cognitive representation of semantic categories-rosch-1975

Journal ol Experimental Psychology: General197S, Vol. 104, No. 3, 192-233

Cognitive Representations of Semantic Categories

Eleanor RoschUniversity of California, Berkeley

SUMMARY

The technique of priming was used to study the nature of the cognitiverepresentation generated by superordinate semantic category names. In Experi-ment 1, norms for the internal structure of 10 categories were collected. InExperiments 2, 3, and 4, internal structure was shown to affect the perceptualencoding of physically identical pairs of stimuli, facilitating responses to physicallyidentical good members and hindering responses to physically identical poor mem-bers of the category. Experiments 5 and 6 showed that the category name didnot generate a physical code (such as lines and angles) but rather affected per-ception of the stimuli at the level of meaning. Experiments 7 and 8 showed thatwhile the representation of the category name which affected perception containeda depth meaning common to words and pictures which enabled subjects to preparefor either stimulus form within 700 msec, selective reduction of the intervalbetween prime and stimulus below 700 msec revealed differentiation of the codingof meaning in preparation for actual perception. Evidence from these experi-ments indicated that pictures could be prepared for more rapidly than words.Experiment 9 demonstrated that good examples of semantic categories do notappear to be physiologically determined, as the effects of the internal structure ofsemantic categories on priming (unlike such effects for color categories) couldbe eliminated by long practice.

From this series of experiments, four basic findings can be derived. First, theinternal structure of superordinate semantic categories was clearly a pervasiveaspect of the way in which those categories were processed in the tasks used, aresult which does not depend for its validity on any particular interpretation ofthe meaning of internal structure. Second, the results supported the claim thatthere are different levels of processing in perception. The meaning of super-ordinate category terms appeared to affect the perception of some pairs of stimuli,but not at the level of concrete features. Third, the set of experiments allowedsome conclusions about the underlying representation of superordinate categories.A depth meaning of these categories did not appear to be specifically coded in termsof words or pictures; rather, that depth meaning appeared to be translated intoa format in preparation for actual perception which differed slightly for wordsand pictures. The fact that less time was required to prepare for pictures suggeststhat pictures may be closer to the nature of the underlying representation than arewords. Fourth, the research made possible a comparison of representations ofsuperordinate categories with other kinds of categories (such as color which iscoded quite concretely) and with categories at other levels of abstraction, therebyproviding a method for collecting data concerning many unresolved issues in thenature and development of abstraction,

192

COGNITIVE REPRESENTATIONS OF SEMANTIC CATEGORIES 193

When we hear a category word in naturallanguage such as jurniture or bird andunderstand its meaning, what sort of cogni-tive representation do we generate? A listof features necessary and sufficient for anitem to belong to the category? A concreteimage which represents the category? Alist of category members ? An ability to usethe category term with no attendent mentalrepresentation at all? Or some other, lesseasily specified, form of representation?

The nature of cognitive representationsof semantic categories has direct relevanceto two important areas of psychological in-quiry : One concerns the structure of cate-gories and concepts and has implications forthe way in which concepts and conceptattainment should be studied in psycho-logical research (Rosch, 1973, in press-a,in press-c). The other area is the natureof mental representations in general, aprominent concern in the postbehavioriststudy of mental events (Cooper & Shepard,1973; Posner, 1973; Segal, 1971).

The present research was designed toinvestigate both issues within the sameseries of experiments. The first four experi-ments focused on the question of the natureand structure of categories and the effectof that structure on the tasks used. Thesucceeding five experiments used the effectsof internal category structure found in thefirst experiments (reliable empirical effectswhich did not depend on any particular,possibly controversial, interpretation of in-ternal structure) to explore several moregeneral questions concerning the nature ofmental representations of categories: ques-tions regarding the type and degree of con-creteness of the representations, and ques-tions regarding the presence or absence ofdifferentiation of the representations intopictorial and verbal forms.

With respect to the issue of the natureof categories, Rosch has previously argued(Rosch, 1973, 1975, in press-c) that manytraditions of thought in philosophy, psy-chology, linguistics, and anthropology implythat categories are Aristotelian in nature—that is, that categories are logical, clearly

See p. 228 for acknowledgments and requests forreprints footnote.

bounded entities, whose membership isdefined by an item's possession of a simpleset of criterial features, in which all in-stances possessing the criterial attributeshave a full and equal degree of member-ship. While such a structure may, in fact,characterize the artificial categories employedin much concept formation research (Bourne,1968) and research on artificial languages(Ginsburg, 1966), many natural categories,that is, concepts designatable by words innatural languages, appear to possess struc-tures of a quite different character (Labov,1973; Lakoff, 1972; Rosch, 1973). Thedomain which has most readily lent itselfto the demonstration of a type of categoricalstructure contradictory to the Aristotelianis that of color. There is now considerableevidence that color categories are processedby the human mind (learned, remembered,denoted, and evolved in languages) in termsof their internal structure; color categoriesappear to be represented in cognition not asa set of criterial features with clear-cutboundaries but rather in terms of a proto-type (the clearest cases, best examples) ofthe category, surrounded by other colors ofdecreasing similarity to the prototype andof decreasing degree of membership (Ber-lin & Kay, 1969; Heider, 1971, 1972;Mervis, Catlin, & Rosch, 1975 ; Rosch, 1973,1974, 1975, in press-c, in press-d).

Color, however, is a perceptual domain;its categorization is probably physiologicallybased (McDaniel, 1972; Rosch, 1974, inpress-b, in press-d). Not all categorieshave an obvious perceptual basis and manycategories may be culturally relative. Is theconcept of internal structure applicable toother types of categories?

At this point, it is necessary to elaborateon the logic of the concept of internal struc-ture as it is used in the present research.The hypothesis that categories have internalstructure is not a theory which specifies, inadvance of the collection of data, a precisemodel (with flow charts, steps, etc.) whichis then tested by means of data from par-ticular tasks. Rather, internal structurerefers to that general class of conceptionsof categories in which categories are notrepresented only as criterial features with

194 ELEANOR ROSCH

clear-cut boundaries and in which itemswithin categories may be considered differ-entially representative of the meaning of thecategory term. Progressive clarification ofthe meaning of internal structure is in-tended to occur through operational defini-tions based on converging experiments.Furthermore, initial experiments are in-tended to be in the form of "class questions"(Are categories of the general form x or y?)rather than in the form of tests of specificstimulus or processing models. The researchreported in this article (and in Rosch'sother articles on the nature of human cat-egorization which are referred to through-out this article) provides illustrations ofthis strategy of conducting research.

This article is concerned with the ques-tion of whether the general concept of inter-nal structure, previously specified only forperceptual domains such as color and form,is applicable to other types of categories;specifically, whether it is applicable to thesemantic classifications of common objects ineveryday use. Applicability refers to twogeneral issues: Can subjects make meaning-ful judgments about internal structure—thedegree to which instances are good or poormembers of categories; and can a reason-able case be made that internal structureaffects cognition with respect to categories?

Some evidence that semantic categoriesfulfill both criteria for internal structurealready exists. Rosch (1973) showed thatsubjects reliably rated the extent to whichan item represented their idea or image ofthe category name for six items from eachof eight categories. Such ratings were pre-dictive of subjects' reaction times in a sen-tence verification task of the form "A(member) is a (category)" (Rosch, 1973).Rips, Shoben, and Smith (1973) confirmedand extended Rosch's (1973) findings, andSmith, Shoben, and Rips (1974) haveargued that a model based on ratings andsentence verification tasks such as these canaccount for most of the findings in theliterature on retrieval from semanticmemory.

With regard to the nature of categories,the purpose of the present research was toextend the range of processing tasks shown

to be affected by category structure and,thus, elaborate further the concepts ofcategory structure and prototypes. Itshould be emphasized that the researchwas designed to investigate the nature (thestructure and content) of cognitive repre-sentations of semantic categories, not thedynamic processes (activation, spread ofenergy, evaluation of paths, etc.) by whichthose representations are formed. That is,the studies were designed to focus on thewhat and not the how of cognitive repre-sentations of semantic categories. Specifi-cally, the present research was not anattempt to construct a model of semanticmemory; the fitting of dynamic memorymodels to the data of the present research isseen as a further stage of investigation. InExperiment 1, norms were collected forratings of the goodness of example of alarge number of instances from 10 sys-tematically chosen classifications of commonconcrete objects. Succeeding experimentsexplored the manner in which those ratingsof internal structure affected outcomes intasks which were designed to explore aspectsof the structure and content of cognitiverepresentations of categories.

With regard to more general questionsabout the nature of mental representations,these issues have been the subject of philo-sophical and psychological debate since theinitiation of both fields of study. The pres-ent research was concerned with a specifictype of category—the higher-level (super-ordinate) classifications into which thecommon concrete objects of; our culture aredivided—and with some of the classic ques-tions concerning the nature of mental rep-resentations as these can be applied to suchcategories. Do our cognitive representa-tions involve concrete imagery as Titchener(1909) claimed, or, as the Wurzburg schoolcountered, is thought essentially abstractand imageless; do all meanings have acommon abstract form regardless of theircontent, or are meanings represented in avariety of modes, for example, by differentforms of representation for verbal materialsand for pictures (Boring, 1950) ? Thepurpose of the present studies was to obtaindata relevant to these interesting issues.

COGNITIVE REPRESENTATIONS OF SEMANTIC CATEGORIES 19S

The questions that can be asked and thekinds of answers that can be obtained aboutthe nature of mental representations ofcategories depend upon the particular re-search procedures by which representationsare studied. The present studies are basedon a method of priming (Beller, 1971) inthe type of matching paradigm describedby Posner and Mitchell (1967), whichRosch (in press-d) had already demon-strated to be particularly suitable to a studyof the nature of the mental representationof color categories.

In the original matching research, sub-jects were required to decide as rapidly aspossible whether two simultaneously pre-sented visual letters were the same or dif-ferent; under some conditions same wasdenned as physical identity (e.g., AA), andunder others as possession of the same name(e.g., Aa). If a subject is provided withsome of the information which he needs tomake the match in advance of presentationof the pair, matching speed should be facil-itated. Beller (1971) primed subjects witha letter (the letter was presented 2 sec inadvance of the pair) and found that phys-ically identical, as well as same-name, pairsimproved—even when physically identicalpairs were in the opposite case from thecase presented in the prime (e.g., if theprime was A, the pair aa, as well as AA,was facilitated). Posner (1969) arguedthat this showed that subjects were notsimply retaining a literal representation ofthe presented letter but were generating anabstract expectation or representation whichdid not depend on case. It should be noted,however, that, as in early studies of set(Haber, 1966), this reasoning assumes that,because the prime is presented prior to thestimulus, it must have its influence on cog-nitive processes which occur prior to—rather than during or after—presentationof the stimulus.

The basic logic of the priming techniqueas it was used by Rosch (in press-d) andas it is used in the present experiments, isthat a prime can only facilitate a responseif it contains some of the information neededto make the response. If the prime facil-itates a response when the prime is pre-

sented prior to the stimulus but not whenpresented simultaneously with the stimulus,it can be inferred that the information whichfacilitates the response is located in the rep-resentation generated by the prime in ad-vance of information provided by thestimulus. If, under such conditions (i.e.,a prime is effective when presented prior to,but not simultaneously with, a stimulus),a prime affects reaction time to physicallyidentical pairs of stimuli, it may be con-cluded that the representation generated bythe prime contains some of the informationused by subjects in the perceptual encodingof the stimulus.1 Since it is assumed thatno facilitation occurs if the representationhas not been allowed sufficient time todevelop, it is possible to assess the amountof time taken to generate the representationby systematically reducing the length of timebetween the prime and the stimulus untilthe prime loses its effectiveness.

There is another possible outcome inwhich the effects of priming may include adifferent process. A prime may be effectivewhen presented simultaneously with a stim-ulus as well as when presented prior to thestimulus. In such cases, it can be inferredthat at least part of the relevant informationin the prime acts on decision processesregarding the stimulus after the stimulus isavailable to the subject. However, becauseof possible interactions between stages ofprocessing (see, e.g., Sternberg, 1969), itmay be impossible to determine, from apriming experiment alone, which processingevents are affected and in what manner bythe prime.

A study of colors (Rosch, in press-d)1 We may make such an assumption because the

only operations which a subject can perform upona physically identical pair are (a) the perception(visual encoding) of the stimuli and (b) the per-ception that the members of the pair are identical.Even if the perception of physical identity is adecision separate from perceptual encoding of thestimulus itself, a prime consisting of a categoryname does not contain within it information aboutthe nature of identity as such (an especially un-tenable assumption if, as was found to be thecase, the prime differentially affects differenttypes of stimuli). Thus, the only possibility,under these conditions, is that the prime affectsthe visual encoding.

196 ELEANOR ROSCH

illustrates how results found by primingcan be of direct relevance both to the natureof color categories and to the nature ofcognitive representations. Advance prim-ing with the category name facilitatedresponses to good examples of color cate-gories and hindered responses to poorexamples for physically identical pairs ofcolors both under instructions which definedsame as belonging to the same category andunder those which defined same as physicalidentity. The selective effect of the primewas entirely eliminated by simultaneouspresentation of prime and stimuli. Follow-ing the previous logic, it was concluded thatthe cognitive representations of color cate-gory names contained information used inthe perception of actual color stimuli.Furthermore, since the effect of the primewas selective, it could be concluded that theinternal structure of color categories affectedthe representation generated by the categoryname; that is, the representation appearedto contain information used in the percep-tion of good members of color categoriesbut which was sufficiently unlike poor mem-bers of the category as to interfere withtheir perception. Representations of colorcategories were generated quite rapidly—selective effects of priming were alreadysignificant at 200 msec, which is less thanhalf the time usually estimated to be requiredfor the formation of a visual image (Posner,Boies, Eichelman, & Taylor, 1969). Insum, the effectiveness of the priming tech-nique for an investigation of cognitive rep-resentations demonstrated by Rosch's (inpress-d) studies of color categories madereasonable the present attempt to use asimilar logic and similar procedures for aninvestigation of cognitive representations ofsemantic categories.

The nine experiments of the presentresearch were designed to form a logicalsequence. In Experiment 1, reliability ofsubjects' ratings of internal structure wasverified, and norms were collected for rat-ings of the goodness of example of 50-60members of 10 systematically chosen cate-gories. In Experiment 2, items (repre-sented by pictures in one condition and bywords in another condition) of high,

medium, and low goodness of example,according to the norms provided by Experi-ment 1, were used in a priming study witha 2-sec interval and same-category instruc-tions. The conditions of Experiment 2were designed to determine whether, undermaximally favorable conditions, any effectof internal category structure on the repre-sentation of the category could be observed.Experiment 3 was performed to checkwhether effects obtained in Experiment 2were the artifactual result of the fact thatsubjects did not consider the poor examplesof the categories members of those cate-gories . at all. Because an interaction be-tween goodness of example and primingwas achieved in Experiment 2 which didnot appear to be an artifact, Experiment 4was performed in which essential aspects ofExperiment 2 were replicated under condi-tions in which the prime and stimulus werepresented simultaneously. Following thelogic previously outlined, the results ofExperiments 2 and 4 indicated that theeffect of internal structure on same categoryand different responses lay partially in thedecision processes after the stimulus pairwas presented and that, in those conditions,effects of the prime on encoding and onsubsequent decisions could not be unequivo-cally disentangled. However, the effectsof the prime on physically identical pairs inthese experiments showed a selective effectof the representation generated by the cate-gory name prior to presentation of thestimulus—an effect very similar to thatwhich had been found for colors by Rosch(in press-d).

Subsequent experiments focused on ob-taining further information about the na-ture of the representation as it appeared toaffect physically identical pairs when theprime was presented prior to the stimulus.In Experiment 5, physical identity instruc-tions were employed to test the concrete-ness of the representation; the disappearanceof the effects of advance priming under suchinstructions indicated that different levelsof encoding of the stimuli were being used.This hypothesis was further verified by arecall task in Experiment 6. Experiment 7approached the question of whether the


meaning which appeared to be part of therepresentation generated by the categoryname was differentiated into pictorial andverbal form. In this experiment, both pairsof words and pairs of pictures were used asstimuli for the same subject, presented in aformat such that subjects were uncertainabout the form of the stimulus pair withwhich they would be presented on eachtrial. The results for Experiment 7 did notindicate a differentiation of the code forwords and pictures. However, since thelength of the 2-sec prime interval may havemasked subtle degrees of such differenti-ation, Experiment 8 was performed to chartthe time course of generation of word andpicture representations both under condi-tions in which the form of the ensuing stim-ulus was certain and uncertain. This experi-ment produced some evidence of differenti-ation of representations required in process-ing verbal and pictorial form, and someevidence that representations for pictureswere generated more rapidly than those forwords. The final experiment was designedto study the effect of long practice andthorough memorization on a small set ofgood and poor examples of categories.

