Memory &: CognitionJ974, Vol. 2,No. JA, 130-138

Context effects in sentence comprehension:A study of the subjective lexicon*

CAROL CONRADtUniversity ofOregon, Eugene, Öregon 97403

These studies explore the role of context in determining what information about the meanings ofwords is activatedin memory at the time a word is encountered in a sentence. Using a color-naming paradigm, it was shown that bothmeanings of a word that has two distinct meanings are activated in rnemory at the time the word is heard in a sentence,This activation occurs even when there is sufficient contextual information to indicate which meaning was intended bythe speaker. These results support the hypothesis that there exists in memory an isolable subjective lexicon. Theysuggest that context which is effective in disambiguating lexical ambiguities in the language has its effect only at arelatively late stage in the cognitive processing involved in language comprehension.

In his 1970 presidential address to the AmericanPsychological Association, George Miller posed aproblem to psychologists interested in studying humanlanguage. The problem was basically this: Can we showthat there exists a subjective lexicon which is an isolablesubsystem in memory? The subjective lexicon, in thesense that Miller speaks of, refers to our intrinsicknowledge of the meanings of words. It is a set ofconcepts which have been acquired and refined over arelatively long period of time, a set of dictionary entrieswhich define words, specify their selection restrictions,and perhaps provide inforrnation about their possiblesyntactic roles.

The notion of an isolable subsystem is a complex one:Posner (1973) has presented evidence which suggeststhat a stimulus may have several different codes and thatdifferent classes of mental operations are performedupon these codes. Thus, a written word may have both aphysical code and a name code. It also has a meaning,which may be represented in a lexical systern inmemory. To be isolable, each system or code must haveits own properties (see Katz & Fodor, 1963, for an earlytheoretical atternpt at defining the properties of a lexicalsystem) and each must make a separable contribution toa complex task such as language comprehension. Oneway of establishing experimentaIly that a coding systemis isolable is to show that manipulating the availability ofone coding system does not affect the other codes forthe same stimulus. An example rnight be the visual and

*This research was based on a dissertation submitted in partialfulfillment of the requirements for the PhD degree at theUniversity of Oregon and was conducted while the author heldanNSF predoctoral traineeship. Funds for Ss and equipmentwere provided by NSF Grant GB 21020 and by USOE GrantOEG-Q-n-0717 from the U.S. Department of Health, Educationand WeIfare. The author thanks Michael Posner, Ray Hyrnan,Gerald Reicher, and Merrill Garrett for their helpful commentsand criticism on this research.

fRequests for reprints should be sent to Carol Conrad,Department of Psychology, New School for Social Research, 66West 12th Street, New York, New York 10011.

auditory systems; brain damage to one of these systemsmay not affect the other.

To answer Miller's question, it seems necessary todemonstrate at least two things. First, to show that alexicon exists, it must be demonstrated that, at somelevel of sentence comprehension, Ss activate in memorythe meanings of the individual words in a sentence.Second, to show that the lexiconis an isolable system, itmust be demonstrated that the lexicallook-up process isnot influenced by either the linguistic or nonlinguisticcontext provided by the sentence or the environment orby the listener's more general knowledge of the world.

The experiments reported here represent an attemptto demonstrate that there does exist a lexical look-upprocess which is not influenced by linguistic context.The methodology used is a variation on a technique usedby Warren (1972). Warren aurally presented Ss withthree words which all had the same category name (e.g.,robin, canary, sparrow); this was followed by the visualpresentation of a single word which was in the list (e.g.,robin), the category common to aIl of the words in thelist (e.g., bird), or an unrelated word. The words wereprinted in colored ink, and Ss were required to name thecolor of the word as quickly as possible after the wordappeared.

Color-naming RTs showed a significant interferenceeffect for the category names as weIl as for the words inthe list. Warren concluded from this that the categorynames for words are activated in memory along with therepresentation of the word itself at the time that thewords are heard.

Warren's technique provides one way of studying theactivation of meanings for the individual words in asentence. By using sentences instead of word lists, it ispossible to manipulate the linguistic context for a wordso that the activation of the meanings for that word canbe studied under varying contextual conditions. Theexperiments reported here use contextual constraints onambiguous words in order to test the notion that thereexists an isolable lexicon which plays a role in sentencecomprehension.

130

CONTEXT EFFECTS IN SENTENCE COMPREHENSION 131

Recently, a number of people have tried to takeadvantage of ambiguities in the language to studysentence comprehension. Methodologies have includedsentence completion tasks, probe techniques to measureinterference, picture-sentence :natching tasks whichprovide reaction-time measures, and dichotic listeningtasks. They have been used to study the role of contextin disambiguating sentences and to determine how theprocessing demands of ambiguous sentences vary as afunction of context. Several of these studies have shownthat the probability of a particular interpretation of anambiguous sentence can be influenced by the form ormeaning of the sentences or words which precede it(Marshall, 1965; Carey, Mehler, & Bever, 1970; Lackner& Garrett, 1973; MacKay, 1973). These studies indicatethat the interpretation of an ambiguous sentence may beinfluenced by linguistic context. In addition, arecentstudy by Bransford and Johnson (1973) suggests thateven the comprehension of unambiguous sentences maydepend on having available prior Iinguistic context.Thus, it is clear that context plays an important role inthe cornprehension of language and that any theory ofcomprehension must be able to specify how theprocessing of contextual information contributes to ourunderstanding of the language we hear or read. Thepresent studies evaluate one hypothesis concerning thisprocessing: that prior linguistic context influencessentence comprehension by suppressing the activation ofinappropriate meanings for words in a sentence.

