overestimated the number of others whocould identify that object. Even items that were identified by10% or less of the subjects were estimated by those who knew itsname to be identifiable to 40-80% of the population (also seeTable 2). Thus, although subjects who knew an item's namewere aware that certain items are less likely to be known thanothers, they substantially overestimated the proportion of peo-ple who would know the names of the lesser known items.

To examine whether subjects attributed knowledge differen-tially to the two sexes, we plotted the mean difference betweenestimates for men and women against the actual difference inproportions correct for men and women (see Figure 3 and Table2). The values are again plotted separately for named and un-named observations. Both groups of subjects appear to be sen-sitive to gender differences in item knowledge, and equally so(z = .32); their judgments also do not appear to be biased to-ward one or the other sex, as indicated by the zero intercept.Thus, although subjects who did not know what something wascalled were poor judges of the relative proportion of peoplewho would know the name of that item, they were nonethelesssensitive, as a group, to which, if either, sex would be better atidentifying it.6

In contrast to estimates of overall expertise, there was sub-stantial variability in individual subjects' sensitivities to genderdifferences. Individual correlations ranged from. 14 to .64 (M=.44), and, although all but three of these reach the minimumrequired (.22) for significance at the .05 level, many were low.Subjects' perceptions of the rating task may have contributed tothis variability: Some stated that they thought it was a stereo-typing experiment and gave the same estimates for men andwomen to avoid appearing sex-biased.

When subjects did not know an item's name, they rated theirfeeling of knowing. For most subjects, these ratings were highlycorrelated with estimated percentages correct for both maleand female targets. Mean individual correlations for malejudges were .64 and .46 for male and female targets, respec-tively; the corresponding values for female judges were .26 and

6 Although it appears in Figure 3 that subjects may be sensitive to thesize of the discrepancy in performance by men and women, this is trueonly to a limited extent. When estimated versus actual absolute valuesof the difference between sexes were plotted against one another, thecorrelation for named observations was .63 and that for unnamed ob-servations was .35. When these correlations were obtained separatelyfor male- and female-advantage stimuli, there was little effect of theabsolute size of the discrepancy for female-advantage items (r = .01 fornamed and .33 for unnamed) and a slightly larger effect for the male-advantage items (r = .56 for named and .37 for unnamed). Hence, acautious interpretation of the findings would be that subjects are pri-marily sensitive to which sex would be better at identifying the stimuli,rather than to the size of the gender advantage.


Estimated 70-PercentCorrect:Women

Estimated 70-PercentCorrect: 6QJMen

• Named

° Unnamed

10 20 30 40 50 60 70

Actual Percent Correct: Women

80

10 20 30 40 50 60 70

Actual Percent Correct: Men

90 100

Figure 2. Mean estimated percentage of correct identifications of objects versus actual percentage ofcorrect identifications for women (top) and men (bottom). (Solid regression lines represent the best fit toeach set of estimates. The broken lines represent the unit lines.)

.67. Although the majority of these individual correlations weresignificant at the .05 level, there was substantial variabilityamong subjects' individual correlations, with some subjects'correlations negative or near zero. Individual correlations be-tween feeling of knowing and estimated male-female differ-

ence were even more variable. It appears that some judges relymore heavily on subjective feelings of recognition than othersbut that all consider additional information when determiningothers' knowledge.

In summary, people's estimates of others' knowledge of every-

EstimatedDifferencein PercentCorrect(Men minusWomen)

•30

50,

40.

30.

20.

10.

0

10.

20.

-30

-40

• NamedO Unnamed

4

° o

o o

i o ° o ^-^*^**

o

-20 -10 0 10 20 30 40

Actual Difference in Percent Correct (Men minus Women)

Figure 3. Mean estimated difference in percentage of correct identifications of objects by men andwomen versus actual difference in percentage of correct identifications by men and women. (Solid regres-sion lines represent the best fit of each set of estimates.)


day objects are reasonably accurate, and these assumptionsabout the way such knowledge is distributed appear to beshared, as evidenced by the fact that most individuals' esti-mates were highly correlated with actual values. In Experiment4, these judgments were used to estimate what speakers assumeabout their addressee's knowledge.

