controlling the intelligibility of referring expressions ...€¦ · ellen gurman bard human...

Journal of Memory and Language42, 1–22 (2000)doi:10.1006/jmla.1999.2667, available online at http://www.idealibrary.com on

Controlling the Intelligibility of Referring Expressions in Dialogue

Ellen Gurman Bard

Human Communication Research Centre and Department of Linguistics,University of Edinburgh, Edinburgh, United Kingdom

Anne H. Anderson

Human Communication Research Centre and Department of Psychology,University of Glasgow, Glasgow, United Kingdom

Catherine Sotillo and Matthew Aylett

Human Communication Research Centre and Department of Linguistics,University of Edinburgh, Edinburgh, United Kingdom

and

Gwyneth Doherty-Sneddon and Alison Newlands

Human Communication Research Centre and Department of Psychology,University of Glasgow, Glasgow, United Kingdom

If speakers articulate clearly enough to meet the perceptual needs of their listeners, clarity shoulddepend on what listeners know about (listener-Given) rather than on what speakers know about(speaker-Given). For words excerpted from spontaneous speech, however, intelligibility to naiveadult listeners showed only effects of the speaker’s knowledge. Words introducing labeled maplandmarks to two successive listeners were less clear on repetition even though the second listenerhad not heard the original mention (Experiment 1). Repeated mentions became less clear even afterthe listener reported inability to see the landmark (Experiment 2). Speakers were affected by whatthey had heard listeners mention: Intelligibility fell equally in coreferential repetitions across andwithin speakers (Experiment 3), whether or not the repeater could see the referent (Experiment 4).The results are explained via fast priming processes dependent on the speaker’s knowledge and slow,optional processes drawing inferences about the listener’s.© 2000 Academic Press

Key Words: intelligibility; referring expression; dialogue; spontaneous speech; hearer-Given;speaker-Given.

enmebi

m.heticghts as

tenin-

theand74;it,tionake

edaryer-ew

andt o

essofdin

No two spoken instances of a word are idtical, even if they are produced by the saspeaker in the same conversation. Yet varia

This work was supported by the ESRC (UK)-fundHuman Communication Research Centre. A preliminreport of this work was presented by Bard, Sotillo, Andson, Doherty-Sneddon, and Newlands (1995). Alison Nlands is now at Strathclyde University, Glasgow, U.K.Gwyneth Doherty-Sneddon is now at the DepartmenPsychology, University of Stirling, Stirling, U.K.

Correspondence concerning this article should be addrto Dr. E. G. Bard, Department of Linguistics, UniversityEdinburgh, Adam Ferguson Building, George Square, Eburgh EH8 9LL, U.K. E-mail: [email protected].

1

-

l-

ity in pronunciation is by no means randoAmong the many controlling factors is tavailability of information beyond the acoussubstance of the word token itself which mihelp a listener to decipher the speech soundthe correct word. For this reason, it is ofthought that the psychological processesvolved in speech production must includespeaker’s model of what the listener knowsperceives (Bolinger, 1963, 1981; Chafe, 19Lindblom, 1990). As Lindblom (p. 405) put“the speaker estimates the running contributhat signal-complementary processes will m

-

f

ed

-

0749-596X/00 $35.00Copyright © 2000 by Academic Press

All rights of reproduction in any form reserved.

inuomthfa

blethetions-

otheheich

y

st-of

ad-on,ts-

wilnkso

enl-t th

thwts o

b

eff n-s ncc 3)si un-db orer hat

n-

theot

g ivec ligi-b ralm &T n-n isk fort ts,w ordi inr bef 3)r ersc ms

thefi er.T ectst ou-s ofs e-t -t is-t stb enp ce,1 ay,w thet inP tu-a onf tlym ’sc r-m thati e ar-t

enm rea thanw . InT thee ce,G ved.S tra-l fert ce

2 BARD ET AL.

during the course of an utterance” and contously adjusts articulatory exactness to accmodate listeners’ residual dependence onspeech signal. Lindblom suggested that in soas other sources of information are availaspeakers “hypoarticulate,” but whereverspeech sound is the sole basis for recognithey “hyperarticulate.” Since we will be dicussing changes in pronunciation in termstheir effects on a listener’s ability to recoverword from its spoken form, rather than tphonetic or phonological processes whachieve this result, we will use the termintelli-gibility for the variable that is controlled bhypo- and hyperarticulation.

This paper will test the proposal that adjuments to intelligibility are based on a modelthe listener’s, or more precisely, of thedressee’s knowledge. Within this introductiwe will first exemplify some of the adjustmento word intelligibility which are thought to represent accommodation to the listener. Weshow, however, that the existing evidence lithem only to the speaker’s own knowledgethe linguistic context, not to the listener’s. Thwe will explain why modeling listeners’ knowedge is a very demanding task and presencurrent views about how speakers managemodeling problem. Finally, we will explain hoour experiments test for the effects of aspeclisteners’ knowledge that speakers shouldreadily able to notice.

CHANGES IN WORD INTELLIGIBILITYWITH CONTEXT

The best-known examples of contextualects on intelligibility depend on the relatiohip between a lexical item and the senteontext in which it is uttered. Lieberman (196howed that tokens of words likenineproduced

n (1) below were longer, had higher peak famental frequencies, and were moreintelligi-le, in the operational sense of being mecognizable when excerpted from context, tokens produced in contexts like (2).

1. The word you are about to hear isnine.2. A stitch in time savesnine.

Intelligibility seemed to respond to redu

--er,

,

f

l

f

ee

fe

-

e

n

dancy: normal adults could easily supplyword nine given the remainder of (2), but n

iven the remainder of (1). Reliable negatorrelations between redundancy and intelility have also been reported in more natuaterials (Bard & Anderson, 1983; Fisherokura, 1995; Fowler & Housum, 1987; Huicutt, 1985; Samuel & Troicki, 1998). Thind of relationship also seems to accounthe careful articulation of words read in lishere there are no contextual clues to w

dentity, and the much less careful deliveryunning speech. While citation forms tend toully recognizable, Pollack and Pickett (196eported that, on average, only 50% of listenould identify individual words isolated frourreptitiously recorded conversation.A second kind of evidence, clearly akin to

rst, will provide the focus of the current paphis work shows that discourse status aff

he way a word is articulated. Fowler and Hum (1987) described this effect as partpeakers’ ability to signal the distinction bween “New” and “Old” or “Given” informaion. In the simple form of the Given/New dinction, New information is information jueing introduced. Given information has berimed, or readied for use in discourse (Prin981), by being made prominent in some whether by obvious physical presence at

ime of speaking (“situationally evoked”rince’s account), by previous mention (“texlly evoked”), or by implication or associati

rom information which has been explicientioned (“inferrable”). Fowler and Housum

laim is that words which mention Given infoation can be interpreted in the context of

nformation and, as a consequence, can biculated less clearly.

Though Given information often has beentioned before, effects on intelligibility associated with Givenness per se, ratherith the mechanics of repeated articulationable 1, the findings which characterizeffect are summarized in terms of refereniven status, and how that status is achieince Given status is a characteristic of ex

inguistic information, a word token must reo such information to induce the effect. Hen

nor,anhelly

han,

refelythngm,orrdao

cem)ns&re

e-

ocths-

vi-ki,ene o1;

x,hethin

ne

eneli-ionr or

eanesbo-rrention

of

n-lli-mtesAson

efor

pre-orr’stheionslis-ax-as a

fromess.ak-ingad-of

betheons

)

3THE INTELLIGIBILITY OF REFERRING EXPRESSIONS

words merely read repeatedly in lists showreliable order effects on intelligibility (Fowle1988), though two tokens of a word in a meingful text do differ, with the second token, tone that refers to a Given entity, generashorter and less intelligible when isolated tthe first (Fowler, 1988; Fowler & Housum1987; Hawkins & Warren, 1994). Repeatederential use itself does not suffice. A relativdegraded word token is produced whenword refers to an entity Given by virtue of beimentioned previously (Fowler & Housu1987), but not when the second use of the wrefers to a New item of the same sort (BaLowe, & Altmann, 1989). Nor is repetitionnecessary condition. Introductory mentionsitems Given by virtue of physical presen(situationally Given in Prince’s 1981 systeare less intelligible than introductory mentiowhich are truly New to the context (BardAnderson, 1994). Thus, Givenness without pvious mention degrades intelligibility but rpeated use without Givenness does not.

These effects appear to parallel the well-dumented influence of discourse status onlexical and syntactic form of referring expresions (Ariel, 1990; Chafe, 1974; Clark & Haland, 1977; Gundel, Hedberg, & Zachars1993; Prince, 1981, 1992). The two phenomalso appear to be sensitive to the structurextended discourse (Fowler & Levy, 199Fowler, Levy, & Brown, 1997; Vonk, Hustin& Simmons, 1992). It is not surprising that tsame explanation has been offered forchanges in intelligibility and the changesreferring expression: the needs of the liste

TAB

Summary of Conditions under Which Articulato

Wordrefers

ReferentGiven

How Given status ach

Has said Can see

2 n.a. 1 21 1 1 21 2 1 21 1 2 1

Note.1, condition holds in critical cases;2, condition

-

-

e

d,

f

-

-e

af

e

r.

Prince (1992), for example, referred to Givstatus as “Hearer-old.” Ariel described a dcate adjustment of forms of referring expressto the antecedent’s accessibility to the readelistener. Gundel et al. described a Gricmaxim of quantity (Grice, 1975), which guidspeakers and writers in choosing a form elarate enough to meet the addressee’s cuneeds. The effects of context on articulatmight be ascribed to an analogous maximarticulatory quantity.

In all these cases—the intelligibility–redudancy trade-off, the Givenness effect on integibility, and the effects of accessibility on forof referring expression—the theory implicaspeakers’ models of listeners’ knowledge.Table 1 indicates, however, the experimentsarticulatory clarity manipulated only thspeaker’s knowledge: The effect was foundmentions of entities which the speaker hadviously mentioned, had heard mentioned,could see at the time. Neither the listeneknowledge nor the speaker’s observation oflistener was examined. Instead two assumptwere made: first, that even if speakers andteners have different knowledge, Gricean mims are accurately observed and, second,consequence, that markers of Given statusa speaker will be reflexes of listener-Givenn

This pair of assumptions attributes to speers and writers a remarkable feat of maintaincontinuous and accurate models of theirdressees’ knowledge. Yet this interpretationthe “Gricean burden” is far too demanding totaken for granted. More strictly construed,Gricean maxims prescribe only assumpti

1

larity Degrades (“YES”) or Fails to Degrade (“NO”)

ed: Speaker

DegradationHas heard

1 NO: Fowler (1988)1 YES: Fowler & Housum (1987)1 NO: Bard, Lowe, & Altmann (19892 YES: Bard & Anderson (1994)

s not hold in critical cases; n.a., does not apply.