The research on the effects of context on ambiguityarose from the question of whether or not peopleprocess both meanings of an ambiguous sentence. Thisresearch has given rise to several theories of howambiguities are processed , The first of these, the gardenpath theory, was initially suggested by Lashley (1951),who proposed that people process onJy one meaning orstructure of a sentence at a time and that if the originalinterpretation later becornes inappropriate, we thenprocess the other meaning. Lashley offered the followingsentence as an example: "Rapid / rajtin I with hisuninjured hand saved from loss the contents 01' thecapsized canoe." Almost everyone tends to interpret Irajtin I as "writing"; however, the context which followsthe word forces us to reinterpret it as "righting." Thetheory fits nicely with the observation that althoughambiguities are quite frequent, at least in the Englishlanguage, we seldom are aware of them.

An alternative to the garden path theory, theexhaustive computation hypothesis, suggests that allpossible interpretations of an ambiguous sentence arecomputed at an early stage of processing and, at somelater point, a single reading is selected for attention. Amore explicit version of this hypothesis is MacKay'sperceptual suppression theory, which proposes that allof the meanings of an ambiguous sentence are processedin parallel at an unconscious I level. In order to perceiveone of the meanings, the Iother meaning must besuppressed, and the time for suppression dcpends on the

apriori likelihood of the suppressed meaning,independent of context. Context does have an effect,however, by strengthening the activation of theappropriate meaning.

At the present time, the data that might allow us tochoose between these theories are themselvesambiguous. Several experiments argue strongly for thegarden path theory (Foss, Bever, & Silver, 1968; Carey,Mehler, & Bever, 1970), while others suggest that atsome point, both meanings of an ambiguity areprocessed (MacKay, 1966; Fodor, Garrett, & Bever,1968; Lackner & Garrett, 1973) and that processingdemands differ for ambiguous and unambiguoussentences (Foss, 1970; Garrett, 1970). Garrett (1970), inreviewing these studies, has suggested that whether ornot effects of arnbiguity are found depends on whethertesting occurs during the processing involved incomprehending the sentence or after the sentence hasalready been comprehended. The latter measures tend tosupport the garden path theory, while the formermeasures indicate that ambiguous sentences are moredifficult to process than are unambiguous ones.

The study of lexical ambiguities may provideinformation about the subjective lexicon. Bothambiguous and unambiguous sentences can contain alexical item with two or more different meanings. Thetwo types of sentences differ in that the unambiguoussentences contain contextual information which teils thelistener which meaning of the word is appropriate, whilethe ambiguous sentences do not contain this type ofcontextual inforrnation. If there are differences ininterference on the color-naming task for ambiguous andunambiguous sentences which may be attributed tothese contextual differences, then we must concludethat the subjective lexicon is not context-independent. Afailure to find these differences, however, would suggestthat, at some stage, the lexicon is not influenced bycontextuai inforrnation. In addition, a study of this typernay provide inforrnation to help us choose between theseveral existing theories of ambiguity.

EXPERIMENT I

This experiment was designed to test the effects ofcontext on the activation of the category names forambiguous words which were part of a sentence.Context which scrved to disambiguate the word wasincorporated into the sentence and occurred either priorto or subsequent to the ambiguous word. The paradigmused was a modification of the color-naming task usedby Warren (1972).

The goal of the experiment was to deterrnine(1 ) whether context places constraints on the wordmeanings which are activated in memory or (2) whcthermore than one meaning is initially activated and contexthas its disambiguating effect subsequent to a lexicallook-up process. Finding 1 would suggest that there doesnot exist a lexical look-up process which is independent

CONRAD

Table 1Examples of the Sentences and Colored Words Used in Each Condition in Experiment I

Display Type Sentence Display

Context After WordAmbiguous Word The nickel is not enough to buy the candy bar. Nickel (Red)Appropriate Category That nickel is from her coin coUection. Money (Red)Inappropriate Category The nickel was lost from her purse. Metal (Red)

Context Before WordAmbiguous Word The girl found a nickel. Nickel (Red)Appropriate Category The toy costs a nickel. Money (Red)Inappropriate Category He bought a cornic book for a nickel. Metal (Red)

ControlsAmbiguous Word He brought a bottle of wine to dinner. Nickel (Red)Appropriate Category The craftsman made the chair. Money (Red)Inappropriate Category The boy next door mows our lawn. Metal (Red)

of contextual influence, whereas Finding 2 wouldprovide some support for the not ion of an isolablelexical system in memory.

