Pragmatics in memory: A study of natural conversation

JOURNAL OF VERBAL LEARNING AND VERBAL BEHAVIOR 16, 549-560 (1977)

Pragmatics in Memory: A Study of Natural Conversation

JANICE M. KEENAN, BRIAN MACWHINNEY, AND DEBORAH MAYHEW

University of Denver

Sentence processing in the context of natural, purposeful communication is said to differfrom sentence processing in laboratory experiments in that pragmatic information isinvolved. Included in pragmatic information are the speaker's intentions, beliefs, andattitude toward the listener; such information is referred to as the interactional content ofan utterance. Recognition memory for statements made during a luncheon discussiongroup was tested in an incidental learning paradigm following a retention interval of 30hours. Statements which were high in interactional content yielded excellent memory forsurface form, as well as meaning; statements low in interactional content showed nomemory for surface form, and less memory for content. Three control studies demonstratethat this difference in memory for high and low interactional content statements cannot bedue to (a) differences in the textual properties of the sentences; (b) differences in thequality of the distractors; or (c) reconstruction based on knowledge of the speaker'sstylistic habits.

In his most recent book Neisser (1976)argues that "the study of information proces-sing ... has not yet committed itself to anyconception of human nature that could applybeyond the confines of the laboratory" (p. 6).With respect to memory, in particular, hestates, "Until we know more about memory inthe natural contexts where it develops and isnormally used, theorizing is premature" (p.142). However, elsewhere, he cautions that"Demands for ecological validity are onlyintelligible if they are specific. They must pointto particular aspects of ordinary situationsthat are ignored by current experimentalmethods, and there must be good reason tosuppose that these aspects are important" (p.34).

The first section of this paper discusses atype of sentential information—pragmaticinformation—which is rarely involved in

This research was supported by a grant from theUniversity of Denver Faculty Research Council toJanice M. Keenan. Reprints may be obtained from Dr.Janice M. Keenan, Department of Psychology, Univer-sity of Denver, Denver, Colorado 80208.

laboratory experiments on sentence memory,but which appears to be important whensentences are processed in the context ofnatural, purposeful communication. This isfollowed by a series of experiments investiga-ting memory for statements occurring in anatural, conversational setting.

Pragmatic Information in NaturalisticSentence Processing

A number of the current models of humanmemory (e.g., Anderson & Bower, 1973;Kintsch, 1974) describe the process of encod-ing a sentence as drawing upon two types ofinformation: (1) syntax—knowledge of thestructure of linguistic expressions, gram-matical as well as lexical information; and (2)semantics—knowledge of objects and theirrelations used to determine the referents of theexpression. These models may adequatelydescribe the encoding of linguistic informationin the laboratory. But when a sentence isspoken by a real person in a real situation, itappears that additional information is brought

Copyright © 1977 by Academic Press, Inc.All rights of reproduction in any form reserved.Printed in Great Britain

549ISSN 0022-5 371

550 KEENAN, MACWHINNEY, AND

MAYHEW

to bear in the encoding of that sentence. Thisadditional information concerns knowledgeabout the context of the linguistic expression,the intentions of the speaker, the speaker'sexpectations of the hearer, and the formalidentity of the speech act. Linguists haveutilized such information in determining theconditions governing the well-formedness ofsentences and texts (e.g., Stalnacker, 1972).Philosophers have studied the role of suchinformation in determining truth values (e.g.,Montague, 1972). Both groups refer to it asthe study of pragmatics.

The term pragmatics is usually taken torefer to information about the communicativesituation, the temporal position of the ut-terance in the communication, and varioustypes of information about the speaker, e.g.,his intentions, beliefs, and knowledge of andattitude toward the listener. Our experimentsfocus on the latter type of pragmatic infor-mation—information about the speaker's inten-tions, his beliefs, and his relations with thelistener. We refer to this information as theinteractional content of an utterance. Theinteractional content of an utterance can becontrasted with its propositional content, sincethe latter conveys information about objectsand events in the world and is essentially inde-pendent of personal knowledge of the speaker.

In assessing the role of pragmatics insentence memory, we examined memory forindividual sentences which differed in theamount of interactional content they conveyor elicit in interpretation. Typically, theamount of interactional content conveyed by astatement is critically dependent on one'sknowledge of the speaker. However, in orderto illustrate what is meant by the term inter-actional content, we present examples ofsentences which do not depend on one'sknowledge of the speaker, such as (l)-(4)below.

(1) Low Interactional Content: Do youalways use CRT displays?

(2) High Interactional Content: Do youalways put your foot in your mouth?

Or, to use an example which does not employa cliche:

(3) Low Interactional Content: I think thereare two fundamental tasks in this study.

(4) High Interactional Content: I think you'vemade a fundamental error in this study.

