JOURNAL OF MEMORY AND LANGUAGE 27, 143 - 165 (1988)

Cognitive Penetration of the Mechanisms of Perception:Compensation for Coarticulation of Lexically Restored Phonemes

JEFFREY L. ELMAN

University of California, San Diego

AND

JAMES L. MCCLELLAND

Carnegie Mellon University

An important question in language processing is whether higher-level processes are ableto interact directly with lower-level processes, as assumed by interactive models such as theTRACE model of speech perception. This issue is addressed in the present study by exam-ining whether putative interlevel phenomena can trigger the operation of intraleve! pro-cesses at lower levels. The intralevel process involved the perceptual compensation for thecoarticulatory influences of one speech sound on another. TRACE predicts that this com-pensation can be triggered by illusory phonemes which are perceived as a result of top-down, lexical influences. In Experiment 1 , we confirm this prediction. Experiments 2 to 4replicate this finding and fail to support several potential alternative explanations of theresults of Experiment 1. The basic finding that intralevel phenomena can be triggered byinterlevel processes argues against the view that aspects of speech perception are encapsu-lated in a module impervious to influences from higher levels. Instead , it supports a centralpremise of interactive models, in which basic aspects of perceptual processing are subject toinfluences from higher levels. (f) 1988 Academic Press, Inc.

How far down into the mechanisms ofperception do higher- level contextual influ-ences reach? There has been considerabledebate on this issue and the more generalquestion of the degree to which cognitiveprocesses interact with one another. Sometheorists (e. , Fodor, 1983) have proposedthat processing is essentially modular innature , and that the notable feature of cog-nition is the autonomy of mental faculties.This is the autonomous or modular view.Others have argued that the flow of infor-mation is rather freer, and that top-downas well as bottom-up interactions are pos-sible. This is the interactive view. The issueis a difficult one to decide , because there is

Requests for reprints should be addressed to Dr.Elman at the Department of Linguistics , University ofCalifornia , San Diego , La Jolla, CA 92093.

considerable overlap in the predictions

made by the two accounts. Many of thephenomena which have been cited as evi-dence for either the autonomous or interac-tive theories are in fact compatible withboth.

In this paper we consider the question asit arises in the perception of speech. describe a technique which we believe pro-vides a rigorous test of the existence of truetop-down interactions in processing. Thetechnique is applied here to the domain ofspeech perception, but we believe it canalso be used in other areas where theseissues arise. Using this technique, we findthat higher-level contextual factors cantrigger compensatory processes that arebasic to speech perception. These findingsdemonstrate that a more direct effect of

143

0749-596X188 $3.Copyright (f) 1988 by Academic Press , Inc.All rights of reproduction in any form reserved.

144 ELMAN AND MC CLELLAND

higher-level knowledge on perception ispossible than has been thought.

EVIDENCE FOR INTERACTIONS

The degree to which cognitive processesare modular has been an issue of great in-terest and controversy. At first glance, thecase for interactions, and , in particulartop-down effects , might seem to be ratherstrong. Marslen- Wilson , Tyler, and theircolleagues (Marslen-Wilson & Tyler, 1980;Marslen-Wilson & Welsh , 1978) haveshown that language comprehension ap-pears to involve the simultaneous pro-

cessing of information at the semantic

syntactic , and phonological levels (atleast). Furthermore , this processing seemsto be not only parallel but interactive; thatis, processing at each level appears to influ-ence , and be influenced by, processing atother levels (cf. Rumelhart , 1977). The re-sult is that many tasks, such as error detec-tion and shadowing, are accomplishedmore quickly when information from thesevarious levels is simultaneously available

(Cole , 1973; Cole & Jakimik, 1978, 1980;

Cole, Jakimik , & Cooper, 1978). Warrenand Sherman (1974) have demonstratedthat lexical information appears to play anactive role in the perception of phonemes;they obtain a " phoneme restoration" effect

in which obliterated phonemes are per-ceived as being present , when the contextis a lexical item. Ganong (1980) has shownthat the perceptual boundary betweenphonemes can be shifted as a function oflexical context; thus, a sound acousticallymidway between (g) and (k) is perceived asa (k) in the context" - iss, " but as a (g) inthe context " ift." Other evidence sup-porting the interactionist position on lexicalprocessing has been presented by Morton(1969), Grosjean (1980), and others (seealso Samuel , 1986 , and McClelland &Elman, 1986 , for extensive considerationsof additional literature).

. On the other hand, it has been claimedthat processing is actually autonomous andthat bottom-up processes make their

output available to higher-level processesbut that the latter do not affect the opera-tion of the former (e.g., Forster, 1979; Tan-enhaus , Carlson

, &

Seidenberg, 1984).

this view, lower-level processes would ex-tract cues to the identity of speech soundsfrom the acoustic signal. This part of themechanism is sensitive to local acousticcontextual influences but not to higher-

level contextual factors , such as whether asequence of phonemes forms a word or isconsistent with the more global context. Asecond stage would take the output of thefirst stage and combine this output withhigher-level factors , such as knowledge ofwhat phonemes make a word , to make de-cisions about the identity of phonemes.

There are several ways in which the au-tonomous model can explain the effectsthat would seem to argue for freer interac-tions. Norris (1982) has proposed that theoutput of the lower-level processes may bequite detailed. If the early stages of pro-cessing are unable to resolve ambiguousinput , they simply need to provide a richenough output that later stages can makethe correct interpretation (based on infor-mation which is available to them , but notto the lower stages). Thus , for example, thelexical bias on phoneme identification dis-covered by Ganong (1980) would be ex-plained as the effect of a postperceptual re-sponse strategy which results in responsesthat are more consistent with lexical items.

Another account of how some apparenttop- down interactions can arise in astrictly bottom-up model is suggested byFodor (1983) and by Forster (1981). It isplausible that frequent sequences of eventswould result in stronger intramodule asso-ciations for these sequences. Thus, fre-quent sequences of words or frequent se-quences of phonemes might facilitate fillingin of degraded or missing input; but thefilling in would arise from intramodule dy-namics and not from higher-level informa-tion.

It is thus difficult to know whether, in agiven instance, the effects of context are

COGNITIVE PENETRATION OF PERCEPTUAL MECHANISMS 145

due to intralevel dynamics or to higher-level postperceptual adjustments or to truetop-down interactions. What would such atest involve?

In order to answer that question , let usfirst describe a model of speech perceptionwe have developed , called the TRACEmodel (Elman & McClelland , 1986;

McClelland & Elman, 1986). This modelholds that there are interactions betweenthe syntactic/semantic level of processing

and the lexicon and between the lexiconand phonetic analysis , with informationflowing in both directions between adjacentlevels. TRACE accounts for a large body ofempirical data. For current purposes, how-ever, we want to describe how the modelsuggests a method for addressing the issueof interactions between levels. In so doing,TRACE not only provides an account forexisting data , but it also serves as thedriving force for new empirical findings.

THE TRACE MODEL OF SPEECHPERCEPTION

The TRACE model is based on an inter-active-activation model of processing(McClelland & Rumelhart, 1981; Rumel-hart & McClelland , 1981). Processing

carried out in a network consisting a largenumber of interconnected elements calledunits. Different classes of units represent

acoustic/phonetic features , phonemes , andwords. At the feature level , there is a unitfor each feature at each of a large numberof time slices relative to the onset of an ut-terance. At the phoneme level , there is aunit for each phoneme in each time slice.At the word level , there is a unit for eachword , starting in each time slice.

Each unit has an activation value whichis taken to be translatable into an estimateof the strength of the hypothesis that the

concept the unit represents is present in thesignal in the time slice or slices that the unitcovers. Thus, the pattern of activation overthe units in the network is the system s rep-resentation of the content of the utteranceit is currently processing. This representa-

tion evolves through time , as activationschange in the course of processing.

Activations of units are continuously up-dated in TRACE , based on activations ofunits that project to them. When a unit'activation exceeds a threshold value , it ex-cites or inhibits other units in a way whichreflects the mutual consistency or inconsis-tency of the concepts represented by eachunit. Influences are bidirectional; featureunits excite units for the phonemes thatcontain them, and phoneme units exciteunits for the features they contain. Simi-larly, phoneme units excite units for thewords that contain them , and word unitsexcite units for the phonemes they contain.There are also inhibitory connections be-tween mutually incompatible units on thesame level. Thus , units for different

phonemes in the Same time slice are mutu-ally inhibitory, as are units for different

words that span overlapping ranges of timeslices (here the extent of inhibition is pro-portional to the degree of overlap).

The activations of units are determinedby a simple two-part activation function.

The first part consists of adding together allof the inputs to the unit , to obtain what iscalled its net input. This net input is thenused to update the activation of the unit. Ifthe net input is positive (excitatory), ittends to increase the activation of the unit;if it is inhibitory, it tends to decrease theunit's activation. The activation function isapproximately linear near the middle of itsrange , but is nonlinear at the extremes , sothat activations of units receiving stronglyexcitatory input level off below the max-imum activation of 1. , and activations ofunits receiving strongly inhibitory input

level off above the minimum activationwhich is set at a value slightly below theunit' s threshold.

Context Effects in TRACE

Let us consider for a moment how theTRACE model accounts for context effectson phoneme identification , since the detailsof this process will be relevant when we

146 ELMAN AND MC CLELLAND

consider how one might test the assump-tion that influences really do feed backdown from higher levels within the speech-processing system. For concreteness, con-

sider the identification of an ambiguoussegment , providing equal bottom up inputto Ib/ and a /p/, followed by /rg/ so that thewhole input is between blug and plug.