The position of the context relative to the ambiguousword was varied, because if context does influence thelexical look-up process, it would do so only when it isalready available at the time the word is heard. Thus,given the assumptions of the color-naming task, contextcould only have an effect when it preceded the word. Inaddition, if context has an effect on the lexicallook-upprocess, the manipulation of the position of the contextwould provide a test of the assumption that thecolor-naming task reflects activation at the time that aword is heard. A result indicating that contextsubsequent to the ward can influence lexical look-upwould strongly suggest that activation does not takeplace at the time that the word is heard. Category nameswere used as representations of the meanings of anambiguous word because previous work (Conrad, 1971)had indicated that the category name for a word is anintegral part of the reported meaning of the word.

Method

Subjects

Thirteen male University of Oregon students served as Ss inthe experiment. They were obtained from the 5 pool for theCenter for Perceptual and Cognitive Studies and were paid $3 fortheir participation. All Ss were native English speakers, had goodvision and hearing, and, when questioned, all reported that theywere not color-blind.

Procedure

For each trial in the experiment, Ss listened to a singletape-recorded sentence under instructions to try to rememberthe sentence. Immediately foUowing the sentence, a word,printed in colored ink, appeared on a rear projection screen infront of the Ss, The words were displayed using a KodakCarou sei slid e projector equipped with a Lafayettetachistoscopic shutter. They were printed in capital letters andwere centered on a pale gray background. Four colors ofink-red, blue, green, and purple-were used. Ss were instructedto name aloud the color of the word as quickly as possible andthen to repeat the sentence verbatim. The word remained on thescreen until the 5 had named the color aloud. Naming the color

activated a voice-operated relay which simultaneouslyterminated the display, stopped a Hunter millisecond timerwhich had been started at the onset of the display, andpositioned the slide tray for the next display. Both color-namingreaction times and Ss' recall were recorded by E. Ss received 25practice trials followed by 96 experimental trials on each of 2consecutive days. Each session lasted approximately 40 min.

Words

Eight ambiguous words having two or more differentmeanings, e.g., pot, were used. They were selected from a set of35 words which had been presented to 25 Ss in apretest. TheseSs were asked to name two categories for each word and toindicate which category they had thought of first. The goal wasto find a set of words which would fit into two differentcategories, with each category corresponding to one of themeanings of the word. In addition, the categories had to be giventhe same name by a majority of the Ss to insure that thestructures selected for testing would be common to as many Ssas possible, The eight words selected met the following criteria:at least 40% of the Ss in the pretest mentioned the samecategory name for a word and at least 40% indicated that theyhad thought of that category first. In addition, the results of apilot study, similar to Experiment 11, indicated that all of thecategory narnes for the eight words showed some interferencewith the color-narning task.

Sentences

The eight ambiguous words were incorporated into twoexperimental sets of 48 unambiguous sentences, In the first set,the context after set, the word was always the second word inthe sentence and was made unambiguous by the words whichfollowed it. In the second set, the context before set, theambiguous word was always the last word in the sentence andwas disambiguated by the words which preceded it. Each of theeight words appeared in six sentences in each set so that bothmeanings of the word were represented three times in each set.In addition, 48 unambiguous sentences that did not contain anyof the ambiguous words were used as a control set.

Conditions

Each of the experimental sentences was followed by a visualdisplay of the ambiguous word used in the sentence (ambiguousword), or by the category name for that word which wasappropriate to the meaning expressed by the sentence(appropriate category), or by the category name for the wordwhich was inappropriate to the meaning expressed by thesentence (inappropriate category). Both the context before andcontext after sentences occurred in each of the three

CONTEXT EFFECIS IN SENTENCE COMPREHENSION 133

Results

Note-R Ts are expressed in milliseconds. Difference equalsexperimental mean minus control mean.

Table 2Mean RTs and Their Differences for Correct Color Responses

in Each Condition in Experiment I

Context Before WordAmbiguous Word 1032 932 100Appropriate Category 990 884 106Inappropriate Category 936 872 64

Context After WordAmbiguous Word 977 958 19Appropriate Category 917 915 2Inappropriate Category 913 878 35

98131

62978167

72-959

-3412Il

980935897936895859

999938920870829823

Experi- Diffe-mental Control renceDay Display Type

----------------

Context Before Word1 Ambiguous Word 10971 Appropriate Category 10691 Inappropriate Category 9822 Ambiguous Word 9672 Appropriate Category 9102 Inappropriate Category 890

Context After Word1 Ambiguous Word 10521 Appropriate Category 9261 Inappropriate Category 9562 Ambiguous Word 9022 Appropriatc Category 9072 Inappropriate Category 870

Notc -R'Ts are expressed in millisecands. Difference equalsexperimental mcan minus control mean.

Trial Type

Table 3Mean RTs and Their Differences for Each Condition

on Each Day in Experiment I

condition where the visual display was a category whichcarrespanded to the meaning of the ambiguous ward inthe sentence, and the context came after that word inthe sentence, does not show this interference effect. Inaddition, there was a significant Trial Type by Contextinteraction [F(1,12) = 23.6, p< .01] . This reflects thefact that there were larger differences between the meanRTs for experimental and cantral trials when thecontext preceded the ambiguaus word in the sentencethan when the context followed the word. Thus, Ssdemonstrated larger interference effects when thecontext came before the ambiguaus word.