Experiment 4

To examine the effects of addressee gender on referring ex-pressions, we crossed two classes of addressees (men andwomen) with three types of stimuli (female-oriented, male-or-iented, and neutral). By using this paradigm, the audience de-sign hypothesis could be tested both within subjects (i.e., acrossitems of greater or lesser recognizability) and between subjects(i£., across categories of addressees who are more or less famil-iar with a particular item category).

As in Experiment 2, we expected that speakers would includemore descriptive information with the name of a stimulus asthe perceived recognizability of that stimulus declined. We alsoexpected the amount of descriptive content accompanying aname to affect the matcher's ability to identify the referent, asindicated by the total number of turns required to establishreference.

Finally, the hypothesis that people's intuitions about whatothers know are socially shared, so that an individual speaker'spresumptions about the audience's expertise can be approxi-mated by group means, was tested by asking each subject toperform, after completing the referential communication task,a knowledge attribution task very similar to that in Experiment3. If agreement between studies is good, then the use of meanjudgments from Experiment 3 as an estimate of speakers' pre-sumptions about their addressees is further validated.

Method

Materials. Two sets of 21 pictures were selected from those used inExperiment 3. Each set consisted of 7 items in three categories, oneperceived by subjects in Experiment 3 to be more familiar to men, onemore familiar to women, and one equally familiar to both: Set A con-sisted of tools, kitchen utensils, and musical instruments; Set B con-sisted of car parts, vegetables, and electronic equipment. Mean differ-ence in estimated percent correct by gender (men minus women) forthe male- female-, and neutrally oriented categories were .20, —. 11, and.01 for Set A; and. 18, -.08, and .01 for Set B. The direction and size ofdifference was consistent within each category. Pictures were mountedon notecards, the color of which indicated their category.

For the identification questionnaire that followed the referentialcommunication task, the response sheet used in Experiment 3 wasmodified to include two additional ratings for subjects who knew thename of an item: a 7-point scale on which they rated their confidencein the name and a place to indicate whether they knew the name priorto the experiment.

Follow-up questionnaires tailored to the subjects' experimentalroles were designed to elicit reactions to the referential communicationand knowledge judgment tasks and to obtain pertinent backgroundinformation.

Subjects and procedure. Forty pairs of subjects participated in theexperiment, 10 of each sex by experimental role combination. Subjectswere seated at small tables facing their partners and were separated bya cardboard barrier 53 in. long by 23 in. high (134.6 X 58.42 cm). Theirconversations were tape-recorded.

Instructions were identical to those used in Experiment 2, exceptthat directors were not allowed to refer to the notecard's color. Pairsrepeated the task three times with each set of cards (half received Set Afirst, and the other half received Set B first). As in Experiment 2, thecards were presented in a different random arrangement on each trial,but the order within a trial was kept constant for all subjects. After-ward, subjects independently performed the identification task andcompleted the postexperimental questionnaire.

Audiotapes were transcribed and punctuated, and directors' first-turn utterances were divided into idea units and coded, in the samemanner as in Experiment 2 (interrater reliability exceeded 90% foreach of these tasks). The content units and words of description in thedirector's first messages were counted (because very few subjects usedcategory information in this study, it was included in the descriptionword totals), as were the additional speaking turns by both partnersrequired to establish reference. Values were truncated to three stan-dard deviations above the mean.


Communication task. As in Experiment 2, the overall pat-tern of results replicated those reported in the other studies andhence are described briefly (complete analyses are available inFussell, 1991). Communicators became more efficient at de-scribing the stimuli over trials, both in terms of the length ofinitial messages (mean content words for Trials 1-3 were 1.7,1.3, and 1.2, respectively) and the number of turns required toestablish reference (mean turns for Trials 1 -3 were 2.9,2.2, and2.1, respectively), and this was true regardless of whether theyknew the name of the target.

To test our hypothesis that prior beliefs about addressees'knowledge affect message production, we examined speakers'Trial 1 messages that contained the name of the target. To en-sure that speaker uncertainty was not confounded with attrib-uted knowledge, we excluded all messages in which the namewas preceded or followed by a hedge orstated in terms of resem-blance.7 Our hypotheses predicted two effects of attributedknowledge: a general effect of overall identifiability of the tar-get (to men and women combined) and an effect of the lis-tener's gender.