LE

ry C

iev

doe

bechityfulmsuatha

ndnce.

er-Asar

kertuachacheto

thahededuld

nnsanings, ts a

isrs

it in,siongt tofauusosarerorlis

ge

the

taketer-r-ofnd7),

onren-9;toAl-

ow-r,e toressduc-res-deltingoretheestsandouldiveWee-

ion, a, andses, orthe

rstcon-e foraree ifsite.d inner.be-

tin-Ford re-

4 BARD ET AL.

that listeners should have about speakers’havior if they are to interpret material whidoes not initially seem to be suitable in quantquality, or the like. Though they provide usestarting points for interpretation, the maxicannot guarantee that the speaker can actread the listener’s mind. Nor can they easeconsiderable difficulty that a speaker of normabilities would have in keeping track of amaking proper inferences from all the eviderevealing the listener’s pertinent knowledge

DEFAULTING UNDER THEGRICEAN BURDEN

Maintaining an accurate model of the intlocutor is a potentially insuperable problem.Clark and Marshall (1981; see also Keys1997; Stalnaker, 1978) explained, speaneed to maintain an internal account of muknowledge, the collection of information whiboth speaker and listener know and which eknows they both know. For both to know tsame things is difficult enough, but for eachknow that both know the same things andthey understand the same thing by what tknow, and so forth, is a potentially open-entask, for every confirmation of agreement woneed to be checked to be sure it maskeddeeper disagreement. To have conversatioreal time, Clark and Marshall suggested,speaker must default to an optimistic workassumption about shared knowledge, that ithe assumption that their own knowledge igood model for the listener’s. The questionwhen this default takes place. Three answethis question have been offered.

The strongest position seems to be implicthe notion of tailoring intelligibility on-linethough we have been unable to find discusof this issue in the literature on controlliarticulation. To achieve genuine adjustmenlisteners’ needs, speakers must seldom deto their own view, and only after strenuoefforts to model listeners as veridically as psible. To support continuous adjustment ofticulation, speakers should observe listencontinuously for signs of misunderstandingdisagreement. Wherever speaker’s andtener’s knowledge differ, listener’s knowled

-

,

llyel

,sl

h

ty

oiny

o

to

n

lt

--s

-

should take precedence. We might call thisno default hypothesis.

The second alternative is that speakersspecial note of middle- or long-term characistics of their listeners which affect likely ovelap with their own knowledge. Various kinds“copresence” in social or regional backgrou(Fussell & Krauss, 1992; Isaacs & Clark, 198physical location during the interacti(Schober, 1993), or recent experiences (Bnan & Clark, 1996; Schober & Clark, 198Wilkes-Gibbs & Clark, 1992) are taken inaccount in adjusting the default assumption.though this work is usually interpreted as shing that the “initial design” (Horton & Keysa1996) of conversational speech is sensitivthe listener’s needs, it does not directly addon-line processes throughout speech protion. The original purpose of assessing copence (Clark & Marshall, 1981) was to provispeakers with reasonable grounds for defauto their own knowledge as a proxy for the melusive facts about the listener’s. Applied toproduction of speech, this approach suggthat speakers should attend to evidence foragainst copresence and that defaulting shgo on for some undefined time after positevidence, without continuous recalculation.might call this thecopresence default hypothsis.

The third proposal makes a modular divisbetween the initial formulation of utterancesprocess based on the speaker’s knowledgethe monitoring and revision of output, procesbased on a model of the listener’s knowledgemore precisely, of common ground. Calledmonitoring and adjustment hypothesisby Hor-ton and Keysar (1996), this model defaults fiand pays later. It offers the advantages of eomy and speed. Faultless utterances, thoswhich speaker’s and listener’s knowledgealike, are produced faster than they would baccurate listener modeling were a prerequiPoorly designed utterances can be reviseresponse to explicit requests from the listeThose explicit requests achieve realignmenttween interlocutors without the need for conuous modeling of the listener’s knowledge.this reason, postfeedback utterances shoul

ichowge

nyfirswlT

n-rma

iap

ondeskarer, oan-rsoeto

reerier’s otheeiar

yertneon

atneonoinsel 1

inis-nd

theverboha

In-emus,asak-nd

lesse

el-es-

r-tedpusak-In-onr’s.De-ere

aryap-l toedatetch-erenottenis-, 16

theirschin

halfthervedpsap,er.edlarnd


flect any aspects of listener knowledge whthe feedback has added to the speaker’s knedge. Otherwise, the listener’s knowledshould be irrelevant to production.

To decide which of these hypotheses, if aaccounts for speakers’ behavior, we mustexamine situations where the listener’s knoedge is not confounded with the speaker’s.do this, we will draw words from running spotaneous speech produced while pairs of noadults communicated a route best definedterms of labeled landmarks on schematic mof imaginary locations. The map design ctrolled speaker’s and listener’s knowledge inpendently. In four experiments, we will awhether the Givenness effects in Table 1sensitive to what listeners have heard (Expment 1), what they can see (Experiment 2)what they have mentioned (Experiments 34). By examining word intelligibility and duration, we will try to discover whether speakeadjust to what the listener plainly does and dnot know. The pattern of results will help usreassess the hypotheses outlined above.

EXPERIMENT 1

We begin by examining how speakersspond to disparities between their own expence of what has been said and the listenWe use introductory mentions of the namelandmarks produced by Instruction Givers,participants who were relating the route to thpartners. Because some landmarks appeonly on one participant’s map, neither placould be sure that they saw what their parsaw. We compare those introductory mentiwith introductory mentions produced whenspeaker who had already led one parthrough a particular map began to lead a secpartner through the same map. From the pof view of the Instruction Giver, the discourstatus of any landmark mentioned in Triachanged between trials. On its introductionthe first trial, it was genuinely New to the dcourse. On the initial mention in the secotrial, it represented Given information. Fortwo successive Instruction Followers, howethe status of the landmark was the same atintroductory mentions. Because neither

l-

,t-o

lns

--

ei-rd

s

--s.f

red

rs

rdt

,thd

worked with the map before or heard thestruction Giver mention the landmark, the itwas New for both Instruction Followers. Ththe introductory mention on Trial 2 wspeaker-Given and listener-New. If the speer’s knowledge controls articulation, secotrial speaker-Given introductions should beintelligible than first. If the listener’s knowledgcontrols articulation, however, no loss of intligibility should be observed between succsive listener-New introductions.

Method

Corpus.All materials in this and later expeiments were drawn from the 128 unscripconversations of the HCRC Map Task Cor(Anderson et al., 1991), in which pairs of speers, designated the Instruction Giver or thestruction Follower, collaborated to reproduceone player’s map a route printed on the otheNeither speaker could see the other’s map.spite their designations, no restrictions wplaced on what either speaker might say.

Because the maps portrayed imaginplaces, all information relevant to the taskpeared on the maps. Of the landmarks criticathe reproduction of the route, only half matchexactly between partners’ maps, with alternlandmarks encountered on any route mismaing in name, number, or location. Speakers wwarned in advance that their maps wouldmatch exactly but they were not told how ofto expect mismatches or what sorts of mmatches there were. Over the whole corpusdifferent basic maps were used.

The participants, all undergraduates atUniversity of Glasgow, were recruited as paof friends, with two such pairs producing easet of eight dialogues. Half the dialogueseach set were between pairs of friends andbetween individuals who had not met beforerecording session began. Every speaker seas Instruction Follower on two different maand as Instruction Giver twice for a single meach time with a different Instruction FollowAll “quads” of participants followed a balancdesign which put all first trials with a particumap in the first four of the eight dialogues, aall second trials in the second four.

in-anrs

ced

tlyht

wacean

s,me

rkn e top

mere

ATre

nsrucbul-s

arker-uld

hisreeretheehen inal

tededPS

etsandandintsund-re

-ordHz

uesgth.of

inalsam-laiddios.ttomsbybythe

ard,7)

208 inwass,

ets,reehanedign.ms

fnts.

thist theadear-nes

reriteere

91S anm d vl /i

6 BARD ET AL.

Eight groups of four speakers participatedthe face-screenedcondition with a flimsy barrier preventing partners from seeing oneother. Within this group, familiarity of speakeusing a particular map was counterbalanAnother eight quads worked in theface-visiblecondition with no barrier, but otherwise exacreplicated the design used for the first eigThus, over the whole design, each mapused the same number of times with fascreened and faces visible and by familiarunfamiliar pairs.

After participating in a series of dialogueeach speaker read a list of landmark nacovering the maps just used. Thesecitationforms provide control tokens of landma

ames in a condition which does not give risragmatic effects (Fowler, 1988).1

All materials were recorded on the saDAT (Sony DTC1000ES) using one ShuSM10A close-talking microphone and one Dchannel per participant. All recordings wemade under the same studio conditions.

Stimuli.The stimuli were introductory tokeof the landmark names uttered by the Insttion Givers in two trials using the same mapdiffering in the identity of the Instruction Folower. All were taken from full literal mentionof the landmark labels. Thus, if the landmwas labeled “Site of Forest Fire,” only a refring expression containing all four words coprovide tokens for this study.

As in all the experiments described in tpaper, disfluent items and those which suffefrom cross-talk from the other speaker wexcluded. Selection of materials followedmultiple criteria for the experiment until thmaximal design-conformant size was reacfor the cell with the sparsest representatiothe corpus with the minimal amount of lexicoverlap within the cell.

For all experiments, words were excerpfrom digitally recorded materials and digitizat a sampling rate of 16 kHz for use with ES

1 For other details of design see Anderson et al. (19ample maps and all transcriptions can be viewedaterials used in the present experiments can be playe

inks given on http://www.ltg.ed.ac.uk/;amyi/maptaskntelligibility_materials.html.

-

.