An apriori test of paired comparisons, using themethod suggested by Winer (1962), was done on themeans for each pair of experimental and control trials.For the context after sentences, only RTs for theambiguous ward candition were significantly larger thantheir contral group (p< .01). For the context befaresentences, however, mean RTs for all of theexperimental conditions were larger than theirappropriate contrals (p < .025).

Overall mean RTs were faster fallowing sentences inwhich the context came after the ambiguous word thanfollowing sentences in which the context preceded theambiguous word, producing a significant main effect ofcontext [F(1 ,12) = 10.1, p< .01] . Thus, in the contextafter condition, Ss were bath faster at color naming andshawed less inteference from the meaning of theambiguaus word.

Finally, there was a significant main effect of displaytype [F(2,24) = 26.0, p< .01] . Mean RTs were fastestfar the inapprapriate category displays and slowest forthe ambiguous word displays. An apriori test of pairedcomparisons (Winer, 1962), done on the experimentalmean RTs far the context before sentences, indicatedthat there was no significant difference between the

Experi- Diffe-mental Control rence

Trial Type

Display Type

experimental conditions. In the control conditions, the samewords, printed in the same color of ink, were paired with thecontrol sentences. Table 1 shows examples of the sentences andcolored words used in each condition.

On each day, Ss received half of the experimental trials andtheir appropriate controls. The sentences were divided in half sothat on each day, Ss received six sentences containing eacharnbiguous word, three for each meaning, with one sentence ineach condition for both the context before and context aftersets,

Mean RTs for correct color responses for each S oneach day were calculated. The means of these means,collapsed over days and Ss, are shown in Table 2.

A five-way analysis of variance was done on themeans. The five factors were Ss, number of days ofpractice (days), the context position with respect to theambiguous word (context), the type of visual display(display type), and the type of trial, i.e., experimental vscontrol (trial type).

As might be expected, there was a significant maineffect of days [F(I,12) = 5.7, p< .05]. This reflects adecrease in overall RTs from Day 1 to Day 2. Inaddition, days interacted with both trial type [F(1 ,12) =

16.4, p<.OI] and context [F(1,12) = 7.9, p<.OI].There was a larger decrease in mean RTs farexperimental trials from Day 1 to Day 2 (106 msec)than for control trials (39 msec) and a larger decrease inmean RTs, over days, for trials in which the contextpreceded the ambiguous word in the sentence(1 19 msec) than for trials where the context came afterthe ambiguous word in the sentence (63 msec).

Because of these interactions, the data werereanalyzed separately for Days 1 and 2, using a four-wayanalysis of variance. The means for Days 1 and 2 areshown in Table 3.

The analysis af the Day 1 data indicates a significantmain effect of trial type [F(1,12) = 28.0, r < .01]. Alook at the data shows that for five of the sixcomparisons, RTs far the experimental trials were langerthan thase for the appropriate controls. Only the

134 CONRAD

means for the ambiguous word and appropriatc categorydisplays [F(I,96) = 1.15] but that there was asignificant difference between the appropriate categoryand inappropriate category means [F(I,96) = 11.11,p< .0I]. However, a test of the difference scores forthese means and their controls indicated that there wasno significant difference between the scores forappropriate and inappropriate categories, suggesting thatthe control means were also different for the twoconditions. Thus, the evidence with respect to whetheror not appropriate category names produced moreinterference than inappropriate category names isinconclusive. The same apriori test, done on theexperimental means for the context after conditions,indicated a significant difference between the means forthe ambiguous word and appropriate category displays[F(I ,96) = 27.1, p < .01] but no significant differencebetween the means for the appropriate category andinappropriate category displays [F(1,96) = 1.3]. Thus,for the sentences in which the context followed theambiguous word, Ss were slower at color naming whenthe visual display was a word used in the sentence thanwhen the display was either an appropriate orinappropriate category for that word.

An analysis of variance done on the Day 2 datashowed a significant main effect of trial type [F(I ,12) =15.0, n< .01] and of display type [F(2,24) = 5.0,p < .05]. Again, for five of the six comparisons, RTswere larger for experimental trials than for control trials,with only the ambiguous word display followingsentences in which the context came after theambiguous word, showing no interference effects. Meansfor the ambiguous word displays were the slowest andthose for the inappropriate category displays werefastest. In addition, there was a significant Trial Type byContext interaction [F(I,12) = 5.4, p< .05], reflectingthe larger differences in RTs between the experimentaltrials and their appropriate controls for the sentences inwhich the context preceded the ambiguous word.

An apriori test of paired comparisons (Winer, 1962),done on the mean RTs for each experimental conditionand its control, indicated that, for the context aftersentences, none of the mean RTs for the experimentalconditions was significantly different from the meanRTs for the control groups. For the context beforesentences, both the ambiguous word and appropriatecategory conditions had mean RTs significantly largerthan those for their control groups.