To examine the effects of overall identifiability, we calculatedmeans for each dependent variable by item, collapsing over sexof speaker and sex of addressee. (Names from both stimulussets were combined for the analyses.) These values were con-trasted with the mean ratings of recognizability for each stimu-lus, averaged across male and female targets, provided by sub-jects who knew the correct name in Experiment 3 (which weterm attributed knowledge in the discussion below).8

As anticipated, speakers provided more information as the

7 Because a speaker's uncertainty about the identity of an object islikely to be reflected in the use of hedges and other devices, which inturn result in a lengthier message, we excluded these hedged messagesfrom the analysis. This provided the strongest test of the hypothesisthat speakers use longer utterances when they think their addressee isunlikely to recognize the target; however, results were nearly identicalwhen all messages containing names were used in the analysis.

8 We also performed the analyses using mean judgments fromcorrect observations in the current study, with the same results. How-ever, because estimates from the current study could have been af-fected by the process of communicating about the stimuli, we usedthose from Experiment 3 instead.


judged recognizability of the stimulus declined, when we usedboth content units and words of description as dependent mea-sures (rs = —.62 and —.66, respectively, both ps < .01). Thesefindings were confirmed by multiple regression of each depen-dent variable onto set (A or B), presentation order, two categoryvariables (hereinafter to be called category as a group), andattributed knowledge. Messages contained significantly lessidentifying information with the name of the target as per-ceived recognizability declined; for content units, r(34) = 4.38,p < .001, B = -2.76; for words of description, /(34) = 4.91, p <.001, B= —14.93. None of the other factors showed any sign ofan effect in either analysis, including presentation order (allts<\).

To examine the effects of gender category membership onmessage construction, means for each dependent measure werecalculated by stimulus and partner sex. All stimuli with at leasttwo observations per partner sex were analyzed in a repeatedmeasures analysis of variance; however, no differences werefound. The absence of an effect may have been due to the gener-ally small differences between estimates for male and femaletargets (see Figure 3), or to the limited number of observationsin each cell.

The effects of the amount of descriptive information addedto a name on matchers' abilities to identify the target was exam-ined by calculating the number of additional turns it took a pairto settle reference and regressing this value onto content units,presentation order, category, and stimulus recognizability (theproportion of correct identifications in Experiment 3). Maineffects were found for order, i(34) = -2.56, p < .02, B = -.05;and for stimulus recognizability, r(34) = —2.96, p < .006, B =-1.72: Fewer turns were required later in the sequence and formore recognizable stimuli. However, contrary to our predic-tions, the number of content units provided had no effect on thenumber of turns required to establish reference. Results aremuch the same when number of words of description is substi-tuted for number of content units in the regression.

As in Experiment 2, shared assumptions about stimulus rec-ognizability were evidenced in subjects' messages for each stim-ulus: Speakers provided more identifying information as theperceived probability of the addressee's knowing the name ofthe referent declined. However, the proportion of variance ex-plained by this factor was relatively small, and no effects attrib-utable to the matcher's gender were found. Most curious wasthe fact that messages for only 20% of the least recognizabletargets, and even fewer of the others, included any descriptiveinformation in addition to the name. It appears that our de-scribers' preferred initial strategy was to present the name andsee what happened.

Knowledge judgment task. The results of the knowledgejudgment task, performed by all subjects after they had com-pleted the referential communication task, provided good sup-port for the assumption that one can model speakers' beliefsabout their addressees' knowledge using judgments from anindependent group of subjects. Correlations between mean es-timated proportions correct by men and women in the twostudies were .92 or greater for named observations and .68 orgreater for unnamed observations.

As in the previous experiment, mean estimated identifiabil-ity was highly correlated with actual proportions correct for

both male and female targets (both rs = .79, p < .001 ).9 As Table2 indicates, the intercepts and raw correlations obtained byregressing estimates onto actual values were quite comparableto those of Experiment 3. When subjects could correctly namethe stimulus, they were significantly more accurate at assessingidentifiability both to men and to women (both zs > 4.0, ps <.001); however, sensitivity to gender differences in performancewas unaffected by whether they knew the name of the object(z < 1). Individual subjects' judgments showed the same patternas the group estimates. All but six of the individual correlationsfor male targets, female targets, and the male-female differ-ence were greater than the minimum required for significance(.39, df= 40) at the .05 level. These results provide further evi-dence to suggest that subjects who know what a stimulus objectis have socially shared assumptions about that item's identifia-bility to others.