.ssd

s

-t

d

d

WAVES software. Word onsets and offswere determined by examining spectrogramtime-amplitude waveform representationslistening to the results of excerptions. Cut powere set at 0-crossings and the segment boary conventions of Laver et al. (1989) weused.

To avoid ceiling effects in intelligibility experiments, the speech file for each original wwas multiplied, sample by sample, by a 16 kfile of random noise (where all sample valwere in the range 0.5 to 1.5) of the same lenIn each resulting stimulus, the amplitudeeach sample was related to that of the origspeech and each had the same sign as thepled data value it replaced. The noise-overspeech files were downloaded to digital autape with an interstimulus interval (ISI) of 8

Design.Forty-eight item triples (Trial 1 firsmention to first listener, Trial 2 first mentionsecond listener, citation form) were used. Itewere restricted to those mentioned bothspeakers working with faces visible andthose with faces screened, to conform todesign of another experiment (Anderson, BSotillo, Newlands, & Doherty-Sneddon, 199run in the same sessions.

Items were taken from the speech ofspeakers in the face-visible condition and 1the face-screened condition. Because theresome lexical duplication among the 48 itemthe materials were divided between two seach distributed by Latin square among thgroups so that no group contained more tone token of a lexical type and all containequal representation of each cell of the desThe stimuli were then randomized with itebelonging to another experiment.

Participants and procedure.Six groups onine listeners were paid to serve as participaIn this and all other experiments reported inpaper, participants were undergraduates aUniversity of Glasgow, where the corpus hbeen recorded. No listener had any known hing loss. Stimuli were played over headphoin a sound-proofed room.

Participants were told that all stimuli wereal spoken words and were instructed to wdown the words they heard. Responses w

).dia

iceefo

helea-

ofn

el-n-onmein-ra

i-meonorein-

r-x-by(byd basnd

ateeaeriadsse

er’sssal,

th2

onsrial

FF

. thee g-n n-t buts ns( ce-v thef .12),b tlyl rial2 eep

byed

lins

icalers.odellthe

fortheentthat

t 1):M r-e nso rs,G


typed directly into the computer. Four practitems preceded test materials. Participants wable to ask questions about the procedure betest materials were presented.

Results

One item was lost by experimenter error. Tanalysis was therefore based on 47 item tripWe report analyses of intelligibility and durtion for the data summarized in Table 2.

Intelligibility. We report direct measuresintelligibility, that is, the proportion of trials owhich a word is correctly identified, andintel-ligibility loss, the difference between the intligibility of a word token excerpted from runing speech and the intelligibility of the citatiform of the same word uttered by the saspeaker. Statistics on raw intelligibility andtelligibility loss measures are always compable. We will regularly report only the intelligbility loss analyses, where the critical outcowill more often depend on a main effect thana comparison within an interaction. Raw scanalyses will be cited only where baselinetelligibility is at issue.

Analyses of variance (ANOVAs) were peformed using participants in intelligibility eperiments (by participants), word triples (items), and participants in the map taskspeakers) as cases. Both by-participants anitems have halves (1, 2) of the experimentsgrouping variable and trial (first, second) aface condition (screened, visible) as repemeasures. Because of the distribution of spers’ contributions across halves of the expment, the by-speakers ANOVA simply hspeakers nested in face condition and crowith trial.

An egocentric adjustment to the speakexperience should show more intelligibility lovis-a-vis citation forms on the second triwhen landmarks are speaker-Given, than onfirst, where they are speaker-New. Tableshows that instead of a uniform effect of trialintelligibility loss (all F-values, 1), there waan interaction between face condition and tF1(1,52) 5 16.44,MSE5 0.0413,p , .0002;

2(1,45) 5 12.35,MSE 5 0.0492,p , .001;(1,36) 5 6.21, MSE 5 0.0286, p ,
speakers
rere

s.

-

y-a

dk--

d

e

,

0174. Face-screened dialogues showedgocentric pattern, with intelligibility loss siificantly greater in Trial 2 introductory me

ions (.18), when the term is listener-Newpeaker-Given, than in Trial 1 introductio.07), when it is New to both participants. Faisible Trial 2 items were as unclear asace-screened tokens of the same words (ut now Trial 1 introductions were significan

ess clear (.23) than either the face-visible Tor the face-hidden Trial 1 tokens. (All thr

airwise comparisons are atp , .05 or better inNewman–Keuls tests by participants anditems. Only the trial effect within face-screendialogues is significant by speakers.)

Duration. Millisecond durations of stimuwere normalized in order to allow comparisobetween cells not containing the same lexitems or not produced by the same speakNormalization was carried out in a methwhich owes much to the findings of Campband Isard (1991). These authors describeddifferent distributions of segment lengthsdifferent phonetic segments as belonging tosame mathematical type but having differmeans and variances. They hypothesized

TABLE 2

Effects of What the Listener Has Heard (Experimenean Intelligibility, k-Normalized Duration (and Diffence from Citation-form Control) for Introductory Mentiof Landmark Names on Trials with Different Listenerouped by Face Condition

Face condition

Form

Trial

Citation1 2

Intelligibility

Screened 0.746(0.072)

0.636(0.182)

0.818

Visible 0.578(0.230)

0.693(0.115)

0.808

Duration

Screened 0.703(0.606)

0.659(0.650)

1.309

Visible 0.646(0.655)

0.588(0.713)

1.301

a-gthuafecanis as ntheplistrlogaminer oe

ranlogcho it

nionofg-sss

f s foi sw ink

,rac

t oni bys

inein

dycona-a

er-ledlisro

torde

amethisn-ere

i-eacht al.ityrsallytheialtoseey or

cor-er-at

ners’vertsume

inritytheper-thee-isi-red

that

ellofseeretheoryialsualferondak-

ers’o-undueon-

8 BARD ET AL.

within linguistic domains like words or syllbles all the items might be shortened or lenened by the same proportion of their individvariance. Though this system is not a perpredictor of word duration, it is better thsimpler normalizations. The current methodversion of the system, which, in pretests, waless accurate than the original in modelingdurations of the present materials. Two simfying assumptions were made: that each dibution of segment length was Gaussian forduration and that all the segments had the smean and standard deviation. Using a machreadable dictionary to determine the numbesegments, number of syllables, and presencstressed syllables in the citation phonetic tscription of a word, we generated potentialtotal duration for each stimulus word with easegment taking on the value corresponding tmean log duration plus some commonz-value,which we call k. The k-transformed duratioeffectively provides a measure of compressexpansion relative to the expected lengthword with the given number of phonetic sements, number of syllables, and lexical stre

Cell means fork appear in Table 2. ANOVAollowed the by-items and by-speakers designntelligibility loss. Although running speech formere predictably shorter than citation forms-units,F2(1,46)5 355.04,MSE5 0.2327,p ,

.0001; Fspeakers(1,36) 5 224.14,MSE 5 0.1389p , .0001, there were no main effects or inteions of interest. The tendency for shorter secntroductions was not significant by items orpeakers (F2 , 1; Fspeakers, 1).

Discussion

Experiment 1 was intended to determwhether speakers adjust clarity of articulationthe light of what their listeners have alreaheard. Results differed across the two faceditions. With no visual channel for communiction, speakers reduced clarity on introducingitem which was listener-New but speakGiven. For this group, behavior was controlby the speaker’s experience and not thetener’s. With a visual channel, speakers pduced speaker-Given second-trial introducmentions which were comparable to the

-lt

o

-i-

e-fof-

s

/a

.

r

-d

-

n

--y-

graded speaker-Given mentions of the swords in face-screened dialogues. Forgroup, however, first-trial introductory metions, New for both speaker and listener, weven less intelligible.

Why were initial introductions so unintelligble when speakers and listeners could seeother? Using the HCRC Corpus, Anderson e(1997) have shown that lowered intelligibilin initial mentions occurs when interlocutocan look at one another but are not actudoing so. When the listener is looking atspeaker (about one word in five or six), initmentions rise in intelligibility. Such tendoccur at points where players who cannotone another would ask for assurance that thetheir listeners have understood instructionsrectly (Doherty-Sneddon et al., 1997). Andson et al. suggested that intelligibility risessuch points because speakers interpret listegaze much as they would interpret the ofeedback. Otherwise, speakers seem to asthat they and their listeners are fully alignedtheir views of the dialogue and so reduce claaccordingly. There are no indications thatspeakers’ optimism is warranted: map taskformance—as measured by deviations ofFollower’s drawn route from the Giver’s prprinted version—was no better with faces vble than with faces screened. All that appeato differ was the speakers’ expectationsthey were being understood.

If Anderson et al. (1997) are correct, this cof the design is not in conflict with the restthe results. Although speakers who cannottheir interlocutors may initially be the mocautious about introducing New entities,next introduction is made with less articulateffort. The actual level of effort in second trintroductions seems to be independent of vicues from the listener, for it does not difsignificantly between face conditions. Sectrial introductions, then, depend on the speer’s experience, not the listener’s.

It is possible that speakers ignored listenknowledge, or more correctly, listeners’ ignrance, because it was covert, the backgrorather than the text of a fairly complex dialogtask. Both Anderson et al. (1997) and the m

necaee

e tothesigon-cend, bedTh

isre

le,

-ac

heer

tededrk,ap

theno

tenm.

toin

rea

anatety,er. Othee

wahe

are

t-theondpro-

dia-d-ndorm

ofal).f theen-

lic-of

on-

by

i ol-l ithn

.”

urrti-

ah.

got

ntm 23on-

rip-thinniald 28ia-ion,, andfre-pos-


itor and adjust hypothesis suggest that listecan affect speakers by providing overt inditions of what they know. In the next threxperiments, we examined such cases.

EXPERIMENT 2

This experiment examines the importancthe speaker of the listener’s ability to seeobject under discussion. The Corpus deforced participants to be continuously ccerned with this kind of perceptual copresenNot only were routes designed around lamarks rather than coordinates or distanceshalf the landmarks critical to the route differin some way between participants’ maps.frequency of accurate feedback about mmatching landmarks correlates with accurateproduction of the route (Anderson & Boy1994).

Accordingly, it is difficult to see how a cooperative speaker could manage without ancount of items which are visually Given for tlistener. To test for effects of this sort of listenmodeling, we contrast two kinds of repeamentions. In both, a single speaker repeatmentioned the name of an unshared landmalandmark appearing only on that speaker’s mIn one case, just after that first mention,speaker learned that the landmark wasshared between the maps, because the lispromptly and explicitly denied having the iteIn the other case, the listener either failedsupply this feedback or supplied misleadingformation, and the original speaker had noson to believe that there was a problem.