In constructing these examples of high inter-actional content statements, we simply madethe propositional content of the utteranceconvey interactional information. It should benoted that the difference between high and lowinteractional content in the actual sentencesused in this experiment was quite differentfrom that illustrated by the above pairs. In theexperiment, both high and low interactionalcontent statements conveyed similar amountsof propositional content, and, unlike the pairsabove, the difference in interactional contentcould only be detected by someone familiarwith the basic sociolinguistic (Ervin-Tripp,1969) components of the communication, thatis, someone who was familiar with the speaker,the topic, the setting, and the purpose of thecommunication.

In formulating hypotheses for the role ofpragmatics in sentence memory, we found butone study which directly addresses this issue(Jarvella & Collas, 1974). These authorsexamined memory for sentences which canconvey either of two intentions of the speaker.An example is:

(5) I've never seen you wear that before.

This statement may be either a compliment oran insult, depending on the context. Jarvellaand Collas found superior recognition forthese statements when the recognition contextinduced the same interpretation of sentenceintention. These results indicate that prag-matic information, such as the speaker'sintention, is encoded and stored.

Our experiment on the role of pragmatics insentence memory goes beyond the Jarvella andCollas study in two respects. First, we testedmemory for sentences which occurred spon-

PRAGMATICS IN MEMORY 551

taneously in a naturalistic conversation, thatis, the experimental materials were generatedby the subjects themselves, and consequently,were more personally meaningful than Jarvellaand Collas' experimenter-constructed sen-tences presented in a laboratory. Our experi-ment therefore allows us to examine (a) thegeneralizability of laboratory experiments onsentence memory to real-world settings and (b)the possible existence of differences in memoryfor sentences that are personally meaningfulversus memory for sentences that are not.Second, we are not simply testing whetherpragmatics affects memory, rather we aretesting two very specific hypotheses con-cerning the role of pragmatic information inmemory.

First, we hypothesized that the meanings ofsentences high in interactional content wouldbe more memorable than those which are low.This hypothesis is based on the belief that sub-sequent interactions with the speaker oftendepend upon the contents of earlier inter-actionally significant events. Second, we hypo-thesized that memory for the surface structureof high interactional content statements wouldexceed memory for the surface forms of lowinteractional content statements. This hypo-thesis is based on the fact that frequently theway a statement is made, the choice of wordsand the style with which it is executed, is anintegral part of the content of high inter-actional content statements. Just consider thedifference between (6) and (7):

(6) Would everyone please stop talking?(7) Would everyone please shut up?

Here the prepositional content of these twosentences may be equivalent in the sense thatany situation fulfilling (6) would also fulfill (7).However, the surface structure of the twosentences reflects marked differences in inter-actional content.

Method

Subjects. Nineteen members of the Psy-chology Department (faculty and graduate

students) at the University of Denver served assubjects. They were participating in a researchdiscussion and were unaware that they wouldlater be given a memory test on the discussion.

Procedure and materials. The proceduralformat for a research luncheon discussion atthe University of Denver consists of oneparticipant presenting a set of ideas or experi-ments and others freely interrupting to askquestions, expand ideas, give criticism, ormake humorous remarks. The research discus-sion under study was held in the usual mannerwith one exception: it was tape-recorded.Participants were aware of the tape recorderbut were not aware of the purpose for taping,nor did anyone inquire as to the purpose. Thespeaker for the day was one of the authors. Hedid not attempt to inject statements particu-larly suited to the experiment; he simply spokein his usual manner.

After the discussion we played the tape,selecting statements for the recognition test.Two classes of statements were defined: highinteractional content statements and low inter-actional content statements. High inter-actional content statements typically conveyedwit, sarcasm, humor, or personal criticism. Weselected, through mutual decision, 15 suchstatements for the recognition test. To matchthese 15 statements, 15 statements which werelow in interactional content were selected fromthe tape in a random fashion. Eight of the 15high interactional content statements wereuttered by the main speaker and the remaining7 were uttered by other participants. The 15low interactional content statements werematched with the high interactional contentstatements in terms of speakers.

Table 1 presents one of the high and one ofthe low interactional content statements. Bothwere uttered by the same speaker. Thestatement about Italians is classified as high ininteractional content because, for someonewho knows the speaker, this statement con-veys not only information about Italians butalso information about the speaker. It was thespeaker's witty way of excusing himself for nothaving completed running the Italian subjects

552 KEENAN, MAcWHINNEY, AND MAYHEW

TABLE 1

EXAMPLES OF A HIGH AND A Low INTERACTIONAL

CONTENT STATEMENT AND THEIR DISTRACTORS"

High interactional content

Target: Italians, you know what Italians are like, they hada strike, they had a heat wave.