Figure 1 illustrates simulated activationsof units at the phoneme and word levels atseveral points in this process. The input isassumed to unfold in time, with features ofthe ambiguous segment arriving first. Thisinput activates feature level units , which inturn provide bottom-up input to phonemeunits in the corresponding time slices. Thegreatest amount of excitation is received byunits for /b/ and /p/, so the activations

both units begin to build up. By the timethe /1/ in the input is coming in (first panelof the figure) these two phonemes havereached an activation of about 0. , and arebeginning to excite units at the word levelthough none have exceeded threshold (ac-tivation of 0) at this point. As the speech

0.45+

'"(j...

0;=

r::

:.:;(\j::.:.:;"'" (j)

0.4

(j)

r::

..c:Il.

input continues to unfold , detectors forother phonemes become activated. Asphoneme level activations build up, the ac-tive phonemes begin to excite units at theword level. At first , all words beginningwith /b/ and /p/ receive equal excitation; asthe /ll in the input becomes active wordsbeginning in Ibl/ and /pl/ begin to dominateother words, then those beginning in Ibr/and /pr/ (including " blood

" "

blushplug, " and "plum ) begin to win out , as

is illustrated in the second panel of thefigure. When the final /g/ becomes active atthe phoneme level (third panel), /prg/ winsout over the others because it receives themost bottom-up support. Up to this pointboth the /b/ and the /p/ have been receivingtop-down activation from the active unitsthat contain them at the word level , but as/prg/ gradually establishes its dominance atthe word level , it supports the /p/ at thephoneme level and inhibits all other wordunits, thereby removing the top- downsupport for the /b/. The result is eventualdominance of /p/ over /b/ (final panel of the

IRmI ~

g 1

ill

I ~

FIG. !. The time course of the build up of strength of the /p/ response based on activations ofphoneme units in Slice 12 , in processing an ambiguous /b/-/p/ segment followed by Irg/ (as in "plugand the same segment followed irS? (as in "blush"). The ambiguous segment is indicated by the "Also shown is the build up of response strength for an unambiguous /p/ segment in /pl' g/. The verticalline topped with " ?" indicates the point in processing corresponding to the center of the initial seg-ment in the input stream. Successive vertical lines indicate centers of successive phonemes. (FromThe TRACE model of Speech Perception , by J. L. McClelland & J. L. Elman, 1986 Cognitive Psy-chology, 18, 1-86. Reprinted by permission.

COGNITIVE PENETRATION OF PERCEPTUAL MECHANISMS 147

figure). The effect is not terribly dramaticin this case, in large part because the con-text does not come in until after the ambig-uous segment; and the effect does not be-come apparent until after the final phonemehas been heard.

Like other models , TRACE does requirea readout process to translate activations ofunits into overt responses. TRACE' s as-sumptions about the readout process areshared with the interactive activationmodel , and derive ultimately from thechoice model of Luce (1959; see McClel-land & Elman, 1986 , for more details). Forpresent purposes, it is sufficient to notethat the readout assumptions of TRACEand the interactive-activation model allowthe models to approximate the quantitativepredictions of models in which informationfrom lexical and featural inputs is inte-grated in a postperceptual decision mecha-nism of the kind described by Massaro(1979, 1988).

We have examined this example in somedetail , not only to give the reader a sense ofthe course of events in TRACE , but also tobring out two features of its operation thatare relevant to the issue of testing theclaim , inherent in the model , that there isfeedback from the word level to thephoneme level. These features are asfollows: (1) Top-down input influencesprocessing by combining additively withbottom-up input in determining the totalor what is often called the net input to the

unit. Thus activation from the word level istreated as just another source of excitatoryinput , indistinguishable from bottom-upinput in kind. (2) The top- down informa-tion process influences the activation of aphoneme unit takes time. This is particu-larly true when the context that determinesthe identity of the initial phoneme comes

I The correspondence to Massaro s model is notexact. It can be quite close in idealized cases- indeedthe mathematical formulations are closely related-but interactions among units between and withinlevels tend to distort the ideal somewhat.

after the phoneme , but even when the con-text comes before, it can take a while forthe effects of this context to percolate up tothe word level and back down ' again (seeMcClelland , 1987).

With these two points in mind , we canevaluate two different sorts of evidencethat have been offered for assessingwhether an effect is truly top-down or not.One of these is based on signal detectiontheory (SDT; Green and Swets, 1966).Within the context of signal detection

theory, one may ask , does context affectsensitivity (d' or does it simply bias re-sponses in one direction or another

without altering the inherent perceptibilityof speech sounds? At first glance , it wouldseem that if context only influences biasthat is , the r3 parameter of SDT, then con-text is simply operating at a response selec-tion or decision stage. It is true that

r3 canbe affected by influences operating on theprocess of deciding what overt response tomake based on the results of perceptualprocesses. However, it should be notedthat in TRACE , the top-down effects we

. have been examining show up primarily inthe bias measure , rather than in eventhough they operate within the processing

system itself, before the response stage.The reason is that an active word unitsimply adds its contribution to the net inputof a phoneme unit that it supports. The ef-fect is a true top-down effect , in that con-text is influencing the very same detectorsthat are influenced by bottom-up input , ina way that is indistinguishable from theway in which perceptual input influencesthese detectors. Yet the result does not or-dinarily involve any change in sensitivity,but simply pushes the subject along the

likelihood dimension in the direction ofchoosing the contextually appropriate re-sponse.

Based on the foregoing, it should be ap-parent that it is completely consistent withthe assumptions of the TRACE model thatcontextual cues combine with bottom-upcues to phoneme identity in just the same

148 ELMAN AND MC CLELLAND

way that bottom-up cues combine witheach other (Massaro, 1979, 1988). Essen-tially, each cue can be seen as contributingto the tqtal number of votes for one candi-date response or another. In some models,such as Massaro , these votes are com-bined in a decision process that reads outthe results of processing. In TRACE , theyare combined within the processing mecha-nism itself. Indeed , it is possible to viewTRACE and related interactive models asincorporating into the internal workings ofthe perceptual mechanism the fundamentaldecisional processes described by Massaroand others.

Thus far we have seen that TRACE isconsistent with evidence that lexical con-text often operates like other cues to

phoneme identity, producing an effectwhich can show up as a bias effect in asignal detection analysis. However, thisleaves us no better off than we were beforesince it indicates that one cannot use theabsence of a d' effect to argue against aninteractive position.

Another line of argument that has beenoffered in support of a postperceptual , re-sponse-stage view of context effects inphoneme perception is the fact that theyoften take time to build up. Fox (1982)demonstrated that subjects who wereforced to respond within 500 ms of theonset of an ambiguous phoneme occurring

2 For completeness we note that there are circum-

stances in which context does produce sensitivity ef-fects; the word superiority effect (Reicher, 1969) is acase in point, and Samuel (1981 , 1986) has exploredcases of sensitivity effects in speech perception. Adiscussion of the conditions under which these effectsoccur, and how they may be accounted for in interac-tive-activation models , is beyond the scope of thepresent paper (but see McClelland & Rumelhart, 1981for a full account for the Reicher effect in the interac-tive-activation model of visual word processing). Suf-fice it to say that sensitivity effects cannot be taken asunambiguous evidence of top-down effects. For ex-amples of models that account for sensitivity effects ofcontext on letter identification in the Reicher taskwithout invoking feedback from higher levels to lowerlevels, see Johnston and McClelland (1980), andThompson and Massaro (1973).

at the beginning of a word, as in our I?rg/example above , did not show a context ef-fect; that is, they did not show a bias infavor of the phoneme that makes a word inthe context. A context effect occurred onlywhen subjects were not forced to respondunder time pressure. On the basis of thesefindings, Fox argued that the effect was alate , postperceptual one. We have alreadyseen , however, that in TRACE , the effectcan be slow to develop, without being anydifferent in kind from the initial bottom-upactivation process. Of course, this viewpredicts that context effects will operatemore quickly for word-final than word-ini-tial ambiguities , and though Fox s experi-

ment has not been repeated exactly for thiscase, this prediction has been confirmed ina slightly different paradigm (Marslen-Wilson, 1980).

We see , then , that previous evidencetaken in support of models in which the ef-fects of context operate on decision mecha-nisms does not in fact constitute evidencethat is inconsistent with the TRACEmodel. In fact , these previous experimentsunderscore what we take to be a positivefeature of TRACE and other interactivemodels; these models embed the informa-tion integration processes traditionallytaken as the hallmark of postperceptual de-cision mechanisms into the perceptual ma-chinery itself. However, we are still leftwithout a definitive test that distinguishesthe interactive models from accounts inwhich context effects arise in a postpercep-tual decision mechanism operating on theoutputs of a strictly bottom-up perceptualprocessing system.

However, there is one more aspect ofTRACE which we have not yet mentionedthat provides a basis for finding just such atest. This is the mechanism that TRACEuses to deal with coarticulatory influences

in speech. We first describe the coarticula-tory influences , and the mechanismTRACE provides to accommodate them.Then we consider how these influences canbe used to test TRACE' s assumption that

COGNITIVE PENETRATION OF PERCEPTUAL MECHANISMS

higher levels to indeed feed activation backto lower levels.