A speaker who takes note of feedbackadjusts to listener knowledge should mitigthe effect of repeated mention on intelligibilirestricting intelligibility loss when the listenhas declared that the referent is inaccessiblethe other hand, should speakers depend onown knowledge, intelligibility loss ought to bthe same in both cases, for in both, the itemalways visually and verbally Given for tspeaker.

Method

Materials and design.All stimuli were wordtokens uttered as part of the names of unsh

rs-

n

.-ut

e--

-

lya.

ter

--

d

nir

s

d

landmarks. All tokens occurred within full lieral mentions of the landmark as labeled onmap where it appeared. The first and secrunning speech tokens of each name wereduced by a single speaker within a singlelogue with no intervening mention of the lanmark in any form. For each of the 120 first asecond mentions, there was also a citation fread by the same speaker.

Of the 120 triplets, 60 belonged to eachtwo feedback conditions (denial, no deniFeedback was determined by the behavior oother participant between first and second mtions. In the denial condition, listeners expitly—and correctly—denied the existencethe feature on their map; in the no-denial cdition, there was no such feedback.

Example (3) illustrates a self-repetitionthe Instruction Giver (G) ofpelicanswith anntervening denial from the Instruction Fower (F). An example of a self-repetition wo denial is presented in Example (4).3. G: Stop, um, beside the “s” of “Saxon

F: Okay.G: And, have you gotpelicans?F: No.G: No. Um, go down about three to fo

centimeters from the Saxon barn vecally downwards.

F: So I’m above the rope bridge?G: Just a bit above the rope bridge, yeF: Okay.G: And then go underneath where I’ve

pelicanstowards the left-hand side.4. F: What aboutthe banana tree?Does it/

G: Oh.F: come beforethe banana tree?

Materials were selected from 21 differespeakers in face-screened dialogues and froin face-visible dialogues. Only 12 speakers ctributed triples to both feedback conditions. Tlets of words were balanced for feedback wieach face condition: 32 denial and 32 no-detriples were selected from face-screened andenial and 28 no-denial from face-visible dlogues. Of the 60 triplets per feedback conditeight were matched verbatim across feedbackthe rest were matched for word length andquency as far as possible. As many items as

ial)hin

rel),kete

twotharenoordde

sin

eth

ity

nintioialan

sseusb

s nlts

el-nion

FF ,b di-c list enw ort e-n nsd 6w mei In-s itioo ns

w eres ossf hs(. nsw tionsw ar-t tef ored tici-p

inec el-l on( calt mee back( byp0 nceb ,

ionby

dis-to

2):M r-e s asa edL

10 BARD ET AL.

sible (47 of the no-denial group, 48 of the denwere selected from first trial dialogues. Witfeedback conditions, some landmark namescurred (10 within denial and 12 within no-deniabut never within the speech of a single speaBecause of the duplication of word types creain these ways, materials were divided intohalves, so that no lexical item appeared moreonce in either and each half was divided into thgroups to provide a Latin square in whichsubgroup heard more than one token of any wPreparation of materials was otherwise asscribed for Experiment 1.

Participants and procedure.The participantwere 54 Glasgow University students, nineeach of the six groups. Other details of the modology were as described for Experiment 1.

Results

Table 3 contains cell means for intelligibiland normalized durations.

Intelligibility. ANOVAs were performed ointelligibility loss. Participants were nestedhalves of the design and crossed with men(first, second) and feedback (no-denial, denItems were nested in halves and in feedbackcrossed with mention. Speakers were crowith mention but nested in feedback, becathe design was not fully crossed for feedbackspeakers. Because the speaker design wareflected fully in the ANOVA, speaker resuare included only for interest.

Instead of a main effect of mention on intligibility loss (all F-values, 1), there was ainteraction between feedback and mentF1(1,52) 5 21.96,MSE5 0.0156,p , .0001;

2(1,116)5 5.30,MSE5 0.0718,p 5 .0231;speaker(1,54)5 5.77,MSE5 0.0319,p 5 .0198ut not the interaction which would have inated a cooperative consideration for theener’s predicament. A cooperative adjustmould have given less intelligibility loss f

okens following denial than following no-dial, but intelligibility loss for second mentioid not differ significantly with feedback (.1hen listeners apparently could see the na

tem, .21 when they apparently could not).tead, there was an effect of feedback condn the introductory mentions. First mentio

-

r.d

ne

.-

-

n).ddeyot

,

-t

d

n

hich failed to elicit accurate feedback wignificantly more degraded (no-denial lrom citation5 .24) than first mentions whicucceeded (denial loss from citation5 .13)Newman–Keuls test by participants atp ,01). In fact the no-denial introductory mentioere more degraded than the second menhich followed them (Newman–Keuls by p

icipants atp , .05). In contrast, where accuraeedback occurred, second tokens were megraded than first (Newman–Keuls by parants atp , .01).This effect was not an artifact of basel

itation measurements. ANOVAs on raw intigibility scores, using three levels of mentifirst, second, citation) but otherwise identio intelligibility loss designs, revealed the saffects as the derived measure. The Feed2) 3 Token (3) interaction was significantarticipants, F1(2,104) 5 11.76, MSE 5.0154,p , .0001, and approached significay items, F2(2,232) 5 2.80, MSE 5 0.0720

p 5 .063. Neither second mentions nor citatforms differed across feedback conditionsNewman–Keuls tests. The denial materialsplayed the usual reduction in clarity from first

TABLE 3

Effects of What the Listener Can See (Experimentean Intelligibility, k-Normalized Duration (and Diffence from Citation-form Control) for Repeated MentionFunction of Feedback on Ability to Find the Mention

andmark

Feedback

Form

Running speech mention

Citation1 2

Intelligibility

Denial 0.631(0.128)

0.554(0.205)

0.759

No-denial 0.491(0.240)

0.572(0.159)

0.731

Duration

Denial 0.600(0.519)

0.461(0.658)

1.119

No-denial 0.675(0.617)

0.601(0.691)

1.292

ei thea n-i l-l rdet en(

nm ake l-lf tok eco.

uce-an

anirs

ineandil-re-ingivit.

ecc-te

ot,Weormac-thisun

ichentha

recvessre

g t

ess

kenessas”/e-ere

dsurs as

nalarks. Inoor

ts insectedid-

ide,thatreven

anters’herng

aremseen.lis-oned

ofakerond

akerre-arenowetionus.oter-d a


second mention (p , .01). The source of thnteraction was what it appeared to be innalyses of intelligibility loss, the unusual u

ntelligibility of those initial mentions not foowed by accurate feedback. They were hao identify than any other running speech tokp , .01).

Duration. ANOVAs for normalized duratioeasures followed the by-items and by-spers designs used for intelligibility. Unlike inte

igibility, duration loss vis-a`-vis the citationorm changed significantly over successiveens (0.568 for first tokens and 0.675 for snd), F2(1,116) 5 6.77, MSE 5 0.1001,p 5

0105;Fspeaker(1,54)5 7.28,MSE5 0.0683,p 5.0093. There were no signs of curtailing redtion in the denial condition: the interaction btween mention and feedback was not signific(0.519 and 0.658 for mentions precedingfollowing a denial; 0.617 and 0.691 for pawith no denial intervening;F2 , 1; Fspeaker, 1).

Discussion

Far from mitigating the repetition effectresponse to the listener’s overt feedback, spers continued to reduce intelligibility. Secomentions were equally degraded in intelligibity and duration regardless of listeners’sponses to the initial mention. There is nothin these results to indicate speakers’ sensitto feedback about what the listener can see

The results were complicated by an unexpedly extreme loss of clarity in those introdutory mentions which failed to elicit accurafeedback. The effect on intelligibility was nhowever, matched by an effect on duration.have extensively reexamined the data—by fof clause, face visibility, and inherent charteristics of the particular samples used inexperiment, for example—and have beenable to find any confounding variable whcould explain the particular lack of care takby speakers here. It is, of course, possiblethe causal relationship runs in the other dition. Poorly articulated tokens may not haelicited good feedback because the addresimply did not hear or recognize them. Mosubtly, the association may have somethin

rs

-

--

-

td

k-

y

t-

-

t-

ee

o

do with how listeners interpret more and lclear tokens.

Fowler and Housum (1987) (see also Ter& Nooteboom, 1987) found that second, lintelligible mentions were correctly classedlater mentions in a forced choice (“Old“New”) task. More interestingly, in probed rcall these less intelligible second tokens wmore likely to prime the recognition of worrelated to the original mention. When ospeakers happened to produce poor tokenfirst mentions, they might have failed to sigto their listeners that these were New landmwhich needed to be located on their mapsfact, the probed recall results suggest that ptokens will instigate searches for antecedenthe listener’s model of the dialogue. In thecases, the fruitless search may have distralisteners from more pertinent tasks like proving a suitable reply to their partner.

Poorly designed introductory mentions asthe results of Experiment 2 seem to indicateadjustments of intelligibility with repetition anot sensitive to what listeners can see. Eovert feedback indicating a tactically importform of copresence had no effect on speakcontrol of clarity. Those speakers were eitnot monitoring for feedback or not adjustitheir messages in response.

EXPERIMENT 3

We have so far asked whether speakerssensitive to two of the means by which itebecome listener-Given, by being heard or sNow we turn to situations where items aretener-Given because the listener has mentithem. To do this, we compare the effectsrepeated mentions of two kinds. Same-speitems were introduced and mentioned a sectime by the same person. Between-speitems were introduced by one player andpeated by the other. If repeated mentionsdegraded relative to introductory mentionsmatter who provided those original tokens,might have evidence that both listener menand speaker mention can confer Given stat

Why should we expect speakers to take nof what their interlocutors mention while ignoing their feedback? First, the listener playe

tingenryt bInunnoot2

thentlis

ca-sentiverermonwbe

faof

givs illi-

ker

arelo-Fondvigleto

tedd b

iny

hea

en

kef-

ssi-me-hileas

ordpro-akerkers

erethee-

un-thetherem:oryta-1;tionionorita-

wosed

ndignleteri-oupatinical

fat-

neneandEx-ad-Af-

12 BARD ET AL.

role close to the speaker’s own in constructhe discourse, and speakers may be more stive to such contributions than to commentaSecond, what the other speaker said mighless critical in itself than in what it implied.dialogues where the participants were encotering a map for the first time, there weregrounds for introducing an item other than spting it on the map. Although Experimentindicated insensitivity to what listeners saw,case here is somewhat different. Experimecontrasted overt feedback about what thetener could not see with overt or covert inditions of shared visual resource. In the preexperiment, introductory mentions are posievidence, overt indications of shared visualsource. Finally, if speakers do not add the teintroduced by their interlocutors to a commset of Given entities, it is difficult to see hoany approximation to common ground cancomposed.