As in the previous experiments, subjective feelings of recogni-tion affected subjects' estimates of stimulus recognizability.Rated confidence in one's response was correlated significantlywith estimates for male and female targets, but not for themale-female difference (mean individual rs = .48, .46, and .03,respectively). Likewise, correlations between feeling of know-ing ratings and estimates were fairly high for male and femaletargets (rs = .43 and .55, respectively) but negligible for themale-female difference (r = -.11).

General Discussion

The studies we have presented address two issues: the natureof communicators' prior beliefs about the social distribution ofknowledge and the effects of these prior beliefs on messageproduction in interactive communicative contexts. Overall, theresults suggest that people are fairly good at making judgmentsof others' knowledge and that speakers in interactive contextsdraw upon these prior beliefs, to a limited extent, in construct-ing their messages. We address each of these points in turnbelow.

Beliefs About the Social Distribution of Knowledge

Our studies show that people can estimate others' knowledgewith quite good accuracy. The ability appears to be general—ap-plying to public figures (Experiment 1), everyday objects (Ex-periments 3 and 4), and, from a recent experiment (Fussell &Krauss, 1991), New York City landmarks.

Two lines of evidence suggest that people agree in their per-ceptions of the social distribution of knowledge. First, almostall individual subject's correlations between estimated and ac-

9 All analyses were performed using as estimates of populationknowledge both proportions correct in the current sample (excludingsubjects who indicated that they learned the name during the commu-nication task) and proportions correct in the earlier study (in whichidentification performance could not be influenced by participationin the communication task). Because these two sets of accuracy scoreswere highly correlated (rs = .96 for both men and women and .67 for themale-female difference, all ps < .01), there was virtually no differencein the results using the two samples, and only those using the currentsample are discussed in the text.


tual values were statistically significant, and most individualsshowed the same biasing effects of knowing the name that wasfound in the group data. Second, mean estimates by subjects inExperiment 4 were highly correlated with estimates for thesame stimuli made by subjects in Experiment 3. Thus it appearsthat, on the basis of these group estimates, we can model withsome accuracy the assumptions speakers make about what then-addressees know.

As hypothesized, subjects' estimates of others' knowledgewere biased in the direction of their own knowledge. In Experi-ments 1, 3, and 4, estimates of others' ability to identify a givenstimulus were higher for subjects who knew the name of thestimulus than for subjects who did not. In fact, the knowledge-able subjects substantially overestimated the identifiability ofeven the most unrecognizable stimuli. The same pattern ofresults was found by Fussell and Krauss (1991) using landmarksas stimuli and appears similar to the estimates of correct re-sponses to general knowledge questions in the Nickerson et al.(1987) study.

The idea that individuals take others' knowledge, beliefs, andfeelings into account in formulating their own behavior haswide currency in social psychological theory. To cite just oneexample, social comparison theory (Festinger, 1954) postulatesthat people evaluate their own abilities and beliefs by compar-ing them with the abilities and beliefs of others—typically withabilities and beliefs that are normative for relevant categories ofothers. To make such comparisons, the individual must esti-mate how these abilities and beliefs are distributed in thosepopulations. Our findings suggest that these perceptions ofothers' beliefs and abilities are likely to be biased by what theperceiver him- or herself believes and can do.

Although the present studies do not directly address the issueof how people estimate others' knowledge, they do suggest sev-eral processes that may be involved:

Reasoning from one's own memory or cognitive processes. InExperiment 1, subjects' estimates of a public figure's identifi-ability to others were highly correlated with estimates of hisidentifiability to themselves. Similarly, in Experiments 3 and 4,estimates of the identifiability of everyday objects were corre-lated with feeling of knowing when subjects did not know thename and with confidence in the response when they did. Simi-lar results are reported by Fussell and Krauss (1991) and Nick-erson et al. (1987). Nevertheless, there is evidence to suggestthat these subjective feelings are not subjects' only sources ofinformation about others' knowledge. In the Fussell and Kraussstudy, subjects given the names of the stimuli (and thus notuncertain about them) were as good at assessing proportions ofcorrect identifications as those who generated the names them-selves. It is possible that an anchoring and adjustment process(Tversky & Kahneman, 1974) is used, in which subjects startwith their own feeling of knowing and then revise up or downbased on estimated recognizability.

