a spreading-activation theory of semantic...

Psychological Review1975, Vol. 82, No. 6, 407-428

A Spreading-Activation Theory of Semantic Processing

Allan M. CollinsBolt Beranek and Newman Inc.,

Cambridge, Massachusetts

Elizabeth F. LoftusUniversity of Washington

This paper presents a spreading-activation theory of human semantic pro-cessing, which can be applied to a wide range of recent experimental re-sults. The theory is based on Quillian's theory of semantic memory searchand semantic preparation, or priming. In conjunction with this, several ofthe miscondeptions concerning Qullian's theory are discussed. A numberof additional assumptions are proposed for his theory in order to apply itto recent experiments. The present paper shows how the extended theorycan account for results of several production experiments by Loftus, Juolaand Atkinson's multiple-category experiment, Conrad's sentence-verificationexperiments, and several categorization experiments on the effect of semanticrelatedness and typicality by Holyoak and Glass, Rips, Shoben, and Smith,and Rosch. The paper also provides a critique of the Smith, Shoben, andRips model for categorization judgments.

Some years ago, Quillian1 (1962, 1967)proposed a spreading-activation theory ofhuman semantic processing that he tried toimplement in computer simulations of mem-ory search (Quillian, 1966) and com-prehension (Quillian, 1969). The theoryviewed memory search as activation spread-ing from two or more concept nodes in asemantic network until an intersection wasfound. The effects of preparation (or prim-ing) in semantic memory were also ex-plained in terms of spreading activation fromthe node of the primed concept. Ratherthan a theory to explain data, it was a the-ory designed to show how to build humansemantic structure and processing into acomputer.

This research was supported by the NationalInstitute of Education, U.S. Department of Health,Education, and Welfare, Project 1-0420, underContract OEC-1-71-0100(508) with Bolt Beranekand Newman Inc. Revision of the paper was sup-ported by a grant from the John Simon Guggen-heim Memorial Foundation to the first author. Wewould like to thank Stephen Woods, Colin Mac-Leod, Mark L. Miller, and the reviewers for theircomments on previous drafts of this paper.

Requests for reprints should be sent to Allan M.Collins at Bolt Beranek and Newman Inc., SOMoulton Street, Cambridge, Massachusetts 02138.

1 Quillian's theory of priming appeared in theunpublished version of the 1967 paper (i.e., CIPPaper No. 70, Carnegie Institute of Technology,

Since the theory was proposed, there havebeen a number of experiments investigatingretrieval and priming in semantic memory.In the present paper, we attempt to showhow an elaboration of Quillian's basic theorycan account for many of the results. In thefirst section, we briefly review the originaltheory while trying to correct a number ofthe common misunderstandings concerningit. In the second section, we extend thetheory in several respects, and in the thirdsection show how the extended theory dealswith some recent experimental findings. Inthe fourth section we compare the theory tothe model of Smith, Shoben, and Rips(1974).

QUILLIAN'S THEORY OF SEMANTIC MEMORY

The fact that Quillian's theory was devel-oped as a program for a digital computerimposed certain constraints on the theory,which Quillian felt were psychologically un-realistic. We will recount the theory as heproposed it, and then elaborate the theoryin psychological terms. The theory madea number of assumptions about structure andprocessing in human semantic memory. Abrief discussion of these assumptions follows.

People's concepts contain indefinitely largeamounts of information. Quillian used theexample of a machine. If one asks peopleto tell everything they know about machines,

407

408 ALLAN M. COLLINS AND ELIZABETH F. LOFTUS

they will start off giving obvious properties,for example, that machines are man-madeand have moving parts. But soon peoplerun out of obvious facts and begin givingfacts that are less and less relevant, for ex-ample, that a typewriter is a machine or eventhat the keys on the IBM electric type-writers select the position of a ball thatstrikes the ribbon against the paper. Theamount of information a person can gen-erate about any concept in this way seemsunlimited.

In these terms concepts correspond to par-ticular senses of words or phrases. For ex-ample, not only is the noun "machine" aconcept, but the verb "to machine" is a con-cept ; the "particular old car I own" is aconcept; the notion of "driving a car" is aconcept; even the notion of "what to do ifyou see a red light" has to be a concept.Thus, people must have a very large num-ber of concepts, and concepts must have verycomplicated structures.

A concept can be represented as a nodein a network, with properties of the conceptrepresented as labeled relational links fromthe node to other concept nodes. Theselinks are pointers, and usually go in bothdirections between two concepts. Links canhave different criterialities, which are num-bers indicating how essential each link is tothe meaning of the concept. The criteriali-ties on any pair of links between two con-cepts can be different; for example, it mightbe highly criteria! for the concept of a type-writer that it is a machine, and not verycriterial for the concept of machine that onekind is a typewriter. From each of thenodes linked to a given node, there will belinks to other concept nodes and from eachof these in turn to still others. In Quil-lian's theory, the full meaning of any con-cept is the whole network as entered fromthe concept node.

The links are not simply undifferentiatedlinks, but must be complicated enough torepresent any relation between two concepts.In the original theory, Quillian proposed fivedifferent kinds of links: (a) superordinate("isa") and subordinate links, (b) modifierlinks, (c) disjunctive sets of links, (d) con-

junctive sets of links, and (e) a residualclass of links, which allowed the specifica-tion of any relationship where the relation-ship (usually a verb relationship) itself wasa concept. These different kinds of linkscould be nested or embedded to any degreeof depth, so that the format was designed tobe flexible enough to express anything, how-ever vague or specific, that can be expressedin natural language.

The search in memory between conceptsinvolves tracing out in parallel (simulatedin the computer by a breadth-first search)along the links from the node of each con-cept specified by the input words. Thewords might be part of a sentence or stimuliin an experimental task. The spread of ac-tivation constantly expands, first to all thenodes linked to the first node, then to allthe nodes linked to each of these nodes, andso on. At each node reached in this pro-cess, an activation tag is left that specifiesthe starting node and the immediate prede-cessor. When a tag from another startingnode is encountered, an intersection betweenthe two nodes has been found.2 By follow-ing the tags back to both starting nodes, thepath that led to the intersection can bereconstructed.