(a) Everyone knows what happens in Italy, firstthey had a strike, then they had a heat wave.

(b) If in Madagascar they're weird enough to putthe verb first, you can guess what their eyemovements must be like.

(c) They're so weird in Madagascar they put verbsfirst; you can imagine what they do with theireyes.

Low interactional content

Target: You put a little morpheme that says you'regoing to choose the Object as Subject.

(a) When you get an Object topic, you add onanother little morpheme.

(b) What Hungarian does is agree with the Object,if the Subject isn't salient.

(c) If the Hungarian Subject isn't salient, the Objectagrees.

" (a) Paraphrase of target, (b) new content distractor,(c) paraphrase of (b).

in his study. The interactional content of thisstatement might thus be paraphrased as: Thespeaker is witty and believes that he canconvince the listeners it was not his fault thatthe study was not yet completed. In contrast,the low interactional content statement simplyconveys a fact about the world. Notice that interms of rhythmic pattern and use of thesecond person, it is just as conversational asthe high interactional content statement, but itconveys relatively little information about thespeaker or his relation to the listener.

The classification of statements as high orlow in interactional content was verified by in-formally asking three psycholinguists to rateeach statement as high or low. These psycho-linguists were not from the University ofDenver, but were familiar with the mainspeaker and the topic of the discussion. Theirratings agreed completely with our classifi-cation.

For each of the 30 target statementsselected, three distractors were constructed.All distractor items preserve the illocutionaryforce of their targets in the sense that all fouritems were jokes, sarcastic remarks, descrip-tive statements, etc. They differ in terms ofpreservation of prepositional content. Thethree distractor types are: (1) a paraphrase ofthe target; (2) a statement which differs inpropositional content, but whose prepo-sitional content was plausible in terms of thediscussion; and (3) a paraphrase of (2). Thislatter distractor is a procedural necessity tomake (2) a plausible alternative. Examples ofthe distractors are also shown in Table 1.

The recognition test was thus a 30-itemfour-alternative multiple choice test. Each itemidentified the speaker, e.g., "Brian said: (a) (b)(c) (d)." Instructions informed the subject thatonly one of the four statements was actuallysaid. They were told, however, to examineeach of the alternatives and to rate eachalternative using the following scale:

OLD = I'm sure the statement was said.GUESS = I cannot remember anything about

this statement. NEW = I'm sure thisstatement was not said.

This 3-point confidence scale was employed,rather than the typical OLDI NEW responsecategories, in order to separate correct guessesfrom true hits. It thus provides a clearerinterpretation of hits and false alarms. The testwas administered by giving each participant inthe discussion a booklet and asking him toreturn it as soon as possible. The intervalbetween research discussion and therecognition test ranged from 27 to 48 hours,with the average at approximately 30 hours.

Results

Table 2 presents the overall percentages ofOLD responses for each type of test item. Thefirst thing to note is that, even though thesubjects were not aware that they would begiven a memory test and even though the test


the rate of false alarms. In calculating d'm,paraphrases of targets, together with thetargets, were counted as correct responses andonly the two distractor types which differed inprepositional content were used to calculatethe rate of false alarms.

Each of the analyses to be reported wasperformed twice: once treating subjects as arandom factor while collapsing over state-ments, and once treating statements as arandom factor while collapsing over subjects.Results are reported using min F' whichcombines the two analyses (Clark, 1973).

The first row of Table 3 presents theaverage d's[ and d'm for both high and lowinteractional content statements. Analysis ofvariance confirmed that the meanings of highinteractional content statements were recog-nized better than low interactional contentstatements, min F'(\, 37) = 16.27, p < .001.More importantly, memory for surface formwas significantly better for high interactionalcontent statements than for low interactionalcontent statements, minF'(l,41) = 10.61,;? <.005.

An analysis was also performed on theitems to determine whether the 16 statementsuttered by the main speaker were morememorable than the 14 statements uttered bythe other participants. Since both high and lowinteractional content statements were balancedfor speaker, this analysis had two mainfactors: Speaker and Interactional Content.There was no effect of Speaker on the

TABLE 3 MEASURES OF

DISCRIMINABILITY BETWEEN TARGETSAND DlSTRACTORS (d'sf) AND BETWEEN TARGETS PLUSTHEIR PARAPHRASES AND OTHER DISTRACTORS (rf'm)

FOR THE CONVERSATION EXPERIMENT ANDLIST-LEARNING CONTROL EXPERIMENT

High interactional Low interactional

was given 30 hours after the discussion,subjects could still discriminate between themeanings of utterances which actually occur-red and those which did not occur but wereequally plausible. The mean percentage ofOLD responses to targets and paraphrases oftargets, pooling over high and low inter-actional content items, was 28% contrastedwith a mean percentage of 5% for statementsexpressing new content.