EXPERIMENT 1

Before attempting to examine the effectof internal structure on mental representa-tions of semantic categories, it was necessaryto establish the reliability of judgments ofinternal structure for these categories. Thatis, it was necessary to gather reliable norm-ative data on subjects' ratings of the extentto which instances of semantic categoriesrepresent their idea or image of the mean-ing of the category name. In addition, ifthe nature and extent of concreteness ofrepresentations were to be explored, it wasnecessary to use concrete, picturable cate-gories in the studies. Previous subjects'ratings of internal structure (Rosch, 1973)were performed on only six items per cate-gory which had been prechosen to representan expected wide range in goodness ofexample. Furthermore, the categories usedin that study had not been systematicallychosen and included abstracts as well as

concrete categories. Rips, Shoben, andSmith (1973) used 12 instances of threecategories for their basic material. Thepurpose of the present experiment was todetermine reliability of ratings and gathernormative data for a systematically sampledset of concrete noun categories in commonuse in English. If reliable, these normswere intended for use in the subsequentexperiments on the nature of mentalrepresentations.

Method

Stimuli

The categories for which ratings of instanceswere to be gathered were chosen in the followingmanner: The population of categories of concretenouns in common use in English was determinedby drawing all concrete nouns with a word fre-quency of 10 or greater from the Kucera andFrancis (1967) sample of written English. Acategory was considered in common use if at leastfive items from the category appeared in that list.A category was considered concrete only if theitems in it could be unequivocally representedby pictures—thus, common categories such asrelative or number were not considered concrete.Categories were eliminated if (a) the items borea part-whole relationship to the only reasonablesuperordinate (e.g., parts of the body, parts ofbuildings), (b) if there was linguistic ambiguityamong possible superordinates (e.g., animal iscommonly used as a synonym for mammal), and(c) if the superordinate crosscut a large numberof other taxonomic structures (e.g., food). Sur-prisingly, only 17 concrete categories met theinitial frequency requirement, 7 of which wereeliminated by the other criteria. The remaining10 categories which were used in the normativestudy were: fruit, bird, vehicle, vegetable, sport,tool, toy, furniture, weapon, and clothing.

All of the categories which met the criteriawere categories which had been included in theBattig and Montague (1969) normative tabula-tions of the frequencies with which instances wereproduced in response to the category name (e.g.,all of the tools appear in the Battig & Montaguenorms under carpenter's tool). Thus, the Battigand Montague lists could be used as a basis forselecting those members of the categories whichwere to be rated in the present experiment. Foreach category, all of the items which had beenproduced by 10 or more subjects in the Battigand Montague study were included. Additionalitems to make a total of 50-60 instances weretaken randomly from the items produced by fewerthan 10 subjects in the Battig and Montaguenorms. Items that were jokes or obvious mis-hearings of the category name were excluded.

198 ELEANOR ROSCH

SubjectsSubjects were students in three psychology

classes who filled out the rating forms as partof their class work. On a front sheet, subjectslisted their country of birth and the country(ies)or state(s) in which they had lived since birth.Subjects who were not native speakers of Englishor who did not complete the forms were eliminatedfrom the study. Analysis was based on theforms of the 209 remaining subjects.

ProcedureAll members of a category were listed below

the category name. Subjects were asked to rate,on a 7-point scale, the extent to which eachinstance represented their idea or image of themeaning of the category term. Specific instruc-tions were:

This study has to do with what we have inmind when we use words which refer to cate-gories. Let's take the word red as an example.Close your eyes and imagine a true red. Nowimagine an orangish red ... imagine a purplered. Although you might still name the orangered or the purple red with the term red, theyare not as good examples of red (as clear casesof what red refers to) as the clear "true" red.In short, some reds are redder than others. Thesame is true for other kinds of categories. Thinkof dogs. You all have some notion of what a"real dog," a "doggy dog" is. To me a re-triever or a German shepherd is a very doggydog while a Pekinese is a less doggy dog.Notice that this kind of judgment has nothingto do with how well you like the thing; youcan like a purple red better than a true red butstill recognize that the color you like is not atrue red. You may prefer to own a Pekinesewithout thinking that it is the breed that bestrepresents what people mean by dogginess.

On this form you are asked to judge how goodan example of a category various instances ofthe category are. At the top of the page is thename of a category. Under it are the names ofsome members of the category. After eachmember is a blank. You are to rate how goodan example of the category each member is ona 7-point scale. A 1 means that you feel themember is a very good example of your ideaof what the category is. A 7 means you feelthe member fits very poorly with your idea orimage of the category (or is not a member atall). A 4 means you feel the member fitsmoderately well. For example, one of themembers of the category fruit is apple. Ifapple fit well your idea or image of fruit, youwould put a 1 after it; if apple fit your idea offruit very poorly you would put a 7 after it; a4 would indicate moderate fit. Use the othernumbers of the 7-point scale to indicate inter-mediate judgments.

Don't worry about why you feel that some-

thing is or isn't a good example of the category.And don't worry about whether it's just you orpeople in general who feel that way. Just markit the way you see it.

Approximately half (116) of the subjects re-ceived one random order of -• instances under thecategory name; the other half a different ran-dom order. Categories were typed on separatepages and each subject received a different orderof categories (taken randomly from the 10'I pos-sible orders of 10 categories).

Results and Discussion

Rank orders and mean ratings of good-ness of example for all instances of all cate-gories are shown in Table Al. To test thereliability of ratings both Spearman rank-order correlations and Pearson product-moment correlations of the mean ratingswere obtained (a) between split halves ofthe sample of subjects divided at random,(b) between subjects who rated the twodifferent item orders, and (c) between sub-jects who had lived predominantly on thewest (n = 131) versus east (n = 78) coastsof the United States. Consistency wasextremely high. All split-half correlationswere .97 or higher, and all correlationsbetween the west and east coast sampleswere .92 or higher. Agreement betweensubjects was particularly high for the itemsrated as very good examples of the category;for example, for 9 of the 10 categories, 95%of the subjects agreed in giving the itemwith the mean best example rating the samescore, that of 1. In addition, 6 categoriesoccurred in both the present study and inRosch (1973), making it possible to com-pare the rank order of the ratings whichitems received when they were the only 6instances of the category present and whenthey were embedded in almost the totalBattig and Montague (1969) list; for all 6categories, all rank orders were identical.

The results of this study clearly indicatethat semantic categories do have internalstructure: (a) Subjects consider it a mean-ingful task to rate members of such cate-gories according to how well they fit thesubjects' idea or image of the meaning ofthe category name and (b) there is highagreement between subjects concerningthese rankings.


Two precautions are necessary in theinterpretation of the data, however. First,the internal .structure that exists in theabove senses does not entail any particularinterpretation of the meaning of that struc-ture. Nor was the present researchdesigned to examine the relationships be-tween the structure of categories obtained bythe present instructions and structures ob-tained by other methods such as those ofBattig and Montague (1969). Experiment1 simply provides a reliable measure ofinternal structure (derived from instruc-tions given to the subjects which incor-porated the view of the nature of internalstructure elaborated in the author's pastwork) which can be used in the subsequentstudies of mental representations of cate-gories. Second, the norms were collectedon a somewhat restricted sample of subjects:college students enrolled in psychologyclasses. However, no claim is made that theinternal structure of semantic categoriesshould be universal for all cultural groups.What is important for the present study(and for other studies that might wish touse the present normative data) is that thepopulation on which these norms werecollected is the same population on whichfurther experiments using the norms as anindependent variable were performed. Thenormative data of Experiment 1 can thusbe used as the basis for examining whetherinternal structure of semantic categoriesaffects mental representations of the cate-gories.

EXPERIMENT 2

The most basic rationale for use of thepriming technique is that a prime can onlyfacilitate a response when it makes possiblegeneration of a mental code which containswithin it some of the information needed tomake the response. If the representationgenerated by a semantic category name issimilar to a list of criterial attributes or insome other way represents only a commondenominator of category membership (ofinclusion within the boundaries of thecategory), same responses to pairs of itemsof any degree of category membership shouldbe facilitated equally by priming with the

category name. On the other hand, if themental code is more like a representation ofa prototype or best example of the categoryor in some other way includes informationabout the internal structure of the categoryas part of the mental representation of thecategory, same pairs of good examplesshould be more strongly facilitated by theprime than same pairs of poor examples.

Experiment 2 was performed to testwhether there are any differential effects onpriming ascribable to level of goodness ofexample—operationally, whether any sig-nificant interaction would be achieved be-tween priming and level of goodness ofexample. The instruction condition usedwas that in which same was denned as be-longing to the same category since itappeared likely that for semantic categoriesthe greatest processing of meaning, andhence the greatest likelihood of the desiredeffect, would occur in this condition.

MethodSubjects

Subjects were 60 students in introductory psy-chology courses who received course credit fortheir participation. In this and all subsequentexperiments, only native speakers of English wereused.

StimuliA pilot study indicated that nine categories was

the largest number that could be tested withinthe 2-hr time period allotted for the experiment.Thus, one category was eliminated randomly: thecategory of clothing.

Four high-, medium-, and low-rated membersof each of the remaining nine categories wereused in the basic set of experiments. Constraintson choice of items were that the items not beconfusable between the nine categories used inthe study (i.e., items were pretested to assurelow error rates) and that the items chosen werethose which could be represented by readilyrecognizable simple line drawings. This latterconstraint prevented close matching of goodnessof example ranks across categories; however,choice of items which could be represented byunambiguous, traditionally representational linedrawings was felt to be of greater importance thanmatch in ranks. All of the items used in thenine categories are shown in Table 1.

Different groups of subjects were tested on theitems in Sets 1 and 2, which were intended asreplications over stimuli using the same categoriesand design. In addition, a third group of sub-

200 • ELEANOR ROSCH

TABLE 1SEMANTIC CATEGORY MEMBERS USED

Set Type Level of goodness of exampleHigh Medium Low

Furniture

1

2

Physically identicalSame category


chairdresser

tablebed

lampdesk

stooltelevision

rugstove

fancounter

Fruit

1

2



applebanana

pearorange

grapefruitcherries

strawberriespineapple

watermelonboysenberry

lemonnut

Vehicle

1

2



carbus

truckambulance

airplanebicycle

boattrolley car

sledwheelchair

tractorwagon

Weapon

1

. 2



pistolsword

kniferifle

arrowtomahawk

cannonclub

slingshotwhip

fistbow

Vegetable

1

2



peascorn

carrotsgreen beans

celeryturnips

tomatoesartichoke

mushroomspotatoes

green onionspumpkin

Carpenter's tool

1

2



hammerdrill

sawnail

wrenchbolt

ladderlumber

hatchetanvil

soldering ironscissors


TABLE I—(Continued)

Set TypeHigh

Level of goodness of exampleMedium Low

Bird

j Physically identicalSame category

2 Physically identicalSame category

robinsparrow

canaryblackbird

owleagle

parrothawk

penguinturkey

peacockchicken

Sport

< Physically identicalSame category


footballtennis

baseballswimming

archeryfishing

fencingping pong

chesshorseback riding

jump ropehunting

Toy

j Physically identicalSame category


dollball

teddy bearblocks

balloonskates

swingtricycle

mitttennis racket

animalssandbox

Clothing

« Physically identicalSame category

pantsshirt

stockingssandals

glovespurse

jects was tested on a third set of items; the thirdset was considered necessary because the cate-gory membership of some of the items in the firsttwo sets, particularly when presented in the formof words rather than pictures, could be con-sidered ambiguous. Especially problematical wasthe category toy, because virtually any concreteobject in any category can occur in a toy version.Set 3 was designed to be free of the possibleartifactual ambiguities in the first two sets; thisset consisted basically of the items in Set 1 butwith those items in the category toy replaced byitems in the category clothing (shown in Table 1)and with the single other Set 1 item whose cate-gory membership could be considered ambiguouswithin the context of the experiment (hatchet wasconsidered a tool but is also a reasonable weapon)replaced by an item which was unambiguous asa tool (soldering iron).

Each set of stimuli received by a subject con-tained three types of items: physically identicalpairs (e.g., the same word chair or the same pic-ture of a chair simply repeated) ; pairs whichwere not physically identical but which belongedto the same superordinate category (e.g., chair

and table), which under the instruction conditionsof the present experiment were to be consideredsame by the subject; and pairs of items whichbelonged to different superordinate categories(e.g., chair and apple). Items designated phys-ically identical in Table 1 were the items used toform the physically identical pairs. Each itemdesignated same category in Table 1 was com-bined with the item of that set and rank labeledphysically identical to form the same-categorypairs.

Different groups of subjects were tested on theitems in Sets 1, 2, and 3. These three sets werereplications of the design over different items.Each set of stimuli consisted of equal numbersof same and different items, half the sames (1/4the total pairs) consisted of physically identicalitems; half the sames (1/4 of the total pairs)belonged to the same category but were not thesame item; and half the pairs belonged to differ-ent categories. A set contained 108 pairs ofstimuli: (a) one high-, one medium-, and onelow-ranked physically identical pair from eachof the nine categories, (b) one high, medium, andlow same-category pair from each category, and (c)

202 ELEANOR ROSCH

two high-, medium-, and low-ranked differentpairs from each category. High-, medium-, andlow-ranked items were always paired with otheritems of the same rank . Different pairs, even oflow rank, were always selected so as to be easilyand unambiguously recognized as belonging todifferent categories.

Pairs of words were typed in IBM Executiveletters centered on 12.70 X 20.32 cm white cards.Under viewing conditions in the experiment, anaverage pair of words subtended a visual angleof 7°. Pictures were simple line drawings takenfrom various elementary reading instruction ma-terials. They were mounted, centered on 12.70 X20.32 cm white cards, spaced .635 cm apart, and,on the average, subtended a visual angle of 7°.No pair of words or pictures exceeded a visualangle of 9°. Both members of all pairs could beeasily perceived without eye movements.

While words differed in length and frequencyand pictures differed in size and complexity, allof which would be expected to affect reactiontime to a specific pair, these factors did not needto be carefully controlled in the present primingdesign. One of the advantages of the primingtechnique is that it contains a built-in controlfor factors such as familiarity or visibility whichcan affect reaction time to a pair as a whole.Each pair' occurred both in the unprimed andprimed condition with the variable of interest thedifference between them; other effects were notof interest unless they affected that difference.

Procedure

Six groups of 10 subjects performed the experi-ment; different subjects received each set ofstimuli, and different subjects received words andpictures.

An interval of 2 sec occurred between the primeand presentation of the stimulus, the interval usedby Seller (1971) and used in the basic experi-ments by Rosch (in press-d). Pairs were pre-sented in a two-field, Harvard-type tachistoscope.Each pair occurred twice, once preceded (primed)by the superordinate semantic category name 2sec in advance of the pair, once preceded by theword blank (to perform the function of a warn-ing signal, part of the function of a prime, seePosner & Boies, 1971). Primed and unprimedtrials alternated (this procedure was used bySeller, 1971). The category term or blank wasspoken by the experimenter who initiated the 2-sec interval as he terminated the word. Subjectswere required to repeat the category name orblank after the experimenter, a procedure whichfollowed Seller (1971) and Rosch (in press-d).In this, and all subsequent experiments, duringthe period preceding presentation of the stimulus,the subject was asked to fixate a cross whichoccurred in the center of a 12.70 X 20.32 cm whitefield.

Subjects were given the names of the ninecategories from which all items were drawn and

told that their task was to press one key toindicate same if the two items of a pair belongedto the same category and another key to indicatethat they belonged to different categories asrapidly as possible without error. Subjects wereexplicitly told that physically identical items (thesame word or picture repeated) were to be con-sidered same. Subjects used their dominant handfor the same key.

A practice session preceded the beginning oftesting. Practice items were drawn from thesame nine categories but were not items usedin the test sets. Practice sets contained 32 stim-uli : 9 physically identical items, 3 from eachrank; 9 same category items, 3 from each rank;and 18 different items, 6 from each rank. Eachpair was presented once primed' and once unprimedfor a total of 72 responses. The practice sessionfamiliarized subjects with the nature of the stim-uli, the judgment required of them, and themechanics of the apparatus and of making theresponse. Subjects received feedback on cor-rectness and response time during practice.

For the test trials, each subject received inrandom order (a different random order for eachsubject) all pairs, primed and unprimed, from allnine categories, a total of 216 items. A testsession lasted 2 hr.