When an intersection has been found, it isnecessary to evaluate the path to decide if itsatisfies the constraints imposed by syntaxand context. The. complicated kinds of de-cision rules that are invoked for comprehen-sion of sentences in this evaluation phaseare described in Quillian (1969). For cate-gorization tasks these rules are describedby Collins and Quillian (1972b) and in thenext section of this paper. As an example,in a phrase such as "the fall leaves," a pathfound between the concept "to fall" and the

2 There can be intersections with more than twostarting nodes, but we have limited our discussionin this paper to the case of two nodes (as didQuillian, 1966, initially). The basic assumptionsin Quillian's theory and our elaboration can applyto intersections with more than two starting nodes,but this leads to complications in the evaluation ofintersections. Some of these were discussed byQuillian (1969) in regard to comprehension, butfor the experiments considered in this paper, onlythe case of two starting nodes needs to be con-sidered.

THEORY OF SEMANTIC PROCESSING 409

concept "tree leaf" would be rejected as awrong interpretation, because syntax re-quires a participial form of "fall" to fit thatinterpretation. If a path found is rejected,other paths are considered in the order inwhich they are found.

Priming (or preparation) involves thesame tracing process that was described formemory search. When a concept is primed,activation tags are spread by tracing an ex-panding set of links in the network out tosome unspecified depth. When another con-cept is subsequently presented, it has tomake contact with one of the tags left earlierto find an intersection. One of the non-obvious implications of this view of primingis that links as well as nodes will be primed.This is because Quillian treated links them-selves as concepts (see above). Thus prim-ing a node such as "red" will prime the linksinvolving the relation "color" throughout thenetwork. This provides a very powerfulcontext mechanism.

Common Misinterpretations ofQuillian's Theory

There is a rich variety of misinterpreta-tions of Quillian's theory, many of them de-riving from Collins's (Collins & Quillian,1969, 1970a, 1970b) simplifications of thetheory. The problem arose because Collinsand Quillian were investigating specific as-pects of the theory and only describedenough of the theory to motivate a particu-lar experiment. In turn experimenters madeinterpretations of these simplified versions,which did not fit with the theory as de-scribed elsewhere (Bell & Quillian, 1971;Quillian, 1966, 1969).

Perhaps the most prevalent misinterpreta-tion of Quillian's theory concerns the ideaof cognitive economy (Anderson & Bower,1973; Conrad, 1972). In this regard, it isimportant to distinguish the strong theory ofcognitive economy, which Conrad takes issuewith in her attack on Collins and Quillian(1969), and the weak theory of cognitiveeconomy, which Collins and Quillian weretesting (though they did not spell it outclearly enough). As Conrad (1972) states,she rejects the "hypothesis that all proper-

ties are stored only once in memory andmust be retrieved through a series of infer-ences for all words except those that theymost directly define" (p. 153). This is astatement of the strong theory of cognitiveeconomy, Undoubtedly the Collins andQuillian (1969) paper gave rise to this no-tion, but the authors cautioned against mak-ing that interpretation of the theory. Asthey said, "people surely store certain prop-erties at more than one level in the hier-archy" (p. 242), and they cited the mapleleaf as an example of this general rule.

The strong theory requires erasing infor-mation whenever it applies at a more gen-eral level. If a person learns a robin canfly and then later that birds fly, the strongtheory implies that "flying" must be erasedfrom "robin." The weak theory of cogni-tive economy merely assumes that everytime one learns that X is a bird, one does notat that time store all the properties of birdswith X in memory. Thus, an inference willbe necessary to decide that X can fly, unlessone encounters this fact directly. Hence,Collins and Quillian, in testing the weaktheory of cognitive economy, picked in-stances where people were not likely to haveencountered the general property with thespecific instance (e.g., "A wren can fly").The point of the experiment was to testwhether it was possible to measure infer-ence time, when the weak theory of cogni-tive economy implies that an inference islikely to be necessary for most subjects.

Another assumption sometimes made aboutQuillian's theory is that all links are equal(Anderson & McGaw, 1973; Rips, Shoben,& Smith, 1973; Wilkins, Note 1). In Quil-lian's original theory, there were criterialitytags on links, as we described earlier. InCollins and Quillian (1969, 1972b) linkswere assumed to have differential accessi-bility (i.e., strength or travel time). Theaccessibility of a property depends on howoften a person thinks about or uses a prop-erty of a concept. Whether criteriality andaccessibility are treated as the same or dif-ferent is a complex issue, but network mod-els allow them to be treated either way.Thus for example, even though "lungs,"


"hands," and "warts" are all linked directlyto the concept "human," these links neednot be in any sense equal. The same istrue for the links between "bird" and itsexemplars, such as "robin," "chicken," or"penguin." Rips et al. (1973) suggest thatintermediate nodes are necessary for a net-work model to explain the reaction time dif-ferences they find in categorizing differentbirds. This makes the mistaken assumptionthat all links are equally criterial or accessi-ble in any network model. It turns out,however, that differences in links are crucialto many different aspects of human seman-tic processing as Carbonell and Collins(1973) point out in their discussion of im-portance (or criteriality) tags.

A related implication of the Rips et al.(1973) paper and also a more recent paperof Smith et al. (1974) is that feature modelscan account for data that network modelscannot. A feature model posits that a con-cept consists of a set of values on a largenumber of semantic dimensions (e.g., ani-mateness, color, etc.). What is strangeabout this argument is that network modelswere developed as a method of representingfeatures in a computer. Any process thatcan be represented in a feature model isrepresentable in a network model; in par-ticular, the Smith et al. model itself couldbe implemented in a semantic network (Hol-lan, 1975). In fact, network models areprobably more powerful than feature mod-els, because it is not obvious how to handleinferential processing or embedding in fea-ture models.

Smith et al. (1974) argued in favor offeature models because their data for com-parison of concepts seemed to fit a featurecomparison process. What should be em-phasized about Quillian's theory is that theparallel search would inevitably lead to justsuch a feature comparison process, thoughthe process would take place over a periodof time as different connections are found.One way that Quillian's theory is differentfrom the Smith et al. models is that super-ordinate connections, if they exist, wouldalso be found and evaluated. The distinc-tion between these two theories is so crucial

that we will discuss it at length in conjunc-tion with the spreading activation theory'sexplanation of the Rips et al. (1973) results.

Another common misconception of Quil-lian's theory shows up in Juola and Atkin-son's (1971) work on categorization judg-ments. In a categorization task, responsetime is measured for a subject to decidewhether or not a particular instance (e.g.,"car") is a member of one or more cate-gories (e.g., "flower" or "vehicle"). Juolaand Atkinson assume that in Quillian's the-ory the memory search to make a categoriza-tion judgment proceeds from the instanceto the category. In fact, the wording inCollins and Quillian (1969) mistakenly givesthat impression. But Quillian's theory(1966, 1969) assumes the search proceedsfrom both the instance and category in par-allel. However, if one or the other is pre-sented first, this gives the search from thatnode a head start, which is the notion ofpriming. Juola and Atkinson's experimentinvolves priming in a complicated way,which we will discuss below.