The interactional content of an utterancewas an important factor in determining thememorability of its meaning. There were 37%OLD responses to high interactional contenttargets and their paraphrases, and only 18.5%OLD responses to low interactional contenttargets and their paraphrases.

Not only was there better memory formeaning when the interactional content washigh, but there was also the predicted findingof better memory for the surface form of theutterance. When the interactional content washigh, there were 56% OLD responses totargets compared to 18% for paraphrases. Onthe other hand, when the interactional contentwas low, subjects responded OLD equallyoften to targets (19%) and paraphrases (18%).

Because of some dependencies in the res-ponses to the four alternatives, all effects weretested using d' scores. Two d' scores werecalculated. One measures memory for surfaceform, d'sf; the other measures memory formeaning, d'm. In calculating d'st, only targetswere counted as correct responses, and allthree distractor types were used to calculate

TABLE 2: OVERALL PERCENTAGES OF OLDRESPONSES TO TARGETS, PARAPHRASES OF

TARGETS, AND NEW CONTENT DISTRACTORS FOR THE

CONVERSATION EXPERIMENT

Target Paraphrase New

High int. content 56 18 2

Low int. content 19 18 8

High interactional Low interactional content contentd's! d'm d'sf d'mConversation 1.73 2.81 0.47 1.18List-learning 1.12 1.96 1.20 1.72

554 KEENAN, MACWHINNEY, AND

MAYHEW

memorability of utterances, F(l, 24) < 1; norwas the interaction significant, F(l, 24) < 1.The only significant effect was the effect ofInteractional Content, F(l, 24) = 20.29,;? <.01.

Discussion

The results of this experiment show adramatic difference in memorability betweenstatements classified as high and low in inter-actional content: There is greater accuracy inrecognizing the meanings of high interactionalcontent statements, and, also, there is excellentmemory for the surface forms of high interac-tional content statements while virtually nomemory for the surface forms of low interac-tional content statements. We would like toclaim that these differences in memorabilityare due to differences in the amount of prag-matic information which supports the encod-ing and storage of sentences in a conversa-tion. However, alternative explanations exist.

First, the difference in memorability bet-ween high and low interactional content state-ments could be due to differences between thesentences besides the difference in the amountof interactional content they convey. Becausethe sentences were uttered spontaneously, theycould not be matched on factors known toaffect memorability. It is possible that the twosets are not equal in terms of imagery, wordfrequency, syntactic structures, or ease of inte-gration. Moreover, they differ in semanticcontent. One or more of these other factorscould be responsible for the difference inmemorability between high and low inter-actional content statements. Control Study 1tests this possibility.

Second, even if the target statements wereequally memorable, differences in memory forsurface form could still obtain if theparaphrase distractors for the high inter-actional content statements did not function aswell as the distractors for the low interactionalcontent statements. Control Study 2 tests thispossibility.

Finally, the results for high interactionalcontent statements may simply reflect recon-struction on the basis of knowledge of thespeaker rather than memory for particularutterances. Since we claim for these sentencesthat the choice of words conveys informationabout the speaker, it may be the case thatanyone who knows the stylistic habits of thespeaker could determine whether he said:"Would everyone please stop talking?" or"Would everyone please shut up?" ControlStudy 3 tests this possibility.

Control Study 1: Are theredifferences in sentence memorabilitywhen the interactional component is

removed?

In an attempt to rule out nonpragmaticfactors as an explanation of the difference inmemorability between high and low inter-actional content statements, we performed atypical sentence memory experiment using thetarget statements of the first experiment. Thesentences were presented in a list-learningfashion, and memory was tested by the samerecognition memory test employed in theconversation experiment. If differences inlexical information, syntax, or semantics areresponsible for the superior memory for highinteractional content statements in theconversation experiment, then we shouldobtain the same pattern of results in thiscontrol experiment. On the other hand, if inter-actional factors were primarily responsible forthe results, then, in this second experiment,where interactional factors have beeneliminated by presenting the sentences as itemson a list, there should be no difference inmemorability between high and low inter-actional content statements.

Method

Subjects. Thirty-four Introductory Psy-chology students from the University ofDenver volunteered to serve as subjects forcourse credit.

PRAGMATICS IN MEMORY

Procedure and materials. Subjects listenedto a tape consisting of 100 conversational-typesentences. The list of sentences included the 15high and 15 low interactional content state-ments from the conversation experiment and70 filler sentences. Fillers were included toprevent ceiling effects on the memory test. Allof the sentences were recorded by one speakerwho spoke in a conversational tone. The orderof sentence presentation was randomized withthe constraint that the first 10 and last 10sentences of the list be filler items to controlfor primacy and recency effects. Subjects wereinstructed to listen carefully to the sentences inorder to do well on a subsequent memory test;the nature of the test was not specified.