Inferring from a small number ofindividuals. In a pilot study,subjects were asked how they made knowledge judgmentsabout how much students in different majors would knowabout a variety of topics. The most common responses werethat they used a few people they knew as models, used a stereo-type of the category, or used both. Although there is reason tobe skeptical about self-reports of cognitive processes (cf. Nisbett

& Wilson, 1977), a preference for reasoning from small num-bers of individual cases rather than from statistical data is con-sistent with results of several studies of social judgment (e^,Borgida & Nisbett, 1977; Hamill, Wilson, & Nisbett, 1980; Nis-bett, Borgida, Crandall, & Reed, 1976).

Inferences from others' likely behaviors. People may also as-sess what others know indirectly, for example by consideringthe opportunities to acquire various sorts of information thattheir typical activities and behaviors provide. Some subjectsreporting on their own judgment processes in the pilot studyclaimed that they considered each major's likely courses ofstudy Nisbett and Kunda (1985) have shown that subjects canestimate the frequency with which other students participate invarious types of activities with some accuracy, but further workis needed to determine whether people can accurately inferwhat others would know on the basis of their participation inthese activities. It is possible that judges use something resem-bling Kahneman and Tversky's (1982) simulation heuristic; thatis, they use the ease with which they could generate a scenarioby which the target could acquire the requisite knowledge toassess others' likelihood of knowing.

Beliefs About Others' Knowledge and ReferentialCommunication

We have argued that speakers' beliefs about their addressees'knowledge play an important role in message construction,even in interactional contexts in which feedback from the lis-tener is readily available. The present studies support this hy-pothesis: The tendency to add identifying information to thenames of both people (Experiment 2) and objects (Experiment4) increased significantly as the perceived likelihood of the lis-tener recognizing these stimuli declined.

Contrary to expectations, however, in Experiment 4 thematcher's gender had no effect on the amount of informationprovided. This in part may be due to the fact that althoughsubjects perceived some objects to be differentially familiar tomen and women, the size of the perceived discrepancy wasoften small. It would be interesting to perform the experimentagain using items of greater differential recognizability to menand women or using a set of social categories for which greaterdifferences in knowledge can be found.

Although we found significant effects of prior beliefs on mes-sage construction, models of language use (e.g., Clark & Mar-shall, 1981; Grice, 1975, Mead, 1934; Rommetveit, 1974) andprevious research (e.g, Danks, 1970; Fussell & Krauss, 1989a,1989b; Isaacs & Clark, 1987) led us to anticipate that the effectswould be larger. Communicators in our studies frequently pro-vided little or no identifying information with the name of astimulus, regardless of its recognizability. Several factors mayaccount for these discrepancies. One is a technical consequenceof the fact that the analyses rest on messages in which subjectsuse the name of the stimulus. It is difficult to determine howmany such observations will be obtained from a set of dyadsbefore the transcription of all their dialogues, and this is partic-ularly problematic with items from the low end of the recog-nizability continuum, which by definition few speakers arelikely to know, but which are essential for the analysis. Thus, it


is possible that we simply lacked enough observations to pro-vide good tests of the hypotheses.

A second factor is that the type of knowledge judged—abilityto name a stimulus from its picture—is somewhat differentfrom the type of knowledge required by the matcher in thereferential communication task. The matcher's task, matchinga stimulus with a name, is easier than identifying and namingthe stimulus, and it may well be that all speakers assumed thattheir listeners could identify the target under these constrainedcircumstances.

It is also possible that subjects were influenced by otheraspects of the communicative situation (Higgins, 1981). Evenwithin the constraints of the task used here, speakers may strivefor different goals (e.g., to amuse their partners, to create a favor-able impression, etc.), and these goals can influence the formtheir referring expressions take.