Compensation for Coarticulation

Coarticulatory influences in speech re-sult from vocal tract dynamics and are amajor source of variability in the speechsignal. It is well known that listeners com-pensate perceptually for the coarticulatoryinfluences of one phoneme on the produc-tion-and hence the acoustic realization-of its neighbors (Repp & Liberman, 1984).

Indeed , adjustment to such coarticulatoryinfluences is thought to be among the mostessential tasks of the early stages of pro-

cessing in speech perception (Fowler, 1985;Liberman , Cooper, Shankweiler, & Stud-dert-Kennedy, 1967).

One example of this is the effect of thephonemes Isl and Isl on the pronunciationand perception of neighboring stop conso-

nants, particularly Id/, It/, Ig/ , and Ik/.There are two sorts of coarticulatory ef-fects. First , the phoneme Is/ , the " sh"sound in ship, is produced with a roundingof the lips , which causes an elongation ofthe vocal tract that may persist throughneighboring phonemes. Thus in saying

foolish dancer" the rounding of the lipsextends from the Isl into the following Idland lre/, thereby coloring the sound that isproduced in pronouncing these phonemes.In contrast , the phoneme Isl is produced byretracting the lips , thereby shortening thevocal tract; this retraction may persist intoneighboring phonemes as well. Thus the Idlin "fearless dancer is produced with ashorter vocal tract than the Idl in " foolishdancer. " The effect of this on the acousticproperties of the Idl is to shift the distribu-tion of acoustic energy into a lower energyrange when the Idl follows Is/ , comparedwith when it follows Is/. A second factor isthat Isl has an alveolar place of articulationwhereas IS! is a palatal sound. As a conse-quence, the front cavity is shorter in thecase of Isl than Isl and its spectrum ishigher. The pronunciation of the alveolarIs/ may pull the place of articulation of

149

nearby velar and , perhaps , alveolar stopsforward , causing them to have higherspectra than in other phonetic environ-

ments (Repp & Mann, 1980).

Listeners compensate for these effectsby adjusting the boundary between pho-netic categories which are distinguished bythe frequency distribution of acoustic en-ergy (Mann & Repp, 1981; Repp & Mann1981). This compensation can be demon-strated in the following way. It is possibleto construct a sequence of sounds thatranges from a It I to a Ik/, or from a Idl to aIg/, by progressively lowering the distribu-tion of acoustic energy contained in thenoise burst which occurs on release of theconsonant. This sequence of stimuli formsa graded continuum of sounds , and there isa boundary between the two percepts. Thecompensation for the coarticulatory influ-ence of the Isl on the following phonemetakes the form of a shift in the perceptualboundary between the It I and Ikl sounds , asillustrated in Fig. 2. The figure shows that asound that is ambiguous between It I and Ik/will be identified as a It I more often whenfollowing a Is/ , but as a Ikl more often whenfollowing a Is/.

In TRACE, we have assumed that thesecoarticulatory influences occur in the basicperceptual mechanisms that identify thephonemes of speech. This assumption isquite generally shared. Such influences are

1.00

8. 0.II)~ 0.

0.70

'" 0.

as 0.

'" 0.

:;; 0.11.= 0.E 0.I/) 0.

after ' SH'

Stimuli

FIG. 2. Effects of a preceding fricative on identifi-cation of stimuli which range perceptually from /t/ to/k/. The graph at the left shows responses when thestimuli are preceded by Is/; the graph at the rightshows responses when the stimuli are preceded by Is/.(Figure adapted from Fig. 1 of Mann & Repp, 1981).

150 ELMAN AND MC CLELLAND

ubiquitous, and it has been suggested thatone of the essential functions of the appa-ratus used for the perception of speech is tofactor out the contextual influences and re-cover the underlying phonetic code(Fowler, 1985; Liberman et aI., 1967).

The TRACE model accounts for theseeffects by allowing active phoneme units tomodulate the connections between the fea-ture level and phoneme level in adjacenttime slices in just such a way as to com-pensate for their coarticulatory influences.That is , to account for the specific effectmentioned above , the units for Isl and Islwould modify the strength of the connec-tions between the feature units that feedinto the phoneme units for Idl and Ig/ , aswell as It I and Ik/, in earlier time slices to

undo" the coarticulatory effect that Island Isl had on the stops. These modulatoryconnections provide the model with a pow-erful tool for adjusting perception in accordwith context , and improve its performancerelative to the case when the connectionsare disabled , as we have demonstrated inapplications of the model to the perceptionof real speech (Elman & McClelland1986).

In the past , we have studied the behaviorof TRACE by always testing the modelunder conditions where the compensationfor co articulation was triggered by the un-ambiguous presence of some phonemewhich was known to have coarticulatoryeffects on adjacent sounds. However, inTRACE , phonemes receive input not onlyfrom the acoustic/phonetic level , but alsofrom word units. Thus it is possible that ac-tivations at the phoneme level might triggercompensation for coarticulation whetherthey are purely determined by bottom-upinput from the feature level or by top-down influences from the word to thephoneme level in addition to these bottom-up influences.

Triggering Inter/evel Effects viaIntraleve! Influences

It is this possibility within TRACE for

top-down processing to have indirect ef-fects on lower levels which suggests a wayto answer the question: Are top-down ef-fects real or only apparent? The logic is totest for the presence of a top-down effectnot by seeing whether a higher level canalter an overt decision about a lower level,but , rather, by seeing whether the higherlevel can actually reach into the lower leveland affect some intralevel process- in thiscase , compensation for coarticulation.

The following two simulations fromTRACE illustrate the model' s predictionthat such effects will occur. In the first sim-ulation , the model was given mock-speechstimuli consisting of the word "abolish" or

progress " followed by one of sevensounds which formed a continuum from anunambiguous Igl at one extreme to an un-ambiguous Idl at the other. The model lo-cated a perceptual boundary roughly in themiddle; however, the precise location ofthis boundary varied depending on whetherthe preceding input was " abolish" or

progress. " The boundary was shifted inaccordance with TRACE' s compensatorymechanism, as shown in Fig. 3a. In thesecond simulation, similar inputs were pre-sented; however, the final fricative fromboth context words (the (s) in "abolish"and the (s) in "progress ) was deleted andreplaced in both cases with an input whichwas exactly half-way between the twosounds. As can be seen in Fig. 3b, essen-tially the same boundary shift occurs. Thisis because the lexical information in the

context word causes the model to perceivethe ambiguous fricative in the appropriateway, so that the compensatory processoccurs even though the bottom-up inputhas not specified the triggering phoneme.

It is important to note that the effect isweaker for perceptually restored phonemesthan it is for phonemes that are actuallypresent. The reason for this in the simula-tion is that both Isl and Isl are partially acti-vated when the input is ambiguous due tothe balanced bottom - up support for eachphoneme; the top-down effect strengthens

Q) 1.00

g 0.g. 0.

~ 0.~ 0.Q) 0.~ 0.

....

I: 0.

0 0.

...

/f. 0.

Q) 1.00

I: 0.

a. 0.

Q) 0.

...

. 0..:s:.~ 0.~ 0.~ 0.Q) 0.

Q) 0.D.. 0.

COGNITIVE PENETRATION OF PERCEPTUAL MECHANISMS 151

after PROGRESS

Stimuli

after PROGRE?

StimuliFIG. 3. Results of a computer simulation of the

TRACE model. In a, the model identified sounds

ranging from to k, preceded either by "abolish" (leftcurve) or "progress " (right curve). In b , the modelidentified the same stimuli as shown in a; however, thefinal sound of the preceding the context word was re-placed by a sound intermediate between Isl and Isl(represented as "

the activation of the contextually appro-

priate phoneme over the other, but thisdoes not eliminate the activation of theother entirely. Compensation for coarticu-lation is graded in the model (see Elmanand McClelland , 1986 , for details), and sothere is greater compensation when lexicaland acoustic information are consistent.

. In the following series of experiments wereport the results of our test of the model'spredictions. As with the simulations , thestrategy was to test the status of the ap-parent top-down interaction by seeingwhether it can trigger a second processwhich operates wholly within the lower(phoneme) level. If these lexical influencesactually feed back information to the basicperceptual mechanisms that interpretspeech sounds, then they ought to induce

co articulatory compensation. On the otherhand , if these lexical influences come onlyat a decision stage at which outputs of theperceptual mechanisms are interpreted,and do not feed back their results into themechanisms responsible for coarticulatorycompensation , we should not see coarticu-latory compensation for phonemes whoseidentity is determined by lexical influences.

The critical point is that the task makes itpossible to decouple the effect of a top-down interaction from the diagnostic usedto detect it. In previous tests of top-downeffects , lexical interactions have been mea-sured through subjects ' responses in-volving the target words themselves or

. phonemes contained within them. Theseresponses might reflect either perceptual orpostperceptual decisions , and so do notprovide reliable evidence for the interac-tion. In the experiments reported here , thetest of lexical interaction did not require aresponse involving the lexical item itself.Instead , the test was to see whether thelexical item could trigger a perceptual ef-fect in an adjacent word. Because the re-sponse involved an unrelated word , andbecause the effect (compensation for coar-ticulation) is uncontroversially perceptualwe believe the test rules out a postpercep-tual decision interpretation. This was thelogic for the following set of experiments.