The overall error rate for color naming in theexperiment was 1.24%. The types of errors and theirfrequency of occurrence were as follows: (I) Ss namedthe wrong color (.16%), (2) Ss named the word whichwas displayed instead of the color (.12%), (3) Ss recalledthe sentence without first narning the color (.16%), and(4) equipment or E errors (.80%). Verbatim sentencerecall was quite good, with only a .16%error rate.

Discussion

One rather unexpected finding was that there was a

marked decrease in interference effects on the secondday of practice. Unlike Warren's findings for categorynames, only the mean RT for appropriate categories inthe context before condition showed an interferenceeffect on the second day. This may be due to thedecrease in RTs for Day 2, since we know the Ss havingvery fast RTs often do not show even the usual Stroopeffect.

The lack of interference effects for the context afterconditions, even on Day 1, may be attributed to thelonger delay between aural presentation of theambiguous word and the visual display. Since theambiguous word appeared at the beginning of thesentence in these conditions and the visual display wasnot presented until Ss had heard the entire sentence,there was a variable delay interval during which Ss werelistening to and processing the remainder of thesentence. Thus, the lack of interference may be dueeither to loss of activation over the interval, attributableto a fast decay rate, or to interference from theprocessing which occurred during the interval. A thirdexplanation might be that there was no activation of themeanings of ambiguous words in the context aftercondition. However, this seems unlikely, since there isevidence for activation in the context before condition,the condition where it rnight least be expected. Inaddition, an experiment by Foss and Jenkins (I972)also suggests that activation does take place. Foss foundthat there was significant interference in RTs to aphoneme-monitoring task immediately following thepresentation of an ambiguous word and that thisinterference occurs both when the word has beendisambiguated by prior context and when it has not yetbeen disambiguated. Thus, it seems probable that wordmeanings were activated in the context after condition.However, it is not c1ear whether the activation wasterrninated 'because of a fast decay rate or as a result offurther processing.

The most interesting results are those for the contextbefore conditions on Day 1. To insure that the resultswere not due to the repetitions of the ambiguous wordsin the sentences, the data for only the first occurrence ofeach ambiguous word in a sentence and its appropriatecontrol were analyzed. It was found that, even on itsfirst occurrence, an ambiguous word activated bothappropriate and inappropriate category names. The factthat all the appropriate and inappropriate categoriesshow interference effects indicates that both meaningsof the word are activated in memory, even when thepreceding context indicates which meaning is intended.Context may be having some effect on this activation,since the experimental means for appropriate categoriesare significantly longer than the means for theinappropriate ones. However, this seemsunlikely, since amore appropriate test, that of the difference scores,indicates that there is no significant difference in theamount of interference for the two category conditions.Thus, the results do not unequivocally support or refutethe notion of an isolable subjective lexicon. They do,

CONTEXT EFFECTS IN SENTENCE COMPREHENSION 135

however, support the existence of a subjective Iexiconby showing that, during sentence comprehension, Ss doactivate in memory the meanings of individual words inthe sentence.

EXPERIMENT 11

The previous experiment suggested that Iinguisticcontext might have some minimal effect on the strengthof activation of category names for ambiguous wordswhich occur in a sentence. It wouId seem desirable toreplicate that finding and to compare the results withthose for ambiguous sentences containing the sameambiguous words. This would provide a condition inwhich there is no context to indicate which meaning ofthe ambiguous word is intended. If context is effectivein reducing the strength of activation of inappropriatecategory information, then one might expect a largerinterference effect of both categories for ambiguoussentences than for the inappropriate category forunambiguous sentences. In addition, the data forambiguous sentences may provide further evidence withwhich to choose among the current theories ofambiguity. The garden path theory would predict that,for the ambiguous sentences, only a single meaning of anambiguous word is activated in memory. Both theexhaustive computation hypothesis and the perceptualsuppression hypothesis, on the other hand, predict thatall meanings of the word will be activated. The latteralso predicts an inequality of strength of activation forthe various meanings of a word.

The present experiment, then, attempts to replicatethe findings of Experiment I for sentences containingambiguous lexical items which are disambiguated byprior context, to extend these findings to sentenceswhich contain no disambiguating context, and to decidebetween the various theories of ambiguity.

Method

Subjects

Fourteen University of Oregon students served as Ss in theexperiment. Ten were obtained from the S pool and were paid$3 for their participation. The remaining four were volunteerswho received extra credit for a psychology dass. All had goodvision and hearing; none reported being color-blind or hadparticipated in Experiment I or the pretest for Experiment 11.Ten of the Ss were male and four were female.

Procedure

The procedure and equipment were the same as inExperiment I. On each trial, Ss Iistened to a single sentencewhich was followed immediately by the visual presentation of aword printed in colored ink. Ss named the color of the word asquickly as possible and then recalled the sentence aloud.Color-naming RTs and recall errors were recorded by E. Fourcolors-red, blue, green, and purple-were used. Ss received 25practice trials, followed by 96 experimental trials on each 01' 2consecutive days. Each session lasted approximately 30 min.