Although these factors probably played some part in atten-uating the effects of attributed knowledge on communication,the most important determinants probably stemmed from theinteractive nature of the task itself. As we argued in the intro-duction, audience design in conversational contexts has at leasttwo components: prior beliefs about one's communicativepartners and feedback from the ongoing conversation. Al-though we hypothesized that speakers would draw on bothsources of information in performing our task, it is likely thatthe interactive nature of the exchange increased their relianceon feedback and lessened their reliance on prior beliefs. Onereason for this may be the "real-time" nature of conversation:Speakers in interaction may not have enough time to considertheir addressee's perspective in the way that subjects perform-ing a paper-based judgment task can. Instead, they may beforced to use simplified judgment heuristics (e.g., determiningwhether the stimulus exceeds some arbitrary threshold of recog-nizability and, if so, using its name without any supportinginformation), or they may fail to take the listeners' knowledgeinto consideration at all.

More importantly, the availability of feedback may leadspeakers to rely less on prior beliefs than they would in nonin-teractive contexts, because they know they can revise their mes-sages on the basis of the addressee's responses (cf. Clark &Wilkes-Gibbs, 1986). This may be particularly true in thecurrent studies, in which the social consequences of producingan inappropriately detailed description are likely to be mini-mal. In cases where these consequences may be greater (e.g.,talking to a superior), we would expect speakers to rely moreheavily on prior beliefs. However, although the effects of priorbeliefs on message construction were not as strong as we hadanticipated, the fact that we found any such effects in an inter-active communicative context attests to their important role inthe process of audience design.

Although we have discussed prior suppositions and interac-tionally provided feedback as though they were independentcomponents of the audience design process, they are likely tobe dynamically related. On the one hand, prior expectationsmay guide message production and the elicitation and interpre-tation of feedback. On the other hand, feedback can lead tomodifications of one's prior beliefs about the addressee's cate-gory memberships and about the kinds of knowledge that arecharacteristic of a given social category. The roles of feedback

and prior beliefs, both separately and in interaction, can only bedisentangled when experimental paradigms permit one to as-sess speakers' prior beliefs independent of their talk.

References

Borgida, E., & Nisbett, R. E. (1977). The differential impact of abstractversus concrete information on decisions. Journal of Applied SocialPsychology, 7,258-271.

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

Brown, R., & McNeill, D. (1966). The "tip-of-the-tongue" phenome-non. Journal of Verbal Learning and Verbal Behavior, 4, 325-337.

Butterworth, B. (1975). Hesitation and semantic planning in speech.Journal of Psycholinguistic Research, 4, 75-87.

Cantor, N., Mischel, W, & Schwartz, J. (1982). Social knowledge: Struc-ture, content, use, and abuse. In A. Hastorf & A. Isen (Eds.), Cogni-tive social psychology (pp. 33-72). New \brk: Elsevier Science.

Clark, H. H. (1985). Language use and language users. In G. Lindzey &E. Aronson (Eds.), Handbook of social psychology (pp. 179-231).New York: Random House.

Clark, H. H., & Carlson, T. B. (1982). Speech acts and hearers' beliefs.In N. V Smith (Ed.), Mutual knowledge (pp. 1-36). San Diego, CA:Academic Press.

Clark, H. H., & Marshall. C. E. (1981). Definite reference and mutualknowledge. In A. K. Joshi, B. L. Webber, & I. A. Sag (Eds.), Elementsof discourse understanding (pp. 10-63). Cambridge, England: Cam-bridge University Press.

Clark, H. H., & Murphy, G. L. (1982). Audience design in meaning andreference. In J.-F. L. Ny & W Kintsch (Eds.), Language and compre-hension (pp. 287-299). Amsterdam: North-Holland.

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

Cohen, J., & Cohen, P. (1983). Applied multiple regression/correlationanalysis for the behavioral sciences (2nd ed). Hillsdale, NJ: Erlbaum.

Danks, J. H. (1970). Encoding of novel figures for communication andmemory. Cognitive Psychology, 1, 179-191.

Dawes, R. M. (1989). Statistical criteria for establishing a truly falseconsensus effect. Journal of Experimental Social Psychology, 25,1-17.