Anderson and Bower (1973) reject aQuillian-like model of a parallel search,while acknowledging that their data arecompatible with "a parallel model whosesearch rate is slower in proportion to thenumber of paths that must be searched" (p.371). Anderson and Bower's argumentimplies wrongly that Quillian has made theindependence assumption for his parallelsearch. An independent parallel search islike a race where the speed of each runneris independent of the other runners. Thisis a common assumption in psychology, be-cause it makes it possible to assign an upperbound to reaction time (see Sternberg,1966). But there is no difficulty for Quil-lian's theory if the parallel search rate de-pends on the number of paths searched.Hence, Anderson and Bower's data areperfectly compatible with Quillian's parallelsearch.

The above discussion, then, shows whatQuillian's theory is not, or at least some ofwhat it is not. Several other misconcep-tions are discussed in Collins and Quillian(1972b), in particular the notion that Quil-


Han's theory of memory is rigidly hierarchi-cal, which Anderson and Bower (1973, p.379) still believe, and Schaeffer and Wal-lace's (1970) argument that Quillian's the-ory predicts it will always take less time tocompare concepts that are close together inthe semantic network. We will return tosome of these same papers below, in orderto describe how the extended version ofQuillian's theory accounts for some of theresults these experimenters have used toreject Quillian's theory.

THE EXTENDED THEORY

In order to deal with the specific experi-mental results that have appeared in recentyears, several more processing and struc-tural assumptions must be added to the basicQuillian theory. These do not bend thetheory, but merely elaborate it in such away that it can be applied to the kinds ofexperiments on semantic memory that havebeen performed recently. The elaborationmay itself be wrong, so our mistakes shouldnot be held against Quillian's theory.

Local Processing Assumptions

There are four local processing assump-tions in the extended theory. These fourassumptions transform the theory from com-puter terms to quasi-neurological terms, a laPavlov. But all the assumptions of the origi-nal theory should be preserved despite thetransformation, except that activation tagsare to be considered as source-specific acti-vation (i.e., activation that is traceable toits node of origin).

1. When a concept is processed (or stim-ulated), activation spreads out along thepaths of the network in a decreasing gradi-ent. The decrease is inversely proportionalto the accessibility or strength of the linksin the path. Thus, activation is like a sig-nal from a source that is attenuated as ittravels outward.

2. The longer a concept is continuouslyprocessed (either by reading, hearing, orrehearsing it), the longer activation is re-leased from the node of the concept at afixed rate. Only one concept can be ac-tively processed at a time, which is a limita-

tion imposed by the serial nature of thehuman central process (Collins & Quillian,1972b). This means that activation canonly start out at one node at a time. Butit continues in parallel from other nodesthat are encountered as it spreads out fromthe node of origin.

3. Activation decreases over time and/orintervening activity. This is a noncommittalassumption that activation goes away gradu-ally by some mechanism. Assumptions 2and 3 impose a limitation on the amount ofactivation that can be allocated in primingmore than one concept, because the moreconcepts that are primed, the less each willbe primed.

4. With the assumption that activation isa variable quantity, the notion of intersectionrequires a threshold for firing. The assump-tion is that activation from different sourcessummates and that when the summation atthe point of intersection reaches threshold,the path in the network producing the in-tersection will be evaluated.

Global Assumptions About MemoryStructure and Processing

There are three assumptions in the ex-tended theory concerned with the globalstructure of memory and its processing.These are generalizations of Loftus's (Note2) arguments that semantic memory is or-ganized primarily into noun categories andthat there is a "dictionary" (or lexicalmemory) separate from the conceptual net-work.

5. The conceptual (semantic) network isorganized along the lines of semantic simi-larity. The more properties two conceptshave in common, the more links there arebetween the two nodes via these propertiesand the more closely related are the con-cepts. This means that different vehiclesor different colors will all be highly inter-linked through their common properties.This also implies that red things (e.g., fireengines, cherries, sunsets, and roses) arenot closely interlinked, despite the one prop-erty they have in common. In these termssemantic relatedness is based on an aggre-


AMBULANCE/y-F|RE

ENGINE

FIGURE 1. A schematic representation of concept relatedness in a stereo-typical fragment of human memory (where a shorter line represents greaterrelatedness).

gate of the interconnections between twoconcepts.3

3 Semantic relatedness is a slightly different no-tion from semantic distance, though the two termsare sometimes used interchangeably. Semantic dis-tance is the distance along the shortest path, andsemantic relatedness (or similarity) is an aggre-gate of all the paths. Two concepts may be closein distance, say by a path through "red," and stillnot be closely related because that is the onlypath. Our use of close to refer to both relation-ships is admittedly confusing. In this paper weshall use close to refer to relatedness or similarity,though in sortie tasks (Quillian, 1966) it is onlydistance that matters.

Figure 1 illustrates this aggregate notionof concept relatedness for a hypotheticalhuman memory. (It is the kind of diagramthat the scaling techniques of Rips et al.,1973, would produce.) In the figure thevarious vehicles are shown as closely re-lated, because of the numerous individualconnections that are assumed to exist be-tween them. Conversely, the concepts asso-ciated with "red" are shown as less related,because of the presumed paucity of inter-connections between them.

From the assumption that memory is or-ganized according to semantic similarity, to-


gethcr with earlier assumptions, it followsthat if "vehicle" is primed, activation at anytype of vehicle will accumulate from manyneighboring nodes. That is to say, to thedegree that "fire engine" is primed by "ve-hicle," it will in turn prime "ambulance,""truck," "bus," etc., and each of these inturn will prime the others. On the otherhand, if "red" is primed, the activation thatspreads to "fire engine" will not prime"cherries," "roses," or "sunsets" to anygreat extent, because there are so few con-nections between these concepts. Instead,"fire engine" will tend to prime other ve-hicles, and "cherries" to prime other fruits.Hence, the same amount of activation willbe diffused among a greater number of con-cepts.

6. The names of concepts are stored in alexical network (or dictionary) that is or-ganized along lines of phonemic (and tosome degree orthographic) similarity. Thelinks from each node in the lexical networkare the phonemic properties of the name,specified with respect to their position in theword. The properties stored about namesare assumed to be the properties that Brownand McNeill (1966) found people couldidentify about words on the "tip of theirtongue." Each name node in the lexicalnetwork is connected to one or more conceptnodes in the semantic network.