Set 1 is the basic set of stimuli used inthe experiments throughout this program ofresearch. Set 2 was used in this experi-ment to provide a replication using differentitems with the same categories and proced-ures. Set 3 was used, not as a replicationper se, but as an additional test of whetherthe results achieved with Set 1 were alsoachieved when ambiguities possibly arisingfrom the inclusion of the category toy (andthe item hatchet in the category tools') wereeliminated. Thus, the basic analysis ofresults was performed on Set 1 with itsreplication Set 2. Results were then com-pared with the results for Set 3. If resultsdid not differ between Sets 1 and 3, thiswould constitute evidence that the findingsfor Set 1 and, thus, the basic findings forall of the following experiments based onSet 1 items were not simply an artifactengendered by the inclusion of one possiblyambiguous category and item in Set 1.

Results for both words and pictures forSets 1 and 2 combined are shown in Figure1. The overall error rate was 4%. Al-though error rates varied slightly by condi-tion in the directions reported by Seller


(1971), none of the tests of significance ofthe differences between error rates wassignificant. The most reasonable explana-tion for the lack of significant error ratedifferences in the present series of experi-ments is the lengthy practice session duringwhich incorrect responses were correctedby the experimenter. It is for that reason,probably, that subjects came to emphasizecorrectness in their responses with—as thespeed-accuracy trade-off would predict(Smith, 1968)—corresponding increases inreaction time for the more difficult items.Thus, reaction time appears to be the appro-priate variable for analysis; error rates arenot discussed further.

The predictions of primary interest werethe effect of priming and the interaction ofpriming with level of goodness of examplewithin the physically identical, same-cate-gory, and different conditions, rather thanpossible differences between those threeconditions. Therefore, the results for wordsand for pictures were analyzed separatelyfor responses to the physically identical,same-category and different stimuli; this isthe same procedure of analysis that hadbeen followed by Rosch (in press-d). Afour-way mixed-model analysis of variance(ANOVA) was performed for each of thethree types of stimulus pairs for words andfor pictures. Within-subjects effects werepriming, level of goodness of example, andcategory. Sets 1 and 2 were treated as areplication, a between-subjects effect. Be-cause no effects of replication proved signifi-cant, data for the two sets were combinedin Figure 1 and will not be discussed sep-arately. Category was treated as a fixedvariable since the nine categories used werevirtually exhaustive of the population ofcategories of concrete nouns in common usein English as defined in Experiment I.2

2 Since Clark (1973), there has been someconfusion about the exact conditions under whichit is appropriate to treat linguistic and other vari-ables as random rather than as fixed effects. Inthe present series of experiments, there was rea-son to argue that either form of analysis was themore appropriate. The categories used werevirtually exhaustive of the population to which theauthor wished to claim the results were appli-cable; less common categories could not neces-

For the physically identical pairs, theANOVA results confirmed that what ap-peared to be a large interaction between theeffects of priming and level of goodness ofexample of the pair, was in fact significantfor both pictures and words: pictures, F(2,36) = 7.95, p < .01; words, F(2, 36) =14.31, p < .001. The t tests verified thatfor both pictures and words, primed re-sponses were significantly faster than un-primed for the good examples: pictures,f(19) = 3.76, p < .01; words, f(19) =11.41, p < .001; and that primed responseswere significantly slower than unprimed forthe poor examples: pictures, £(19) = 2.51,p < .05; words, *(19) = 10.68, p < .001.Primed and unprimed responses did notdiffer significantly for the intermediateexamples for either stimulus type. Thus,as Rosch (in press-d) had found for colors,for physically identical stimuli in same-category instruction conditions, priming

sarily be expected to yield the reliable goodnessof example norms obtained in Experiment 1, and,as noted in the general discussion of this article,more abstract categories or categories at a dif-ferent level in the class inclusion taxonomies ofEnglish were predicted to yield systematicallydifferent results, in certain respects, from thoseof the present experiment. For all these reasons,it appeared mathematically unjustified to treatcategory as a random effect. On the other hand,the study carries with it the implication that itconcerns the processing of superordinate categoriesother than the specific ones used in this specificresearch. The following procedures of analysiswere therefore used: Category was treated asa fixed effect in the basic analyses of results, andit is these results which are reported in the bodyof this article. However, in addition, pseudo Fs(Clark, 1973; Winer, 1971) in which categorywas treated as a random effect were computedfor all of those effects which attained significancewhen category was treated as a fixed effect. F'was computed by the conservative minimummethod described in Clark; min F' is a statisticalways smaller than the Fs from which it iscomputed. In actual fact, all of the effects ofinterest on which the theoretical conclusions ofthe research are based remained significant undermin F' computation. In the body of this article,any reported significant effect which was not sig-nificant (p < .05) by the min F' computation isso noted in parentheses after the report of the Fvalue. (It should be noted that category effectscannot be computed at all when category istreated as a random effect.)

204 ELEANOR ROSCH

PHYSICALLY SAME DIFFERENT

1300

1200

1100

"5 1000t/>5CD

•| 900cO

0

8. 800

700

600

500

IDENTICAL

_Pictures :

CATEGORYAy

.A^A

> Unprimed s0--- Primed /**

- Words: /A— Unprimed /&--- Primed

f///

/ /A / /

>/

-/

/»

*^~^r<y~

i i i

/

Q

f

~ ^''

I 1

-

_ ^^^

^^^

*~~~ -~t

Cf"~^_

-

J "*

•— "*.0— 5— -C

-

1 1 |

High Med Low High Med Low High Med LowLevel of goodness-of-example

FIGURE 1. Two-sec prime interval, same-categoryinstructions, words and pictures separate.

with the category name facilitated responsesto good members of the category andactually hindered responses to poor ex-amples. Intermediate examples appearedtruly intermediate in that priming neitherfacilitated nor hindered them.

Some other effects for the physicallyidentical pairs also attained significance.The main effect for level of goodness ofexample was significant for both picturesand words: pictures, F(2, 36) = 9.87, p <.001; words, F(2, 36) = 19.64, p < .001.As can be seen from Figure 1, the poorexample pairs required somewhat longerreactions when unprimed as well as whenprimed than did the good or intermediateexamples. These longer reactions to un-primed poor example pairs than to goodexamples are not surprising in the light ofthe fact that, for the stimuli used in thisseries of experiments, word length, fa-miliarity, visibility, and complexity werenot equated between levels of goodness ofexample, but rather differed slightly (as

can be partly observed from Table 1) infavor of the good examples. The intentionwas to control such factors, not in thestimuli, but in the experimental design, bythe occurrence of each pair both in the un-primed and primed condition and by thearrangement of the difference between thosetwo conditions as the variable of primaryinterest. The differential effect of primingon goodness of example (the interactionbetween the two variables), not the possibleeffects of goodness of example itself, was ofmajor interest for the physically identicalpairs.

The main effect for category for thephysically identical pairs was significantfor both pictures and words: pictures, F(8,152) = 3.07, p < .001; words, F(8, 152)= 3.59, p < .001. However, since none ofthe interactions with category reached sig-nificance, the effect is not of interest to thepresent study. The main effect of primingwas not significant for either pictures orwords: pictures, F(l, 19) = 1.57; words,F(l, 19) = 2.06.

In order to test whether the results forphysically identical pairs were artifacts ofthe inclusion of some possibly ambiguousitems, the above analyses were also per-formed on Set 3 alone, and on Sets 1 and 3together, with set treated as a between-subjects effect. The results for Set 3followed the identical pattern as thosereported above: The Priming X Level ofGoodness of Example interaction was sig-nificant for both words and pictures, maineffects of goodness of example and of cate-gory were significant, and t tests were sig-nificant in the same directions as thosereported above. Furthermore, the effectof set when Sets 1 and 3 were analyzedtogether was not significant. Thus, themain results for the experiment did notappear to be a simple artifact of the fewpossibly ambiguous items in Sets 1 and 2.

The basic finding for the physically iden-tical stimuli of interest to the presenthypothesis was, thus, the interaction be-tween effect of priming and level of good-ness of example of the stimulus.

The results for the same-category pairspresented a quite different picture. The


Priming X Level of Goodness of Exampleinteraction was not significant either forpictures or words, while the main effect ofpriming was significant for both types ofstimulus: pictures, F(l, 19) = 21.90, p <.001; words, F(l, 19) = 20.04, p < .001.Thus, for same-category pairs, primingfacilitated responding in general and wasas helpful for the intermediate and poorexamples of the category as for the goodexamples.

Other effects which attained significancefor the same-category pairs were similar tothose significant for the physically identicalpairs. The main effect of level of goodnessof example was significant for both picturesand words: pictures, F(2, 36) = 9.24, p <.001; words, F(2, 36) = 8.90, p < .01.The main effect of category was also sig-nificant for both pictures and words: pic-tures, F(8, 152) = 4.97, p < .001; words,F(8, 152) = 5.18, p < .001. None of theother effects for the same-category pairs wassignificant. Effects for Set 3 were analyzedin a manner identical to that described forthe physically identical pairs. As in thecase of the physically identical pairs, therewere no differences between the effects ofSet 1 and Set 3.

For different pairs, the main effect ofpriming was significant for both types ofstimuli: pictures, F(l, 19) = 8.09, p < .05;words, F(l, 19) = 7.58, p <.05 (min F'not significant). None of the interactionswith priming was significant. The effect ofcategory was significant for words, F(8,152) = 3.96, p < .01, but not for pictures,F(8, 152) = .94, ns. None of the inter-actions with category was significant. Levelof goodness of example was significant forboth pictures, F(2, 36) = 3.51. p < .05(min F' not significant), and for words,F(2, 26) = 3.82, p < .05 (min F' not sig-nificant) . As in the case of physically iden-tical and same-category pairs, effects of Set3 did not differ from those of Set 1.

Of themselves, the results of Experiment2 establish two points. First, they showthat priming with superordinate semanticcategories can have significant effects in asame-different matching task; that is, theresults establish the possibility of using this

technique to investigate these types of stim-ulus materials. Second, the results suggestthat the representation of superordinatesemantic categories is something more thana list of criterial attributes which, by defini-tion, must be common to all members of thecategory and which should thus be affectedequally by priming for all members of thecategory. The actual findings were that forone type of stimulus, physically identicalpairs, priming speeded responses to goodexamples of the category and slowed re-sponses to poor examples.

However, two additional tests wereneeded before these results could be furtherinterpreted. First, the present experimentswere claimed to be investigations of theinternal structure of semantic categories;one of the basic assertions is that all of theinstances of a category are indeed membersof the category, differing only in the extentto which they represent subjects' ideas orimages of the meaning of the category name.However, it is possible to argue that thepoor examples of categories used in Experi-ment 2 were, for subjects, actually mem-bers of different categories entirely. Posnerand Snyder (1975) have shown that when,in a priming task, subjects are "fooled" on20% of the primed trials (i.e., are shown apair which should be designated sameaccording to the instructions, but which ismisnamed by the prime), priming facili-tates response time on the trials for whichthe prime is correct and inhibits responsetime for the trials on which the prime isincorrect. Results of the present experi-ments would be far from interesting werethey obtained simply because the poorexample items were actually members ofother categories and thus incorrectly primedby the category name used (as is suggestedin the Loftus commentary). It appearednecessary, therefore, to compare the resultsobtained by use of the present experimentalparadigm with those obtained when actualincorrect primes were administered. Thatcomparison is carried out in Experiment 3.

The second test which is needed beforethe results of Experiment 2 can be furtherinterpreted is inherent in the logic of thepriming technique, as discussed in the intro-

206 ELEANOR ROSCH

duction to this article and as demonstratedby the pattern of results in the experimentsin Rosch (in press-d). Interpretation ofthe results of priming with a 2-sec delay(either of the predicted results for thephysically identical pairs or the anomalousresults for the same-category pairs) canonly be made when compared to results ofa prime administered simultaneously withthe stimulus. Experiment 4 was designedto provide the necessary comparison.

EXPERIMENT 3

Keele (1973) and Posner and Snyder(1975, in press) make a distinction betweenthe effects of conscious attention—whichdirects scarce processing capacity to anattended item, thus producing facilitationof performance for that item and a decre-ment in performance for unattended items—and the effects of unconscious activation—which can produce facilitation for one itemwithout a corresponding decrement forother items. In one set of experimentsdesigned to support the distinction, a prim-ing technique was used in which the pro-portion of trials in which the prime pro-vided the subject with correct informationwas systematically varied. When the prob-ability that the prime was informative washigh (e.g., 80% of the primes were correct;20% incorrect), subjects showed a gain inboth speed and accuracy of responses tocorrectly primed items and a loss in speedand accuracy of responses to incorrectlyprimed items. On the other hand, when theprobabilities were reversed (20% of theprimes were correct, 80% incorrect), therewere small but significant facilitation effectson the correctly primed trials but no cor-responding decrements on the incorrectlyprimed trials. These differences wereassumed to support the differential effectsof conscious and unconscious activation.

The empirical pattern of results in thecondition in which the majority of trials(80%) were correctly primed and a minor-ity (20%) incorrectly primed was similarto that obtained for the physically identicalitems in Experiment 2 of the present re-search. Were the results of Experiment 2obtained simply because the poor example

items (33% of the items) were actuallymembers of other categories and, thus, in-correctly primed by the category name used ?The purpose of Experiment 3 was to testsuch a possibility by use of an experimentalparadigm similar to that of Experiment 2 butin which, on 33% of the primed trials, acategory name was given which was knownto be an incorrect designation of the items.That is, in Experiment 3, the basic designof Experiment 2 was repeated with theexception that the poor example items werereplaced by items from a category whichwas unequivocally not named by the prime.If Experiment 3 is a true analogue of Ex-periment 2—if it embodies the elements ofdesign which led to the results of Experi-ment 2—the pattern of results obtained inExperiment 3 should be very similar to thoseobtained in Experiment 2.

Method

Subjects

Subjects were 20 native English-speaking stu-dents in introductory psychology classes who hadnot participated in any of the other experiments.They received course credit for their participation.

Stimuli and Procedure

Sets 1 and 3 of Experiment 2 were used asstimuli; the single change in stimulus materialswas that all of the poor example pairs werereplaced by good example pairs taken from cate-gories other than the one which was primed. (Theunprimed pairs corresponding to the changedprimed pairs were, of course, also changed.) Tensubjects received Set 1, 10 subjects Set 3. Pro-cedures of administration and practice wereidentical to those of Experiment 2.


Results for correctly and incorrectlyprimed good example items, for the two setscombined, are shown in Figure 2. Theresults were analyzed in the same manneras described for Experiment 2; the effectof set was treated as a between-subjectsvariable. Incorrectly primed items filled thesame position in the design (one level ofthe goodness of example variable) that thepoor example items had filled in Experi-ment 2. The results of the ANOVAs weresimple: For none of the types of pairs(physically identical, same-category, or dif-


ferent) did any of the effects (priming,goodness of example, or the Priming XGoodness of Example interaction) achievesignificance. The only significant effectsof interest were those of two t tests. Thesetests indicated a significant facilitation effect(i.e., decrease in reaction time) for cor-rectly primed good example pairs of physi-cally identical pictures, f(19) = 2.66, p <.05, and significant facilitation of reactiontime for correctly primed good examplepairs of same-category words, t(\9) = 2.28,p < .05. There were no significant effectsfor error rates, perhaps, as in Experiment2, because of the emphasis on correctnessapparently conveyed by the feedback in thepractice session.

The results of Experiment 3 were inmarked contrast to those of Experiment 2.None of the effects of interest in Experi-ment 2 (the Priming X Level of Goodnessof Example interaction for physically iden-tical pairs and the general facilitation bypriming for both good and poor examplesame-category pairs) was obtained in Ex-periment 3. In particular, the decrement(i.e., increase) in response time for primedphysically identical poor example pairs, theeffect which had appeared possible to at-tribute to the fact that those items were notreally members of the primed category, wasnot obtained in Experiment 3 when thoseitems were unequivocally not members ofthe primed category. Thus, the results ofExperiment 2, particularly the Priming XGoodness of Example interaction for phys-ically identical pairs, do not appear to beartifacts of the possible fact that poor ex-ample pairs actually belonged to categoriesother than the ones named in the prime.