On the other hand, since our results soshow no sensitivity to fairly obvious aspectslistener knowledge, it may be that speakersa special status to their own discourse. If thithe case, then the effect of repetition on integibility loss will be greater for same-speathan for different-speaker examples.

Method

Stimuli and design.All stimuli were wordtokens uttered as part of the names of shlandmarks, that is, of landmarks identicallycated, depicted, and labeled on Giver’s andlower’s maps. All occurred in both the first athe second mention of the same landmarkthe same referring expression within a sindialogue and with no intervening referencethe landmark in any other form. Of the repeamentions, 48 were introduced and repeatethe same speaker, while the other 48 weretroduced by one speaker and repeated bdifferent speaker. All items came from tspeakers’ first trial with a map: none wassecond introduction in the sense of Experim1. In Example (5) below,the extinct volcanoisrepeatedly mentioned by the same speawhile the tribal settlementis repeated by a diferent speaker.

si-.e

-

-

2-

t

-s

r

es

d

l-

a

y-a

t

r,

5. G: Start atthe extinct volcanoand godown round the tribal settlement.AndthenF: Whereabouts isthe tribal settlement?G: It’s at the bottom. It’s to the left oftheextinct volcano.

Lexical identity was matched as far as poble: 18 lexical items appeared as both saspeaker and different-speaker repetitions, wthe remaining word types were matchedclosely as possible for syllable length and wfrequency. Where feasible, speakers whoduced both same-speaker and different-spetokens were selected in preference to speawho contributed to a single cell. Materials wdrawn from the speech of 12 participants inface-visible condition and 14 in the facscreened condition.

For each of the word tokens uttered in rning speech, a citation form produced bysame speaker was used as a control. Thuswere four stimuli for each experimental iteRunning Speech Mention 1, the introductmention from the dialogue; Citation 1, the cition form read by the speaker of MentionRunning Speech Mention 2, the second menfrom the same dialogue; Citation 2, the citatform read by the speaker of Mention 2. Fsame-speaker repetitions, Citation 1 and Ction 2 are the same word token.

With 48 word types nested in each of tlevels of repeater (same, different) and croswith two of form (running speech, citation) atwo of mention (first, second), the desyielded a total of 384 stimuli. To avoid multipencounters with the same lexical items, maals were divided into two groups and each grwas distributed among four subgroups by Lsquare with the same random order of lexitems across each Latin square.

Participants and procedure.Eight groups o10 University of Glasgow undergraduatestempted to identify the words, with only otoken of any lexical item heard by any olistener. Materials preparation, instructions,response collection were as described forperiment 1. Stimuli were presented over hephones in individual sound-proofed booths.

ere

i-n-lveanatkel oco

n rde

e-toen-assig

0 dt ens tera

resra

rm

,

7 int anr ;

otty bere

reoneeof

os

onre-ter’sre-me-arksnd-nt-epe-toenheren

inst

en,iven

nedec-

Ex-thanntro-d we

ent3 r-e s asa

n

S

D

S

D


ter the introductory phases, the 48 stimuli wpresented with no further break.

Results

Intelligibility. ANOVAs again used intelligbility loss relative to citation form as the depedent variable. Participants were nested in haof the experiment and crossed with mentionrepeater. Items were nested in levels of repeand halves and crossed with mention. Speawere crossed with mention; but since not althe 26 speakers appeared in both repeaterditions, speakers were treated as nested ipeater level, and by-speaker results are incluonly for interest.

Table 4 shows mean intelligibility andk-normalized duration for all cells. Overall, reptition increased intelligibility loss relativecitation form (first mentions .15, second mtions .23). The main effect of repetition wsignificant by participants and approachednificance by items,F1(1,78) 5 6.17, MSE 50.0856,p 5 .0151; F2(1,92) 5 2.97, MSE 5

.1069,p 5 .0884;Fspeaker, 1. Repetition hahe same effect for same-speaker and differpeaker repetitions (all token by repeater inctions,F1 , 1).Duration.ANOVAs followed the designs fo

by-item and by-speaker intelligibility analysand showed the same pattern. Again overeductions in duration relative to citation fowere greater for second tokens (byk of 0.801)than for first (by 0.581),F2(1,92) 5 14.22MSE 5 0.1638, p 5 0.0003; Fspeaker(1,28) 5

.50, MSE 5 0.0909,p 5 0.0106, and agahere was no interaction between tokenepeater condition,F2(1,94) 5 2.67, n.s.

Fspeaker, 1.

Discussion

The results of Experiment 3 suggest that bspeakers can confer Given status on an entimentioning it, for repeated mentions wshorter and less clear than first mentions,gardless of who produced the first mentiSpeakers did not seem to discriminate betwtheir own introductory mentions and thoseanother speaker.

The design of Experiment 3 leaves the p

sderrsfn-e-d

-

t--

ll

d

hy

-.n

-

sibility, however, that the results were basedan egocentric consideration. All materialsferred to landmarks appearing on the repeamap. Since introductory mentions werestricted to items on the introducer’s map, saspeaker words had to name shared landmwhich the repeater could see. To match lamark names across conditions, differespeaker items also had to be shared. The rtition effect in both conditions might be duethe fact that both items were situationally Giv(Prince, 1981) for the repeater. The otplayer’s introductory mention might have beirrelevant. Experiment 4 offers a control agathis interpretation.

EXPERIMENT 4

To interpret the results of Experiment 3, thwe need to be sure that repeaters assigned Gstatus to items which they had heard mentiobut could not see. We cannot provide the nessary controls by applying the design ofperiment 3 to unshared landmarks, rathershared, because speakers seldom, if ever, iduced landmarks they could not see. Instea

TABLE 4

Effects of What the Listener has Mentioned (Experim): Mean Intelligibility,k-Normalized Duration (and Diffence from Citation-form Control) for Repeated MentionFunction of Repeating Speaker

Speakers

Mention

First Second

Runningspeech Citation

Runningspeech Citatio

Intelligibility

ame 0.500(0.150)

0.650 0.379(0.231)

0.610

ifferent 0.560(0.146)

0.706 0.371(0.227)

0.598

Duration

ame 0.706(0.625)

1.331 0.535(0.796)

1.331

ifferent 0.762(0.537)

1.299 0.448(0.805)

1.253

forhil

onicher

r. Icaenitynd

eaareuaenare

anerandioneti

lls ker ot dinc uchs d:a ano arw ths

f a

ath

gond

reenoor

omce-ia-s ofred,

ia-es.th asossi-con-of

) andrminonsbothe ex-ourere

fthents.

ces

ac-sual

thekers,d inen-

eti-

1F. arda nd-m l-y ala kers,FF

14 BARD ET AL.

examine only different-speaker repetitionsthe effects of what the repeater could see wuttering the second mention.

Experiment 4 uses introductory and secmentions of the names of the landmarks whwere either shared and visible to both speakor unshared and not visible to the repeatespeakers assign Given status to what theysee but not to what their listeners have mtioned, then they should reduce intelligibilonly when repeating the names of shared lamarks. This is the outcome which would revan artifact in Experiment 3. If speakerssensitive to both listener-mention and viscontext, they should reduce intelligibility whrepeating the names of shared and unshlandmarks, but more so when mentioninglandmark which they can see. Finally, if listemention is sufficient to convey Given statusvisual context is unimportant, then reductshould be the same with both kinds of reption.

Method

Materials and design.The materials were apoken words excerpted from different-speaepeated mentions which used the literal formhe landmark name and from the corresponitation form from each speaker. For 48 sets, the landmark referred to was unshareppeared only on the introducer’s map. Forther 48, the landmark was shared and appeith identical form, location, and label on bopeakers’ maps. In the examples below,the

rope bridge provides repeated mentions oshared landmark, andthe macheteprovidesmentions of an unshared item.

6. F: Right. How far?G: Um, at the opposite side.F: To the opposite side. Is it underne

the rope bridgeor to the left?G: It’s underneaththe rope bridge.And

then from the tribal settlementstraight up towards the rope bridge aover the rope bridge. Then down thsteps and along to above the volca

F: Eh, d . . . Isdown three steps belowabovethe machete?

G: Ah. The machete’s not on my map.

e

d

s,fn-

-l

l

d

-

rfg

it-ed

.

Items were contributed by 20 speakers frface-visible dialogues and 13 from fascreened. All items were from first trial dlogues. There were 48 repeated mentionshared landmark names and 48 of unshawith half of each group from face-visible dlogues and half from face-screened dialogu

Because few landmark names appear boshared and unshared features, it was not pble to match word types across sharednessditions. Items were thus nested in levelsspeaker’s visual access (shared, unsharedcrossed with mention (first, second) and fo(running speech, citation), giving 384 stimuliall. Twelve of the 33 speakers whose repetitiwere used contributed second mentions toshared and unshared categories. Items wercerpted and distributed by Latin square into fgroups. Methods for excerption and ISIs was described for earlier experiments.

Participants and procedure.Four groups onine participants each were drawn frompopulation sampled in the other experimeProcedure remained the same.

Results

Table 5 shows raw means and differenfrom citation form.

Intelligibility. In analyses of intelligibilityloss, participants were crossed with visualcess and mention. Items were nested in viaccess and crossed with mention. Becausedesign was not completely crossed by speaa by-speaker analysis, with speakers nestelevels of visual access and crossed with mtion, was performed for interest only.