7. Loftus's (Note 2) data lead to the fur-ther assumption that a person can controlwhether he primes the lexical network, thesemantic network, or both. For example, aperson can control whether to prime (a.)words in the lexical network that sound like"bird," (b) concepts in the semantic net-work related to "bird," or (c) words in thelexical network corresponding to the con-cepts in (b). This control over primingcan be thought of in terms of summation ofdiffuse activation for an entire network (per-haps in a particular part of the brain) andsource-specific activation released from aparticular node. Thus, (a) would derivefrom activation of the lexical network to-gether with the word "bird," (b) would de-rive from activation of the semantic networktogether with the concept "bird," and (c)

would derive from activation of both net-works together with the concept "bird."

Assumptions About SemanticMatching Process

There are a number of assumptions aboutthe decision process for evaluating whetheror not two concepts match semantically.This is a fundamental process that occursin many aspects of language processing,such as matching referents, assigning cases,and answering questions (Collins & Quil-lian, 1972b; Collins, Warnock, Aiello, &Miller, 1975). Categorization tasks, whichask "Is X a Y?" (where X and Y are con-cepts), directly investigate this process. Thedecision process described here is a more ex-plicit and somewhat revised version of theprocess postulated by Collins and Quillian(1972b), with additions to encompass theresults of Holyoak and Glass (1975).

8. In order to decide whether or not aconcept matches another concept, enoughevidence must be collected to exceed eithera positive or a negative criterion. The evi-dence consists of various kinds of intersec-tions that are found during the memorysearch. Evidence from different paths inmemory sum together. Positive and nega-tive evidence act to cancel each other out, asshown by dialogue excerpts in Carbonell andCollins (1973). Failure to reach either cri-terion before running out of relevant evi-dence leads to a "don't know" response(Collins et al,, 1975). This process is es-sentially the Bayesian decision model that iscommon in the reaction time literature (see,for example, Fitts, 1966; Stone, I960).

There are a number of different kinds ofpaths between the two concepts that consti-tute positive or negative evidence. Any ofthese types of evidence might contribute toa particular decision. The different typesare listed in Table 1 and described below inAssumptions 9-13.

9. If the memory search finds that thereis a superordinate (or a negative superordi-nate) connection from X to Y, that factalone can push the decision over the posi-tive (or negative) criterion. Superordinatelinks act like highly criterial property links


TABLE ITYPES OF PATHS FOUND IN MEMORY

THAT CONSTITUTE POSITIVE ORNEGATIVE EVIDENCE

Positive evidence Negative evidence

Superordinate connection

Property comparison,matching property

Wittgenstein strategy,matching property

Negative superordinateconnection

Property comparison,distinguishing property

Wittgenstein strategy,distinguishing property

Mutually exclusivesubordinates

Counterexamples

(see below). For example, it is conclusivepositive evidence that a mallard is a bird, ifsuperordinate links are found between "mal-lard" and "duck" and between "duck" and"bird." Similarly, if a negative superordi-nate link is found between "bat" and "bird,"it is conclusive evidence that a bat is nota bird.

10. If the memory search finds propertieson which X and Y match (i.e., commonproperties), this is positive evidence propor-tional, to the criteriality of the property forY. If the memory search finds propertieson which X and Y mismatch (i.e., distin-guishing properties), this is negative evi-dence proportional to the criteriality of theproperty for Y. There is an asymmetry inthe weighing of positive and negative evi-dence in a property comparison, because amismatch on just one fairly criterial prop-erty can lead to a negative decision, whereasmost of the highly criterial properties mustmatch in order to reach a positive decision(Collins & Quillian, 1972b).

It Is important to note that property com-parisons and superordinate connections sumtogether in reaching either criterion as thememory search finds them. Thus, distin-guishing properties make it harder to reachthe positive criterion when there is a super-ordinate connection and therefore slow downthe process.

As an example of property comparison,suppose there is no superordinate connectionin a particular person's memory between"mink" and "farm animal" and between"cat" and "farm animal." Then the deci-sion as to whether minks or cats are farm

animals might be based on a comparison ofthe properties of minks or cats on one hand,and farm animals on the other. The mostcriterial properties of farm animals are pre-sumably being animate and being kept onfarms, but other less criterial properties in-clude being domesticated, being raised forsome purpose, or being kept in barns oroutside. How a particular person wouldweigh the various properties of minks andcats to decide whether they are farm ani-mals would vary from person to person (ourintuition is that minks are farm animals andcats are not, even though both have the twomost criterial properties of farm animals—what Smith et al., 1974, called definingproperties). This decision strategy is simi-lar to that proposed by Smith et al., as wewill discuss later.

11. The Wittgenstein strategy is a variantof the property comparison strategy. It ispostulated on the basis of Wittgenstein's(1953) observation that to decide whethersomething is a game (for example, frisbee),a person compares it to similar instances thatare known to be games. Our assumption isthat if any properties of X are found thatmatch properties of another instance whosesuperordinate is Y, these constitute posi-tive evidence. Similarly, any distinguishingproperties constitute negative evidence. Inthe Wittgenstein strategy, unlike the prop-erty comparison strategy, matching proper-ties count just as much toward a positivedecision as distinguishing properties counttoward a negative decision.

To illustrate the Wittgenstein strategy,Collins and Quillian (1972b) pointed outthat in deciding whether a stagecoach is avehicle, it might be compared to a car. Themany properties that a stagecoach has incommon with a car constitute strong posi-tive evidence that a stagecoach is a vehicle.But notice that a stagecoach does not havea motor, which is highly criterial for beinga car. Though this is strong evidence thata stagecoach is not a car, it is only weakevidence that it is not a vehicle. This illus-trates how the same evidence is weigheddifferently in the property comparison strat-egy and the Wittgenstein strategy. The


final decision that a stagecoach is a vehiclemight depend both on matching propertiesbetween a stagecoach and vehicles in gen-eral (conveyance, motion, etc.) and match-ing properties between a stagecoach and par-ticular vehicles like a car (seats, doors, etc.).Thus the property comparison strategy andthe Wittgenstein strategy might combine todetermine a person's response.

12. The mutually exclusive subordinatesstrategy was necessary for programming acomputer to answer questions (Collins et al.,1975). Holyoak and Glass (1975) arguethat this strategy accounts for some of theirreaction time data (they call it a contradic-tion). The assumption is that if two con-cepts have a common superordinate withmutually exclusive links into the commonsuperordinate, then this constitutes strongnegative evidence, almost comparable to anegative superordinate link.