Why should priming with an incorrectcategory name not have suppressed reactiontime for the physically identical pairs in Ex-periment 3 ? Such a finding appears to fitthe Keele (1973) and Posner and Synder(1975, in press) models of attention. Theresults of Experiment 3 (for both physicallyidentical and same-category items) aresimilar to those obtained by Posner andSnyder (1975) in the condition in whichthe majority of primed trials (80%) wereincorrectly primed. Posner and Snyder

PHYSICALLYIDENTICAL

SAMECATEGORY

DIFFERENT

1200

1100

1000

os5 900

• 800

700

600

500

Pictures:. • Unprimed

O--- PrimedWords;

A— Unprimed. A-— Primed

Match Mismatch Match Mismatch Match MismatchMatch to prime

FIGURE 2. One third of items incorrectly primed,2-sec prime interval.

attribute the small facilitation and lack ofdecrement obtained in this condition to thefact that subjects came not to attend con-sciously to the prime. Subjects in the pres-ent experiment commented during practiceon the disruptive effects of being "fooled" bythe prime and reported ignoring the prim-ing when performing the experiment. Thedifference between the present experimentand Posner and Snyder (1975) is in thelow percentage of incorrectly primed trialswhich appeared able to produce withdrawalof conscious attention from the prime in thepresent experiment. The design of thepresent experiment differed from that ofPosner and Snyder (1975) both in that thematching task was more complex than theirtasks and in that the prime and stimuli werepresented in different modalities (the primewas auditory, the stimuli visual), whereasin Posner and Snyder's tasks both prime andstimuli were visually presented. Either orboth factors may have contributed to sub-jects' inattention to the prime under thepresent experimental conditions. (Otherexamples of lack of processing of the primeby subjects in the present research designoccurred in Experiments 5 and 6).

208 ELEANOR ROSCH

While Experiment 3 has indicated thatthe effects of the internal structure of cate-gories on priming cannot be dismissed as anartifact of the fact that poor examples ofcategories are actually members of othercategories, it is undeniable that the poorexamples of the categories used in the pres-ent research were "closer" to other cate-gories than were the good examples. How-ever, this was not due to an artifactualchoice of particular items; rather, it is anecessary aspect of the structure of realworld categories. That is, categories appearto form in the real world in a manner whichrenders them maximally discriminate fromeach other, and, thus, the best examples ofcategories are those items both with themost attributes in common with other mem-bers of the category in question and withthe least attributes in common with, or leastpossibility of membership in, other cate-gories; such facts have been demonstratedfor both natural and artificial categories(Rosch & Mervis, in press; Rosch, Mervis,Gray, Johnson, & Boyes-Braem, in press;Rosch, Simpson, & Miller, Note 1). Thesefacts are some of the structural bases whichmay lie behind the norms of Experiment 1 ;the purpose of the present research, how-ever, was not to investigate the formationof categories but to examine the nature (i.e.,the structure and content) of the cognitiverepresentation of the category (as it is gen-erated from hearing the category name).What appears to be the case is that thenatural structure of categories (which isquite different from a list of attributes neces-sary and sufficient for category membership)is represented in the cognitive representa-tion of the category.

Experiment 3 has clarified one aspect ofthe meaning of the results of Experiment 2 ;namely, it has shown that the results of Ex-periment 2 were not an artifact of the useof poor example items which were actuallymembers of categories not named by theprime. However, there remains a secondaspect of Experiment 2 for which compar-ison data must be obtained before the resultsof Experiment 2 can be interpreted. Ac-cording to the logic of priming outlined inthe introduction, interpretation of the results

of priming with a 2-sec delay (the conditionof Experiment 2) can only be made whencompared to results obtained when the primeis administered simultaneously with thestimulus. Experiment 4 was designed toprovide the necessary comparison.

EXPERIMENT 4

In Experiment 2, priming selectivelyfacilitated and impeded responses to goodand poor examples of categories for phys-ically identical pairs, and it had a constantfacilitatory effect on same-category and dif-ferent pairs, regardless of level of goodnessof example. Any of these outcomes, how-ever, could as well have been the result ofeffects on decision processes after the stimuliwere presented as the result of generation ofa representation of the category prior topresentation of the stimuli. The presentexperiment was designed to test the locus ofthe facilitation effect by simultaneous pre-sentation of the prime and the stimuluspair. Priming effects which are the resultof generation of a representation prior toviewing the pair should disappear if theprime and the visual stimulus are presentedsimultaneously. On the other hand, primingeffects which are at least partly the resultof decision time after the pair is presentshould still occur when the prime and visualstimulus are presented simultaneously.

Method

SubjectsSubjects were 20 native English-speaking stu-

dents in introductory psychology classes who hadnot participated in the previous experiments.They received course credit for their participa-tion.

Because in the previous experiments, effects ofSet 1 had been shown not to differ from those ofSet 2 (or Set 3), it appeared justifiable to useonly the Set 1 stimuli in the present experiment.

ProcedureWith two exceptions, procedures were identical

to those in Experiment 2. In the present experi-ment, instead of a 2-sec interval between theprime or blank and presentation of the stimuluspair, the pair of stimuli were presented to thesubject simultaneously with the experimenter's


utterance of the category name or blank. In orderto assure that utterance of the prime or blankand onset of the stimulus were, in fact, simul-taneous, the experimenter first engaged in twopractice sessions, the second of which was tape-recorded. The recording was checked to ascer-tain that the click denoting the relay whichinitiated stimulus onset invariably occurred duringthe interval of sound denoting the experimenter'sutterance. This was, in fact, found to be the case.The 20 test sessions were also tape-recorded andthe first and last sessions were similarly checked.Of the 432 trials represented by the two sessions,there was difficulty on only one trial in which thebeginning of the experimenter's word was ob-scured by a cough.

The second change in procedure was elimina-tion of the requirement that the subject repeatthe prime or blank after the experimenter, a changemade necessary by the fact that when the primeand stimuli were presented simultaneously, a sub-ject's repetition of the word interfered with hisability to make a response. In Rosch (in press-d)the patterns of results with and without subjectrepetition were found to be identical. In orderto assure that the omission of subject repetitiondid not change the effects found for semanticcategories in Experiment 2, five subjects weretested with pictures and five with words usingSet 1 stimuli and Experiment 2 procedures modi-fied by the omission of the repetition requirement.As in Rosch (in press-d) the same pattern ofresults was obtained from these subjects as fromthe original groups in Experiment 2. This testingserved both as a replication of the Experiment 2findings and as a pilot study making possible theprocedures of Experiment 3.


Results for reaction times are shown inFigure 3. Overall error rate was 6%. Asin previous experiments, there were no sig-nificant changes of error rate in any of theconditions. Results were analyzed as inExperiment 2. For the physically identicalcondition, the Priming X Level of Good-ness of Example interaction was not signifi-cant for either pictures or words. Thus, asfor colors (Rosch, in press-d), presentationof the prime simultaneously with the phys-ically identical stimuli eliminated both thefaciliatory and inhibitory effects of priming.The only remaining significant effects underthis condition were the main effects of levelof goodness of example: pictures, F(2, 38)= 6.03, p. < .01; words, F(2, 38) = 5.85,p < .01; and the main effect of category:pictures,"F (8, 152) = 2.60, p < .05 (min

PHYSICALLY)400 IDENTICAL

BOO

SAME._ CATEGORY

DIFFERENT

600 L_L

——If'


FIGURE 3. Simultaneous prime, same-categoryinstructions, words and pictures separate.

F' not significant) ; words, F(8, 152) =2.71, p < .01 (min F' not significant).

According to the logic of the presentresearch, if a prime facilitates a responseprior to but not simultaneously with astimulus, the representation generated bythe prime in advance of the stimulus canbe inferred to affect the response. Thephysically identical stimulus condition thusis clearly a condition which allows us toinvestigate the primary topic under study:the nature of the representation generatedwhen we hear a category name and under-stand its meaning. The import of thefindings for physically identical stimuli willbe discussed further below.

For the same-category condition, theresults were quite different. The main effectof priming was significant for both picturesand words: pictures, F(l, 19) = 23.80, p< .001; words, F(l, 19) = 22.29, p < .001.The Priming X Level of Goodness ofExample interaction was not significant. Asin the physically identical condition, maineffects of level of goodness of example weresignificant: pictures, F(2, 38) = 6.47, p <

210 ELEANOR ROSCH

.01; words, F(2, 28) = 7.10, p < .01. Andcategory was significant: pictures, F(8, 152)= 4.01, p < .001; words, F(8, 152) = 4.77,p < .001. No interactions were significant.

For different responses, effects were verysimilar to those found for the same-categoryresponses. For both pictures and words,there was a significant main effect of prim-ing: pictures, F(l, 19) = 4.52, p < .05(rain F' not significant); words, F(l, 19)= 8.26, p < .01. There was no Priming XLevel of Goodness of Example interaction.As in previous conditions, the main effect oflevel of goodness of example was significant;pictures, F(2, 38) = 5.20, p < .01; words,F(2, 38) = 4.13, p < .05 (min F' not sig-nificant). And the main effect of categorywas significant: pictures, F(8, 152) = 2.78,p < .01 (min F' not significant); words,F(8, 152) = 4,17, p < .001. Interactionswere not significant.

The patterns of results for the same-category and different stimuli were quitedifferent and appear to warrant a quite dif-ferent interpretation from the results forphysically identical stimuli. In these cases,priming affected reaction time both whenpresented prior to and simultaneously witha stimulus. The following discussion isdesigned to demonstrate that in such a case,due to the possible interaction between stagesof processing (which rendered Donders',1969, additive model of reaction time in-operable; see Sternberg, 1969), it is im-possible to verify any specific model of theprocessing of items in the task. Let us takeas an example the case of the same-categoryitems.

There are three major ways in which theresults for same-category items differ fromthose for physically identical items: (a)The overall reaction time for same-categoryitems, both unprimed and primed, wasgreater than for the physically identicalitems, (b) There was uniform facilitationof response time both for good and poorexamples of the category for the same-category items; whereas for physicallyidentical items, there was facilitation onlyfor primed good examples and a decrementin response time for the primed poorexamples, (c) For same-category items,

the effect of advance priming was not elim-inated when the prime was presentedsimultaneously with the items; whereas inthat condition it was eliminated for phys-ically identical items.

The difference in overall reaction time tothe two types of items is relatively trivial.To respond to same-category items, the sub-ject must not only go through the encodingprocess required for all items but must, inaddition, perform judgments not requiredfor physically identical items; that is, thesubject must, in some form, determine cat-egory membership for the items and judgewhether or not the category of both items isthe same. Same-category items have con-sistently been found to require longer reac-tion times than physically identical itemsunder a wide variety of conditions (Posner& Mitchell, 1967). This effect is not ofinterest per se for the present study.

It is the second difference which is ofmajor interest: Why should advance prim-ing facilitate response time equally for goodand poor same-category pairs when it doesnot have such an effect on physicallyidentical pairs? The third difference, theoccurrence of priming effects for same-category items under conditions of simul-taneous priming, is of interest primarily forits ability to cast light on the seconddifference.

If it is assumed that there are two basic,separate stages in the processing of same-category items, an encoding stage in whichthe subject perceives the stimulus item anda decision stage in which the subject makesdecisions about the category membership ofthe items, then there are two basic types ofstage models which can characterize theeffect of the prime on same-category items.In a two-stage model, the observable effectof priming is attributed both to the effectof the prime on the encoding stage and toits effect on subsequent decision processes.In a one-stage model, the observable effectof the prime is attributed only to thedecision stage.

Consider the general form of a two-stagemodel. When the prime is presented priorto the stimulus pair, it should affect encod-ing of the pair in an identical manner tQ it§


effect on physically identical pairs, speedingresponses to good examples and slowingresponses to poor examples (as shown inFigure 1). That is, since the subject doesnot know in advance what type of pair hewill see, priming should affect encoding ofall pairs equally. For same-category items,the subject perceives (or judges) that themembers of the pair are not physicallyidentical (which may or may not be aseparate step from perceiving the itemsthemselves), and must now proceed toconsider category membership. If the trialis not primed, the subject must retrieve thesuperordinate category names from mem-ory. (He may either retrieve the super-ordinates for each member separately, com-paring them to determine whether they arethe same, or, as suggested by Smith, Shoben,& Rips, 1974, he may retrieve the super-ordinate name for the first member of thepair and then confirm or disconfirm thatthe second member belongs to that samecategory; which process is used is irrelevantto the present argument). Since poor mem-bers of categories are slightly more difficultto perceive than good members (see Experi-ment 2) and since superordinates for poormembers are probably more difficult toretrieve from memory than for good mem-bers (indirect evidence in Loftus & Scheff,1971, and Rosch & Mervis, in press), longerresponse times should be expected for poormembers of unprimed same-category pairsthan for good members. Such an effect ofgoodness of example on unprimed pairs was,in fact, obtained (see Figures 1 and 3).

The two-stage model must now explainhow the Priming X Level of Goodness ofExample interaction obtained in the encod-ing stage is reversed in the decision stagesuch that the final effect is an equal facilita-tion of primed responses for good and poorcategory members. Priming provides asuperordinate category name, thereby elim-inating the need for the memory retrievalof a superordinate (or superordinates) andchanging the retrieval task into a verificationtask of the type in which subjects answertrue or false to statements of the form "A(member) is a (category)." Previous workhas already established that category mem-

bership of good members of categories isverified faster than that of poor members(Rips, et al, 1973; Rosch, 1973); that factwould of itself tend to increase, rather thaneliminate, the Priming X Level of Goodnessof Example interaction. In order to elim-inate the interaction, it must be assumedthat: (a) retrieval of superordinates of poormembers of categories is a very much moredifficult task than retrieval of superordinatesof good members, (b) because retrieval ofpoor category members is so difficult, thegain in reaction time obtained by turning theretrieval task into a verification task isgreater for poor members than for goodones, and (c) in fact, this relatively greatergain in time for poor members obtained inthe change from retrieval to verification isgreat enough to offset both the decrement inresponse time for poor members created bypriming in the encoding stage and the rela-tively slower verification times themselvesfor poor than for good members. Thatretrieval of poor members is indeed con-siderably more difficult than for good mem-bers can be argued from the finding in Roschand Mervis (in press) that, due to thestructural nature of categories, poor mem-bers have more attributes in common withand belong to more superordinate categoriesthan good members. For the sake of argu-ment, the offset of this retrieval difficultyby priming of poor members may be as-sumed to be just great enough to counter-act the interactions expected from encodingand verification times. Only such a quan-titative equalization could produce the uni-form facilitation of good and poor membersof same-category pairs shown in Figures 1and 3. Such a two-stage model is quitesimilar to that offered by Loftus (in hercommentary on this article) with the excep-tion that the present account makes explicitthe processes which must be assumed if theencoding stage Priming X Level of Good-ness of Example interaction is to be offsetin the decision stage.

The problem with such a model is that itdoes not fit the data; that is, it does not fitExperiments 2 and 4 considered together.If the reaction times for same-category pairsin Experiment 2 are due to the effect of the

212 ELEANOR ROSCH

prime on encoding plus the effect of theprime on decision processes, there shouldbe systematic differences between the casein which the prime is presented 2 sec in •advance of the stimulus and thus affectsencoding (Experiment 2), and the case inwhich prime and stimulus are presentedsimultaneously and the prime does notappear to have sufficient time to affect en-coding (Experiment 4). If there is noPriming X Level of Goodness of Exampleinteraction during encoding, primed goodexamples would expectably be relativelyslower and primed poor examples would berelatively faster than in the case where thereis an encoding stage interaction to counter-balance. Thus, for the same-category pairs,responses to primed good examples wouldexpectably be slower in Experiment 4 thanin Experiment 2 and responses to primedpoor examples would be faster in Experi-ment 4 than in Experiment 2.

To test possible differences for same-category items in Experiments 2 and 4, amixed-model, four-way ANOVA (Prim-ing Interval X Priming X Level of Good-ness of Example X Category) was per-formed separately for words and for picturesfor the same-category responses of Experi-ments 2 and 4. Priming interval (2 secvs. simultaneous), the variable of greatestinterest, was a between-subjects variable;the other variables were within-subjects.The results show that there were no effects,either main effects or interactions, of prim-ing interval for either pictures or words.As an additional check on the existence ofpossible differences between the effects ofpriming in the 2-sec and simultaneous con-ditions for both pictures and words, t testswere performed on the differences betweenprimed and unprimed responses for the 2-secand simultaneous conditions separately foreach level of goodness of example. Noneof the six t tests was significant. Thus, thetwo-stage processing model fails to fit thefacts of lack of difference between same-category pairs under 2-sec delay and simul-taneous priming conditions.

The data from Experiments 2 and 4 leadto consideration of a one-stage model of theprocessing of same-category pairs, a model

in which priming is considered to affectonly the decision stage for such items. In aone-stage model, the encoding stage neednot be considered, and, thus, a theory whichgenerates the reversal (counterbalancing)of the encoding stage Priming X Level ofExample interaction is not needed.