Sentences

The eight arnbiguous words used in Experiment I wereincluded in each of two types of sentences, and in a11 cases, theambiguous word was the last word in the sentence. Forty-eightof the sentences, 6 for each word, contained a lexical ambiguity(e.g., "She has lost her bat. "), Apretest had indicated that all ofthese sentences were behaviorally as weil as logically ambiguous.That is, both meanings of each sentence were plausible enoughso that Ss could find both meanings when asked to do so, andnot a11 Ss gave it the same meaning initially. The remaining 48were unambiguous (e.g., "She fed the bat. "). Within theunambiguous set, there were three sentences for eacn meaning ofeach of the ambiguous words. In addition, there was a controlset of 48 unambiguous sentences which did not contain any ofthe ambiguous words.

Conditions

Each of the sentences was followed by a visual display of theambiguous word used in the sentence (ambiguous word), by thecategory name for that word which was appropriate to themeaning expressed by the sentence (appropriate category), or bythe category name for the ambiguous word which wasinappropriate to the meaning expressed by the sen­tence (inappropriate category). In reality, the appro­priate-inappropriate category distinction could not bemade for the ambiguous sentences, since it was impossible withthis experimental design to know what interpretation each Sgave to the sentences. For these sentences, then, there was nodifference between the appropriate category and inappropriatecategory conditions. However, the distinction was used tomaintain a balanced design for purposes of data analysis.

In the control conditions (control), each of the words used ineach experimental condition was paired with a control sentence.Experimental and control words were printed in the same colorofink.

Table 4 shows examples of the experimental sentences andcolored words for each condition.

On each day, Ss received half of the experimental trials andtheir appropriate controls. The sentences were divided so thateach day, Ss heard three ambiguous and three unambiguoussentences containing each ambiguous word, with one sentence ofeach type in each condition.

Results

Mean RTs for correct color responses for each S oneach day were calculated. The means of these means,collapsed over days and Ss, are shown in Table 5.

A five-way analysis of variance was done on themeans. The five factors were Ss, number of days ofpractice (days), type of sentence, i.e., ambiguous vsunambiguous (sentence type), the type of visual display(display type), and the type of trial, i.e., experimental vscontrol (trial type).

As expected, there was a significant main effect ofdays [F(l,13) = 25.0, p< .01], reflecting the decreasein RTs from Day I to Day 2. Since there were nointeractions with days as a factor, means for both dayscombined were calculated for each condition for each S,and these were analyzed using a four-way analysis ofvariance.

The analysis showed a significant main effect of bothtrial type [F(l,13) = 35.8, p< .01 J and display type[F(2,26) = 15.8, p< .01], but no main effect of

136 CONRAD

Table4Examples of the Sentences and Colored Words Used in Each Condition in Experiment 11

Display Type Sentences Display

Ambiguous SentencesThe cupboard contains the pot. Pot (Red)

Ambiguous Word I traded rny desk for the pot. Pot (Blue)

lohn bought the pot. Utensil (Red)Category Word He finished the pot. Marijuana (Blue)

Unambiguous SentencesBeans are cooked in the pot. Pot (Red)

Ambiguous Word He was arrested for possessing pot. Pot (Blue)

We boiled water in the pot. Utensil (Red)Appropriate Category lt is illegal to smoke pot. Marijuana (Blue)

Inappropriate CategoryWe made tea in the pot Marijuana (Red)The field was full of pot. Utensil (Blue)

Control SentencesThe truck collided with the wall. Pot (Red)

Ambiguous Word She made a pie for dinner. Pot (Blue)

The race track was thoroughly prepared. Utensil (Red)Appropriate Category The man owns a sailboat. Marijuana (Blue)

Inappropriate CategoryThe television station covered the political campaign. Marijuana (Red)He bought a flowered shirt. Utensil (Blue)

sentence type (F = 1.0). In addition, there were nosignificant interactions.

An apriori test of paired comparisons, using themethod suggested by Winer (1962), was done on each ofthe experimental means and their appropriate control todetermine whether or not there were significantinterference effects for each experimental condition. Allof the comparisons were significant (with p < .0 I).

Although there was a significant main effect ofdisplay type, a comparison of the difference scores(experimental means minus control mean) for theambiguous word and category word conditions for boththe ambiguous and unambiguous sentences suggests thatthere are no differences in degree of interference fordifferent types of displays. In addition, the lack of arnain effect for sentence type indicates that there wereno differences in amount of interference for ambiguousand unambiguous sentences, although, in a11 cases, the

Table5Mean RTs and Their Differences for Correct Color Responses

in Each Condition in Experiment 11

Trial Type

difference scores are longer for ambiguous sentencesthan for unambiguous ones.