EXPERIMENT 1

We had two goals for Experiment First, we wanted to replicate the phenom-enon initially reported by Mann and Repp(1981), namely, that listeners perceptuallycompensate for the coarticulatory effect ofIsl or Isl on a following sound. Second , wewanted to see whether the same effectcould be obtained , but with stimuli inwhich the identity of the Isl of Isl was de-termined by lexical influences. Accord-ingly, we constructed stimulus pairs inwhich a context word (e.

g., "

Spanish" orridiculous ) was presented either intact

or with its final sound replaced by the sameintermediate sound, which was chosen to

152 ELMAN AND MC CLELLAND

be perceptually intermediate between Island Is/. From the TRACE model, we pre-dicted that the word-final fricative in thecontext word would bias the perception ofa subsequent ambiguous It/-/kl or Id/-/gldiscrimination, whether the identity of thatfricative were determined either by lexicaland acoustic influences (in the intact condi-tion) or by lexical influences alone (in thereplaced condition).

Method

Stimuli. We began by preparing threesets of target stimuli for the It/-/k/ or Id/-/gldiscriminations. One set of stimuli variedperceptually from " tapes" to "capes, " an-other varied from " dates" to "gates " anda third from "deer" to "gear. " The con-tinua were created by first carrying out anacoustic analysis of naturally producedtokens of " tapes,

" "

capes

" "

dates,gates,

" "

deer, " and "gear. " A computerprogram was used to edit these tokens inorder to produce seven stimuli whichvaried in equal acoustic increments from It Ito Ikl and Idl to Ig/. The parameters whichwere manipulated were (a) the frequencyand amplitude of the noise burst followingrelease of the stop consonant: (b) the onsetfrequency and trajectory of formant transi-tions; (c) the formant transition durations;and (d) the voice onset time of the stops

since this is known to covary with place ofarticulation. It is important to note that thestep size-both acoustic and perceptual-between stimuli was deliberately madesmall. Even the first and last stimuli in eachcontinuum were not unambiguous exem-

plars of their categories. In each case , eachof the seven stimuli could be preceded ei-ther by an intact or by an altered version ofa context word. For the intact conditions,the stimuli on the " tapes capes" con-tinuum were preceded by "Christmas" or

foolish" ; the " dates

" -

" gates stimuliwere preceded by "copious " or "En-glish" ; and the "dear gear stimuliwere preceded by " Spanish" or " ridicu-lous. " For the replaced conditions , the

final sounds of these context words werereplaced by sounds that were perceptuallyintermediate between Isl and Is/. These re-placement sounds were constructed sepa-rately for each of the three continua by dig-itally excising the final Isl and Isl from theoriginal naturally produced tokens. Thesesounds were modified with a computer pro-gram to create a continuum of soundsranging from Isl to Is/; then a separategroup of subjects was asked to identify thesounds. That sound which was closest tobeing identified as Isl 50% of the time waschosen as the replacement sound, to be

used in place of the final sound from bothof the context words used with each of thetarget continua.

Thus , instead of hearing, for example,English (d/g)ates " or " copious

(d/g)ates " subjects in the replaced condi-

tion heard "EngliX (d/g)ates " or "copiouX(d/g)ates " in which both the final sound ofthe context word (here shown with an X)and the initial sound of "dates gateswere ambiguous.

Procedure. Three separate groups of 1 0

subjects were used in the experiment, witheach group of subjects being tested ontarget stimuli from a different target con-tinuum. Each subject participated in boththe intact and the replaced versions of theexperiment , with order of condition coun-terbalanced across subjects. In a given

condition , a particular subject heard eachof the seven stimuli on the target con-tinuum 20 times , 10 times with one or thetwo context strings (e.

, "

English" orEngliX") and 10 times with the other

(e.

, "

copious " or copiouX"

).

Thetarget words followed the context wordswithout any intervening pause, as though

they had been pronounced together.At the beginning of the experiment , sub-

jects were told that they would hear a se-quence of word pairs , and that they shouldtry to identify the initial sound of thesecond word. Responses were indicated bypressing the appropriate key on the key-board. The alternatives allowed were re-

COGNITIVE PENETRATION OF PERCEPTUAL MECHANISMS 153

stricted to the two endpoints of the appro-priate continuum , either /t/ or /k/, or /d/ or/g/, as appropriate. Subjects were told thatthe first sound of the second word wasoften not pronounced clearly, and that theyshould do their best to indicate which ofthe two alternatives it sounded most like,even if they felt they were guessing. A re-sponse was required on every trial. The in-terval between trials was approximately 4s. The instructions also indicated to thesubjects that none of the word pairs madesense, and that they should not try to thinkof them as sensible phrases. The word pairswere described simply as pairs of unrelatedwords.

Before beginning the experiment propereach subject received a sequence of prac-tice trials with a different continuum thanthat used for the experiment proper for thatsubject. Following this practice , each sub-ject received the 140 trials from either theintact or replaced condition; the other con-dition was run in a second session, on thenext day or shortly thereafter.

Subjects. Thirty undergraduates at theUniversity of California, San Diego, servedas subjects. All were native speakers of En-glish with no known speech or hearing dis-orders.

Results

Our analysis of the results looks first tosee if a reliable context effect occurred ineach of the two conditions. Following thiswe consider the relative size of the contexteffect in the intact and replaced conditionsand check for possible order effects.

The results for the intact conditions ofExperiment 1 are shown separately foreach target continuum in Fig. 4. Each panelof the figure displays two labeling func-tions. In one case, the percentage of /k/ (or/g/) responses is shown when the targetstimuli were preceded by a word ending inIs/; in the second case, the target stimuliwere preceded by Is/. As expected, thecurves rise from left to right , reflecting thefact that the target stimuli vary along a /t/-

StimuliFIG. 4. Identification curves for three sets of exper-

imental stimuli using intact context words. In astimuli ranging from " tapes" to "capes" were pre-ceded by the word " Christmas (left curve) or

foolish" (right curve); in b , stimuli ranging fromdeer" to "gear" were preceded by the word "co-

pious" (left curve) or "English" (right curve); in c,stimuli ranging from "dates" to "gates" were pre-ceded by the word " ridiculous (left curve) or

Spanish" (right curve).

c::

g. 0.

....

~ 0.c::

....

1.00

en 0.c::00.~ 0."- 0.

01 0.

Q) 0.

~ 0.c::Q) 0.

Qj 0.a. 0.

1.00

en 0.c::

&. 0.

~ 0.

01 0.

Q) 0.

(\! 0.

~ 0.e 0.

a. 0.

;,

Stimuli

after ENGLISH

Stimuli

after RIDICULOUS

/k/ or /d/-/g/ continuum , as intended. Themain effect of the target stimulus washighly significant in all three cases, with

-:::: .

001 (F(6 54) = 82. 881) for "dates/gates,

-:::: .

001 (F(6, 54) = 87. 172) for

deer/gear " and

-:::: .

001 (F(6, 54)

154 ELMAN AND MC CLELLAND

95.629) for "tapes/capes. " Of central im-portance here is the fact that the precedingword affected perception of the followingtarget sounds as follows: When a soundwhich is ambiguous between /t/ and /k/ (or/d/ and /gl) is immediately preceded by Is/it is more likely to be perceived as a /t/ (or/d/). When the ambiguous sound is pre-ceded by Is/, it is more likely to be heard asa /k/ (or /gl). The responses were subjectedto an ANOVA, and the two conditions

target following /s/ vs following Is/, weresignificantly different , with

-:::: .

001 (F(1,69.456) for "dates/gates

-:::: .

001(F(1, 9) = 26. 938), for "deer/gear " andp-::::

005 (F(1, 9) = 13. 654), for " tapes/capes.Note that the result of this shift in bound-

aries is to undo the coarticulatory effectthat /s/ and /s/ ordinarily have on the fol-lowing stop consonant. Since /s/ makes afollowing /k/ sound more It/-like , the per-

ceptual mechanisms compensate for this,and move the boundary between /k/ and /t/.In our experiments , the stimuli on the /t/~/k/ continuum were originally produced inisolation; the shift in the boundary reflectswhat would have been an appropriate com-pensation for coarticulation , had the /t/-/k/stimuli actually been produced followingthe /s/ Of /s/ context.

Having validated our methods by veri-fying that listeners do indeed compensatefor coarticulation , we are now in a positionto ask whether or not this phonologicalcompensation process can be triggered byinformation from the lexical level. Thisquestion can be addressed by consideringthe results obtained from the replaced con-ditions of the experiment. These results,shown graphically in Fig. 5 , indicate that

perceptual compensation does occur. As inFig. 4 , there are two identification func-tions for each of the three /t/-/k/ or /d/-

/g/

continua. One curve shows the percentageof (or

g)

identifications when the target ispreceded by a context word which origi-nally ended in /s/, and the second curveshows the responses when the target is pre-

ceded by a context word which originallyended in Is/.

The results are striking. Even thoughboth context words end in a sound which isphysically identical , listeners perceive thefollowing sounds as if they were compen-sating for coarticulatory influences of thecontextually appropriate interpretation

the preceding sound. An analysis of vari-ance revealed that the two conditions

(target following a context word whichoriginally ended with /s/ vs /S/) were signifi-cantly different , with

-:::: .

001 (F(1, 9)

42.882) for "dates/gates

-:::: .

04 (F(1,= 6. 191) for "deer/gear, " and

-:::: .