How might a one-stage model work?For unprimed trials, poor examples wouldbe expected to require longer responsetimes than good examples for the samereasons enumerated in describing the two-stage model. To explain the effect of prim-ing, a one-stage model need only explainwhy primed responses are generally fasterthan unprimed and why primed responsesto good examples are faster than primedresponses to poor examples. Primed re-sponses would be generally faster than un-primed if priming can be assumed to turn amemory search plus verification task into apure verification task, and if verificationalone is faster than search plus verification.As explained earlier, both are reasonable as-sumptions. Primed responses to good ex-amples would tend to be faster than primedresponses to poor examples both as a re-sult of the fact that, in the present ex-periment, unprimed good examples havefaster response times than unprimed poorexamples and as a result of the knownfact that verification of good examples ofcategories is faster than verification of poorexamples (Rips et al., 1973; Rosch, 1973).By these few assumptions, the uniformlybeneficial effect of priming for same-cate-gory pairs is accounted for. The one-stagemodel has the advantage that it fits thedata for Experiments 2 and 4 consideredtogether. It also has the advantage thatit is simpler than the two-stage model andneed not make "strained" quantitative as-sumptions about the relatively greater ad-vantages of verification over search for poorthan for good category members.

The one-stage model has the disadvan-tage that it is a thoroughly unreasonableaccount of the processing of same-categorypairs under conditions in which the primeis presented 2 sec in advance of the stimuluspair. Given that under these conditionspriming affects the encoding of physically


identical pairs, and given that subjects donot know in advance whether they will seea physically identical or a same-categorypair, it is unreasonable to assume that prim-ing does not also affect the encoding ofsame-category items. Thus, neither modelappears to offer a viable account which fitsall of the data.

A very similar logic can be applied to thedifferent pairs. A two-stage model predictsdifferences between Experiments 2 and 4which failed to occur. A single-stage modelis an unreasonable account of the results ofExperiment 2.

The only reasonable conclusion appearsto be that we cannot have an additive modelof encoding and decision time stages for apriming task of this nature. Encoding anddecision stages appear to interact in thosecases in which decisions about categorymembership are required (same-categoryand different items). Given present con-ceptions of cognitive processes and the lackof quantification of the stages hypothesizedfor those processes, we cannot make infer-ences about the occurrence or nonoccur-rence of processes when interactions suchas these appear to be present. Thus, theconclusion from the results of Experiments2 and 4 taken together must be that it isonly in the case of physically identical itemsthat we can make inferences about the basicquestion of the study: the nature of thecognitive representation generated whenwe hear a category name and understandits meaning.

What have the results of Experiments 2and 4 shown about the nature of that rep-resentation? An assumption of the prim-ing method is that a prime can only facil-itate a response if it contains within it someof the information needed to make thatresponse. Because of the selective effectsof internal structure on priming, the facili-tation of responses for good and inhibitionof responses for poor category members, itis clear that the representation containsmore of the information needed to respondto good than to poor category members.But it is not clear in what form that infor-mation is given. Because advance but notsimultaneous priming affected physically

identical pairs, it can be inferred that therepresentation generated by the super-ordinate category name contained some ofthe information used by subjects in theactual perception of the stimulus. But thatdoes not tell us much about the form ofthe information. The information must beabstract enough to have similar facilitatoryand depressive effects on very different itemsbelonging to similar levels of goodness ofexample of a category and similar effects onstimuli presented in the very different visualforms of words and pictures; yet the infor-mation must be concrete enough to affectthe actual perception of specific items. Thefollowing two experiments were designed toinvestigate the issue of the concreteness ofthe representation and its effect on differentlevels of encoding.

EXPERIMENT 5

How concrete, how specifically visual, amental representation can be produced fromhearing the superordinate semantic categoryterms used in the present series of studies?Can subjects actually generate visual con-figurations which are aspects of possible pic-tures and words?

Experiments 2 and 3 used instructionswhich defined same as belonging to thesame category. A series of experiments(summarized by Posner, 1969; Beller &Schaeffer, Note 2) have demonstrated thatstimuli in a matching task are processed ata far more concrete and physical level wheninstructions define same as physical identitythan when instructions define same as samecategory. Rosch (in press-d) found thatthe representation generated by color cat-egory names produced the same selectiveeffects on physically identical pairs underboth types of instruction, arguing that colorterms can generate a quite concrete mentalrepresentation. If the representationsgenerated by superordinate category namescontain some concrete elements (e.g., rep-resentations of features such as possiblelines and angles of pictures or letters inwords), differential effects of priming forphysically identical pairs should be foundunder physical identity instruction condi-tions which are similar to those found for

214 ELEANOR ROSCH

PHYSICALLY IDENTICAL DIFFERENT

Pictures: Words:•—— Unprimed A UnprimedO-— Primed A--- primed

70°

I 600

500High Med Low High Med

Level of goodness-of-exampleLow

FIGURE 4. Physical identity instructions, 2-secprime interval, words and pictures separate.

physically identical pairs under same-category instruction conditions. The pres-ent experiment used physical identity in-structions.

Method

Subjects

Subjects were 40 students in introductory psy-chology classes who had not participated in anyof the previous experiments. All were nativespeakers of English. They received course creditfor participation.

Stimuli

Stimulus pairs were the same as those used inthe physically identical and different pairs ofExperiment 2. In order to have the same numberof pairs as used in the previous experiments andin order to keep the number of same and differentresponses equal, it was necessary to double thenumber of physically identical pairs while remov-ing the same-category pairs. Since Sets 1 and 2had appeared to be completely equivalent (therewere no effects of replication in Experiment 2),the physically identical pairs from Sets 1 and 2(shown in Table 1) were both used in the presentexperiment. Different pairs consisted of half theSet 1 and half the Set 2 different stimuli.

Procedure

Procedures were the same as for Experiment 2with the exception that instructions were to pressthe same key only when the members of a pairwere physically identical and to press the differentkey for all other pairs. The practice set wasadjusted to reflect the new conditions. Twentysubjects received picture stimuli, 20 receivedwords.


Results for both words and pictures areshown in Figure 4. Error rate was 1.5%.As in previous experiments, there were nosignificant differences in any of the analysesof error rate, A within-subjects three-wayANOVA (Priming X Level of Goodnessof Example X Category) was performedseparately for same and different responses.The results were strikingly simple: Therewere no main effects of priming and noPriming X Level of Goodness of Exampleinteractions in either the same or differentconditions. The few significant effects thatdid occur were not relevant to the hypothe-sis: a significant main effect for level ofgoodness of example for same responses tothe word stimuli, F(2, 38) = 4.61, p < .05(min F' not significant), and significantmain effects for category for the wordsstimuli, same, F(8, 152) = 3.58, p < .001;different, F(8, 152) = 2.71, p < .01 (minF' not significant).

The purpose of Experiment 5 was todetermine whether the representationsgenerated by superordinate category namescontained visual elements of sufficient con-creteness to facilitate responses under phys-ical identity instructions. Since there wasno effect of priming at all under physicalidentity instructions, the answer appearedto be that they do not. Such results, how-ever, contain a paradoxical element. Theresults of Experiments 2 and 4 indicatedthat representations generated by super-ordinate category names in advance of pre-sentation of the stimulus contained some ofthe information used by subjects in theactual perception of the stimulus. YetExperiment 5 has shown that what wenormally consider the elements of perception—physical visual representations such aslines, curves, and angles—are not part ofthe information provided by the super-ordinate category name. These apparentlycontradictory results suggest that less con-crete aspects of the representation of thecategory name, aspects of the meaning ofthe name itself, constitute the informationwhich was used by subjects to selectively


facilitate perception of the stimuli in Experi-ment 2.

It is of great interest to demonstrate thatperception of.the meaning (identity) of astimulus is an actual perceptual process, asdenned by the operations and logic of thepresent series of studies. The demonstra-tion depends upon two steps: (a) Underphysical identity instructions in which sub-jects do not process the meaning but onlythe physical configuration of the stimulus,priming with the superordinate categoryname has no effect on perceptual encodingof physically identical pairs, (b) Undersame-category instructions when subjectsmust prepare to process all pairs as mean-ingful stimuli, representations generated bysuperordinate category names in advance ofthe stimulus do affect the perceptual encod-ing of physically identical stimuli. Hence,it must be an aspect of the meaning itselfwhich affects perceptions in the same-category instruction condition.

However, this argument makes oneassumption, which, without proof from somesource other than the present experiment, iscircular. It is assumed that subjects do notencode the identity or meaning of the stimuliunder physical identity instructions becausethe prime, which affected physically identicalpairs under same-category instructions, didnot affect such pairs under physical identityinstructions. The purpose of Experiment5 was to provide converging evidence froma quite different experimental paradigmthat the stimuli used in the present seriesof experiments are encoded as meaningfulitems under same-category but not physicalidentity instructions.

EXPERIMENT 6Research on incidental learning (Hyde &

Jenkins, 1973; Till & Jenkins, 1973; Walsh6 Jenkins, 1973) and on levels of process-ing in memory (Craik & Lockhart, 1972)has shown that the task which subjects aretold to perform on a list of items affectstheir later ability to recall the items. Taskswhich require a "low level" visual process-ing, such as estimation of the number ofoccurrences of the letter e, result in con-siderably lower levels of recall of words than

tasks which require processing of the mean-ing of the words, such as judging each wordfor its pleasantness. In addition, there isevidence that cues that are psychologicallypresent for a subject at the time when he isencoding a word are more effective in en-hancing later recall than cues not present atthe time when subjects are encoding thewords (Tulving & Thomson, 1973).

The results of Experiment 4 suggestedthat it was the cognitive representation ofthe meaning of category names whichaffected the perceptual encoding of phys-ically identical pairs in Experiment 2. Tosubstantiate this argument it is necessaryto show that under physical identity instruc-tions, subjects do not encode the stimuli asidentified meaningful items, Using the logicof incidental learning and of cued recall, itcan be predicted that if, under the physicalidentity instructions of Experiment 5, sub-jects processed items at a purely visual levelwithout identifying them, recall for itemsshould be very poor. Furthermore, if itemswere not processed as meaningful units,even though subjects were continually hear-ing and repeating the category name, thosenames should be ineffective cues for recall.By the same logic, it can be predicted thatif, under the same-category instructions ofExperiment 2, subjects did process thestimuli as identified, meaningful items, re-call of items under those instruction condi-tions should be superior to recall of itemsunder physical identity instructions. Inaddition, under same-category instructionconditions, cuing with the category nameshould produce an improvement in recall.

Method

SubjectsEighty subjects from introductory psychology

classes who had not been subjects in any of theprevious experiments participated. They receivedcourse credit for their work.

StimuliStimuli used in the physical identity instruction

condition were the same as those used in Experi-ment 5; stimuli used in the category identityinstruction condition were the same as those usedin Set 1 of Experiment 2. In each instructioncondition, there were 54 actual items which could

216 ELEANOR ROSCH

TABLE 2NUMBER OF ITEMS RECALLED IN FREE RECALL

Number of items recalled

Instruction condition Words Pictures

Uncued Cued Uncued Cued

Physical identity instructionsCategory same instructions

310

214

616

621

be recalled. There was a small difference in thedistribution of those 54 items as can be seen fromthe methods sections of Experiments 2 and S;however, there is no reason to believe that eitherof the distributions should favor or depress recall.In addition, the cued group had exactly the samedistribution of items as the uncued group for eachof the instruction conditions.

ProceduresThere were four experimental conditions:

physical identity instructions without cued recall,physical identity instructions with cued recall,same-category instructions with cued recall, andsame-category instructions without cued recall.Each of the four conditions was administered toone group of subjects for the picture stimuli andto another group of subjects for the word stimuli.Ten subjects served in each of the resultant eightconditions.

Priming procedures for the same-category con-ditions were identical to those of Experiment 2and for the physical identity condition identical tothose of Experiment 5. At the conclusion of thepriming session, the subject was given a sheet ofpaper and asked to write down all of the itemswhich he had seen. Ten min were allotted forthis task; subjects who claimed to have finishedbefore this time interval ended (and the majorityof subjects did) were required to remain andattempt to recall additional items. In the uncuedcondition, the subject was provided with no infor-mation at the time of recall. In the cued condi-tion, the subject was provided with a list of thenine category names which he could refer to atwill throughout the 10-min period of recall. Infact, when asked to describe how they performedthe task, most subjects in the uncued conditionreported making their own list of category namesas a memory aid; however, they also reportedbeing unable to recall all nine categories. Thus,the effect of cuing was to supply not nine categorynames but the one to four categories which uncuedsubjects appeared unable to provide themselves.


The purpose of Experiment 6 was toshow, by means of converging operations,that in the same-different matching paradigmused in the present series of experiments,

stimuli are encoded as meaningful items" under same-category instructions but notunder physical identity instructions. Impli-cations for the memory task of Experiment6 were that items should be recalled betterunder same-category than under physicalidentity instructions, and cuing with thecategory name should be a greater aid torecall under same-category than underphysical identity instructions.

The mean number of words recalled in theeight conditions are shown in Table 2. Athree-way ANOVA (Instruction ConditionX Cuing X Words vs. Pictures) was per-formed. Of primary interest was the pre-dicted effect of instruction condition andthe Cuing X Instruction condition inter-action. The main effect for instructioncondition was highly significant, F(l, 79)= 20.92, p < .001. As Table 2 shows,memory for items under physical identityinstructions was minimal. The main effectfor cuing was significant, F(\, 79) = 10.75,p < .01, and the Cuing X Instruction con-dition interaction was significant: F(\, 79)= 13.89, p < .001. Cuing significantly im-proved recall only for the same-categoryinstruction condition. Pictures were re-called better than words, F(\, 79) = 7.03,p < .05, an effect which did not interacteither with instruction condition or cuing.

Thus, the main hypotheses of the experi-ment were confirmed. Under physicalidentity instructions, stimuli appear to beprocessed at a level sufficiently lacking inmeaning that such items were difficult torecall. Furthermore, in the physical identityinstruction condition, despite hearing thecategory name in conjunction with eachitem twice (once in the same and once inthe different primed conditions), cuing withthe category name produced no improve-ment in recall. Thus, it is possible to con-clude that stimuli in this condition areprocessed as concrete visual configurationsrather than as identified meaningful objects.

Experiment 6 provided the needed con-vergent evidence that it was not the repre-sentation of visual configurations but ratheran aspect of meaning contained in the rep-resentation generated by category nameprimes which affected the perceptual en-


coding of physically identical pairs in Ex-periment 2. How is that meaning rep-resented? Experiment 7 addressed thisquestion further.

EXPERIMENT 7

One of the more surprising aspects of thefindings of Experiments 2 and 3 is that,while the selective effects of the representa-tion generated by the prime appear to bepresent only in the condition of physicallyidentical stimuli, they show identical pat-terns of results for words and for pictures.While it might be thought that good butnot poor examples of categories that havebeen processed as meaningful items mighthave some very general visual features incommon (such as legs and surfaces forfurniture) which the category name couldarouse and which could act selectively onperception of items, what possible visualfeatures can words (names) for the goodexamples of categories arouse? And if therepresentation were simply a word list ofthe names of category members orderedaccording to goodness of example, how canwe explain the effects of the prime on pic-tures? We come again to the central ques-tion : What can be the nature of the rep-resentation generated by the category namesuch that it affects perception of physicallyidentical pairs in the ways which the pre-vious experiments have shown it does ? Theonly possibility appears to be that the cat-egory name prime arouses an abstractordered set of inclusion probabilities of themeanings of items that might be included.

The present experiment was designed toexplore two possible forms in which theordered set of inclusion probabilities for thecategory might be represented. On the onehand, it could be represented by a generalmeaning common to words and pictures.On the other hand, the list of possibilitiesmight be in the form of a set of very generalpictorially possible items in the picturecondition and very general word possibilitiesin the word condition. Single versus dualcoding hypotheses are an issue of presentdebate in psycholinguistics (Clark, Carpen-ter, & Just, 1973; Paivio, 1971), and themethods of the present study make possible

an investigation of a related, though notidentical, question with regard to the na-ture of the code generated by superordinatecategory names.

If the set of inclusion probabilities arousedby the category name is specific to a word orpicture modality, that is, if a quite differentrepresentation of possibilities is arousedwhen a subject knows he will see pairs ofpictures than when he knows he will seepairs of words, the selective effects of prim-ing physically identical pairs may occur onlyfor conditions in which the subject receivesonly pictures or only words. In fact, in theprevious experiments, only such conditionswere used. If, however, the set of prob-able occurrences for the category is in acommon form representing an underlyingmeaning of the category name, differentialeffects of priming should still occur underconditions in which word and picture pairsare mixed and the subject does not knowfor any given trial whether the pair he isabout to see will be a pair of words or apair of pictures. This condition was ex-plored in the present experiment.

MethodSubjects

Subjects were 30 paid student volunteers whohad not participated in any of the previous prim-ing experiments.

Stimuli and ProcedureStimuli were the words and pictures of Set 1

used in Experiment 2. Because a double set ofstimuli was being used, conditions of testing otherthan the variable of interest (the subject's un-certainty as to whether a pair of words or pic-tures would be presented) were unavoidably intro-duced. To control for these extraneous variations,two types of stimulus set and conditions of testingwere employed.