Overall, there was a robust effect of reption, with intelligibility loss increasing fromfirst (.19) to second (.36) mention,F1(1,35) 528.77,MSE5 0.0343,p , .0001;F2(1,94) 5

2.36, MSE 5 0.1064, p 5 .0007;speaker(1,43) 5 16.10, MSE 5 0.0612, p 5

0002. There was no tendency whatever towmore extreme repetition effect for shared laark names (all Mention3 Visual Access ana

ses,F , 1). A main effect for repeater’s visuccess, significant by participants and spea1(1,35) 5 18.50,MSE5 0.0171,p 5 .0001;(1,94) 5 3.26, MSE 5 0.1295,p 5 .0743,
2

,s dinu ans ).

sf asm i-gs re-d nd(

0ua

a sign al-y

rgers

byarkyes

D

atedr and-m tor nt iths vi-s Thisp con-f cew isuala hee ofE wss dm kersa xt.W en-t

derw oni s-t cellcl t 1s , orr eak-e nd-m ec-o ersr ce.E hatt it-i s.A thei nd,E nerm in-t so ed.E orym vens cure aterc

hat

t 4)M r-e s asa

n


Fspeaker(1,43)5 5.87,MSE5 0.0815,p , .0197howed both introducer and repeater degranshared item names (first mention, .23econd, .41) more than shared (.15 and .30Duration. The ANOVA for duration los

rom citation form to running speech form (easured ink) followed the design for intellibility loss. Like intelligibility, durationhowed an overall effect of repetition, withuction increasing from first (0.601) to seco0.714) mentions,F2(1,94) 5 4.35, MSE 5

0.1398,p 5 .0397;Fspeaker(1,43)5 5.97,MSE5.0841,p 5 .0187).The interaction between mention and vis

ccess, though now more prominent, wasificant only in the imperfect by-speaker ansis, F2(1,94) 5 3.58, MSE 5 0.1398, p 5

.0616;Fspeaker(1,43)5 5.05,MSE5 0.0841,p 5

.0298. There was a tendency toward a laeffect of repetition on names of shared (fimentions by 0.538k-units, second mentions0.753) than on names of unshared landm(first mentions by 0.664k-units, second b0.675). The main effect of shared visual accwas now not significant (F2 , 1; Fspeaker(1,43)51.57,MSE5 0.21881, n.s.).

TABLE 5

Effects of What the Speaker Can See (Experimenean Intelligibility, k-Normalized Duration (and Diffence from Citation-form Control) for Repeated MentionFunction of Repeater’s Visual Access to Landmark

Visual access

Mention

First Second

Runningspeech Citation

Runningspeech Citatio

Intelligibility

Shared 0.632(0.150)

0.782 0.456(0.301)

0.757

Unshared 0.581(0.229)

0.810 0.421(0.410)

0.831

Duration

Shared 0.713(0.538)

1.251 0.430(0.753)

1.183

Unshared 0.544(0.664)

1.208 0.462(0.675)

1.137

gd

l-

rt

s

s

iscussion

In this experiment, cross-speaker repeeference to both shared and unshared larks led to increased intelligibility loss and

eduction in length. Fork-normalized duratiohe effect of mention was less uniform, womething close to an additional effect forual access to what was being named.attern indicates that listeners may indeed

er Given status on items they introduhether or not speakers can see the item. Vccess, if it works at all, works in addition to tffect of mention. Taken with the resultsxperiment 3, however, the outcome shoimilar effects on intelligibility of repeateention by a single speaker or across speand with or without supporting visual contehat is treated as Given is what has been m

ioned.

GENERAL DISCUSSION

Table 6 summarizes the conditions unhich we found an effect of Given status

ntelligibility, including what speaker and liener have said, seen, or heard. Where aontains both a “1” and a “2,” intelligibility

oss occurred in both conditions. Experimenhowed that what the listener had heardather had not heard, was unimportant to sprs who reduced clarity on introducing a laark to a new listener. Instead, in treating snd-trial introductions as Given, speakesponded to their own previous experienxperiment 2 showed that feedback about w

he listener could or could not see did not mgate intelligibility loss in second mentiongain, once the speaker had mentioned

tem, it was treated as Given. On the other haxperiment 3 showed that what the listeentioned was then treated as Given, with

elligibility loss virtually identical to repetitionf names which the repeater had introducxperiment 4 showed that the introductention by the listener sufficed to assign Gi

tatus, because intelligibility loss could ocven for names of objects which the repeould not see.All in all, the results seem to suggest t

:

a-ge

er’sn-o

t rep ant ent thes ionc osi erb eake hai

hana int icas , fon thep pred itorf just ef-e geE mai erim qus anw np geT ger

lis-forts 1,dic-inortoande onthebe-

nt 2ck.theth-andisjusta-ofers’edro-heirfor-nlyel-ro-ds.de-ge,ef-to

aded

heard

16 BARD ET AL.

although listeners may introduce New informtion, no other manifestation of their knowled(the right half of the table) affects the speaksubsequent control of intelligibility. Table 6 idicates that we might state the results even mstrongly. Only one column has the value “1”hroughout. In all the cases we tested, theeating speaker heard the original mention

hen produced relatively degraded second mions. If Given status is assigned to whatpeaker has heard, then listener introductonform to a general rule. Rather than suppng that speakers model one aspect of listenehavior only, we can now assume that sprs can merely fail to note who introduced w

nto a dialogue.This conclusion appears more radical t

ny of the hypotheses characterized in theroduction to this paper. Some more radtance seems to be warranted, howeverone of the hypotheses predicted all ofresent results. The no default hypothesisicted that speakers would continually mon

or evidence of listeners’ knowledge and adheir articulatory effort to such evidence in prrence to any record of their own knowledxperiments 1 and 2 show that speakers

gnore important facts about listeners. Expents 3 and 4 together seem to show e

tatus for what the speaker has introducedhat the listener has introduced, rather thareference for adapting to listener knowledhe copresence hypothesis allows for lon

TAB

Summary of Speaker-Knowledge and Listener-Know

Experiment

H

Speaker

Has said Can see H

1 1 12 1 13 2/1 14 2 2/1

Note.1, condition applies in experimental materials22/1, critical comparison between conditions.

re

-d-

s-

’s-

t

-lr

-

t

.y-alda.-

term substitution of speaker knowledge fortener knowledge but only when evidencecopresence triggers the default. Experimen2, and 4 also largely fail to meet these pretions, for they all show that differencesshared information—experience with a mapability to see a particular landmark—failedchange speakers’ behavior. The monitoradjust hypothesis predicted that dependencspeaker knowledge could be mitigated iflistener supplied feedback on misalignmenttween listener and speaker. Yet Experimereveals no mitigating effects of overt feedba

To explain the results, we turn instead tomodel on which the monitor and adjust hypoesis is based, a model presented in BrownDell (1987) and Dell and Brown (1991). Thmodel goes further than the monitor and adsystem in assigning different kinds of informtion to different stages in the formulationutterances. In an attempt to model speakdescriptions of pictured activities performwith unusual instruments, Brown and Dell pposed that speakers begin by structuring tbasic message around a constellation of inmation from their own personal scenarios. Ogradually, via iterative monitoring and emblishment, do they adjust their output to nonptotypical information or to the listener’s neeConsigning all early stages of propositionsign exclusively to the speaker’s knowledBrown and Dell suggest that the repetitionfects of Fowler and Housum (1987) are due

6

ge Conditions under Which Word Intelligibility Is Degr

Given status achieved

Listener

heard Has said Can see Has

2 2 22 2/1 1

1/2 1 11 1 1

ndition does not apply in experimental materials;1/2 or

LE

led

ow

as

1111

;, co

eirser

on

incntro-ingssionli-

imeng.rinan

perndri-ancalli-of

forn&ano

icition

eedar

fas

im-xt

chdhatac

ua-rdse-63ta-ef

po

&reare

in-Be-en-thekerowbybe

enerentwithwo

ro-at isl ofde-ncehtts.e-in-d inheyAs

sarom-edeare

eiring

letedheey-the

ter-beut-thisget:is inIneAt


intralexical priming of the kind found in thproduction of semantically related word pa(Balota, Boland, & Shields, 1989). Listenmodels are not available at this stage.

The present results seem to conform tvariant of this model. Like Dell and Brow(1991), we propose that at least two distkinds of processes account for speakers’ coof the intelligibility of words in referring expressions. The first are fast automatic primprocesses, the second are slower procewhich may demand elaborate and computatally taxing inference. We add a further compcation: More demanding tasks compete for tand attention with dialogue and task planni

The fast priming processes operate duany attempts to produce spoken utterancesclosely resemble well-established effects inception. As in perception, priming here depeexclusively on what a single individual expeences. Though we cannot discount DellBrown’s (1991) suggestion that lexical–lexipriming is involved, context effects on intelgibility seem to demand two more kindspriming.

The first of these depends on discourse inmation. Since Given information in writtetexts is activated information (see McKoonRatcliff, 1980), all entities introduced intodialogue can be activated, even if they arefully grounded as mutual knowledge by explinteractions between interlocutors. Activatof the representation of referent objects primtheir names (Mitchell & Brown, 1988). Primnames, as in the Balota et al. (1989) study,produced faster, and the usual result ofspeech is decreased articulatory detail.

The second kind of priming depends onmediate linguistic context. Sentence conteprime the recognition of single words whicontinue them (see, for example, O’Seagh1997) to the extent that those words are syntically and semantically appropriate contintions of the context. Once again primed woshould reduce in length and intelligibility. Rdundancy, in the sense of Lieberman’s (19study, translates into priming in this interpretion. Like the repeated mention effects, thefects of sentence context are strongest in s

a

tl

es-

gd

-s

d

-

t

s

et

s

,-

)

-n-

taneous speech (Fowler, 1988; PedlowWales, 1987; Samuel & Troicki, 1998), whesentence structure and word pronunciationbeing planned simultaneously.

Because no inferences or decisions arevolved, the priming processes are quick.cause Given status is the prime for the Givness effect, it is unimportant who makesintroductory mention, so long as the speaobserves it. If articulatory adjustments shany consideration for the listener, it is onlycoincidence. In fact, consideration couldsimulated to the extent that speaker and listshare knowledge to begin with or to the extthat the listener populates the conversationovert references to information which the tdid not initially share.