For example, if the question is whether amallard is an eagle, the fact that a mallardis a duck and ducks and eagles are mutuallyexclusive kinds of birds is rather conclusiveevidence that a mallard is not an eagle.Though Holyoak and Glass (1975) do notmention it, the mutually exclusive restrictionis necessary. For example, the fact thatMike Mansfield is a politician does not ex-clude him from being a lawyer. Although"politician" and "lawyer" are both occupa-tional roles, they are not mutually exclusiveand in fact most politicians are lawyers.But lacking specific information to the con-trary, people may make a default assump-tion of mutual exclusivity when two con-cepts have a common superordinate.

13. Counterexamples also can be used asnegative evidence. This strategy derivesfrom Holyoak and Glass (1975), who arguethat statements of the kind "All birds arecanaries" are disconfirmed by finding acounterexample, such as "robin." If thequestion is of the form "Ts X a Y?" andthere is a superordinate link from Y to X,then finding a counterexample involves find-ing a Z that also has X as superordinate andis mutually exclusive from Y. This is con-clusive evidence that X is not always a Y.

Holyoak and Glass (1975) discusss coun-terexamples in the context of the universalquantifier "all," but the same process wouldoccur for a question of the kind "Is a mar-supial a kangaroo?" In such a case, re-trieving a counterexample (such as a wall-aby) can be used to determine that kanga-roos are a subset of marsupials and notequivalent (e.g., "automobiles" and "cars"are equivalent concepts).

Though these five kinds of evidence (As-sumptions 9-13) are the only ones we havepostulated for the semantic matching pro-cess seen in categorization tasks, there maybe other kinds of evidence of this sort. Weshould stress that there are many otherkinds of evidence people use for answeringmore complicated questions (Collins et al.,1975). It is beyond the scope of this paper,however, to consider all the different wayspeople use evidence to make semantic deci-sions.

RECENT EXPERIMENTS

In this section, we discuss how the theorydeals with some different kinds of recentexperiments. The four types of studies towhich we apply the theory are (a) severalproduction experiments by Loftus (Freed-man & Loftus, 1971; Loftus, 1973a, 1973b,Note 2) ; (b) Juola and Atkinson's (1971)multiple-category experiment; (c) the Con-rad (1972) sentence-verification experiment;and (d) several categorization experimentson the effects of semantic relatedness andtypicality (Holyoak & Glass, 1975; Rips etal., 1973; Rosch, 1973; Smith et al., 1974).We intend to deal with the major kinds ofavailable findings to which the Quillian the-ory has not yet been applied. Our objectiveis to show how a spreading-activation theorycan handle these results, not to consider allthe possible alternative explanations of theexperiments.

Production Experiments of Loftus

There are several Loftus experiments wewant to discuss in terms of the spreading-activation theory. The first of these is anexperiment by Freedman and Loftus (1971),in which subjects had to produce an in-


stance of a category that began with a givenletter or was characterized by a given ad-jective. For example, subjects might beasked to name a fruit that begins with theletter A or a, fruit that is red. On sometrials the category was shown first and onsome trials second. Hence, this was a prim-ing experiment in that one concept was ac-tivated before the other. Reaction time wasmeasured from the onset of the secondstimulus.

Our concern is with the finding that sub-jects were faster when the category (e.g.,"fruit") was given first than when eitherthe letter or the adjective was given first.This basic result was later replicated evenfor cases in which the instance named wasa more frequent associate to the adjectivethan to the category noun (e.g., "lemon" isa closer associate of "sour" than of "fruit").

The explanation in terms of the theoryis as follows: When a noun, such as "fruit,"is presented first, the activation spreads tonodes connected to "fruit," among which areinstances such as "apple," "pear," "peach,""orange," and "lemon." But these conceptsare all highly interlinked with each other(though some, such as "orange" and"lemon," are more closely interlinked thanothers). Thus, the total amount of activa-tion is spread among a relatively small num-ber of closely interlinked concepts (see As-sumption 5). However, when an adjectiveor letter is presented first, say "red" or"A," the activation spreads to a much widerset of concepts, which are not particularlyinterlinked with each other. Thus, thelarge variety of different things that are redor that start with the letter A will receiverelatively little priming when the adjectiveor letter are given first. Because primingthe noun leads to a greater accumulation ofactivation on the instances, these are closerto their threshold for firing, so that it takesless stimulation, and hence less time, to trig-ger an intersection when the second stimulusis presented.

Freedman and Loftus (1971) explainedtheir finding in terms of entering the cate-gory when a noun is presented and enteringa cluster within the category when the adjec-

tive or letter is presented. Thus if thenoun is presented first, the subject can enterthe category immediately and need onlychoose the correct cluster when the adjectiveor letter is presented. But if the adjectiveor letter is presented first, the subject mustwait until the category is presented, becausethe cluster is specific to that category. (How-ever, Loftus, Note 2, has revised this expla-nation for the letter stimulus in her dic-tionary-network model.)

The Freedman and Loftus explanation isnot altogether different from the explanationoffered here, though our theory is less rigidlyhierarchical. The rigid hierarchy gets intotrouble with errors such as one we encoun-tered where a subject produced "Ben Frank-lin," given the stimulus pair "president"and "F," although he later recognized hismistake. In an activation theory, "Frank-lin" is a very likely intersection starting at"president" and "F," because he is so closelylinked with the concept, "president," andsome of its foremost instances, such as"Washington." Such a wrong intersectionwas likely in this case because the correctanswer (prior to Ford) was "Fillmore,"who is rather inaccessible and unlikely to befound quickly enough to preclude finding"Franklin." Once such an intersection isfound, it is only by evaluating the connec-tion between "president" and "Franklin,"that it can be rejected (see Assumptions8-13). It is a general problem of category-search models that they cannot deal withsuch errors.