In the first type of testing, a single subject wasgiven both the Set 1 word and picture stimuli ofExperiment 2. Thus, each subject received onemore replication and twice as many stimuli assubjects had received in Experiments 2 and 4.The second type of testing employed matchedpairs of subjects, each of which received half ofthe double stimulus set. The sets for this testingwere prepared to maximize similarity to the typeof set each subject had received in Experiments2 and 4.

For the first type of testing, a stimulus set con-sisted of all of the words and all of the pictures

218 ELEANOR ROSCH

PHYSICALLYIDENTICAL

1400 -

1300

1200 -

: 1100 -

.i 1000 -

900 -

800

700

600

Pictures: I• UnprimedO-— Prim

Words:A— Unprimed

- A Primed

DIFFERENT


FIGURE 5. Words and pictures together, same-category instructions, 2-sec prime interval.

of Set. 1. These were administered with the sameprocedures used in Experiment 2 with the excep-tion that for each trial the subject did not knowwhether he would see a pair of words or of pic-tures. Because subjects were receiving a doubleset, each pair occurred four rather than twotimes; once as a pair of primed pictures, once asunprimed pictures, once as primed words, andonce as unprimed words. The test session lasted4 hr. Subjects were tested on 2 successive daysand were paid for their services. Ten subjectsparticipated in this testing procedure.

For the second type of testing, 36 subjects wereinitially screened in a 15-min pretesting sessionwhich consisted of a brief test of simple visualreaction time and a brief priming task of the samedesign as the present study but using differentcategories and items. On the basis of scores onthese tests, it was possible to select 10 matchedpairs of subjects whose reaction times had beenmost similar in the initial screening session. Eachmember of a pair received half of the randomlyordered double set of stimuli with the followingconstraints on randomness: Each subject receivedall of the items of Set 1; exactly half of the pairswere in the form of words and half in the form ofpictures. The matched subject of a pair receivedall of the items of Set 1 equally divided into pairsof words and pictures; however, all items which

the first subject had received as words, the secondsubject received as pictures and vice versa. Whichitems were words and which pictures varied foreach of the 10 pairs of subjects. Different randomorders of presentation were used for each pair ofsubjects, but the two subjects in a matched pairreceived the same order.

Instructions and practice trials were modifiedto include the fact that pairs could now be eitherwords or pictures but were otherwise identicalto those used in Experiment 2. Testing proced-ures also were as described in Experiment 2.


Results were analyzed in the mannerdescribed for Experiment 2: separately forthe group of subjects who had received bothsets and for the matched group in whicheach subject had received a half set. In thelatter case, each matched pair was treatedas a single subject in the analysis. Thepattern of results and the effects whichreached significance were identical for bothgroups. The only difference between thegroups was in overall reaction time; sub-jects who had received both sets with 4 hrof testing showed a faster mean reactiontime in all conditions than subjects for whomeach item had not been replicated and whohad been tested for only 2 hr. Because theconditions of testing for each matched-pairsubject approximated those of Experiment2 more closely than the testing conditionsof the double-set subjects, and because thefact of having two subjects per set did notappear to affect results, it is the results fromthe matched pairs which are presented anddiscussed below.

Results for Experiment 6 are shown inFigure 5. Overall error rate was 6%; asin previous experiments, there were no sig-nificant differences in error rates for anycondition. The results are, in all importantrespects, identical to those of Experiment 2.Results were analyzed in the same manneras that described for Experiment 2. Forthe physically identical stimuli, for bothpictures and words, the Priming X Levelof Goodness of Example interaction wassignificant: pictures, F(2, 18) = 8.42, p <.01; words, P(2, 18) = 8.14, p < .05.T tests established that priming significantlyhelped good examples: pictures, t(9) —2.30, p < .05; words, t (9) = 2.47, p < .05;


and hindered poor examples: pictures, £(9)= 3.27, p < .01; words, *(9) = 2.51, p <.05.

For the same-category pairs, results werealso equivalent to Experiments 2 and3. The Priming X Level of Goodness ofExample interaction was not significant.Main effects of priming were significant:pictures, F(l, 9) = 20.67, p < .01; words,F(l, 9) = 10.88, p < .01. The effects oflevel of goodness of example were signifi-cant: pictures F(2, 18) = 18.34, p < .001;words, F(2, 18) =22.58, p < .001. Asusual, main effects of category reached sig-nificance but did not interact with the othervariables.

For the different responses, results were,likewise, very like those of Experiments2 and 4. The main effect of priming wassignificant: pictures, JF(1, 9) = 5.91, p <.05 (min F' not significant) ; words, F(i, 9)= 11.00, p < .01. The main effect of levelof goodness of example was significant:pictures, F(2, 18) = 4.18, p < .05 (min F'not significant) ; words, F(2, 18) = 8.41,p < .01. The Priming X Level of Good-ness of Example interaction was not sig-nificant. Effects of category reached sig-nificance for words and not pictures but didnot interact with the other variables.

In order to test further whether seeingword stimuli only and picture stimuli onlyversus seeing both in the same experimentalparadigm affected any of the factors forphysically identical pairs, a mixed-modelfour-way ANOVA (Words & PicturesSeparate vs. Words & Pictures TogetherX Priming X Level of Goodness of Ex-ample X Category) was performed on theresults for physically identical pairs for Set1 stimuli of Experiments 2 and 7. Wordsand pictures separate (Experiment 2)versus words and pictures together (Ex-periment 7) was a between-subjects vari-able ; the other variables were within sub-jects. Separate analyses were, necessarily,performed on words and on pictures sincethe variable of interest required that wordsand pictures be a between-subjects variablein Experiment 2 and a within-subjects vari-able in Experiment 7. The result of interestfrom this analysis was that none of the inter-

actions between words and pictures togetherversus words and pictures separate reachedsignificance.

The present experiment was not designedto show that there was no possible differ-entiation between pictorial and verbal cod-ing of the underlying meaning of categoryrepresentations. It did demonstrate clearly,however, that within a time interval of 2sec, subjects were capable of generating arepresentation of the type of semantic cat-egory used in the present studies whichshowed selective effects of priming on wordsand pictures in a manner identical to theeffects shown when subjects knew theywould see only pictures or only words.What sort of representation is this? Theresults of the present experiment supportthe idea that the representation is notentirely specific to either a pictorial orverbal mode but is some set of abstractprobabilities of items that can represent themeaning of the category in either mode.Yet there may be some aspect of the repre-sentation which is specific to pictures andto words which was masked in the presentexperiment by a priming interval whichwas long enough for both specific codes tobe generated. The purpose of Experiment8 was the further examination of the timecourse of generation of the representationboth for pictures and words presented aloneand presented together.

EXPERIMENT 8

In Experiments 2 and 7, a 2-sec timeinterval occurred between the prime andpresentation of the stimulus pair. Thisinterval may well have been long enoughfor subjects to generate all aspects of therelevant category representation so thatthey would be prepared for underlyingmeaning represented in the form either ofpictures or words. The purpose of thepresent experiment was to observe theeffects of systematic reduction of the timeinterval between the prime and the stimulus.

MethodSubjects

A total of 200 subjects were tested. Of these,100 were tested in conditions in which words and

220 ELEANOR ROSCH

500 MSEC 400 MSEC 300 MSECINTERVAL INTERVAL INTERVAL

Pic tures :• Unprimed

II001- o--— Primed

1000

(J

$•E 900

•I 800u

700

600

Words:A— UnprimedA—-- Primed


FIGURE 6. Conditions of "uncertainty," varyingprime interval, words and pictures together.

pictures were presented separately; these subjectswere students in introductory psychology courseswho had not participated in previous primingexperiments and who received course credit forparticipation. Subjects who were tested in thecondition where words and pictures were bothpresented were obtained in the same manner asmatched pairs had been obtained in Experiment7. For each group of 10 matched pairs required,40 paid volunteers were pretested as describedin Experiment 7. Twenty students were selectedwho formed pairs best matched in reaction time;these 20 subjects participated as paid volunteersin that part of the priming experiment itself.

Stimuli and ProcedureThe stimuli consisted of Set 1 of Experiment 2.

For the words only and pictures only groups,procedures were as described in Experiment 2with 'the exception that, as in Experiment 4, sub-jects did not repeat the prime or the word blank.Stimulus sets were prepared for the words pluspictures groups as described in Experiment 7.

Five intervals of time between the prime andpresentation of the pair were used: SOO msec,400 msec, 300 msec, 200 msec, and 100 msec. Inorder to provide greater accuracy and consistencyof the intervals, the primes and word blank wererecorded on a Wallensac tape recorder in differentprearranged random orders for each subject.Stimulus cards were rearranged for each subjectto match the order on the tape. The experimen-ter initiated each trial after he had positioned theappropriate card in the tachistoscope by pressinga foot pedal which began the tape. The tapeplayed the prime or blank and, as the word was

terminated, initiated the required time intervalbefore the stimulus appeared. After initiatingthe time interval, the tape stopped until the ex-perimenter initiated the succeeding trial.

At each of the five time intervals, 10 subjectswere tested under the words only condition, 10subjects under the pictures only condition, and 10matched pairs of subjects under the words pluspictures condition.

Results

The results for the physically identicalpairs of each condition ,at each time intervalare shown for conditions of uncertainty inFigure 6 and for conditions of certainty inFigure 7. (Figure 6 does not show resultsfor the 200- and 100-msec conditions becauseof their similarity to the 300-msec condi-tion.) Overall error rate was 5%, and, aspreviously found, did not differ significantlyin any condition. Results were analyzedas previously described in Experiments 2,4, and 7, separately for words and for pic-tures, in each condition, and at each time

500 MSEC 400 MSEC1000

900

800

700

15 600(Ue

o 500

P r INTERVAL

Pictures: 1 f• Unprimed /

_ o Primed /Words /

A Unprimed /A Primed /

~

-

_

i i i

/

- / /

^ / •_ L>r•' ' 8cr""

i i t

P INTERVAL

4f1

1 A1 s

^/',- pf

c/

1 i8& 300 MSEC 200 MSEC 100 MSEC

900

800

700

600

_ INTERVAL

,£</

A *f P£•• /• "*

^5a-''

\ |

INTERVAL

^

^^fr—"JjS,%

,0 X^*0

1 1 1

_ INTERVALJ.

ters**^_

>

f>~-cr

j . . 1 iHigh Med Low High Med Low High Med Low

Level of goodness-of-example

FIGURE 7. Conditions of "certainty," varyingprime interval, words and pictures separate.


TABLE 3TIME COURSE FOR GENERATION OF PHYSICALLY IDENTICAL STIMULI

Stimulusinterval(in msec)

Interaction P*Picture

t testb

High GOE Low GOE

WordInteraction F* t testb

High GOE Low GOE

Certainty

500400300200100

8.16**10,06**5.26*4.19*3.06

2.47*2.42*2.50*2.211.83

3.80**2.56*2.71*2.08.05

20.61***9.84**3.90*.93.78

3.85**2.78*2.17.23.66

5.62***4.02**1.99.09.34

Uncertainty

500400300200100

7.82**5.58*3.202.351.59

4.20*1.882.17.1.511.07

2.84*2.16.19

1.941.23

2.321.962.761.901.86

1.031.14.63

1.04.56

2.01.921.01.82.07

Note. GOE = goodness of example.• df = 2, 18.' df = 9.

* p < .05.** t < .01.

***p < .001.

interval. Results of significance tests ofrelevance to the present experiment areshown in Table 3. Table 3 shows theresults of the ANOVAs for the Priming XLevel of Goodness of Example interactionand the results of t tests for the significanceof .the facilitation (speeding of responses)provided by the prime for good examplesand the interference (slowing of responses)provided by the prime for poor examples.Other tests of the significance of effects inExperiment 8 are not included in Table 3.s

The pattern of relevant results was quiteclear and extremely orderly. It requiredmore time (a longer time interval betweenthe prime and presentation of the stimulus)to generate a representation capable of

8 In fact, consistent main effects of level ofgoodness1 of example were obtained in all cases;however, such main effects are not relevant tothe particular questions which the present experi-ment was designed to test. Similarly, some cate-gory effects were significant but did not interactwith the effects of interest to the present hypothe-sis. Results for the same-category and differentitems (which, it should be remembered, are notaffected by priming interval) were virtually iden-tical, in the present studies, to tbe results foundfor those conditions in Experiments 2 and 4;these data are also irrelevant to the hypothesisof the present study and are not presented.

selectively affecting responses to the phys-ically identical pairs under conditions ofuncertainty than under conditions of cer-tainity. And it required a longer timeinterval to generate a representation capableof those selective effects when the stimuliwere words than when they were pictures.The two effects appeared to be additive (seeSternberg, 1969).

Each of the effects will be summarizedseparately. The effect of uncertainty isapparent: A consistently longer time inter-val was required between prime and stimulusfor effects to be obtained under conditions ofuncertainty than under conditions of cer-tainty (the words only and pictures onlycondition). At a 500-msec interval, underconditions of uncertainty, the expected effectfor the word stimuli (a significant PrimingX Level of Goodness of Example inter-action) was not significant; however, at thissame time interval, under conditions of cer-tainty (words only and pictures only), alleffects were identical to those achieved at a2-sec interval (that is, expected effects forall stimuli were achieved). At the 400-msec interval, in conditions of uncertainty,not only were the effects for the wordstimuli not significant, but the effects for

222 ELEANOR ROSCH

the picture stimuli were diminished; that is,although the Priming X Level of Goodnessof Example interaction reached significance,neither of the t tests for specific effects offacilitation and suppression of response weresignificant. At that same 400-msec interval,effects in the two conditions of stimuluscertainty were still the same as they hadbeen at the 2-sec interval; all expectedeffects were obtained. At the 300-msecinterval and at all shorter intervals, no sig-nificant effects at all were achieved in con-ditions of uncertainty, although a pattern ofsignificant effects continued to be found forsome stimuli under conditions of stimuluscertainty at the 300-msec and 200-msecintervals.

The effect of mode of representation (pic-tures or words) was apparent; it required aconsistently longer time interval betweenprime and stimulus for effects to be obtainedfor word stimuli than for picture stimuli.This difference between words and pictureswas obtained both under conditions of un-certainty and of certainty. Under condi-tions of uncertainty, at the 500-msec inter-val, the expected Priming X Level of Good-ness of Example interaction was not sig-nificant for the words stimuli; however,that interaction was as large for the picturesstimuli as it had been at the 2-sec interval.At the 400-msec interval under conditionsof uncertainty, there was, not surprisingly,also no significant Priming X Level ofGoodness of Example interaction for thewords stimuli; for the pictures stimuli, how-ever, although the effects of priming werediminished (t tests for specific effects offacilitation and suppression of responseswere not significant), the overall PrimingX Level of Goodness of Example inter-action was still significant.

Under conditions of certainty, effects forboth words and pictures at 500 msec and400 msec were as they had been at the 2-secinterval. However, at the 300-msec interval,the effect of priming on the words stimuliwas diminished; that is, the Priming XLevel of Goodness of Example interactionattained significance but neither of the ttests for facilitation of good examples andinhibition of poor examples was significant.

At this same 300-msec interval, the picturesstimuli still exhibited the same effects ashad been found at the longer time intervals.At the 200-msec interval, no effects at allwere significant for the words stimuli; how-ever, at this time interval effects were foundfor the pictures stimuli in their diminishedform.

At the 100-msec interval, no effects eitherof words or pictures under conditions eitherof certainty or uncertainty were signifi-cant, the same results which had been ob-tained for priming of physically identicalpairs in Experiment 4 when the prime andstimulus were presented simultaneously.

In conclusion, the results of Experiment7 had appeared to indicate that representa-tions of words and pictures were derivedfrom a single underlying meaning since,within 2 sec, subjects could generate a rep-resentation of the meaning of the categoryname which was as effective under condi-tions of uncertainty as it was under condi-tions of certainty. However, the presentexperiment has shown that when the timeinterval between the prime and the stimulusis reduced below some minimum (by extrap-olation probably 600 or 700 msec), anelement of differentiation in the representa-tions of words and pictures is revealed bydifferences in the time intervals required togenerate representations. Generation ofboth representations (i.e., generation 'underconditions of uncertainty) requires longerprocessing time than generation of a repre-sentation of either words or pictures alone;and generation of a representation capableof selective effects on word stimuli requireslonger processing time than generation of arepresentation which will affect pictures.Thus, we may tentatively conclude thatwhile there is considerable similarity inthe depth meaning of superordinate cate-gories not specifically coded in terms ofwords or pictures, there is some differenti-ation of the format into which the meaningis translated in preparation for the percep-tion of actual words or pictures. The factthat less time is required to prepare forpictures suggests very tentatively that pic-tures may be closer to the nature of theunderlying representation than are words.