Immediately after listening to the 100sentences, the subjects were given the samememory test used in the conversation experi-ment, except here the items were not identifiedby a speaker. As before, the subjects were toldto rate each alternative using the 3-point scaleof OLD, NEW, and GUESS.

Results

Table 4 presents the overall percentages ofOLD responses for each type of test item. Likethe participants in the discussion, these sub-jects were clearly able to differentiate betweenthe contents of sentences they heard andsentences they did not hear. Collapsing overhigh and low interactional content statements,the percentage of OLD responses to targetsand their paraphrases was 32%, while thepercentage of OLD responses to new content

TABLE 4

OVERALL PERCENTAGES OF OLD RESPONSES TO

TARGETS, PARAPHRASES OF TARGETS, AND NEW

CONTENT DISTRACTORS FOR CONTROL STUDY 1

High interactional Target Paraphraseof target

Newcontent

content 47 21 7Low interactional 47 14 8

distractors was only 7.5%. Unlike the partici-pants in the discussion, however, these sub-jects do not show differences in memorybetween high and low interactional contentstatements. The percentage of OLD responsesto the target was exactly the same, 47%, forboth high and low interactional content state-ments.

The analysis in terms of d' scores wasidentical with that used in the previousexperiment. Mean d' scores are presented inthe lower half of Table 3. Analysis of varianceconfirmed that for these subjects there was nodifference in memorability between the sets ofstatements we termed High and Low in inter-actional content. [Minimum F' was not com-puted, since for the surface form analysisF,(l, 33) = 1.41, p = .24, and F2(\, 28) < 1,and for the meaning analysis Fj(l, 33) = 2.33,^ = .14,andJF2(l,28)< 1.]

The data from both experiments werecombined in two analyses of variance of d'scores, one for surface form and one formeaning. Both analyses yielded significantinteractions between Experiment and Inter-action^ Content. In the analysis ?OT surfaceform, min F'(l, 60) = 7.75, p < .01; in theanalysis for meaning, min F'(l, 60) = 9.66,;?< .005. These interactions confirm the resultsof the separate analyses where differences ininteractional content led to differences inmemory in the conversation experiment, butnot in the list-learning experiment. Eventhough the d' sf values for both sets of state-ments in the list-learning experiment are higherthan the d'sf values for low interactionalcontent statements in the conversation experi-ment, high interactional content statementswere remembered better in the conversationthan in the list-learning paradigm, minF'(l,60) = 4.03, p < .05. This is a ratherdramatic demonstration of the role ofpragmatics in memory since the participants inthe discussion were tested after 30 hours in anincidental learning paradigm, whereas thesubjects in the list-learning experiment weretested immediately in an intentional learningparadigm.


Discussion

The results of this experiment allow us toeliminate the possibility that there were dif-ferences, other than interactional differences,between the sentences of the conversationexperiment which led to differences in memor-ability. This experiment clearly demonstratesthat the high interactional content statementsare not intrinsically more memorable than thelow interactional content statements. Further-more, it demonstrates that for these sentencesthe textual elements of the sentence play aminor role in eliciting pragmatic information.It is one's knowledge of the speaker whichrenders a statement sarcastic, humorous, orwitty, not its linguistic structure.

Control Study 2: A test of the quality of theparaphrase distractors

In interpreting the results of the conver-sation experiment we assumed that the highrecognition rates for high interactional contentstatements reflected subjects' memory for thesurface forms of these statements. An alter-native explanation is that the paraphrasedistractors for these statements were of poorerquality than the paraphrases for low inter-actional content statements. In other words,the subject who remembered the gist of a highinteractional content statement could easilyselect the target over its paraphrase, becausethe paraphrase did not have the propertiesassociated with a conversational statement,e.g., rhythmic pattern, lexical items, etc.

In the following experiment subjects weregiven the 15 high and the 15 low interactionalcontent statements, each paired with itsparaphrase, and asked to choose which of thetwo statements would be most likely to occurin a conversational setting such as the researchluncheon discussion. If the paraphrase distrac-tors of high interactional content statementsare less conversational than the targets, thenthe same pattern of results which obtained inthe conversation experiment should obtain

here: high interactional content targets selec-ted more frequently than their paraphrases andlow interactional content targets selected justas often as their paraphrases. On the otherhand, if the paraphrases for both sets ofstatements are of equal quality, no such inter-action should obtain.

Method

Subjects. Thirty-one students in an under-graduate learning course at the University ofDenver agreed to serve as subjects duringclass time. None of the subjects knew thespeakers from the conversation experiment.