The second major process, or family of pcesses, maintain, elaborate, and exploit whusually supposed to be the speaker’s modethe listener. Such processes might includeciding which kinds and degrees of copreselicense which varieties of default. They miginclude updating memory for dialogue evenThey might also include determining which bliefs and which goals must be attributed toterlocutors on the basis of what they say, anparticular on the basis of any feedback tprovide to the speaker’s own utterances.Brown and Dell (1991) and Horton and Key(1996) and their colleagues suggest, these cplex processes are simply too slow to precevery attempt at speech production. Theytherefore likely to be too slow to make theffects felt on a word-to-word basis in runnspeech.

Whether these processes are ever compwill depend on at least their complexity, ttime at the speaker’s disposal (Horton & Ksar, 1996), and the other demands onspeaker’s attention. Carefully preplanned utances in formal lectures, for example, maytailored over many iterations to match the ocome of these processes. Written output ofkind has the advantage of a stationary tarThe only change in the putative addresseehis or her experience of the growing text.theory, it is possible to allocate sufficient timto assure completion of the revision cycles.

omketh

partheror, thn oel

th. Inen-ayim

thesereornslly

on.ked

inrecia90

ces80m-s t

thraneio6)

cidimurerelredncre

tan2)ire

fig-thef ataskver’sling.inllyhe

ins ofcon-n-

la-oulder’slis-ionse-avecrossThen-bse-hatiatecan

thishyrep-r toria-ifon-lo-odetheant

int ofpe-

tioner-m-er-

18 BARD ET AL.

the other extreme, speakers performing a cplex task with an interlocutor who could tathe floor at any pause may never completecomputation and may not even have the scapacity to acquire the information on whichcomputation is based. Although genuine erof expression and understanding may resultrisk appears to be warranted (for a discussiorisky referential strategies, see Carletta & Mlish, 1996).

This model seems to accommodate bothpresent results and others in the literaturethis study, it helps to explain why the on-licontrol of production is sensitive to what listeers mention, but not to what they offer by wof feedback or copresence: Mentions fuel pring but feedback needs to be interpreted.

The model also helps to explain whyvisual channel is not so much used as abuAlthough speakers took the opportunities psented by a visual channel to become mcareless in articulating introductory mentiothey did not monitor the listener’s face carefufor signs of comprehension or confusiSpeakers with open sight lines actually looat their listeners during only about one wordeight (Anderson et al., 1997). A typical figufor dialogue, this outcome suggests that sorestrictions on interspeaker gaze (Argyle, 19and the detrimental effect of gaze on the proof planning discourse (Beattie 1978, 191981; Exline & Winters, 1965) are more copelling than the need for fine-grained accesthe listener’s reactions.

The model also makes some sense ofdistribution of results across tasks. In genethe easier the task, the more firmly listevariables are found to affect speakers’ behavFor example, Horton and Keysar’s (199speakers had to enable listeners to dewhether a description was true of a single sple shape on each trial. With no time pressspeakers preferred to describe the figurestive to shared context rather than unshaWith time pressure, this cooperative preferedisappeared. Similarly, more cooperativesults are reported for experiments on thegram task (e.g., Wilkes-Gibbs & Clark, 199than on the map task. The tangram task requ

-

ee

sef

-

en

-

d.-e,

l)s

,

o

el,rr.

e-,

a-.

e--

s

matching from a fully shared closed set ofures to a simple array. The main problem forinstructor in this task is labeling the next oshrinking set of candidate figures. The mapis more difficult. Since landmarks may hadifferent names on the Giver’s and Followemaps, there are occasional problems of labeEach player holds unique critical informationthe form of unshared landmarks and initianeither player knows which critical items tother may have or exactly how many remainplay as the game proceeds. The demandaligning partial representations under theseditions should leave little time for computatioally expensive cooperative behavior.

Finally, the model predicts trouble. Articution geared to the speaker’s experience shsometimes be inappropriate to the listenneeds. Since articulatory clarity affects howteners interpret the spoken referring express(Fowler & Housum, 1987; Terken & Nootboom, 1987), these misalignments should hconsequences. We seem to have come aone such consequence in Experiment 2.introductory mentions were extraordinarily uclear in just those cases where listeners suquently failed to report the important news tthey lacked the landmark. The inapproprfirst mentions may not have alerted them to sfor a new item.

We can see three major objections toapproach. First, it is not immediately clear wGiven status should prime more than mereetition. Second, our model does not appeapredict the copious evidence for orderly vation in form of referring expression. Third,adequate communication is the goal that ctrols language use, it is unclear how homoquens could develop such an egocentric mof operation. In all three cases, however,literature allows for interpretations consonwith the present model.

First, the particular status of Given entitiespriming of reduced pronunciations seems oustep with the many cases in which mere exrience of the prime affects latency in perceptasks (for reviews, see Balota, 1994; Zwitslood, 1996). Although Balota et al. (1989) deonstrated priming of word duration in an exp

ofce

amociTondvi-tudingssoredlikeeortedrte

.le.e.l,owat7)ednt-nd

allytinngrimd bivelytin

lat-

thefirss aelylisestheento

ersall

e ofner

buted.s inithst-hen

andthetch-rs’er-fer-orstellneworshenverarseNo

thisu-

s aac-o

ofaye-to

ng,veryed

-net ami-ed

aK ades ew.S fer-r ctedf ona n-t thed at-


imental paradigm which also yielded primingresponse latency, the two effects are not nesarily identical. Balota et al. used delayed ning of words presented with or cued by assated and semantically related words.demonstrate duration priming, Shields aBalota (1991) used modified reports of preously read sentences. The nature of their smakes the demonstration of repetition primless compelling than the demonstration of aciative priming. Shields and Balota measuthe duration of a single word per sentence,cat in (7)–(9) below (p. 49), which might bpreceded by the same word (7), a related w(8), or an unrelated control (9). The repeaand related condition targets were both shothan the control target.

7. Her cat chases our cat under the table8. Her dog chases our cat under the tab9. Her son chases our cat under the tablBecause no target item was coreferentia

might be argued that this experiment did shsimple repetition priming. The difficulty is thit is almost impossible to utter items like (without contrastive stress on the words precing both tokens ofcatand consequent deacceing of the target. Associated items like (8) aunrelated cases like (9) might more naturinclude accented tokens ofcat. The fact thaonly items like (7) differed from controlsamplitude, as would be expected with a chain accenting, bolsters the suggestion that ping and duration effects have been obscurea global prosodic artifact. If only associatand semantic priming of duration are firmdemonstrated, we are not so much contradicthe findings for duration priming as extrapoing from them.

Second, the model is less in conflict withliterature on referring expressions than it atseems. These studies all show that speakerwriters make adjustments. It is not completclear that they make the adjustments whichteners and readers need. The studies of accbility in texts, as we noted earlier, observerelationship between anaphor and antecedbut do not directly test either accessibilityreaders or writers’ attempts to track readneeds. Studies of spoken interactions, typic

s---

y

-

d

r

it

-

e-y

g

tnd

-si-

t,

’y

between an instruction giver and a sequencfollowers, do offer contrasting cases of listeneeds, and contrasting speaker behavior,they do not show that the two are closely linkSchober (1993), for example, found changeostensible viewpoint in speech to listeners wdifferent physical orientations, but less adjument when the listeners were present than wthey were merely described. Wilkes-GibbsClark (1992) found several differences inway tangram directors approached new maers differing in their experience of the directoearlier trials with other matchers. Some diffences of matcher experience induced difences in the time and words which directdevoted to the next trial, but we cannotwhether these adjustments actually met thematchers’ needs. If they did, that is, if directsupplied what each new matcher lacked, twe might expect all new matchers, whatetheir experience, to have to make equally spcontributions to the interaction themselves.matcher figures are given which would testprediction. Directors also varied the distribtions of their referring expressions acrosfour-way accessibility hierarchy. The resultscord nicely with prediction, but there are twdifficulties. First, without contextual supportthe kind that would invite artifacts, a three-wcritical distinction within this hierarchy reprsents one of the most difficult discriminationsmake in recognizing, let alone transcribispontaneous speech. Two metrically weak,short syllables with centralized, often devoicnuclei, one with “real” frication (definitetheman), the other with “nuisance” frication (indefinitea man), must be distinguished from oanother and from something which is nosyllable but just background noise (bare nonal man). With much more easily distinguish

lternatives (e.g.,big and small), Horton andeysar (1996) observed that time pressure mpeakers indifferent to what listeners knecond, we do not know that the suitable re

ing expression has the Gricean effect predior it, easing the matcher’s way. In Andersnd Boyle’s (1992) study of introductory me

ions to naive map task listeners, however,efinite–indefinite choice had no effect wh

riaesanoun

roy anttentaer

, this oosin

lvethemg

ther.ste

du-n-m

spgein-1)unllyt rim-in-rstesy.

o-andi-

G.,J.,.,

or-

, &l of

e.

A -

eyd.),i-

9).ione,

ity

ityity.

B ofachHech

B on,i-udn-.

B aze in

B tion-to-

.

B og-con-

B

ofn-

and-

tonts.

C ions

C ery

20 BARD ET AL.

ever on any subsequent measures of appropfeedback or accurate task completion. At bthen, the literature on referring expressionslistener modeling does not present a clear cterargument to the model outlined above.

Finally there is the issue of how speech pduction could possibly be designed as poorlthe present model suggests, with adjustmelisteners’ needs conditional on time and attion commensurate with a complex computional task. We suggest that the design is vlikely to be optimal because the default casecase in which the speaker proceeds from hher own knowledge, is usually adequate. Mconversations either in individual lifetimes orthe history of the species are likely to invocopresent interlocutors who are close togein time, space, and linguistic and social comunity. Much of what the interlocutors brinwith them to the conversation and much ofway they view its contents will be similaThere may be little pressure to develop facomputation in dialogue because we cansurprisingly well without it. Approximate accracy of this kind is often the end-result of evironmental pressures on perceptual systeFor example, oyster catchers survive as acies despite their tendency to retrieve the larrounded object in their environs when their sgle egg rolls out of the nest (Tinbergen, 195Because the egg usually is the largest roobject visible, “largest-round-object” is usuaan adequate model of an egg. Our presensults point to nothing more radical than a silarly approximate adjustment to the likelyternal state of likely conversational partneAdapting to actual but less likely internal stamay be an evolutionary and cognitive luxur

REFERENCESAnderson, A. H., & Boyle, E. A. (1994). Forms of intr

duction in dialogues: Their discourse contextscommunicative consequences.Language and Cogntive Processes,9, 101–122.