Perhaps the major advantage of thespreading-activation theory over the Freed-man and Loftus (1971) explanation is intying their result to a parallel result ina quite different experiment by Loftus(1973b). In a categorization experiment,Loftus found that the direction of the asso-ciation between the category and the instancedetermined whether subjects were fasterwhen given the category first or the instancefirst. In the experiment she used four kindsof category-instance pairs: (a) pairs whereboth the category and instance evoked theother with high frequency (e.g., "tree-oak") ; (b) pairs where the category evoked


the instance with high frequency, but the in-stance evoked the category with low fre-quency (e.g., "seafood-shrimp"); (c) pairswhere the category evoked the instance withlow frequency, but the instance evoked thecategory with high frequency (e.g., "insect-butterfly") ; and (d) pairs where both thecategory and instance evoked the other withlow frequency (e.g., "cloth-orlon"). Whenthe category was presented before the in-stance, reaction time for Conditions (a) and(b) was approximately equal and signifi-cantly faster than for Conditions (c) and(d). However, when the instance was pre-sented first, reaction time for Conditions(a) and (c) was approximately equal andsignificantly faster than for Conditions (b)and (d). That is to say, subjects are fastwhen the category is presented first, if thecategory evokes the instance with high fre-quency, and subjects are fast when the in-stance is presented first, if the instanceevokes the category with high frequency.The spreading-activation theory explains thepattern of reaction times in the followingway, assuming that production frequency isa measure of the strength or accessibility ofthe path from one concept to another. Whenthe first concept (i.e., the one presentedfirst) evokes the second with a relativelyhigh frequency, this means that more activa-tion spreads to the second, and it takes lesstime to reach the threshold for an intersec-tion. Thus, the amount the first conceptprimes the second concept determines thereaction time.

By comparing this experiment with theFreedman and Loftus (1971) study, it canbe seen that the two results are exactlyparallel. Based on our structural assump-tions, "fruit" primes "apple" more than"red" or the letter "A" primes "apple" inthe Freedman and Loftus study. Hence,the shorter reaction time occurs when "fruit"or "A" is presented first. Similarly in theLoftus (1973b) study, when the categoryprimes the instance most highly, the shortestreaction times occur when the category ispresented first. But when the instanceprimes the category most highly, the short-est reaction times occur when the instance

is presented first. A spreading-activationexplanation is quite compelling to accountfor the Loftus (1973b) results, and thetheory offered here encompasses the ordereffect in both the Loftus study and the ear-lier Freedman and Loftus experiment withina single framework.

Recently Loftus (Note 2) has found twodifferent ways in which presenting a letteracts differently from presenting an adjective,in variations of the Freedman and Loftusparadigm. This has led her to the develop-ment of a dictionary-network model, whichwe will translate into spreading-activationterms. The first difference between pre-senting a letter and an adjective appearedwhen Grober and Loftus (1974) comparedreaction time in two conditions: one wherenoun-adjective (e.g., "fruit-red") and noun-letter (e.g., "fruit-A") trials were ran-domly intermixed, and one where noun-adjective and noun-letter trials were sepa-rated into blocks. In all cases the nounpreceded the adjective or letter. The resultsof this experiment are shown in Figure 2.It is clear that when the subject knows aletter is coming, he can prepare for it. Butin the mixed condition, the subject appar-ently prepares for either kind of trial thesame way he prepares for an adjective trial,since adjective trials take the same amountof time in either case. The theory's descrip-tion of semantic processing on the adjectivetrials is the same as that given earlier forthe Freedman and Loftus (1971) experi-ment, with the amendment that only the se-mantic network and not the lexical networkwould be diffusely primed before the ad-jective is presented. When an intersectionis found in the semantic network, then thesubject must retrieve the name from thelexical network.

Loftus (Note 2) described what musthappen on noun-letter trials in the blockedcondition as follows:

The first step of the process is entering the cate-gory. The next step is a quasi-parallel simul-taneous search towards the Dictionary. That isto say, the subject traces some number of path-ways leading from category instances to the Dic-tionary representations of those instances. Thisstep can be started during the interval between the


o0)(ft

1.60

1.50

1.40

1.30

adject ive

letter

mixed blockedCondition

FIGURE 2. Reaction time for noun-adjective andnoun-letter stimuli in mixed and blocked condi-tions (from Loftus, Note 2).

presentation of the category name and the restrict-ing letter if the subject knows a letter is coming.(P. 13)

This is essentially the spreading-activationexplanation, if the dictionary is taken to bea lexical network. Rather than saying that"the subject traces some number of path-ways," which suggests a conscious tracingprocess, the present theory would say thatactivation spreads along some number ofpathways, because the subject has activatedthe lexical network in addition to the se-mantic network (see Assumption 7). Hence,in the present explanation, the subject's con-trol is reduced to diffusely activating wholenetworks rather than specific pathways (inaddition to the specific nodes activated bythe stimuli in the experiment). The differ-ence in the results for the noun-letter trialsin the two conditions then depends onwhether the subject primes both networks(as in the blocked condition) or only thesemantic network (as in the mixed condi-tion). The reason he only activates thesemantic network in the mixed conditionmay be either because of a principle of leasteffort (hence he could speed up his reac-tion time if he tried) or because there is less

activation available to the semantic networkif both are primed (hence he will be sloweron noun-adjective trials if he primes both).

As can be seen in Figure 2, the subject ismuch slower on noun-adjective trials thanon noun-letter trials in the blocked condi-tion. This is accounted for by the fact thatan intersection on a noun-adjective trialoccurs in the semantic network and requiresthe further step of retrieving the correspond-ing name in the lexical network. On theother hand, the intersection on a noun-lettertrial occurs at the name in the lexical net-work. Therefore, the name does not thenneed to be retrieved.

The second result that shows up the dif-ference between adjectives and letters waspredicted by Loftus from the dictionary-net-work model. In this experiment (Loftus &Cole, 1974) subjects saw three stimuli, or-dered either noun, adjective, letter or noun,letter, adjective. For example, the threestimuli might be "animal," "small," and"M," for which an appropriate response is"mouse." The prediction was that the sub-ject should be faster when the adjective ispresented before the letter, and this was theresult found. The reasoning is as follows:When the adjective appears before the let-ter, activation will spread from a small setof instances in the semantic network to thelexical network where the intersection oc-curs, since the letter can be expected just asin the blocked condition. When the letteris presented before the adjective, activationwill spread from a small set of instances inthe lexical network back to the semantic net-work where an intersection with the adjec-tive will occur. Then the subject must re-turn again to the lexical network to retrievethe name, so there is an extra transit neces-sary in this condition.

Loftus has also run a series of experi-ments in which subjects were asked to pro-duce a member of a category and a short timelater asked to produce a different memberof that category (Loftus, 1973a; Loftus,Senders, & Turkletaub, 1974; Loftus & Lof-tus, 1974). This was accomplished by show-


ing a category-letter pair (e.g., "fruit-P"),which asked the subject for an appropriateinstance, then, following 0, 1, or 2 interven-ing items, showing the same category pairedwith a different letter (e.g., "fruit-A"),which asked for a different instance. Thegeneral finding is that reaction time for thesecond instance is shorter than reaction timefor the first instance and increases mono-tonically with the number of interveningitems. For example, in Loftus (1973a) asubject's baseline time to name a fruit be-ginning with the letter "P" was 1.52 sec.However, it took him 1.22 sec to produce thesame response if he had named a differentfruit on the previous trial and 1.29 sec toproduce the response if he had named a dif-ferent fruit two trials back.