(F(1, 9) 8.442) for " tapes/capes. " Onceagain the main effect of the target stimuluswas overwhelmingly reliable in all cases,with

-:::: .

001 (F(6 54) 97.407) fordates/gates,

-:::: .

001 (F(6, 54) = 72.866)for "deer/gear " and

-:::: .

001 (F(6 54)

129.932) for " tapes/capes.Additional analyses of variance were

carried out for each continuum to deter-mine whether the effect of context was sig-nificantly greater in the intact conditionthan in the replaced condition (as expectedfrom the simulation results) and to deter-mine whether there were any effects ' oforder of the conditions on subjects ' perfor-mance. Table 1 indicates that the effect waslarger in the intact than the replaced condi-tion for all three analyses. However, theIntact-Replaced x Context word interac-tion was only significant for "dates/gates,

= .

038 (F(1, 9) = 5.878), withp

= .

149(F(1, 9) 2.486) for "dear/gear " and

312 (F(1, 9) 1.145) for " tapes/capes.There were no significant effects for orderof presentation of the two conditions or sig-nificant interactions involving this factor.

Discussion

Experiment 1 reveals that a compensa-tory adjustment in the phoneme boundaryfor a target phoneme did occur, based onthe lexically determined identity of the pre-ceding phcmeme. These results suggest that

OJ 1.00

g. 0.

~ 0.~ 0.. 0.

01 0.

(\!

C 0.~ 0.

...

OJ 0.a. 0.

OJ 1.~ 0.&. 0.

~ 0.

...

" 0.

. 0.

~0.~ 0.c::OJ 0.

Qj 0.a. 0.

OJ 1.00

g.

01

OJ

c::

...

. COGNITIVE PENETRATION OF PERCEPTUAL MECHANISMS 155

LoLl?after XMA?

Stimuli

after COPIOU?

Stimuli

after RIDICULOU?

StimuliFIG. 5. Identification curves for three sets of exper-

imental stimuli preceded by context stimuli in whichthe final sound Isl or Isl, was replaced by a sound

intermediate between Isl and Is/; this intermediate

sound is represented as a . In a , stimuli rangingfrom " tapes " to "capes" were preceded by the word

Christma?" (left curve) or "fooli?" (right curve); in, stimuli ranging from "deer" to "gear" were pre-

ceded by the word "copiou?" (left curve) or "Engli?"(right curve); in c , stimuli ranging from "dates" to

gates" were preceded by the word " ridiculou?" (leftcurve) or "Spani?" (right curve).

there are indeed top-down effects on pho-nemic processing. However, there areother possible explanations for these re-sults which would be compatible with abottom-up-only flow of information. Onepossibility is that the perceptual effectarose from acoustic information embeddedin earlier portions of the context stimuli.This might have occurred because thevowel just preceding the ambiguous (s/s)stimuli differed slightly in the two contextwords (e.

, "

Spanish" vs " ridiculousPerhaps the compensation was for long-dis-tance coarticulation between these vowelsand the initial stops in the target words. Al-ternatively, the vowels themselves mightnot have triggered coarticulatory compen-sation in the /d/-/g/ discrimination, but theymay have contained coarticulatory cues tothe identity to the following fricative whichcould have influenced the way in whichthat sound was perceived. These two alter-native accounts share the prediction thatthe context leading up to the vowel shouldnot be necessary to trigger the compensa-tory changes in the target discrimination.We tested this possibility in Experiment 2.

EXPERIMENT 2

This experiment examined the effects ofthe vowels used in the contexts associated

TABLE IEFFECT OF CONTEXT ON PROBABILITY OF Igl

Ik/ RESPONSES

Context condition

Endsin Isl

Endsin Isl Difference

0.410.45

0.500.53

dates-gatesIntactReplaced

dear-gearIntactReplaced

tapes-capesIntactReplaced 0.49

156 ELMAN AND MC CLELLAND

with the " (d/g)ates" targets. In one condi-tion , the context consisted simply of thevowels from the last syllables of the wordsSpanish" and "ridiculous . In another

the vowels were presented together withthe neutralized (s/s) fricative from the re-placed condition of Experiment 1. In bothconditions, the vowel is the only possiblesource of information that differs betweenthe two contexts. If the vowel is indeedtriggering the compensation in the /d/-

/g/

discrimination, then the effect should showup in one or both of these conditions.

Method

Stimuli. The target stimuli for both con-ditions of the experiment were the sevenitems varying from "dates" to "gates " that

were used in Experiment 1. For the vowel-only condition , the context stimuli were theexcised vowels from the "Spanish" and

ridiculous " context stimuli used in Ex-periment 1. The vowels were digitally ex-cised from the stimuli used in Experiment and consisted of that portion of the wave-form in which none of the surrounding con-sonantal information could be perceived.

Each vowel was preceded by a silent inter-trial interval of 4 s, and followed by a silentinterval of 0.250 s, after which the targetstimulus occurred. For the vowel-fricativecondition , the same vowels were used, buteach was followed by the neutralized (s/s)segment used in the replaced conditionswith "Spanish" and " ridiculous" in Ex-periment 1. The CV context was followedimmediately by the target stimulus.

Subjects. Eighteen subjects from thesame sources as Experiment 1 were used inExperiment 2. None had participated inExperiment 1.

Procedure. Nine subjects were assignedto the vowel-only condition and nine wereassigned to the vowel-fricative condition.

Each group received practice with differentstimuli of the same type as those used inthe main experimental trials. The proce-dure was identical to Experiment 1, except

that (a) the context stimuli were different asdescribed above; (b) the instructions weremodified to reflect this difference; and (c)each subject was run in a single sessiononly since each contributed data only toone condition.

Results

N eigher condition of Experiment 2 pro-duced a reliable effect of context (see Fig.6). In the vowel-alone condition , the ex-cised vowels ((I) from "Spanish" and ((ill)from " ridiculous ) had no detectable effecton perception of the initial consonant in thetarget

(p ? .

697 F(1, 8) = 0. 163). In thevowel-fricative condition , there was no re-liable difference when these vowels werecombined with the neutralized fricativeused with "Spanish" and " ridiculous" inthe replaced condition of Experiment

? .

193, F(1, 8) = 2.022). However, as Fig.6 indicates , there may be a nonsignificanttrend in the direction we would expect ifthe vowel-fricative pair were triggeringcoarticulatory compensation.

Discussion

Although we found no reliable effect dueto the excised vowel in Experiment 2 , therewas a trend in the vowel-fricative conditiontoward an effect of the excised vowel. Suchan effect , if reliable , would have indicatedthat the vowel portion of the speech streamactually did contain information that in-fluenced the perception of the subsequent

/d/-/g/ segment. Since such informationwould be part of the acoustic input fromthe context stimuli used to induce the /d/-/g/ boundary adjustment in Experiment 1, itwould represent a confounding of anacoustic cue with what was intended as apurely lexical influence. Although the ef-fect of the vowel was not reliable in eithercondition , we felt that there was enough ofa suggestion of such an effect that it wouldbe important to try to find a way to isolatethe lexical effect.

COGNITIVE PENETRATION OF PERCEPTUAL MECHANISMS

Q) 1.

~ 0.0 0.en 0.

... 0.

01 0.after II.

Q) 0.~ 0.'E 0.

0 0.

...

Q) 0.

Stimuli

c::

...

after II.

c::

...

StimuliFIG. 6. Identification curves for "dates/gates

target stimuli when preceded in a by the final vowelsin "Spanish" and "ridiculous " ((I) and ((Q;i), respec-tively); and in b when preceded by these same vowelsplus a neutral fricative (represented by an X).

EXPERIMENT 3

In this experiment , we devised two newsets of context materials , used in two newcontext conditions. In one of these condi-

tions, called the VC-replaced condition , thefinal VC portions of the context words

Spanish" and "ridiculous" were replacedwith an ambiguous version , pretested tosound halfway between (IS) and ((ills), heredesignated CX). In the other of these con-

ditions, called the syllable-replaced condi-tion , the final syllable of the context wordsfoolish" and' ' ridiculous" was replaced

with an ambiguous version , pretested tosound halfway between (lIs) and (l(ills); thissyllable is here designated as (l"X);

Method

Stimuli. The targets for both conditions

157

of the experiment were the seven "dates/gates" items used in Experiments 1 and 2.In the VC-replaced condition, these targetswere preceded by " ridicul(X)" or

Span(X)" . The neutralized (X) stimuluswas selected from a set of candidate VCsegments as the one judged to be closest tohalfway between (IS) and ((ills). Thus , thefinal fricative and the vowel preceding itwere both ambiguous and identical in bothwords. The pronunciation of the vowel inboth these words in normal speech is actu-ally virtually identical anyway, and so thismanipulation resulted in a natural stimulus.This composite VC segment, designated(X), was then appended to the "Span-and "ridicul-" stimuli , yielding the twocontexts.In the syllable7replaced condition , the

words " foolish" and " ridiculous" werechosen to provide the two contexts. Digitalrecordings of these words were edited todelete the final syllables (" \ish" and "lous " respectively). A composite syllablewas created under computer control whichwas intermediate between these two finalsyllables and was , thus, afso ambiguous asfar ' the final fricative. This composite syl-lable , designated (l"X), was then appendedto the "foo-" and "ridicu-" stimuli,yielding the two contexts. These compositestimuli were not as natural sounding as the

SpanCX)" and " ridicul(X)" stimuli, butthey were still quite recognizable as thewords from which they were derived.