Anderson, A. H., Bader, M., Boyle, E., Bard, E.Doherty, G., Garrod, S., Isard, S. D., Kowtko,McAllister, J., Miller, J., Sotillo, C., Thompson, H. S& Weinert, R. (1991). The H.C.R.C. Map Task Cpus.Language and Speech,34, 351–366.

Anderson, A. H., Bard, E. G., Sotillo, C., Newlands, A.Doherty-Sneddon, G. (1997). Limited visual contro

tet,d-

-sto--yert

r-

ro

s.e-st

.d

e-

.

the intelligibility of speech in face-to-face dialoguPerception and Psychophysics,59, 580–592.

rgyle, M. (1990).Bodily communication.London: Routledge.

Ariel, M. (1990).Accessing noun-phrase antecedents.Lon-don: Routledge/Croom Helm.

Balota, D. A. (1994). Visual word recognition: The journfrom features to meaning. In M. Gernsbacher (EHandbook of psycholinguistics(pp. 303–358). San Dego: Academic Press.

Balota, D. A., Boland, J. E., & Shields, L. W. (198Priming in pronunciation: Beyond pattern-recognitand onset latency.Journal of Memory and Languag28, 14–36.

Bard, E. G., & Anderson, A. H. (1983). The unintelligibilof speech to children.Journal of Child Language,10,265–292.

Bard, E. G., & Anderson, A. H. (1994). The unintelligibilof speech to children: Effects of referent availabilJournal of Child Language,21, 623–648.

ard, E. G., Lowe, A., & Altmann, G. (1989). The effectsrepetition on words in recorded dictations. In J. Tub& J. Mariani (Eds.),Proceedings of EUROSPEEC’89: Vol. 2. (pp. 573–576). Paris: European SpeCommunication Association.

ard, E. G., Sotillo, C., Anderson, A. H., Doherty-SneddG., & Newlands, A. (1995). The control of intelligbility in running speech. In K. Elenius & P. Brander(Eds.), Proceedings of the XIIIth International Cogress of Phonetic Sciences: Vol. 4.(pp. 188–191)Stockholm: Stockholm University.

eattie, G. (1978). Sequential patterns of speech and gdialogue.Semiotica,23, 29–52.

eattie, G. W. (1980). The role of language producprocesses in the organization of behaviour in faceface interaction. In B. Butterworth (Ed.),LanguageProduction: Vol. 1, Speech and Talk(pp. 69–107)London: Academic Press.

eattie, G. W. (1981). A further investigation of the cnitive interference hypothesis of gaze patterns inversation.British Journal of Social Psychology,20,243–248.

olinger, D. (1963). Length, vowel, juncture.Linguistics,1,5–29.

Bolinger, D. (1981).Two kinds of vowels, two kindsrhythm.Bloomington, Indiana: Indiana University Liguistics Club.

Brennan, S., & Clark, H. (1996). Conceptual pactslexical choice in conversation.Journal of Experimental Psychology: Learning, Memory, and Cognition,22,1482–1493.

Brown, P., & Dell, G. (1987). Adapting productioncomprehension—the explicit mention of instrumeCognitive Psychology,19, 441–472.

ampbell, W. N., & Isard, S. D. (1991). Segment duratin a syllable frame.Journal of Phonetics,19, 37–47.

arletta, J., & Mellish, C. (1996). Risk-taking and recov

C ,

C and

.

C ce, &

s.D er-

ofgl: A

D C.,-to-ari-

.E e

ter-gica

F t ine,

F on-on-

ofep-d

msIna-.

F of

F n oas-

G & Jech

G og-s in

H on

H take.n-

rsa--

L lby,E.

ogy

L andtion

L he.),

M ionana-r-

M pe-ionl

O fects-ni-

P t onh,

P x-

P ess,on

al-:

P for-

S isor

d

S rsa-

S by


in task-oriented dialogue.Journal of Pragmatics,26,71–107.

hafe, W. (1974). Language and consciousness.Language50, 111–133.

lark, H. H., & Haviland, S. E. (1977). Comprehensionthe given-new contract. In R. O. Freedle (Ed.),Dis-course production and comprehension(pp. 1–40)Hillsdale, NJ: Erlbaum.

lark, H. H., & Marshall, C. R. (1981). Definite referenand mutual knowledge. In A. K. Joshi, B. WebberI. Sag (Eds.),Elements of discourse understanding(pp.111–222). Cambridge: Cambridge University Pres

ell, G., & Brown, P. (1991). Mechanisms for listenadaptation in language production: Limiting the rolethe “model of the listener”. In D. J. Napoli & J. A. Ke(Eds.),Bridges between psychology and linguisticsSwarthmore Festschrift for Lila Gleitman(pp. 111–222). Hillsdale, NJ: Erlbaum.

oherty-Sneddon, G., Anderson, A. H., O’Malley,Langton, S., Garrod, S., & Bruce, V. (1997). Faceface and video-mediated communication: A compson of dialogue structure and task performance.Jour-nal of Experimental Psychology: Applied,3, 105–125

xline, R. V., & Winters, L. C. (1965).Effects of cognitivdifficulty and cognitive style upon eye contact in inviews. Paper presented at the Eastern PsycholoAssociation.

isher, C., & Tokura, H. (1995). The Given-New contracspeech to infants.Journal of Memory and Languag34, 287–310.

owler, C. (1988). Differential shortening of repeated ctent words produced in various communicative ctexts.Language and Speech,28, 47–56.

Fowler, C., & Housum, J. (1987). Talkers’ signalling‘new’ and ‘old’ words in speech and listeners’ perction and use of the distinction.Journal of Memory anLanguage,26, 489–504.

Fowler, C., & Levy, E. (1991). Some ways in which forarise from functions in linguistic communications.P. Romeas (Ed.),Actes du XIIIeme Congres Interntional des Sciences Phone´tiques: Vol. 1(pp. 279–282)Aix-en-Provence: Universite´ de Provence.

owler, C., Levy, E., & Brown, J. (1997). Reductionsspoken words in certain discourse contexts.Journal ofMemory and Language,37, 24–40.

ussell, S. R., and Krauss, R. M. (1992). Coordinatioknowledge in communication: Effects of speakers’sumptions about what others know.Journal of Personality and Social Psychology,62, 378–391.

rice, H. P. (1975). Logic and conversation. In P. ColeMorgan (Eds.),Syntax and semantics: Vol. 3. Speacts (pp. 41–58). New York: Academic Press.

undel, J. K., Hedberg, N., & Zacharski, R. (1993). Cnitive status and the form of referring expressiondiscourse.Language,69, 274–307.

awkins, S., & Warren, P. (1994). Phonetic influencesthe intelligibility of conversational speech.Journal ofPhonetics,22, 493–511.

l

f-

.

orton, W., & Keysar, B. (1996). When do speakersinto account common ground?Cognition,59, 91–117

Hunnicutt, S. (1985). Intelligibility vs. redundancy—coditions of dependency.Language and Speech,28, 47–56.

Isaacs, E., & Clark, H. H. (1987). References in convetion between experts and novices.Journal of Experimental Psychology: General,116,26–37.

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

aver, J. D., Hieronymus, J., Bennett, C., Cohen, I., DaJ., Davies, D., Hiller, S., McAllister, M., & Purves,(1989).The ATR/CSTR Speech Database Project(Re-port 3). Edinburgh: Centre for Speech TechnolResearch, Edinburgh University.

ieberman, P. (1963). Some effects of the semanticgrammatical context on the production and percepof speech.Language and Speech,6, 172–175.

indblom, B. (1990). Explaining variation: A sketch of tH and H theory. In W. Hardcastle & A. Marchal (EdsSpeech production and speech modelling(pp. 403–439). Dordrecht: Kluwer Academic.

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

itchell, D. B., and Brown, A. S. (1988). Persistent retition priming in picture naming and its dissociatfrom recognition memory.Journal of ExperimentaPsychology: Learning Memory and Cognition,14,213–222.

’Seaghdha, P. G. (1997). Conjoint and dissociable efof syntactic and semantic context.Journal of Experimental Psychology: Learning, Memory, and Cogtion, 23, 807–828.

edlow, R., & Wales, R. (1987). Some effects of contexspoken language production.Language and Speec30, 373–385.

ollack, I., & Pickett, J. (1963). The intelligibility of ecerpts from conversation.Language and Speech,6,165–171.

rince, E. (1992). The ZPG letter: Subjects, definitenand information-status. In W. Mann & S. Thomps(Eds.),Discourse description: Diverse linguistic anyses of a fund-raising text(pp. 295–326). AmsterdamBenjamins.

rince, E. (1981). Toward a taxonomy of given-new inmation. In P. Cole (Ed.),Radical pragmatics(pp. 223–255). New York: Academic Press.

amuel, A. G., & Troicki, M. (1998). Articulation qualityinversely related to redundancy when childrenadults have verbal control.Journal of Memory anLanguage,39, 175–194.

chober, M. (1993). Spatial perspective taking in convetion. Cognition,47, 1–24.

chober, M., & Clark, H. H. (1989). Understandingaddressees and overhearers.Cognitive Psychology,21,211–232.

nd

s oten

d

of

ng-

-

((

22 BARD ET AL.

Shields, L. W., & Balota, D. A. (1991). Repetition aassociative context effects in speech production.Lan-guage and Speech,34, 47–55.

Stalnaker, R. C. (1978). Assertion. In P. Cole (Ed.),Syntaxand semantics 9: Pragmatics.(pp. 315–332). NewYork: Academic Press.

Terken, J., & Nooteboom, S. (1987). Opposite effectaccentuation and deaccentuation on verification lacies for Given and New information.Language anCognitive Processes,2, 145–163.

Tinbergen, N. (1951).The study of instinct.London: OxfordUniversity Press.

f-

Vonk, W., Hustinx, L., & Simmons, W. (1992). The usereferential expressions in structuring discourse.Lan-guage and Cognitive Processes,7, 301–333.

Wilkes-Gibbs, D., & Clark, H. H. (1992). Coordinatibeliefs in conversation.Journal of Memory and Language,31, 183–194.

Zwitserlood, P. (1996). Form priming.Language and Cognitive Processes,11, 589–596.

Received July 13, 1998)Revision received June 1, 1999)

controlling the intelligibility of referring expressions ...€¦ · ellen gurman bard human...

Documents