The spreading-activation theory predictsthese results by assuming that when an itemis processed, other items are activated tothe extent that they are closely related tothat item. That is, retrieving one categorymember produces a spread of activation toother category members, facilitating theirlater retrieval. The assumption (Assump-tion 3) that activation decreases over timeor trials predicts the lag effect.

Meyer and Schvaneveldt (Meyer, 1973;Meyer & Schvaneveldt, 1971; Schvaneveldt& Meyer, 1973; Meyer, Schvaneveldt, &Ruddy, Note 3) have also shown that thetime to retrieve information from memoryis faster if related information has been ac-cessed a short time previously. Their para-digm is somewhat different. Subjects wererequired to classify letter strings as wordsor nonwords. The general finding was thatthe response, time to classify a letter stringas a word is faster if the subject has justclassified a semantically similar word as op-posed to a semantically dissimilar word.Thus, for example, the time it takes to clas-sify "butter" as a word is faster if "butter"is preceded by "bread" than if it is precededby "nurse." Their results have led Meyerand Schvaneveldt to an explanation in termsof spreading activation and illustrate thewidely different paradigms that such a the-ory can encompass.

Juola and Atkinson's Study withMultiple Categories

An increase in reaction time with multiplecategories has been found by Juola and At-kinson (1971) in a task where subjects hadto decide whether a stimulus word belongedto one of a variable number (1-4) of pre-specified (target) categories. They com-pared this task with one where subjects de-cided if the stimulus word was the same asone of a variable number (1-4) of targetwords. Their experiment was designed todistinguish between two kinds of models,one they attribute to Landauer and Freed-man (1968) and one they attribute to Col-lins and Quillian (1970a). In most respects,their results fit the model they derived fromLandauer and Freedman, but since thespreading-activation theory provides an al-ternative explanation for their results, wewant to compare their two models with ourtheory.

The model Juola and Atkinson (1971)derived from Landauer and Freedman(1968) is very similar to what Landauerand Meyer (1972) call the "category-searchmodel." It assumes that the subject searchesthrough instances of the categories in mem-ory seeking a match for the stimulus word.Such a model predicts that as the numberof categories or words in the memory setincreases, reaction time for the category-matching task should increase at a greaterslope than reaction time for the word-match-ing task. This is because each additionaltarget category adds more instances thatmust be searched, whereas each additionaltarget word only adds one, the word itself.This result was essentially what Juola andAtkinson found.

The model they ascribed to Collins andQuillian (1970a) assumed that subjects per-form the category-matching task by retriev-ing their stored category for the stimulusword and comparing this to the given cate-gories to see if it matches one of them. Thismodel would predict that the slope for thetwo tasks should be about the same, andthe intercept for the category-matching taskshould be greater than for the word-match-ing task. Their results clearly reject this


model. Although attributed to Collins andQuillian, this model is quite different fromQuillian's theory, because the semanticsearch in Quillian's (1966) theory is as-sumed to spread in parallel from both cate-gories and instances. When the categoriesare given first, as in Juola and Atkinson'sexperiment, then activation would spread outfrom the categories before the instance evenappeared.

In order to explain our interpretation ofJuola and Atkinson's results, it is necessaryto describe their procedure in more detail.They chose 10 large categories and 12 com-mon instances from each category as stimuli.This makes a total of 120 instances in all.In the word-matching task, they presentedfrom 1 to 4 of the 120 instances as targets oneach trial. In the category-matching task,they presented from 1 to 4 of the 10 cate-gories as targets on each trial. In bothcases the negative stimuli were chosen fromthe same set of 120 instances. In the word-matching task, then, the discrimination nec-essary to categorize the stimulus was be-tween one of the target words or a wordthat had not occurred as a target for a largenumber of trials (on the average about 24trials earlier). In the category-matchingtask, however, the discrimination was be-tween a word in one of the target categoriesand a word in one of the categoriesfrom a recent trial (on the average about 2trials earlier). The discrimination, there-fore, was rather easy in the word-matching task and quite difficult in the cate-gory-matching task.

What we think must be happening in thetask is that the discrimination between posi-tive and negative responses is made (at leastpartly) on the basis of activation level. Inthe category-matching task, as the numberof categories increases, the amount of acti-vation allocated to each category decreases(see Assumptions 2 and 3). Furthermore,activation will be left over from previoustrials on the categories corresponding tonegative instances, though it will have partlydecayed (Assumption 3). For example,suppose "tree" was a target category on aparticular trial and "body part" was not,

but "body part" was a target on a previoustrial. Then "tree" will have a higher activa-tion level than "body part," but the differ-ence will not be very large because "bodypart" was presented so recently. If a posi-tive instance such as "oak" is presented, itwill intersect with "tree"; if a negative in-stance such as "arm" is presented, it willintersect with "body part." The less activa-tion on "tree," which depends on how manyother targets there are, the harder it is todiscriminate that it is in fact a target orthat "body part" is not a target. The moredifficult the discrimination is, the longer ittakes to make, and thus there will be a fairlylarge effect of the size of the target set onreaction time.

In the word-matching task, however, thedifference in activation level as the numberof targets is varied will not be so critical afactor. This is because the absolute differ-ence between the activation level of targetsand nontargets is so much greater in theword-matching task, given Juola and Atkin-son's experimental procedure. That is tosay, each nontarget was presented as a targetso many trials previously (approximately24) that the activation level for a nontargetwould have decayed (Assumption 3) to avery low activation level as compared to anytarget. Hence, the large absolute differencein activation levels between targets and non-targets makes the differences due to target-set size relatively unimportant.

In conclusion, the spreading-activationtheory's explanation of Juola and Atkinson's(1971) results is that the effect of differ-ences in activation level due to target-setsize matter more when the discrimination isdifficult and matter less when the discrimi-nation is easier. Furthermore, as Juola andAtkinson point out, there are two aspects oftheir data (namely, the fact that the datafor positive responses are not linear, and themarked recency effects in the serial positioncurves) that fit much better with a parallelmodel, such as spreading-activation theory,than they do with a serial model, such as theone they derive from Landauer and Freed-man (1968).


There are two implications of this viewthat could be tested fairly easily. One isthat reaction time to decide that an instancesuch as "arm" is a negative instance willdepend on the recency with which "bodypart" was presented as a target category.The more recent its presentation, the longerit will take to say "no." A more global im-plication is simply that the slope of the curvewith respect to target size depends on thosefactors that affect the difficulty of discrimi-nation, such as the recency with which nega-tive instances occurred as targets (and prob-ably as nontargets as well). This has im-portant implications for the memory-searchliterature as a whole.