COGNITIVE REPRESENTATIONS ,OF SEMANTIC CATEGORIES 223

EXPERIMENT 9

In Experiment 1, reliable ratings of thegoodness of example of category memberswere obtained for nine semantic categories,and in subsequent experiments consistentand explainable effects of those norms oflevel of goodness of example were obtained.Similar consistent selective effects of prim-ing on good and poor examples of colorcategories were obtained by Rosch (inpress-d). For color categories, good ex-amples retained an advantage when primedover poor examples even after long practicewith a limited set of stimuli. These resultswere taken as an additional form of evidencethat goodness of example in color categoriesmay be partly physiologically determined bythe salience of certain areas of the colorspace. There is no reason to think thatgood examples of the semantic categoriesused in the present study are determined byperceptual saliency in the same manner asare prototypes for color categories. Thatis, while the internal structure of categoriesmay be determined by psychological prin-•tiiples which are not arbitrary (Rosch,197S), it seems unlikely that good examplesof semantic categories are attached to thecategory name more strongly than poorexamples in a manner which cannot bealtered by sufficient practice. The presentexperiment attempted to duplicate as closelyas possible the practice conditions of Rosch(in press-d).

The purpose of Experiment 9 was two-fold: In part it was intended as a com-parison between semantic categories andcolor categories. It was expected that goodexamples of semantic categories would notretain an advantage over poor examplesgiven the amount of practice in which goodexamples of color categories had retainedan advantage over poor ones. The secondpurpose of the experiment was a methodo-logical point, the demonstration that effectsobtained with artificial categories, even whenthey are comprised of a set of real wordsredefined into a new artificial set, can givemisleading information concerning the na-ture of actual natural categories.

MethodRosch (in press-d) used as stimuli 16 colors—

8 good and 8 poor examples of the 8 basic colorname categories—and used physical identity in-structions. In the present study, it was necessaryto use same-category instructions (as shown bythe results of Experiment 5). In order to havean equivalent 16 stimuli, only 4 categories wereused per subject and only the good and poor(omitting the intermediate) examples of eachcategory were used. Four subjects were selectedfor the task and were paid for their services. Onesubject received word stimuli and one receivedpicture stimuli for the 4 good and 4 poor examplesshown in Table 1 for Set 1 of furniture, fruit,vehicle, and sport. The other two subjects re-ceived the good and poor examples of Set 1 forthe categories: weapon, vegetable, bird, and car-penter's tool; one subject received words, theother pictures.

The instructions and procedures used in Ex-periment 2 were followed. Practice trials weremodified to take into account the reduced set ofstimuli being used. Following the proceduresused for the color categories, each subject wastested 21 times on the same small set of stimuli.After an initial session, subjects made 2 runs perday, 5 days per wk, for 2 wk.


Figure 8 shows the results for the thefour subjects for the physically identicalstimuli of their last test session. Overallerror rate was .02%; there were no sig-

SUBJECT 1-WORDS SUBJECT 3-WORDS

500

83400

A A- &

1

Pictures: Words:• Unprimed A— Unprimedo--- Primed A---primed

SUBJECT 2PICTURES

500 p. f

400

SUBJECT 4PICTURES

o o-—...

j IHigh Med Low High Med

Level of goodness-of-exampleLow

FIGURE 8. Effects of 2-wk practice on reducedset of stimuli.

224 ELEANOR ROSCH

nificant differences in error rate in any con-dition. By this session the Priming XGoodness of Example interaction and themain effect of goodness of example appearto have disappeared. To test whether thatwas in fact the case, following the proced-ures of analysis used by Rosch (in press-d),a within-subjects ANOVA was performedfor each subject with cell variance suppliedby the six different responses for each itemsupplied by the last six test runs of thatsubject. In all cases, the main effect ofpriming was significant—Subjects 1-4, re-spectively: F(l, 5) = 9.23, p < .05; F(l,5) = 10.71, p < .05; F(l, 5) = 6.17, p <.05; F(\, 5) = 5.97, p < .05 (min F' notsignificant). In summary, in no case werethe Priming X Goodness of Example inter-action and the main effect of goodness ofexample significant. This result is not foundfor the physically identical pairs in any ofthe other experiments in this study.

Subjects' descriptions of their behaviorin the experiment were quite similar: Allfour reported that they quickly memorizedthe possible stimuli (words or pictures),and when primed, formed an image of thetwo possible words or pictures. Such anaccount (we may replace image with expec-tation if greater theoretical neutrality isdesired) would lead to the obtained data.As predicted, the names of the categoriesacquired a new meaning; they now namedartificial sets consisting of two specific pos-sible stimuli. It should be noted that thenew categories were now no longer neces-sarily at a superordinate level of abstractionwhich had been the level of their naturalmeaning but now may have been at a basicor subordinate level for which differenteffects of priming can be expected (Roschet al., in press).

This experiment has shown that theinternal structure of semantic categories,although consistent and pervasive in theprocessing of natural categories, can, unlikethe internal structure of color categories, bereadily altered by practice and by taskdemands. The present experiment demon-strates how easily a very general effectobtained with natural semantic categoriescan be eliminated when experimental struc-

ture changes those categories into artificialcategories.

GENERAL DISCUSSION

The present research is part of a generalinvestigation of the nature of human cate-gorization. The part of the work reportedin this paper was concerned with oneimportant type of category, that into whichthe common concrete objects of the realworld are classified, and with one type ofquestion about such categories, the structureof the representations which the humanmind generates when a human subject hearsa category term and understands its mean-ing.

The set of categories which were used inthe investigation represent virtually theentire population of common higher-order(superordinate) classifications of concreteobjects with the restriction that the cat-egories not be embedded in idiosyncratictaxonomic structures (common and tax-onomic structure were operationally definedin Experiment 1). It should be noted thatthese are some of the categories which havebeen most commonly used in psychologicalstudies of learning. This research was spe-cifically not designed to investigate abstractconcepts of the type represented by wordssuch as beauty or causality. The researchalso did not directly investigate representa-tions of specific exemplars (members) ofhigher order superordinate categories, al-though part of the purpose of the presentexperiments was to make possible a com-parison of representations of superordinatecategories with representations of categoriesof other types and at other levels of abstrac-tion (Rosch, in press-d; Rosch et al., inpress).

There were four basic findings from thepresent series of experiments. First, theinternal structure of superordinate semanticcategories was clearly a pervasive aspectof the way in which those categories wereprocessed in the tasks used. For responsesto physically identical stimuli, the internalstructure of the categories appeared to bea part of that cognitive representation ofthe category aroused prior to presentationof the stimuli which affected subjects' per-


ception of the stimuli. In this case, cogni-tive representations of categories clearlycontained more of the information neededto respond to category members which hadbeen rated good examples of the category inExperiment 1 than to respond to categorymembers which had been rated bad examples.In other words, cognitive representationsof categories appeared to be more similarto the good examples than the poor ex-amples.4 An account of the meaning ofsuperordinate category names by means ofa definitive list of attributes or featuresnecessary and sufficient for category mem-bership requires considerable additionalexplanation to account for such findings.While it may still be argued that the "true"meaning of such category names must residein philosophical or linguistic primitives con-sisting of feature lists of that nature, thepresent study offers evidence that suchaccounts do not appear to mirror psycho-logical reality.

In line with that point, it should be notedthat the research reported in this article wasnot intended either as a model of semanticmemory or as a verification or refutation ofany particular theory of semantic memory.Indeed, as was shown in the discussion ofsame-category and different items in Experi-ment 4, the priming technique appears tobe a poor tool for distinguishing amongalternative models of decision processes con-cerning category membership. Nonetheless,the findings of the present research are morecompatible with some classes of semanticmemory theory than with others. It isrelatively difficult to integrate the presentfindings with models of semantic memorywhich depend on criterial features and clear-cut category boundaries (such as that ofGlass & Holyoak, in press). On the otherhand, the attribute theory of Smith et al.

4 Although for same-category and differentstimuli it was not possible to isolate encodingfrom decision effects, it should be noted that effectsof internal structure were also apparent in theresults for these conditions and were in no placecontradictory to effects of category structure pre-viously found in tasks requiring search andretrieval from semantic memory (Collins &Loftus, in press; Rips et al., 1973; Rosch, 1973;Smith et al., 1974).

(1974), which is based on structures suchas those explored in the present research,is necessarily compatible with those struc-tures. And a theory such as that of Collinsand Loftus (in press), which has as itsbasis the analogue process of spreading acti-vation, fits well with the present research,at least in the sense that it is relatively easyto describe many of the "structural" findingsof the present studies using the "process"terminology of spreading activation.

It is important to note that the basicfindings of the present study are not depen-dent for their validity upon any particularinterpretation of the meaning of internalstructure (e.g., of the norms shown inTable Al). It may be argued that suchnorms merely reflect word frequency (whichthe norms in fact, do not—Mervis, Rosch,& Catlin, Note 3), associative strength tothe category name, frequency of the actualobjects in the category, or co-occurrencefrequency with the category name. Thatthese are not viable as complete accountsof the meaning of internal structure andthat, in fact, the actual structure of theattributes of category members is involvedin goodness of example norms is argued byseveral lines of ongoing research (Rosch,in press-a; Rosch & Mervis, in press;Rosch et al., in press; Rosch, Simpson, &Miller, Note 1; Mervis et al., Note 3).However, regardless of its interpretation,the effects of the internal structure of cat-egories in perceptual tasks found by thepresent study both constitute a refutationof the psychological reality of an Aristo-telian view of categories and make possiblethe further investigation of the nature of thecognitive representation generated by thecategory name which affected perception ofstimuli.

Subsequent experiments of the presentresearch focused on the issue of the natureof effects on perception generated by super-ordinate category terms. The second basicfinding concerned levels of processing inperception (Cohen, 1968; Posner & Mitch-ell, 1967; Beller & Schaeffer, Note 2).Experiments 5 and 6 showed that the selec-tive effects of internal structure on percep-tion of physically identical pairs could not

226 ELEANOR ROSCH

be accounted for by concrete physical fea-tures present in the representation generatedby the category name. Therefore, by thelogic outlined in the discussion sections ofExperiments 5 and 6, the effects of therepresentation on perception must have beendue to the abstract representation of themeaning of the category name. Such afinding demonstrates, by means of a spe-cific operation, that, at least at one level ofperception, the meaning of pictures and ofwords is part of the actual perception of thestimuli and not something inferred afterthe perception occurs. The nature of suchmeaning appeared to be in the form of anabstract ordered set of inclusion probabil-ities of the meanings of members of thecategory with the probabilities ordered ac-cording to the internal structure of thecategory.

The fact that meaning itself could beinferred to affect perception of a stimuluspair and that the identical design could beused for pairs of words and pictures madepossible the investigation which led to thethird basic finding of the research whichconcerned the nature of the underlying rep-resentation of the meaning of words andpictures. Experiments 7 and 8 exploredthe question of whether the representationof words and pictures was derived from acommon meaning or from dual codes spe-cific to the modality of the stimulus. Theresults indicated the partial truth of bothformulations. Sufficient underlying meaningappeared to be common to words and pic-tures that, within 2 sec (by extrapolationfrom Experiment 8, within 700 msec),subjects can generate a representation ofthe meaning of the category name which isequally representative of a pair of words orpictures under conditions in which the sub-ject is uncertain as to the type of stimuluswhich is to appear. However, when thetime intervals between the prime and thestimulus are reduced below this minimum,an element of differentiation in the repre-sentations of words and pictures is revealedby differences in the time intervals requiredto generate representations. Generation ofboth representations (that is, generation ofa representation under conditions of un-

certainty) requires a longer time intervalthan generation of a representation for eitherwords or pictures alone (i.e., under condi-tions of certainty), and generation of a rep-resentation capable of selective effects onword stimuli requires a longer interval thangeneration of a representation which willaffect pictures. Thus, we may conclude,tentatively, that there is a depth meaning ofsuperordinate categories not specificallycoded in terms of words or pictures, butthat the depth meaning is translated into aformat in preparation for actual perceptionwhich differs slightly for words and pic-tures. The fact that less time is requiredto prepare for pictures suggests that picturesmay be closer to the nature of the under-lying representation than are words.

The fourth finding of the present re-search is the comparison which it made pos-sible between representations generated bycategory names of different types (such ascolor vs. semantic categories) and of cat-egories at different levels of abstraction.Color category names appeared to generatea concrete physical representation, andthere was no evidence of a semantic mean-ing additional to the physical representation(Rosch, in press-d). Names of the mem-bers of superordinate categories (more spe-cifically basic level and subordinate categorynames, Rosch, in press-a; Rosch et al., inpress) also appear capable of generatingconcrete physical representations at a levelat which superordinate category names didnot generate physical features.

Levels of abstraction have been definedand studied by a variety of different meanssuch as logic (Kay, 1971); a variety ofempirical ratings of levels of abstractness(Kammann & Streeter, 1971; Paivio,1971) ; effects of rated levels on learningand memory (Paivio, 1971); and levels ofabstractness defined in terms of sub- andsuperordinations to basic level classifica-tions (Rosch, in press-a; Rosch et al., inpress). Arguments over whether the lan-guage of a child develops by becoming moreor less abstract have received considerableattention (Anglin, 1970; McNeill, 1966).The present method makes possible a de-tailed investigation of the nature of the


mental representations generated by cat-egory terms at any level of abstraction. Theaddition of such knowledge may makepossible a greater understanding of the na-ture of psychological representation, thenature of taxonomies, and the process ofabstraction.

REFERENCE NOTES

1. Rosch, E., Simpson, C., & Miller, R. S.Structural bases of typicality effects. Manu-script submitted for publication, 1975.

2. Seller, H. K., & Schaeffer, B. On matchingwords: Syllabification and priming. Unpub-lished manuscript, 1974. (Available from H.K. Beller, Department of Psychology, StateUniversity of New York, College at Brockport,Brockport, N.Y. 14420.)

3. Mervis, C. B., Rosch, E., & Catlin, J. Rela-tionships among goodness-of-example, categorynorms, and word frequency. Unpublishedmanuscript, 1975. (Available from E. Rosch,Department of Psychology, University of Cal-ifornia, Berkeley, California 94720).

REFERENCES

Anglin, J. The growth of word meaning. Cam-bridge, Mass.: MIT Press, 1970.

Battig, W. R., & Montague, W. E. Categorynorms for verbal items in 56 categories: Areplication and extension of the Connecticutcategory norms. Journal of Experimental Psy-chology Monograph, 1969, 80, (3, Ft. 2).

Beller, H. K. Priming: Effects of advance in-formation on matching. Journal of Experimen-tal Psychology, 1971, 87, 176-182.

Berlin, B., & Kay, P. Basic color terms: Theiruniversality and evolution. Berkeley: Univer-sity of California Press, 1969.

Boring, E. G. A history of experimental psy-chology (2nd ed.) New York: Appleton-Century-Crofts, 1950.

Bourne, L. E. Human conceptual behavior.Boston: Allyn and Bacon, 1968.

Clark, H. H. The language-as-fixed-effect fallacy:A critique of language statistics in psychologicalresearch. Journal of Verbal Learning andVerbal Behavior, 1973, 12, 335-359.

Clark, H. H., Carpenter, P. A., & Just, M. A.On the meeting of semantics and perception. InW. G. Chase (Ed.), Visual information process-ing. New York: Academic Press, 1973.

Cohen, G. A comparison of semantic, acousticand visual criteria for matching of word pairs.Perception & Psychophysics, 1968, 4, 203-204.

Collins, A. M., & Loftus, E. F. A spreading-activation theory of semantic processing. Psy-chological Review, in press.

Cooper, L., & Shepard, R. N. Chronometricstudies of the rotation of mental images. InW. G. Chase (Ed.), Visual information process-ing. New York: Academic Press, 1973.

Craik, F. I. M., & Lockhart, R. S. Levels ofprocessing: A framework for memory research.Journal of Verbal Learning and Verbal Be-havior, 1972, 11, 671-684.

Donders, F. C., [On the speed of mental pro-cesses.] In W. G. Koster (Ed. and trans.),Attention and performance II. Amsterdam:North-Holland, 1969. (Reprinted from AdaPsychologies, 1969, 30.)

Ginsburg, S. S. The mathematical theory ofcontext-free languages. New York: McGraw-Hill, 1966.

Glass, A. L., & Holyoak, K. J. Alternative con-ceptions of semantic memory. Cognition, inpress.

Haber, R. N. Nature of the effect of set on per-ception. Psychological Review, 1966, 73, 335-351.