Subjects. Twenty subjects from the samesources as before who had not participatedin Experiment 1 or 2 were used in this ex-periment.

Procedure. Ten subjects were used in theVC-replaced condition and 10 more in thesyllable-replaced condition. For eachgroup, each of the two context stimuli ap-peared 10 times in fandom order beforeeach of the seven target " (d/g)ates " stimuliused in previous experiments. Instructionswere identical to those used in Experiment, and the procedure was the same except

158 ELMAN AND MC CLELLAND

that each subject was run in a single ses-sion.

Results

For both conditions of the experimentresults were in accordance with the hy-pothesis that lexical factors can determinethe perception of an ambiguous speechsound in a way that allows it to influencethe perception of other ambiguous sounds.That is, the lexically determined identity ofthe vowel-fricative combination at the endof each context word appeared to producea coarticulatory compensation. The resultsare shown graphically in Fig. 7. In the VC-

c:: 0:90

g. 0.

...

OI 0.

01

c::

...

1.00

c::

0.70

...

c::

...

Stimuli

StimuliFIG. 7. Identification curves for ambiguous

dates gates " stimuli. In a, stimuli were precededby the word "ridicuWX)" (left curve) or " Span(AX)"(right curve), in which the final (AX) corresponded tothe high central vowel (*), followed by a sound inter-mediate between /sl and Is/. In b, stimuli were pre-ceded by the word " ridicu(rX)" (left curve) or

foorrX)" (right curve), in which the final (l'X) corre-sponds to a neutralized syllable beginning with n), fol-lowed by the high central vowel (*), followed by asound intermediate between Isl and Is/.

replaced condition , there was a significantlexical effect in that subjects tended to hear(g) after " ridicul(X)," and (d) after

Span(X)"

(p -:::: .

018 F(1, 9) = 8. 352).Likewise, in the syllable-replaced condi-tion, there was also a significant main ef-fect for context

(p = .

030, F(1, 9) = 6.652)with subjects hearing the ambiguous (d/g)sound as (g) more often following the " ri-dicuWX)" context than after fooWXJ"both cases, there were no significant inter-actions.

Discussion

The results of Experiment 3 appear toestablish that lexical factors can indeed beresponsible for perceptual compensationand support the conclusion that the percep-tual effects are triggered by informationfrom the lexical level. It is worth notingthat the syllable-replaced condition makesit very hard to argue that the source of thecompensation effect is sublexical. Onemight have proposed that simple phoneme-to-phoneme sequential constraints could beincorporated within the phonological level.Possibly, these sequential constraints are

such that they would lead subjects to pre-dict that the final phoneme in " Spanish"was an /s/ but the final phoneme in " ridicu-lous" was an /s/ , quite apart from specificlexical factors; it may be that (nI-) is moreoften completed with (s), while (I?- ) is

more often completed with (s). Howeverin the syllable-replaced condition , the con-text in the replaced items is actually thesame for three phonemes before the finalfricative; the vowel in "foo- " and the

last vowel in "ridicu-" are the samevowel , though they may have slightly dif-ferent acoustic realizations due to coarticu-lation , and the next two sounds in the twocontexts are both acoustically and phoneti-

cally identical in the syllable-replaced

stimuli. Thus , any differential prediction ofthe identity of the final fricative would haveto be based on " " vs "ridic- " and

thus would seem to be attributable to

COGNITIVE PENETRATION OF PERCEPTUAL MECHANISMS

knowledge that is specific to the particularlexical items involved.

The results thus far are compatible withour claim that the shifts we are seeing inthe judgment of ambiguous /d/-/g/ stimuliare due to coarticulatory compensation forthe lexically determined perception of thefinal fricative in the context word. But onefurther alternative interpretation remains tobe ruled out. Possibly, there is a subtle se-mantic bias in our materials that is in-fluencing the results. Perhaps , for example

Spanish gates" seems more plausible tosubjects that" Spanish dates , or perhaps

ridiculous dates" seems more plausiblethan " ridiculous gates ; and perhaps theseplausibility factors are determining sub-jects ' /d/-/g/ choices, rather than any coar-ticulatory adjustment to the word final fri-cative. Although the stimulus set was de-signed so that there would be no tendencyfor the context to actually allow subjects topredict the target words , it is nonethelesspossible that there are subtle but significant

differences in the plausibility of the dif-ferent word pairs that influence subjectsdecisions about the identity of the firstsound in the target word. Experiment 4 wasdesigned to examine this possibility.

EXPERIMENT 4

The two competing explanations for thetendency of subjects to hear sequencessuch as " Spani(s/s) (d/g)ates" and " ridi-culou(s/s) (d/g)ates" as "Spanish gatesand " ridiculous dates" are as follows: Oneis that the lexical level affects perception ofthe ambiguous (s/s) sound so that listenersperceive it in the lexically appropriate way;and then this lexically biased percepttriggers a compensatory perceptual effectin the following sound. The other explana-tion is that our particular choice of contextand target stimuli favored interpretations inwhich , for semantic or pragmatic reasonsthe word containing or (or or k) wasmore appealing or plausible. If this werethe case , then the apparent perceptual bias

159

simply reflects differences in semantic andpragmatic factors , and could be treated as apostperceptual phenomenon.

These explanations compete only in thecase where the stimuli are presented audi-torily. If the stimuli are presented visually,only what we will call the semantic factor ispresent; it is not likely that visual percep-tion would be subject to compensatory per-ceptual mechanisms which owe their exis-tence to facts about human articulation. Tosee if our effects were due to semanticfactors, then, we presented visual analogsof the "dates/gates " target stimuli , pre-ceded by visual versions of the two contextwords, "Spanish" and "ridiculous, " usedin several of the conditions described

above. Two separate groups of 10 subjectswere run with the same stimuli. One groupwas given instructions similar to those usedin the auditory experiments , in which at-tention to the first word was down-playedand the items were described as pairs ofunrelated words. To give any possible ef-fect of semantic and pragmatic factors thegreatest chance of producing a reliable ef-fect , the second group was given instruc-tions which urged subjects to attend toboth words and process the . target stimulias words rather than to focus just on thefirst letter of the target word.

Method

Stimuli and display conditions. Experi-ment 4 consisted of an analog to Experi-ment 1, except that stimulus presentationwas visual rather than auditory. The stimuliwere analogous to the "dates/gates" targetstimuli and " ridiculous " /" Spanish" con-texts used in several of the auditory condi-tions already described. Word pairs weredisplayed on an IBM PC monitor with aspecial low-persistence phosphor. The firstword of each pair was either "SPANISH"or "RIDICULOUS" (in uppercase), andwas presented centered on the displayscreen, following a 500-ms presentation ofthe word READ Y centered at the same 10-

160 ELMAN AND MC CLELLAND

cation. The first word was displayed for175 ms. This word was followed by a 75-msinterval during which an all-white maskwas displayed by turning on all pixels in acentered row of 10 character blocks 7 x 14pixels each. This premask display was fol-lowed by a 25-ms presentation of thesecond target word, in lowercase , also cen-tered in the display area. The target wasfollowed by a 225-ms presentation of the

all-white mask.3 The second word was one

of a set of seven stimuli which ranged vi-sually from "dates" to "gates" (in lower-case). Each consisted of the letters ates,preceded by a specially constructed patternof pixels, consisting of a lowercase andup to three additional pixels which tendedto make the character appear more likeor more like. The particular pixel pat-terns used were selected on the basis ofpilot work to provide a set of closelypacked items giving rise to about the samedensity of steps along a visual con-tinuum as the stimuli used in previous ex-periments did on an auditory continuum.

Generally, successive stimuli along thecontinuum differed by a single pixel. Dueto the pre- and post-masks , the impressionof the second stimulus was rather faint , butthe second stimulus was still visible as evi-denced by the systematic shift from to

responses across the seven stimuli , shownin the results section below.

Subjects. Twenty Carnegie-Mellon un-dergraduates served in the experiment to

fulfill a course requirement. All were nativespeakers of English with normal or cor-rected-to-normal vision.

Procedure. The procedure was set up tobe as analogous as possible to the proce-

3 The all-white mask was intended to serve as a light

mask, reducing contrast of the target. Though a con-toured field would have produced stronger masking,we did not want to introduce possible bias effects offeatures from the mask on identification of the lettersin the target word. The use of the all-white premaskwas intended also to buffer the target from any pos-sible interactions of this kind with the context words.

dures used in Experiments 1 and 3. Eachsubject viewed 30 practice trials followedby 20 trials with each of the seven targetson the "dates-gates" continuum. Ten pre-sentations of each target were preceded byRIDICULOUS and ten were preceded bySPANISH. One group of 10 subjects wastold to focus on identifying the ambiguousletter at the beginning of the second word,and the response buttons were labeledsimply and to encourage focusing of at-tention. These subjects were given the fol-lowing instructions concerning the fact thatthe words occurred in pairs:

None of the word pairs really make sense , andyou should not try to think of them as sensiblephrases. They are simply pairs of unrelatedwords.

These instructions were taken word forword from the instructions used with theauditory stimuli in Experiments 1 and 3.