Procedure and materials. The subjects werefirst given an extensive description of thesetting for the conversation experiment. Thisincluded a description of the people involved(no names were mentioned), the purpose andprocedural format for the discussion, themanner in which people interact in thesediscussions, and some information concerningthe topic under discussion. They were also toldthat we had selected certain statements fromthe discussion and had constructed para-phrases for them, and that we wonderedwhether they could distinguish our para-phrases from the real statements.

Following this description subjects weregiven test booklets containing 30 pairs of state-ments. Each pair consisted of one of the 30target statements from the conversation ex-periment and its paraphrase. The subjects weretold that one of the statements in each pair hadactually occurred in the conversation, and thattheir task was to try to determine which one itwas. The experimenter suggested that inmaking their judgments they consider suchthings as phrase structure, word choice, andoverall "sound" of the statement. It should benoted that all subjects felt that the task waseasy to do and they felt quite confident of theirchoices.

Results and Discussion

The top half of Table 5 presents theproportion of targets and paraphrases selected


TABLE 5

PROPORTION OF TARGETS AND PARAPHRASES SELECTED AS "MOST LIKELY TO OCCURIN A CONVERSATION" FOR HIGH AND Low INTERACTIONAL CONTENT STATEMENTS AND

FOR SUBJECTS WHO KNEW THE SPEAKERS AND THOSE WHO DID NOT

High interactional content Low interactional content

Target Paraphrase Target Paraphrase

Subjects who did not know thespeakers (Control Study 2) 41 59 44 56

Subjects who knew the speakers(Control Study 3) 51 49 48 52

by subjects as most likely to occur in aconversation such as the research luncheondiscussion. The paraphrases were selectedmore often than the targets for both high inter-actional content statements, 59 versus 41%,and low interactional content statements, 56versus 44%. This difference is highly signifi-cant, F(\, 30) =17.7, p < .001. Further, therewas no interaction between target and para-phrase and high and low interactional content,F(\, 30) < 1.

These results clearly show that there wasnothing peculiar about the paraphrase distrac-tors for high interactional content statementsthat could have led subjects in the conversa-tion experiment to reject them in favor of thetargets. The paraphrases for both high and lowinteractional content statements were equallygood candidates for conversational statements,and, interestingly, were even thought to bebetter candidates than the targets themselves.

Control Study 3: Is knowledge ofthe speaker sufficient to distinguish

target from paraphrase?

The subjects in Control Study 2 did notknow the speakers; their judgments simplyreflect the conversational quality of targets andparaphrases. The following experiment em-ploys the same task as Control Study 2, buthere the statements are identified by speaker,and the subjects all know the speakers in thatthey interact with them on a weekly, if not

daily, basis. The purpose of this experiment isto determine whether knowledge of thespeaker's stylistic habits is sufficient to deter-mine which statement occurred in theconversation experiment. We are particularlyinterested in whether knowledge of the speakerallows one to select the target rather than theparaphrase for high interactional contentstatements but not for low interactional con-tent statements. If so, then the surface formeffect in the conversation experiment could notbe interpreted as a result of memory; rather, itwould reflect reconstruction on the basis ofknowledge of the speaker's stylistic habits. Onthe other hand, if subjects who know thespeakers cannot distinguish targets from para-phrases, then the results of the conversationexperiment can be said to truly reflect memoryfor specific statements.

Method

Subjects. Fifteen University of Denverfaculty and graduate students who ratedthemselves as familiar with the speakers andthe setting of the conversation experiment, butwho did not attend that particular discussion,served as subjects.

Procedure and materials. The procedurewas similar to that of Control Study 2. Theonly difference was that each pair of state-ments in this case was identified by thespeaker. In making their judgments of whichstatement occurred, the subjects were asked to


consider phrase structure, lexical items, andoverall "sound" of the statement in terms ofthe particular speaker. It should be noted thatall subjects were highly motivated and nonefelt that their selections were at all random.

Results and Discussion

The bottom half of Table 5 presents theproportion of targets and paraphrases selectedas most likely to have been said. As can bereadily seen, even subjects who knew thespeakers were totally unable to distinguish thetargets from the paraphrases, both in the caseof the high interactional content statements(51% selection of targets and 49% selection ofparaphrases) and in the case of the lowinteractional content statements (48% selec-tion of targets and 52% selection of para-phrases). The analysis of variance confirmedthat there was no main effect of InteractionalContent, or Target—Paraphrase, and there wasno interaction (for all three,/) > .1).

These results clearly support the contentionthat the participants in the discussion actuallyremembered the surface forms of high inter-actional content statements, and did notsimply base their responses on their knowledgeof the stylistic habits of a given speaker. Thepresent experiment shows that knowledge ofthe speaker is not sufficient to distinguish thetargets from the paraphrases. However, know-ledge of the speaker surely plays some role inthe decision process, since subjects who didnot know the speakers selected theparaphrases significantly more often than thetargets (Control Study 2), while subjects whodid know the speakers selected targets andparaphrases equally often (Control Study 3).Thus, for the sentences used in these experi-ments, knowledge of the speaker was onlysufficient to overcome a response bias for theparaphrases and allow responding in accor-dance with chance.