Heider, E. R. "Focal" color areas and the develop-ment of color names. Developmental Psy-chology, 1971, 4, 447-455.

Heider, E. R. Universals in color naming andmemory. Journal of Experimental Psychology,1972, 93, 10-20.

Hyde, T. S., & Jenkins, J. J. Recall for wordsas a function of semantic, graphic, and syntacticorienting tasks. Journal of Verbal Learningand Verbal Behavior, 1973, 12, 471-480.

Kammann, R., & Streeter, L. Two meanings ofword abstractness. Journal of Verbal Learningand Verbal Behavior, 1971, 10, 303-306.

Kay, P. Taxonomy and semantic contrast.Language, 1971, 47, 866-306.

Keele, S. W. Attention and human performance.Pacific Palisades, Calif.: Goodyear, 1973.

Ku£era, H., & Francis, W. N. Computationalanalysis of present-day American English.Providence, R. I.: Brown University Press,1967.

Labov, W. The boundaries of words and theirmeanings. In C. J. Bailey & R. Shuy (Eds.),New ways of analysing variations in English:Washington, D.C.: Georgetown UniversityPress, 1973.

Lakoff, G. Hedges: A study in meaning criteriaand the logic of fuzzy concepts. Papers fromthe Eighth Regional Meeting, Chicago Linguis-tics Society, Chicago: University of ChicagoLinguistics Department, 1972.

Loftus, E. F., & Scheff, R. W. Categorizationnorms for fifty representative instances. Jour-nal of Experimental Psychology, 1971, 91, 355-364.

McDaniel, C. K. Hue perception and hue mean-ing. Uupublished BA thesis, Harvard College,1972.

McNeill, D. A study of word association. Jour-nal of Verbal Learning and Verbal Behavior,1966, 5, 548-557.

228 ELEANOR ROSCH

Mervis, C. B., Catlin, J., & Rosch, E. Develop-ment of the structure of color categories. De-velopmental Psychology, 1975, 11, 54-60.

Paivio, A. Imagery and verbal processes. NewYork: Holt, Rinehart & Winston, 1971.

Posner, M. I. Abstraction and the process ofrecognition. In G. H. Bower & J. T. Spence(Eds.), The psychology of learning and moti-vation (Vol. 3). New York: Academic Press,1969.

Posner, M. I. Cognition: An introduction. Glen-view, 111.: Scott, Foresman, 1973.

Posner, M. I., & Boies, S. J. Components ofattention. Psychological Review, 1971, 78,391-408.

Posner, M. I., Boies, S. J., Eichelman, W. H., &Taylor, R. L. Retention of visual and namecodes of single letters. Journal of ExperimentalPsychology Monograph, 1969, 79, (1, Pt. 2).

Posner, M. L, & Mitchell, R. F. Chronometricanalysis of classification. Psychological Re-view, 1967, 74, 392-409.

Posner, M. L, & Snyder, C. R. Facilitationand inhibition in the processing of signals. InP. M. A. Rabbitt & S. Domic, Attention andperformance V, London: Academic Press, 1975.

Posner, M. I., & Snyder, C. R. Attention andcognitive control. In R. Solso (Ed.), Informa-tion processing and cognition: The LoyolaSymposium. Potomac, Md.: Erlbaum Asso-ciates, in press.

Rips, L. J., Shoben, E. J., & Smith, E. E. Seman-tic distance and the verification of semanticrelations. Journal of Verbal Learning andVerbal Behavior, 1973, 12, 1-20.

Rosch, E. On the internal structure of perceptualand semantic categories. In T. E. Moore (Ed.),Cognitive development and the acquisition oflanguage. New York: Academic Press, 1973.

Rosch, E. Linguistic relativity. In A. Silver-stein (Ed.), Human communication: Theo-retical explorations. New York: Halsted Press,1974.

Rosch, E. Universals and cultural specifics inhuman categorization. In R. Brislin, S. Boch-ner, & W. Lonner (Eds.), Cross-cultural per-spectives on learning. New York: HalstedPress, 1975.

Rosch, E. Classifications of real-world objects:Origins and representations in cognition. Bul-letin de Psychologic, in press, (a)

Rosch, E. Cognitive reference points. CognitivePsychology, in press, (b)

Rosch, E. Human categorization. In N. Warren(Ed.), Advances in cross-cultural psychology(Vol. 1). London: Academic Press, in press,(c)

Rosch, E. The nature of mental codes for colorcategories. Journal of Experimental Psy-

chology: Human Perception and Performance,in press, (d)

Rosch, E., & Mervis, C. B. Family resemblances:Studies in the internal structure of categories.Cognitive Psychology, in press.

Rosch, E., Mervis, C. B., Gray, W., Johnson, D.,& Boyes-Braem, P. Basic objects in naturalcategories. Journal of Experimental Psy-chology: General, in press.

Segal, S. J. Processing of the stimulus in imageryand perception. In S. J. Segal (Ed.), Imagery.New York: Academic Press, 1971.

Smith, E. E. Choice reaction time: An eval-uation of the major theoretical positions. Psy-chological Bulletin, 1968, 69, 77-110.

Smith, E. E, Shoben, E. J., & Rips, L. J. Struc-ture and process in semantic memory: A fea-tural model for semantic decisions. Psycho-logical Review, 1974, 81, 214-241.

Sternberg, S. Memory-scanning: Mental pro-cesses revealed by reaction-time experiments.American Scientist, 1969, 57, 421-457.

Till, R. E., & Jenkins, J. J. The effects of cuedorienting tasks on the free recall of words.Journal of Verbal Learning and Verbal Be-havior, 1973, 12, 489-498.

Titchener, E. B. Lectures on the experimentalpsychology of thought processes. New York:Macmillan, 1909.

Tulving, E., & Thomson, D. M. Encoding spec-ificity and retrieval processes in episodic mem-ory. Psychological Review, 1973, 80, 352-373.

Walsh, E. A., & Jenkins, J. J. Effects of orien-ting tasks on free recall in incidental learning:"Difficulty," "effort," and "process" explanations.Journal of Verbal Learning and Verbal Be-havior, 1973, 12, 481-488.

Winer, B. J. Statistical principles in experimentaldesign. New York: McGraw-Hill, 1971.

(Received December 19, 1974)

This research was supported by grants to theauthor (under her former name Eleanor RoschHeider) by the National Science Foundation(GB-38245X), by the Grant Foundation, and bythe National Institute of Mental Health (1 RO1MH24316-01). Portions of this data were pre-sented in a paper delivered at a meeting of thePsychonomic Society, St. Louis, November 1972.I wish to thank Brian Schmidt, Richard Duran,and R. Scott Miller for assistance in performingthe experiments and Carolyn Mervis for assist-ance with the work on the Kucera and Francisnorms leading to choice of the categories.

Requests for reprints should be sent to EleanorRosch, Department of Psychology, University ofCalifornia, Berkeley, California 94720.

COGNITIVE REPRESENTATIONS OF SEMANTIC CATEGORIES

APPENDIX

TABLE Al: NORMS FOR GOODNESS-OF-EXAMPLE RATING FOR NINE SEMANTIC CATEGORIES

229

Member

Goodne

Rank

>s of exampleSpecific

score Member

Goodnes

Rank

3 of exampleSpecific

score

Furniture

chairsofacouchtableeasy chairdresserrocking chaircoffee tablerockerlove seatchest of drawersdeskbedbureaudavenportend tabledivannight tablechestcedar chestvanitybookcaseloungechaise loungeottomanfootstoolcabinetchina closetbenchbuffet

1.51.53.53.556.56.589

101112131415.515.51718192021222324252627282930

1.041.041.101.101.331.371.371.381.421.44.48.54.58.59.61.61.70

1.831.982.112.132.152.172.262.432.452.492.592.772.89

lampstoolhassockdrawerspianocushionmagazine rackhi-ficupboardstereomirrortelevisionbarshelfrugpillowwastebasketradiosewing machinestovecounterclockdrapesrefrigeratorpictureclosetvaseashtrayfantelephone

313233343536373839404142434445464748495051525354555657585960

2.943.133.433.633.643.704.144.254.274.324.394.414.464.525.005.035.345.375.395.405.445.485.675.705.755.956.236.356.496.68

Fruit

orangeapplebananapeachpearapricottangerineplumgrapesnectarinestrawberrygrapefruitberrycherrypineappleblackberrymelonblueberryraspberrylemonblack raspberryboysenberrywatermeloncantaloupelimetangelo

123456.56.589

1011121314151617181920212223242526

1.071.081.151.171.181.361.361.371.381.521.611.771.821.861.192.052.092.142.152.162.212.382.392.442.452.50

papayahoneydewfigmangoguavapomegranatecranberrypassion fruitprunesgooseberrydatekumquatraisinmuskmelonpersimmonpawpawcoconutavocadopumpkintomatonutgourdolivepicklesquash

27282930313233.533.53536373839404142434445464748495051

2.582.732.862.883.033.053.223.223.303.333.353.393.423.483.634.304.505.375.395.586.016.026.216.346.55

230 ELEANOR ROSCH

TABLE Al—(Continued)

Member

Goodness of example

RankSpecific

score

Goodness of example

Member RankSpecific

score

Vehicle

automobilestation wagontruckcarbustaxijeepambulancemotorcyclestreetcarvanHondacable cartraintrolley (car)bicyclecarriageairplanebikeboatjetshipscootertractorwagon

12.345.55.5789

10111213141516171819202122232425

1.021.141.171.241.271.271.351.621.651.901.952.032.112.152.192.512.592.642.732.752.792.823.243.303.31

subwaytrailercartwheelchairyachttankgo-cartrowboatdogsledtricyclecanoeraftsubmarinesledhorserocketblimpskatescamelfeetskisskateboardwheelbarrowsurfboardelevator

26272829303132333435363738394041424344454647484950

3.323.503.553.683.763.843.853.923.954.004.014.374.514.614.634.744.814.995.225.345.405.545.725.785.90

Weapon

gunpistolrevolvermachine gunrifleswitchbladeknifedaggershotgunswordbombhand grenadeA-bombbayonetspearbazookacannonbow and arrowclublancebrass knucklesbullettomahawkmortararrowblackjacktankteargasmissilewhip

123456789

1011.511.513.513.515161718192021222324252627282930

1.031.071.091.161.171.351.401.411.461.471.611.611.691.691.701.761.961.982.092.142.382.442.482.562.662.682.742.882.903.11

ice pickhatchetslingshotfistsaxebowrazorrazor bladerocketjudostickpoisonrockstonegaschainscissorsbrickspitchforkhammerwordshandpiperopeairplanefootcarscrewdriverglassshoes

313233343536373839404142434445464748495051525354555657585960

3,143.163.293.313.343.443.823.833.903.944.044.074.184.264.314.454.504.644.674.784.935.015.045.085.095.235.315.405.606.23



Member

Goodnes

Rank

3 of exampleSpecific

score

Goodness of example

Member RankSpecific

score

Vegetable

peacarrotgreen beansstring beansspinachbroccoliasparaguscorncauliflowerbrussels sproutssquashlettucecelerycucumberbeetsgreenstomatolima beansartichokesturnipeggplantromainegreen peppersokraradishesonionsbeangreen onion

123456789

1011121314IS1617181920212223.S23.525262728

1.071.151.181.211.221.281.411.551.621.721.831.851.902.052.082.182.232.282.322.372.382.442.492.492.512.522.542.60

potatoparsnipturnip greenscollardwax beanswatercressblackeyed peasleekpepperssweet potatoyamsparsleyendiverutabagamushroomavocadorhubarbkaleescarolesauerkrautpicklesbaked beanspumpkinseaweedgarlicdandelionpeanutrice

29303132333435363738394041424344454647484950515253545556

2.892.912.952.993.023.043.063.153.213.273.313.323.393.423.563.623.663.673.904.184.574.734.745.045.075.205.565.59

Carpenter's tool

sawhammerrulerscrewdriverdrillnailstape measuresawhorsesandpapersanderlevelplanefiletoolboxT-squarechiselrasppencilhacksawsquarebenchplierswrenchladderlathevisescrewswoodwedgelumber

123456789

101112131415161718192021222324252627.527.52930

1.041.341.481.561.591.671.691.771.781.791.821.912.012.122.222.262.372.392.412.442.502.592.602.642.752.762.772.772.802.84

blueprintsbraceawlcrowbarhingemiter boxknifechalknutsboltsplumb linebrushsledge hammergluepaintbrushvarnishapronstaplerextension cordsoldering ironplasterwheelbarrowaxeslide rulecementanvilhatchetragsscissorscrane

31323334.534.5363738.538.5404142.542.54445464748495051.551.55354555657585960

2.902.923.093.123.123.233.443.503.503.633.653.743.743.793.813.914.064.214.294.394.434.434.534.574.915.105.155.205.365.70

232 ELEANOR ROSCH


Member

Goodness of exampleSpecific

Rank score Member

Goodness of exampleSpecific

Rank score

Bird

robinsparrowbluejaybluebirdcanaryblackbirddovelarkswallowparakeetoriolemockingbirdredbirdwrenfinchstarlingcardinaleaglehummingbirdseagullwoodpeckerpigeonthrushfalconcrowhawkraven

123456789

10111213.513.SIS1617.517.5192021222324252627

1.021.181.291.311.421.431.461.471.521.531.611.621.641.641.661.721.751.751.761.771.781.811.891.961.971.992.01

goldfinchparrotsandpiperpheasantcatbirdcranealbatrosscondortoucanowlpelicangeesevulturestorkbuzzardswanflamingoduckpeacockegretchickenturkeyostrichtitmouseemupenguinbat

282930313233343536373839404142434445464748495051525354

2.062.072.402.692.722.772.802.832.952.962.983.033.063.103.143.163.173.243.313.394.024.094.124.354.384.536.15

Sport

footballbaseballbasketballtennissoftballcanoeinghandballrugbyhockeyice hockeyswimmingtrackboxingvolleyballlacrosseskiinggolfpolosurfingwrestlinggymnasticscricketsquashbadmintonracingpole vaultfencingbowlingwater skiingice skating

123456789

10111213141516171819202122.522.524.524.52627282930

1.031.051.121.151.291.411.421.431.441.451.531.611.661.711.721.761.771.801.841.871.881.991.992.082.082.092.132.182.222.29

jai alaiskatingskindivingsailingdivingarcheryjudocar racingping pongrowingfishinghorseback ridingrunninghorse racinghikingweight liftingcroquethorseshoesboatingpoolbilliardshuntingjump ropecampingchessdancingcheckerscardssunbathing

313233343536373839404142434445464748495051525354.554.556575859

2.302.392.402.442.502.542.632.782.802.822.842.853.013.183.503.593.603.723.753.823.954.055.005.075.075.495.645.796.75


TABLE Al— (Continued)

Member

Goodness of example

RankSpecificscore

Goodness of example

Member RankSpecific

score

Toy

dolltopjack-in-the-boxtoy soldieryo-yoblockmarblesrattlestuffed animalwater pistolteddy bearrocking horsedoll houseballjackspaper dollserector sethula hoopjump ropepogo stickclaywagonkitetraintricyclecoloring bookcrayonstruckpuzzlefire engine

123.53.556789

1011121314151617.517.5192021.521.52324252627282930

1.411.481.611.611.621.631.741.791.871.881.901.911.951.962.092.142.162.162.292.442.492.492.512.592.612.682.712.782.822.83

balloonskatesbaseballdrumfootballgameswingboatmonopolysledseesawstiltscheckersbatcarairplanebicycletractorsandboxbikebow and arrowropedishescardsmitthorsegunanimalstennis racketbooks

3132333435363738394041.541.54344454647.547.5495051.551.55354555657585960

3.073.133.183.233.283.353.353.363.423.433.483.483.563.623.633.663.683.683.713.774.204.204.514.564.774.975.365.395.405.91

Clothing

pantsshirtdressskirtblousesuitslacksjacketcoatsweatersweatshirtunderpantssports jacketjumperpantiessocksparkapajamasundershirtovercoatnightgownraincoatbathing suitbathrobeslipbrashoestuxedo

12.52.5456789

1011121314.514.516171819202122.522.52425262728

1.121.141.141.211.271.451.491.681.881.891.952.012.052.082.082.132.192.252.262.292.392.442.442.652.672.682.732.76

stockingsvestnylonscapebootssandalstiegirdlebeltscarfmittensslippershatovershoesglovesapronearmuffshandkerchiefpursehairbandring_earringswatchcuff linksnecklacebraceletcane

29303132333435363738394041424344454647484950.550.552535455

2.792.812.983.383.423.563.713.723.933.963.994.084.204.424.534.575.045.875.925.986.116.156.156.186.216.246.25

cognitive representation of semantic categories-rosch-1975

Documents