In order to try to give any subtle se-mantic effect a greater opportunity tooccur, we altered these instructions for thesecond group of 10 subjects , stressing theimportance of attending to both words andtreating them as wholes. The choice alter-natives were labeled "dates" and "gatesin an attempt to encourage subjects to treatthe stimuli as whole words. These subjectswere told that we wanted to determinewhether the first word would influencetheir judgment , and that they must attendto both words for this to occur:

In this experiment , we are specifically interestedin whether the first word of each pair influencesyour judgment of the identity of the second wordof each pair. Thus, we want you to pay attentionto both words of the pair. Your respOnse shouldbe based on your impression of which word, was

shown as the second word; accuracy is not asimportant as your impression , and your impres-sions are more likely to be affected by the firstword if you attend to both words as wholes, sowe would like you to make an effort to identifyboth the first and second words on each trial.

The wording in these instructions down-playing accuracy was inserted after twopreliminary subjects reported that they felt

COGNITIVE PENETRATION OF PERCEPTUAL MECHANISMS

they had to try to ignore the first word inorder to be as accurate as possible in re-

porting the identity of the first letter of thesecond word. Actually, accuracy is notalways greater when subjects adopt a fo-cused strategy (Jonston & McClelland1974).

Results

There was no main effect of context on/d/-

/g/

judgments in either condition of theexperiment. The results are shown in Fig.8. An analysis of variance of the results forthe condition in which subjects were told tofocus on the first letter in the second wordindicates no significant main effect on per-ception of the ambiguous " (d/g)ates" word

..........

Il.

..........

Il.

Stimuli

after SPANISH

after RIDICULOUS

StimuliFIG. 8. Identification curves for ambiguous visual

dates/gates " stimuli , preceded by the context wordsRIDICULOUS and SPANISH , as indicated. In a, re-sults are shown for subjects who were given instruc-tions stressing attention to the first (target) letter ofthe second word. In b, results are for subjects whowere given instructions stressing attention to bothwords as whole words.

161

as a function of the preceding word

(p

11, F(1, 9) = 2.977) and a marginally sig-nificant interaction between the stimulusand context variables

(p = .

038, F(6 54)

2.418). Even with explicit instructions toattend to both words , there was no maineffect of context on the perception of theambiguous letter in the second word. Therewas no main effect of context

(p = .

950,F(1, 9) = 0.003), and no interaction be-tween stimulus and context

(p = .

786F(6 54) 0.526). There appears to be aslight preference for perceiving afterSPANISH , which is opposite of the predic-tion for the auditory case; but this trend isreversed for the most like stimulus.

Discussion

We did not find an effect of the visualcontext words " ridiculous" and "Spanish"on decisions about a visually ambiguousId/-/gl stimulus. What little tendency thereis toward an effect of these words appearsto be in the direction opposite that of the

effect we obtained in the auditory versionsof the experiment; if anything there is atendency to perceive ambiguous

stimuli as more like following SPANISH,rather than more like as we found in theauditory case. Thus, Experiment 4 gives usno reason to suppose that the results wefound in earlier experiments are the resultof semantic.or pragmatic biases in our ma-terials. On the other hand, it must be con-ceded that Experiment 4 is not completelydefinitive. For one thing Experiment 4 restson a null effect , and it is possible that more sensitive method could be devisedthat would pick up a subtle bias effect. An-other caveat is that we have tested for se-mantic/pragmatic effects using visual an-alogs of the stimuli used in the previous ex-periments. We see no reason to suppose

that such an effect would be more potentwith ambiguous auditory stimuli than withambiguous visual stimuli, but the possi-bility remains that it is.

162 ELMAN AND MC CLELLAND

Though it was only reliable in one case,there was a tendency toward an interactionof the context factor with the main effect oftarget stimulus. The most likely interpreta-tion of the interaction is that RIDICU-LOUS, perhaps due to its greater length-reduced attention to the second word morethan SPANISH did , thereby producing aflatter curve. This would make morelikely than at the most like end of thecontinuum, and more likely than at the

like end of the continuum.

GENERAL DISCUSSION

Let us now ask once again what the basisis for the compensatory adjustment in thereplaced conditions of Experiment 1. Thesound preceding the ambiguous /t/~lkI andId/-/gl stimuli was the same , equidistant be-tween Isl and Is/. If compensation werebased only on an analysis of the acoustic

properties of the context, whatever coar-ticulatory effect this ambiguous context hasshould be the same in all cases. Experi-ment 3 rules out the possibility of covertacoustic cues in the vowel which mighthave triggered the adjustment. Apparently,then , what listeners are compensating forcannot solely be defined in terms of theacoustic stimulus.

In Experiment 4 we considered the pos-sibility that the basis for the differing re-sponses in the two contexts was not per-ceptual, but reflected higher-level knowl-edge about the relative likelihood of oneword pair versus another. The results ofthis experiment provided no support forthis possibility.

The most reasonable interpretation forour basic results is that listeners are com-pensating for phonemes whose identity isdetermined by lexical constraints. Al-though the final sound in each of the con-text words is ambiguous, the context wordsare such that only one of the possibilities isconsistent with the word. Thus

, "

English"is a word, but "Englis" is not. The results

of this lexical influence appear to be madeavailable to the parts of the perceptual

mechanism that are responsible for com-pensating for co articulatory influences onspeech perception. This is indicated by thefact that the restored phonemes are able totrigger another perceptual phenomenonwhich is the adjustment of phoneme bound-aries to compensate for coarticulatory ef-fects. It will be noted that this second phe-nomenon is clearly an influence of theidentity of one phoneme /sl or Isl, on an-other It I or Iki. It cannot reasonably be at-tributed to influences at the lexical level,since the Isl or Is/ sound is in a separateword from the It I or Iki. Since the phenom-enon can be triggered by a lexically re-stored phoneme, this strongly suggests thatthe lexical restoration influences the mech-anisms that compensate for coarticulation.

In motivating our experiments , we ar-gued that if there were true top-down ef-fects as assumed in the TRACE modelthen we should expect to see compensationfor coarticulation with lexically restored.phonemes. But let us look at the questionthe other way around: Can the compensa-tion for lexically restored phonemes thatwe have observed in our simulations be ac-counted for without postulating the exis-tence of a true top-down effect? We thinknot.

Our argument hinges on what it meansfor some influence to be a true top-downeffect. To us , it means that the influencegoes back down and affects the same pro-cessing levels that information flows

through bottom-up. Given this , our dem-onstration that lexical factors can trigger

coarticulatory compensation is convincingevidence of a top-down effect , if it is ac-cepted that the coarticulatory compensa-tion actually occurs in the mechanisms thatprovide the phonemic information thatserves as the basis for word recognition.

Is it true that the output of the mecha-nism that compensates for coarticulationdoes indeed serve as the input to word rec-ognition? We think the answer to this ques-

COGNITIVE PENET~TION OF PERCEPTUAL MECHANISMS

tion is yes. If it were no , then compensa-tion would influence phoneme identifica-tion , but would not influence wordidentification. This would lead to the para-doxical result that the identity of the firstphoneme in the input " (t/k)apes" wouldreflect coarticulatory influences of pre-

vious phonemes but the identity of thissame input as a word would not reflectthese influences. More generally, it wouldlead to a nonsensical state of affairs inwhich lexical access did not benefit fromthe exquisite sensitivity of phoneme identi-fication processes to local context.

The situation, then , would appear to bethis: The mechanisms that provide input toword identification appear themselves toreceive input from word identification. In-fluences run top-down , as well as bottom-Up.

As we have already indicated, ourfindings constitute just the sort of evidencethat counts against the hypothesis that themechanisms that perceive the individualspeech sounds are "encapsulated" (to usethe term introduced by Fodor, 1983, pp.65ff) or isolated from the influence ofhigher levels.

Recently, Connine and Clifton (1987)have also described a very different line ofexperimental results that support the viewthat lexical information feeds back activa-tion to the phoneme level. They contrastedthe effects of a lexical manipulation with

effects of a manipulation of payoffs. Theyfound that both lexical factors and payoffmanipulations produced beta-like effectson phoneme identification functions, butthat lexical effects and payoff manipula-tions influenced reaction times in differentways; as predicted from an interactive ac-count , lexical effects influenced reactiontimes to ambiguous stimuli , but not to un-ambiguous stimuli; in contrast , payoff ma-

4 It is logically possible that there are two separatemechanisms that are capable of compensating forcoarticulation , one that is part of the main speech-pro-cessing stream and one that is used for generating

163

nipulations influenced reaction times forunambiguous stimuli, not for ambiguousones. Their interpretation was that the lex-ical manipulation influenced perceptualprocesses, while the payoff manipulation

influenced the decision processes, as as-sumed in the TRACE model.

Neither our results , nor those of Connineand Clifton, rule out the possibility thatlevels of processing below the phonemelevel are informationally encapsulated; justhow far down higher-level factors can pen-etrate remains to be demonstrated throughfurther experimentation. The TRACEmodel assumes that there are bidirectionalconnections throughout the multilevel pro-cessing system that underlies language per-ception and comprehension , and the modelmakes use of feedback from the phonemeto the model's feature level to account forthe phenomenon of categorical perceptionbut we have not yet established conclu-sively that that feedback extends to thefeature level itself.