CONCLUSIONS

The central finding of this study is theexcellent retention of surface form for state-

ments that have high interactional contentwithin the context of a real-life communica-tion. This finding contrasts sharply with thecommonly held belief that memory for sen-tences involves only memory for their mean-ings and not for their surface forms (James,1890; Pillsbury & Meader, 1928; Stern, 1931;Sachs, 1967; Brewer, 1974). The studies whichhave investigated this issue present a broadrange of results. Some experiments (Wanner,1974) have shown the influence of taskdemands on memory for surface form, andothers (Anderson, 1974; Keenan, 1975) haveshown the influence of text structure onmemory for surface form. Some experimentshave found that memory for surface form isnegligible or, at best, transient (Sachs, 1967;Wanner, 1974); others find that it persists oververy long retention intervals (Keenan, 1975;Kolers, 1975). These apparently contra-dictory findings indicate that the parameterswhich control the retention of surface forminformation in laboratory experiments haveyet to be worked out. The present experimentdemonstrates the importance of a parameter ofa quite different nature. This parameter in-volves the degree to which a statementconveys information about the speaker's inten-tions, beliefs, and attitudes toward the listener.Our results indicate that the impact of thisparameter on memory is more dramatic thanthat of any of the other parameters which havebeen examined to date.

The finding that interactional contentimproves memory can be explained in at leastfour ways. According to one explanation, highinteractional content sentences are remem-bered better because they contain a largerquantity of information than low interactionalcontent sentences. Both types of sentencesinvolve what Searle (1975) calls direct speechacts, that is, literal statements. However, highinteractional content sentences also involvewhat Searle calls indirect speech acts. Forexample, as a direct speech act, the sentenceit's hot in here refers to the temperature in theroom. However, as an indirect speech act, thesentence may be a request to have a window


opened. Since indirect speech act sentencescontain two types of information and directspeech act sentences only contain one,memory will always be best for indirect speechact sentences. The problem with this ex-planation is that there is no such thing as anindirect speech act sentence, since indirectmeaning is not a property of the sentence as anisolated object but only of the sentence as acomponent of a real interaction. In fact, theresults of the list-learning study show that thisexplanation is not tenable, since memory forsentences that did not involve indirect speechacts was as good as memory for sentences thatdid.

A second explanation avoids the errors ofthe first by claiming that the additionalquantum of information contained by anindirect speech act statement derives from therelation of the statement to an interaction.According to this quantitative account, theexact structure of the interaction is irrelevant.The crucial factor facilitating memory forsurface form is that somehow the interactionleads to the perception of the statement as twoacts in one.

A third explanation focuses on the qualityof the information conveyed and emphasizesthe interpersonal nature of the statements. Thisexplanation states that high interactional con-tent statements are remembered better becausethey can be easily assimilated into a richcognitive organization developed by the lis-tener through previous interactions with thespeaker. Bartlett (1932) often employed thisexplanation in his experiments on memory forschematic faces and memory for texts. Hefound that, when a face or some detail in astory reminded subjects of themselves orsomeone they knew, retention was particularlygood. Explaining memory performance interms of the cognitive organization a listenerhas acquired for a speaker extends thetheoretical notions developed by Craik andTulving (1975). These authors stressed the roleof semantic information in rememberingevents; we add to their notions the role ofinterpersonal knowledge. The more com-

patible an event is with the structure andorganization of semantic memory or with thestructure of one's beliefs and attitudes, thegreater its retention because a more elaboratememory trace can be laid down and becausethe existing cognitive organization can bedrawn upon to facilitate reconstructive pro-cesses involved in retrieval. Note that thesenotions accommodate the finding in ControlStudy 3, that subjects who knew the speakersselected target statements more often thansubjects who did not.

A fourth explanation also focuses on thequality of the information conveyed, but herethe emphasis is on the affective nature of highinteractional content statements. This explana-tion states that high interactional content state-ments are remembered better because theyelicit an emotional or affective reaction in thelistener. While most of our high interactionalcontent statements did not elicit a readilyobservable affective response, such aslaughter, it is possible that more sensitiveobservations, such as GSR or heart-raterecordings, would provide evidence in supportof this explanation.

It would seem that explanations three andfour share a further presupposition which isthat the interpersonally or affectively salientfeatures are often properties of surface formsuch as lexical choice, word order stylistics, orparalinguistic indicators. Memory for surfaceform can, therefore, be seen as memory for thelevel of information richest in these markers.