Deaux, K. (1976). Sex: A perspective on the attribution process. InJ. H. Harvey, W J. Ickes, & R. F. Kidd(Eds.), New directions in attribu-tion research (Vol. 1, pp. 335-352). Hillsdale, NJ: Erlbaum.

Festinger, L. (1954). A theory of social comparison processes. HumanRelations, 7, 117-140.

Fiske, S. T, & Taylor, S. E. (1984). Social cognition. New York: RandomHouse.

Fleming, J. H., & Darley, J. M. (1991). Mixed messages: The multipleaudience problem and strategic communication. Social Cognition, 9,25-46.

Fleming, J. H., Darley, J. M., Hilton, J. L., & Kojetin, B. A. (1990).Multiple audience problem: A strategic communication perspectiveon social perception. Journal of Personality and Social Psychology,58, 593-609.

Fussell, S. R. (1991). The coordination of knowledge in communica-tion: People's assumptions about others' knowledge and their effectson referential communication (Doctoral dissertation, Columbia Uni-versity, 1990). Dissertation Abstracts International, 51, 4101B.

Fussell, S. R., & Krauss, R. M. (1989a). The effects of intended au-dience on message production and comprehension: Reference in acommon ground framework. JournalojExperimentalSocialPsychol-ogy, 25,203-219.

Fussell, S. R., & Krauss, R. M. (1989b). Understanding friends and


strangers: The effects of audience design on message comprehen-sion. European Journal of Social Psychology, 19, 509-525.

Fussell, S. R., & Krauss, R. M. (1991). Accuracy and bias in estimatesof others' knowledge. European Journal of Social Psychology, 21,445-454.

Grice, H. (1975). Logic and conversation. In P. Cole & I L. Morgan(Eds.), Syntax and semantics: Speech acts (pp. 41-58). San Diego,CA: Academic Press.

Hamill, R., Wilson, T. D, & Nisbett, R. E. (1980). Insensitivity to sam-ple bias: Generalizing from atypical cases. Journal of Personalityand Social Psychology, 39, 578-589.

Hastie, R. (1981). Schematic principles in human memory. In E. T.Higgins, E. P. Herman, & M. Zanna (Eds.), Social cognition: TheOntario Symposium (pp. 39-88). Hillsdale, NJ: Erlbaum.

Higgins, E. T. (1981). The "communication game": Implications forsocial cognition and communication. In E. T. Higgins, C. P. Her-man, & M. P. Zanna (Eds.), Social cognition: TheOntario Symposium(Vol. 1, pp. 343-392). Hillsdale, NJ: Erlbaum.

Innes, J. M. (1976). The structure and communicative effectiveness of"inner" and "external" speech. British Journal of Social and ClinicalPsychology, 15, 97-99.

Isaacs, E. A., & Clark, H. H. (1987). References in conversation be-tween experts and novices. Journal of Experimental Psychology: Gen-eral, 116, 26-37.

Kahneman, D, Slovic, P., & Tversky, A. (1982). Judgment under uncer-tainty: Heuristics and biases. Cambridge, England: Cambridge Uni-versity Press.

Kahneman, D., & Tversky, A. (1982). The simulation heuristic. In D.Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment under uncer-tainty: Heuristics and biases (pp. 201-208). Cambridge, England:Cambridge University Press.

Kaplan, E. (1952). An experimental study on inner speech as contrastedwith external speech. Unpublished master's thesis, Clark University,Worcester, MA.

Krauss, R. M. (1987). The role of the listener: Addressee influences onmessage formulation. Journal of Language and Social Psychology, 6,81-97.

Krauss, R. M., & Fussell, S. R. (1988). Other-relatedness in languageprocessing: Discussion and comments. Journal of Language and So-cial Psychology, 7, 263-279.

Krauss, R. M., & Fussell, S. R. (1990). Mutual knowledge and com-municative effectiveness. In J. Galegher, R. E. Kraut, & C. Egido(Eds.), Intellectual teamwork: Social and technical bases of collabora-tive work (pp. 111-145). Hillsdale, NJ: Erlbaum.