Since information about the bias of eachinterpretation of the ambiguous sentences was availablefrom apretest, it was possible to rescore the categoryresponses and their controls for the ambiguous sentencesseparately for categories which were appropriate to thatmeaning of the sentence which had a bias of > 50%(likely category) and for the categories whichcorresponded to the meaning which had a bias of ,;;;; 50%(unlikely category). The mean RT, over days, for thelikely category was 1,014 msec, with a control of939 msec, and the mean for the unlikely categories was1,017msec, with a control of 943msec. Bothexperimental means were significantly larger than theircontrols (p< .001) but were not different from oneanother.

The error rate for color naming was 1.02%; namingthe wrong color constituted .18% of the errors; repeatingthe word narne, .03%; color omissions, .18%; andequipment or E errors, .63%. Verbatim sentence recallwas quite good, with an error rate of .18%.

Display TypeExperi- Diffe-mental Control rence

Discussion

Note-RTs are expressed in milliseconds. Difference equa/sexperimental mean minus control mean."Since there was no difference between appropriate andinappropriate category conditions for ambiguous sentences.these were combined into a single mean.

1015*

Ambiguous SentencesAmbiguous Word 1044Appropriate CategoryInappropriate Category

Unambiguous SentencesAmbiguous Word 1059Appropriate Category 997Inappropriate Category 993

971

938*

1001936943

73

77

586150

The resul ts for unambiguous sentences in which theambiguous word is preceded by disambiguating contextagain show significant interference effects for bothappropriate and inappropriate categories, but nodifference in the size of the interference effects for thetwo types of categories. In addition, there were nosignificant differences in the strength of activation of thecategory names for ambiguous and unambiguoussentences. These findings, then, provide further supportfor the idea that there may be a context-independentlexical look-up stage which occurs early in sentenceprocessing and that context has its effect on sentence

CONTEXT EFFECTS IN SENTENCE COMPREHENSION 137

comprehension at some later stage of processing.The results also seem to provide further support for

the exhaustive computation hypothesis, at least to theextent that it was intended to explain lexicalambiguities. This hypothesis, like the perceptualsuppression theary, suggests that at sorne level ofprocessing, all meanings of an ambiguous sentence arecomputed. Clearly, the results suggest that at least twointerpretations of a lexically ambiguous sentence arecomputed at an early stage of processing.

The results do not seem to support MacKay'sperceptual suppression theary. The theory proposes thatin order to perceive one meaning of an arnbiguity, theother meaning must be suppressed. This would seem topredict a difference in the strength of activation for thelikely and unlikely categories in the present experiment.However, the results indicate that there is not adifference in the strength of activation for the twocategory types. Although the present study uses alimited number of words to test this hypothesis, theprobabilities reflecting the likelihood of a meaning didrange from .8 to .2, with a fairly equal distribution ofscores throughout the range.

CONCLUSIONS

The results of these experiments suggest that theredoes exist a subjective lexicon as Miller suggested andthat linguistic context does not have an effect on thelexicallook-up process. However, since there were somesignificant differences in the strength of activation ofappropriate and inappropriate category names inExperiment I, it may be possible that small differencesdo exist. If so, perhaps they could be measured with amore sensitive experimental technique.

It is possible to argue that the color-naming paradigmused in these experiments does not measure theactivation of meaning which occurs when a word isencountered in a sentence. Instead, one might argue thatthe categary names are activated when they arepresented as the visual display and only then are theyassociated with the ambiguous word which occurred inthe sentence. However, there are several reasons forbelieving that a backwards association theory of this sortcannot account for the data. First, a theory of this typewould seem to predict that at least some of thecolor-naming errars would be intrusions of theambiguous word. The fact that there were no errors ofthis type in either of the present experiments or inWarren's (1972) study suggests that this is not areasonable explanation of the effects found in theseexperiments. Second, Ss' reports following theexperiments indicated that they were not aware of therelationship between the last ward in the sentence andthe ward presented visually. This suggests that if Ss wereassociating the two words only at the time the secondward was presented, this was not done consciously.Third, arecent experiment by Wanen (I 973)

provides a direct test of the backwards associationtheory by using words which are highlyassociated in only one direction in thecolor-naming paradigm. For example, we know thatgiven the word "bitter" in a free association test,"sweet" is a highly probable first response. On the otherhand, "bitter" is a very unlikely response to "sweet"(palermo & Jenkins, 1964). Warren aurally presented Sswith words like "bitter" followed by a visualpresentation of words like "sweet" and compared theamount of color-naming interference to a condition inwhich "sweet" preceded "bitter." He found significantinterference for the condition where "bitter" preceded"sweet" but no interference when "sweet" preceded"bitter." These results do not support a backwardsassociation theory; rather , they strongly argue that thevisually presented word was activated in memory priorto its actual presentation. Finally, Cairns (1973), in anexperiment where Ss were forced to reprocessphonological ambiguities (e.g., weight vs wait), hasfound that the time for reprocessing is on the order of350 msec. This time to find the other meaning of anaurally presented word is considerably longer than thedifference scores found in the present experiments. TheCairns experiment, then, suggests that if a backwardsassociation explanation of the present data were correct,one might expect the differences in RT for experimentaland control trials to be three times as large as theyactually were. Thus, the best account of thecolor-naming interference effect found in the presentstudies would seem to necessitate the assumption thathearing the ambiguous word automatically activates twoor more of its meanings in memory and that thisactivation occurs prior to the visual presentation of aword.