It should be said , as well , that we haveonly established that lexical informationcan produce feedback into a lower level- ofprocessing. It remains to be seen whetherconstraints arising from even higher levelscan have effects that percolate far enoughdown to trigger co articulatory compensa-tion. We hope that the technique of exam-ining whether perceptual decisions whoseoutcome is influenced by higher level con-siderations can trigger compensatory ad-justments will prove useful in testing for in-

overt phoneme identity responses that do not feedback into lexical access. Our results would be consis-tent with the possibility that only the latter of thesemechanisms is able to make use of information aboutthe identity of context phonemes that has been arrivedat through influences from the lexical level. We havediscounted this argument because it is completely un-motivated and unparsimonious, and it would still leadto the implausible conclusion that the lexically trig-gered compensation for coarticulation that occurred inour experiment influenced the identification of thefirst sound of the target word , but not the identifica-tion of the target word itself.

164 ELMAN AND MC CLELLAND

teractions among other levels. For ex-ample , one could ask whether semanticcontextual influences on phoneme identitycan trigger coarticulatory compensationby using larger contexts in which the finalcontext word is acoustically ambiguous(e.

, "

mesh" mess ). Or one could askwhether such compensation could be trIg-gered by an acoustically ambiguous wordwhose identity was constrained by the syn-tactic structure in which it occurs.Samuel' s - suggestion (1981) that sentence-level context does not reach down to thephoneme level would predict that such atop- down effect would not occur.Samuel's suggestion is consistent with thefindings of Connine (1987); it would beuseful to address this issue using the para-digm developed here.

Whatever the outcome of these futurestudies, we hope the present paper demon-strates the positive role a modeling ap-proach can play in increasing the empiricalbase of psychological and cognitive scienceand in clarifying exactly which experi-mental tests actually bear on central issuesof general importance beyond the specificsof the particular model under study. In thisinstance, thinking within the context of a

specific simulation model made it clearwhy previous attempts to determinewhether contextual influences really dofeed back into the mechanisms of percep-tion have not been conclusive and lead toan experiment that addresses , not only aspecific prediction of this particular modelbut also an issue that distinguishes interac-tive models as a class from noninteractive models. The specifics of the TRACE modelare only one particular form in which thismore general prediction of interactivemodels could have come to our attention.However, the model served a crucial heu-ristic role in that the prediction became ap-parent as a consequence of the configura-tion of assumptions built into the model.

More generally, this paper illustrateshow modeling can uncover predictions that

depend upon the interplay of the assump-tions that underlie the model and can leadus to ask new empirical questions. It willbe the answers to these questions, togetherwith ongoing attempts to build models thatmake sense of these answers, that ulti-mately lead to deeper understanding of themechanisms of speech processing in partic-ular and of cognitive processes in general.

REFERENCES

COLE, R. A. (1973). Listening for mispronunciations:A measure of what we hear during speech. Per-ception and Psychophysics 13, 153- 156.

COLE, R. A. , & JAKIMIK, J. (1978). Understandingspeech: How words are heard. In G. Underwood(Ed.

),

Strategies of information processing

(pp.

149-211). London: Academic Press.COLE , R. A. , & JAKIMIK, J. (1980). A model of

speech perception. In R. A. Cole (Ed.

),

Percep-tion and production of fluent speech

(pp.

133- 163). Hillsdale , NJ: Erlbaum.COLE , R., JAKIMIK, J. , & COOPER, W. E. (1978). Per-

ceptibility of phonetic features in fluent speech.Journal of the Acoustical Society of America,

44-56.CONNINE, C. M. (1987). Constraints on interactive

processes in auditory word recognition: The roleof sentence context. Journal of Memory and Lan-guage, , 527- 538.

CONNINE , C. M. , & CUff ON , C. (1987). Interactiveuse of lexical information in speech perception.

Journal of Experimental Psychology: HumanPerception and Performance, 13, 291-299.

ELMAN, J. L. , & MCCLELLAND , J. L. (1986). Ex-ploiting lawful variability in the speech wave. In1. S. Perkell & D. H. Klatt (Eds.

),

In variance andvariability in speech processes (pp. 360- 385).Hillsdale , NJ: Erlbaum.

FODOR, J. A. (1983). The modularity of the mind: Anessay on faculty psychology. Cambridge , MA:MIT Press.

FORSTER, K. I. (1979). Levels of processing and thestructure of the language processor. In W. E.

Cooper & E. Walker (Eds.

),

Sentence processing:

Psycholinguistic studies presented to Merrill Gar-rett. Hillsdale, NJ: Erlbaum.

FORSTER, K. I. (1981). Priming and the effects of sen-tence and lexical contexts on naming time: Evi-dence for autonomous lexical processing. Quar-terly Journal of Experimental Psychology A, 465-495.

FOWLER, C. A. (1985). Segmentation of coarticulatedspeech in perception (Status Report on Speech

COGNITIVE PENETRATION OF PERCEPTUAL MECHANISMS

Research, Haskins Laboratories, SR-

, pp.

21).Fox, R. Unpublished manuscript. Vanderbilt Univer-

sity, 1982.

GANONG, W. F. (1980). Phonetic categorization in au-ditory word perception. Journal of Experimental

Psychology: Human Perception and Perfor-mance, 6, 110- 125.

GREEN , D. M.

, &

SWETS, J. A. (1966). Signal detec-

tion theory and psychophysics. New York: Wiley.GROSJEAN, F. (1980). Spoken word recognition pro-

cesses and the gating paradigm. Perception andPsychophysics 28, 267- 283.

JOHNSTON , J. C. , & MCCLELLAND , J. L. (1974). Per-ception of letters in words: Seek not and ye shallfind. Science, 184, 1192- 1194.

JOHNSTON , J. c., & MCCLELLAND, J. L. (1980). Ex-perimental tests of a hierarchical model of wordidentification. Journal of Verbal Learning andVerbal Behavior 19, 503- 524.

LIBERMAN, A. M. , COOPER, F. S., SHANKWEILER, & STUDDERT-KENNEDY, M. (1967). Percep-

tion of the speech code. Psychology Review452-471.

LUCE , R. D. (1959). Individual choice behavior. NewYork: Wiley.

MANN , V. A. , & REPP , R H. (1981). Influence of pre-ceding fricative on stop consonant perception.Journal of the Acoustical Society of America 69,548-558.

MARSLEN-WILSON , W. D. (1980). Speech under-standing as a psychological process. In J. C.Simon (Ed.

),

Spoken language generation andunderstanding (pp. 39-67). New York: Reidel.

MARSLEN-WILSON , w., & TYLER, L. K. (1980). Thetemporal structure of spoken language under-

standing. Cognition, , 1-71.MARSLEN-WILSON , W. D. , & WELSH , A. (1978). Pro-

cessing interactions and lexical access during

word recognition in continuous speech. CognitivePsychology, 10, 29-63.

MASSARO , D. W. (1979). Letter information and or-thographic context in word perception. Journal ofExperimental Psychology: Human Perceptionand Performpnce, , 595-609.

MASSARO , D. (1988). Some criticisms of connectionistmodels of human performance. Journal ofMemory and Language, 27, 213- 234.

MCCLELLAND , J. L , & ELMAN , J. L (1986). TheTRACE model of speech perception. CognitivePsychology, 18, 1-86.

165

MCCLELLAND , J. L , & RUMELHART, D. E. (1981).

An interactive activation model of context effectsin letter perception: Part I. An account of basicfindings. Psychological Review, 375-407.

MORTON , J. (1969). Interaction of information in wordrecognition. Psychological Review 76, 165- 178.

NoRRIs , D. G. (1982). Autonomous processes in com-prehension: A reply to Marslen- Wilson and

Tyler. Cognition 11, 97- 101.REICHER, G. M. (1969). Perceptual recognition as a

function of meaningfulness of stimulus material.Journal of Experimental Psychology, 81,274-280.

REPP , B. H.

, &

LIBERMAN, A. M. (1984). Phoneticcategories are flexible (Status Report on SpeechResearch, Haskins Laboratories, SR-77178, pp.31- 53).

REPP , B. H.

, &

MANN, V. A. (1980). Perceptual as-sessment of fricative-stop coarticulation. Journalof the Acoustical Society of America 69,1154- 1163.

RUMELHART, D. E. (1977). Toward an interactivemodel of reading. In S. Domic (Ed.

),

Attentionand performance VI. Hillsdale , NJ: Erlbaum.

RUMELHART, D. E.

, &

MCCLELLAND, J. L. (1981).An interactive activation model of context effectsin letter perception: Part II. The contextual en-hancement effect and some tests and extensionsof the model. Psychological Review in press.

SAMUEL, A. G. (1981). Phonemic restoration: Insightsfrom a new methodology, Journal of ExperimentalPsychology: General.

SAMUEL, A. G. (1986). The lexicon in speech percep-tion. In E. C. Schwab & H. C. Nusbaum (Eds.Pattern recognition by humans and machines:Vol I. Speech perception. Orlando, FL: Aca-demic Press.

TANENHAUS , M. , CARLSON , G., & SEIDENBERG, M.(1984). Do listeners compute linguistic representa-tion? In D. Dowty & A. Zwicky (Eds.

),

Naturallanguage parsing. Cambridge, MA: CambridgeUniv. Press.

THOMPSON, M. C. , & MASSARO, D. W. (1973). Visualinformation and redundancy in reading. Journalof Experimental Psychology, 98, 49-54.

WARREN , R. M. , & SHERMAN , G. (1974). Phonemicrestorations based on subsequent context. Per-ception and Psychophysics. 16, 150- 156.

(Received August 4, 1987)

(Revision received November 9, 1987)