Krauss, R. M., & Fussell, S. R. (199 la). Constructing shared communi-cative environments. In L. B. Resnick, J. Levine, & S. D. Behrend(Eds.), Socially shared cognition (pp. 172-200). Washington, DC:American Psychological Association.

Krauss, R. M., & Fussell, S. R. (199 lb). Perspective-taking in communi-cation: Representations of others' knowledge in reference. SocialCognition, 9, 2-24.

Krauss, R. M., Vivekananthan, P. S., & Weinheimer, S. (1968). "Innerspeech" and "external speech": Characteristics and communicationeffectiveness of socially and nonsocially encoded messages. Journalof Personality and Social Psychology, 9, 295-300.

Krauss, R. M., & Weinheimer, S. (1966). Concurrent feedback, confir-

mation, and the encoding of referents in verbal communication.Journal of Personality and Social Psychology, 4, 343-346.

Kraut, R. E., Lewis, S. H., & Swezey, L. (1982). Listener responsivenessand the coordination of conversation. Journal of Personality and So-cial Psychology, 43, 718-731.

Markus, H., & Zajonc, R. B. (1985). The cognitive perspective in socialpsychology. In G. Lindzey & E. Aronson (Eds.), The handbook ofsocial psychology (pp. 137-230). New York: Random House.

Mead, G. H. (1934). Mind, self and society. Chicago: University of Chi-cago Press.

Nickerson, R. S., Baddeley, A., & Freeman, B. (1987). Are people'sestimates of what other people know influenced by what they them-selves know? Ada Psychologica, 64, 245-259.

Nisbett, R. E., Borgida, E., Crandall, R., & Reed, H. (1976). Popularinduction: Information is not always informative. In J. S. Carroll &J. W Payne (Eds.), Cognition and social behavior (pp. 227-236). Hills-dale, NJ: Erlbaum.

Nisbett, R. E., & Kunda, Z. (1985). Perception of social distributions.Journal of Personality and Social Psychology, 48, 297-311.

Nisbett, R., & Ross, L. E. (1980). Human inference: Strategies and short-comings of social judgment. Englewood Cliffs, NJ: Prentice-Hall.

Nisbett, R. E., & Wilson, T. D. (1977). Telling more than we can know:Verbal reports on mental processes. Psychological Review, 84, 231—259.

Nyquist, L., Slivken, K., Spence, J. T, & Helmreich, R. L. (1985). House-hold responsibilities in middle-class couples: The contribution ofdemographic and personality variables. Sex Roles, 12,15-34.

Robinson, J. P. (1988). Who's doing the housework? American Demo-graphics, 10, 24-63.

Rommetveit, R. (1974). On message structure: A framework for thestudy of language and communication. New \brk: Wiley.

Ross, L., Greene, D, & House, P. (1977). The false consensus phenome-non: An attributional bias in self-perception and social perceptionprocesses. Journal of Experimental Social Psychology, 13, 279-301.

Schegloff, E. A. (1972). Notes on a conversational practice: Formulat-ing place. In D. Sudnow (Ed.), Studies in social interaction (pp. 75-119). New York: Free Press.

Schober, M. E, & Clark, H. H. (1989). Understanding by addresseesand overhearers. Cognitive Psychology, 21, 211-232.

Taylor, S. E., & Crocker, J. (1981). Schematic bases of social informa-tion processing. In E. T. Higgins, C. P. Herman, & M. P. Zanna(Eds.), Social cognition: TheOntario Symposium (pp. 89-134). Hills-dale, NJ: Erlbaum.

Taylor, S. E., Fiske, S. T., Etcoff, N, & Ruderman, A. (1978). The cate-gorical and contextual bases of person memory and stereotyping.Journal of Personality and Social Psychology, 37, 357-368.

Tulving, E., & Pearlstone, Z. (1966). Availability versus accessibility ofinformation in memory for words. Journal of Verbal Learning andVerbal Behavior, 5, 381-391.

Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty:Heuristics and biases. Science, 185,1124-1131.

Werner, H., & Kaplan, B. (1963). Symbol formation. New York: Wiley.

Received August 1,1990Revision received May 20,1991

Accepted September 11,1991 •

coordination of knowledge in communication: effects of ... · coordination of knowledge in...

Documents