The present experiments suggest that, althoughlinguistic eontext ean be effective in resolvingambiguities in the language, this disambiguating effectdoes not oeeur during the lexical look-up process.Rather , there seems to be a process during sentencecomprehension in which the meanings of words in asentence are activated in memory without regard to theavailable linguistie eontext. This, then, implies thatcontext has its disambiguating effeet at some timesubsequent to the lexieal look-up proeess. Thus, we arestill left with the problem of how eontextualinformation and lexieal information are integrated toproduee a meaning for an ambiguous sentenee. MacKay'spereeptual suppression hypo thesis attempts to accountfor this type of integration, but the results ofExperiment II do not support his hypo thesis. Thesimplest alternative might be a theory in whiehcontextual information (and perhaps other types ofinformation) increases the strength of one or more ofthe activated meanings of a ward to the point that itreaches threshold and becomes conscious. Certainly,other alternative theories are also plausible. A next step,then, in trying to gain an understanding of sentence

138 CONRAD

comprehension might be to formulate and test theseintegration theories. 111e color-naming methodologyused in the present studies may prove to be a useful toolfor these investigations.

REFERENCES

Bransford, J. D., & Johnson, M. K. Considerations of someproblems of eomprehension. In W. G. Chase (Ed.), Visualinformation processing. New York: Aeademic Press, 1973.

Cairns, H. S. Effeets of bias on proeessing and reprocessing oflexically ambiguous sentenees. Journal of ExperimentalPsychology, 1973,97,337-343.

Carey, P. W., Mehler, J., & Bever, T. G. When do we compute allthe interpretations of an ambiguous sentenee? In G. B. Floresd'Arcais and W. J. M. Levelt (Eds.), Advances inpsycholinguistics: New York: American Elsevier. 1970.

Conrad, C. An analysis of a hierarchical model of semanticmemory organization. Unpublished master's thesis, UniversityofOregon, 1971.

Conrad, C. Studies of the subjeetive lexicon. Unpublished PhDdissertation, University of Oregon, 1972.

Fodor, J., Garrett, M., & Bever, T. Some syntaetic determinantsof sentential complexity, 11: Verb strueture. Perception &Psychophysics, 1968,3,453·461.

Foss, D. J. Some effects of arnbiguity upon sentencecomprehension. Journal of Verbal Learning & VerbalBehavior, 1970,9,699-706.

Foss, D. J., Bever, T. G., & Silver, M. The comprehension andverifiea tion of ambiguous sentences. Perception &Psychophysics, 1968,4,304-306.

Foss, D. J., & Jenkins, C. M. Some effects of eontext on thecomprehension of ambiguous sentences, Cited by D. T. Hakes,On understanding sentences: In search of a theory of sentencecomprehension. Final report, October 1972, U.S. Office ofEducation, Washington, D.C.

Garrett, M. F. Does ambiguity complicate the perception ofsentences? In G. B. Flores d'Arcais and W. J. M. Levelt (Eds.),

Advances in psycholinguistics. New York: American Elsevier1970. '

Katz, J. J., & Fodor, 1. A. The structure of a semantic theory.Language, 1963,39, 170-210.

Lackner, J. R., & Garrett, M. Resolving ambiguity: Effects ofbiasing context in the unattended ear. Cognition, 1973, inpress.

Lashley, K. S. The problern of serial order in behavior. In L. A.Jeffress (Ed.), Cerebra! mechanisms in behavior. New York:Wiley, 1951.

MaeKay, D. G. To end ambiguous sentences. Perception &Psychophysics, 1966, 1,426-436.

MacKay, D. G. Mental diplopia: Towards a model of speechperception at the semantic level. In G. B. Flores d'Arcais andW. J. M. Levelt (Eds.), Advances in psycholinguistics. NewYork: American Elsevier, 1970.

MacKay, D. G. Aspects of the theery of comprehension,memory and attention. Quarterly Journal of ExperimentalPsychology, 1973,25,22-40.

Marshall,. J. C. Syn.t~ctie analysis as apart of understanding.Bulletin of the British Psyehological Society, 1965, 18,28.

Miller, G. The subjective lexicon. Presidential address, 78thAnnual Convention, American Psychological AssociationSeptember 1970. '

Palermo, D. S., & Jenkins, J. J. Word assocation norms: Gradeschool through college. Minneapolis: University of MinnesotaPress. 1964.

Posner, M. I. Coordination of internal codes. In W. G. Chase(Ed.), Visual information processing. New York: AeademicPress, 1973.

Wanen, R. E. Association, directionality and stimulus encoding.Journal of Experimental Psycnology, 1973, in press.

Warren, R. E. Stimulus encoding and memory. Journal ofExperimental Psychology, 1972,94,90-100.

Winer, B. J. Statistical principles in experimental design. NewYork: McGraw-HiII, 1962.

(Received for publication May 25, 1973;revision reeeived June 25, 1973.)