Future research will need to distinguishbetween these last three explanations, andpossibly others. In addition, future researchwill need to determine the relative importanceof various factors which can be said to controlthe interactional context. These factors includethe degree of previous involvement with thespeaker, the formal identity of speech actsrepresented by particular statements, theorganizational structure of, the interaction[e.g., lecture format (Kintsch & Bates, 1977)versus group discussion], and the amount ofactive participation on the part of the listener.For example, would the results of the present


experiment obtain if subjects, instead ofparticipating in the discussion, simply viewedthe discussion on videotape? Or, would theyobtain if the subject simply read a personalletter which contained high interactional con-tent statements?

As future research proceeds, an importanttask will be to develop a set of a priori criteriathat will permit one to determine whichsentences are high in interactional content.High interactional content statements for thisexperiment were selected intuitively on thebasis of our knowledge of the speaker and thestructure of the interaction. We know thatthese intuitive judgments are reliable, sincethree psycholinguists, who were not present atthe discussion but who knew the main speaker,were able to sort perfectly the high from thelow interactional content statements. Futureresearch, however, should provide an em-pirical and theoretical analysis of the factorsunderlying these judgments.

This study is a simple demonstration ofdifferences in the operating characteristics ofmemory as studied in the laboratory and innatural situations. The results strongly suggestthat the study of pragmatics in psychologicalprocesses must be undertaken if we are everto apply our work on learning and memory toreal-world situations.

REFERENCES

ANDERSON, J. R. Verbatim and prepositional representa-tions of sentences in immediate and long-termmemory. Journal of Verbal Learning and VerbalBehavior, 1974, 13, 149-162.

ANDERSON, J. R., & BOWER, G. H. Human associativememory. Washington, D.C.: Winston, 1973.

BARTLETT, F. C. Remembering: A study in experimentaland social psychology. Cambridge: The UniversityPress, 1932.

BREWER, W. F. The problem of meaning and the inter-relations of the higher mental processes. In W. B.Weimer & D. S. Palermo (Eds.), Cognition and thesymbolic processes. Hillsdale, N.J.: LawrenceErlbaum Associates, 1974.

CLARK, H. H. The language-as-fixed-effect fallacy: Acritique of language statistics in psychologicalresearch. Journal of Verbal Learning and VerbalBehavior, 1973, 12, 335-359. CRAIK, F. I. M., &

TULVING, E. Depth of processing andthe retention of words in episodic memory. Journalof Experimental Psychology: General, 1975, 104,268-294. ERVIN-TRIPP, S. Sociolinguistics. In L.

Berkowitz (Ed.),Advances in experimental social psychology. NewYork: Academic Press, 1969. Vol. 4. JAMES, W.

The principles of psychology. London:Macmillan, 1890. JARVELLA, R. J., & COLLAS,

J. G. Memory for theintentions of sentences. Memory and Cognition,1974, 2, 185-188. KEENAN, J. M. The role of

episodic information in theassessment of semantic memory representations forsentences. Doctoral dissertation, Department ofPsychology, University of Colorado, 1975.

KINTSCH, W. The representation of meaning in memory.Hillsdale, N.J.: Lawrence Erlbaum Associates,1974. KINTSCH, W., & BATES, E. Recognition

memory forstatements from a classroom lecture. Journal ofExperimental Psychology: Human Learning andMemory, 1977, 3, 150-159. KOLERS, P. A.

Specificity of operations in sentencerecognition. Cognitive Psychology, 1975, 7, 289—306. MONTAGUE, R. Pragmatics and intensional

logic. In D.Davidson & G. Harman (Eds.), Semantics ofnatural language. Dordrecht, Holland: D. Reidel,1972. NEISSER, U. Cognition and reality. San

Francisco:Freeman, 1976. PILLSBURY, W. B., & MEADER, C.

L. The psychology oflanguage. New York: Appleton, 1928. SACHS, J.

Recognition memory for syntactic andsemantic aspects of connected discourse.Perception andPsychophysics, 1967, 2, 437-442.

SEARLE, J. Indirect speech acts. In P. Cole & J. Morgan(Eds.), Syntax and semantics, Volume 3: Speechacts. New York: Academic Press, 1975.

STALNAKER, R. C. Pragmatics. In D. Davidson & G.Harman (Eds.), Semantics of natural language.Dordrecht, Holland: D. Reidel, 1972. STERN, G.

Meaning and change of meaning. Goteborg:Elanders Boktrycken Aktiebolag, 1931. (Reprinted

at Bloomington: Indiana University Press, 1964.)WANNER, E. On remembering, forgetting and under-standing sentences. The Hague: Mouton, 1974.

(Received December 27, 1976)

Documents

Pragmatics in memory: A study of natural conversation