Conrad's Study

Using a true-false reaction time techniquefor sentences (e.g., the task is to decidewhether "A salmon can eat" is true orfalse), Conrad (1972) found results whichshe interpreted as contradictory to Quillian's(1966, 1969) theory of semantic processing.In fact, the results of her study are quiteclose to what Quillian's theory would pre-dict given Conrad's methodology.

In her first experiment, which was likethe Collins and Quillian (1969) study, Con-rad selected 2-level and 3-level hierarchiesfrom the common culture (e.g., salmon —»fish —» animal) and properties associated withthe objects at different levels. Then sheconstructed sentences with instances, such as"salmon," from the lowest level and prop-erties from all three levels. The resultsCollins and Quillian found were that re-action time increased as the property wasfarther removed from the instance in thehierarchy. The reason for the increases inreaction time according to spreading-activa-tion theory is that as the instance and prop-erty are farther apart in the hierarchy, ittakes activation longer to spread betweenthem and to trigger an intersection (andperhaps to evaluate the path found as well).

Unlike Collins and Quillian (1969), Con-rad (1972) broke down the properties inher sentences into three groups on the basisof the frequency (high, medium, low) withwhich people generated each property, given

the different objects in the hierarchies. An-other difference from Collins and Quillian'sstudy is that she collected data over 5 daysby repeating all the sentences each day,

The results of her first experiment weregenerally in the same direction as the in-creases in reaction time that Collins andQuillian found when the property was far-ther removed from the instance in the hier-archy. There was one reversal in her dataout of nine comparisons, and this occurredfor the high-frequency properties, where it.was not unexpected given the weak theory ofcognitive economy (as we will argue below).However, the increases she found were muchsmaller on the average than those of Collinsand Quillian. The weak theory of cognitiveeconomy predicts that people store a prop-erty with whatever instance it is linked toin a sentence, so Conrad's repetition of sen-tences over 5 days should lead to the smallerreaction time increases she found. This isbecause an inference necessary on the firstday would be less likely on the second day,and so on. Conrad, in fact, reports a largeLevel X Day interaction.

In general, Conrad found that the higherthe frequency of the property, the smallerthe increases betwen levels. Given the weaktheory of cognitive economy, we would ex-pect that high-frequency properties are morelikely to be stored at several levels in thehierarchy, because they are more likely to beencountered in contexts involving specificinstances. For example, "leaves" are morelikely to be stored as a property with par-ticular types of trees (such as "maple" and"oak") than is "bark," because leaves are ahigher frequency property. Thus, the effectof property frequency found by Conrad isconsistent with the weak theory.

Conrad (1972) argued that Collins andQuillian's (1969) results could be explainedby a confounding of property level and prop-erty frequency. Her argument was that thesentences Collins and Quillian used mayhave been based on high-frequency proper-ties for Level 1 sentences (e.g., "A salmonis pink"), moderate-frequency properties forLevel 2 sentences (e.g., "A salmon hasfins"), and low-frequency properties for


Level 3 sentences (e.g., "A salmon caneat"). To support her argument, she showedthat if one plots her reaction time data inthe above way, one obtains approximatelythe same slope as Collins and Quillian did,whereas if one plots the slope for low-, me-dium-, or high-frequency properties sepa-rately, one obtains much smaller slopes.

However, there are two weaknesses inConrad's argument. First, she did not useher frequency data to evaluate systematicallythe frequency of the properties in Collinsand Quillian's sentences, so her conjectureabout such a confounding had no empiricalbasis. Collins and Quillian (1969) did ob-tain subject ratings of importance of theproperty for the relevant level concept,which should correlate with Conrad's fre-quency measure. These ratings averaged1.90 for Level 1 sentences, 1.92 for Level 2sentences, and 2.16 for Level 3 sentences(based on a 5-point scale, where 1 = veryimportant and 5 = not important). Theseare small differences and certainly do notsupport the notion that the slope betweenLevel 1 and Level 2 sentences was dueto the confounding Conrad hypothesized.The difference between Level 3 sentencesand the others may have contributed to thegreater slope that Collins and Quillian found,but even that is doubtful. For those sub-jects who had sentences with all 3 levels,the slope was actually larger (approximately100 msec rather than 75 msec), but this wasoffset by a group of subjects who wereslower overall and saw only Level 1 and 2sentences. So the latter group acted tocancel out any exaggeration of the slopedue to the lower importance of Level 3properties.

Second, the comparison Conrad (1972)made in plotting her data against Collins andQuillian's data compared data based on fiveresponses to the same sentence with databased on one response to a sentence. Asindicated above, the weak theory of cogni-tive economy predicts that repetition of asentence makes an inference less likely andshould reduce the slopes in the way Conradfound. A fairer comparison would be be-tween her data on the first day and Collins

and Quillian's data. But even that compari-son has the problem that she may well haveincluded sentences of the kind, "A maplehas leaves," where the property is a generalproperty of trees, but where most peoplewould store it as a property of maples aswell. This suggests that the fairest com-parison is between Collins and Quillian'sdata and her data on the first day for low-frequency properties, where the propertieswere least likely to be stored at more thanone level. But we cannot make this com-parison because she did not break down herdata by days. In conclusion, the differencesbetween the two experiments and the factthat the only relevant data do not particu-larly support the conjecture about a con-founding of property level and property fre-quency make Conrad's argument rathertenuous.

It was Conrad's second experiment thatappears more damaging to Quillian's theory,but here she made a crucial methodologicalchange. She presented the object 1 sec be-fore the property, and this turned the ex-periment into a priming study. In the studyshe presented properties true of the highestlevel nodes, together with objects (e.g., "sal-mon," "fish," or "animal") at different lev-els in the hierarchy. Therefore, she pre-dicted from Quillian's theory that the lowerlevel objects, such as "salmon," would takelonger to confirm, since it would take activa-tion longer to spread between lower levelobjects and higher level properties. But bypresenting the object 1 sec before the prop-erty and by using only high-level proper-ties, she made it possible for her subjectsto prepare during the interval by primingthe object's superordinates. For example,if a subject saw "salmon," his best strategywas to retrieve the superordinates, "fish"and "animal," because the property to ap-pear would be a high-level property, such as"eating." In these circumstances, there islittle reason to expect systematic differ-ences between objects such as "salmon,""fish," and "animal." Thus, this particularexperiment had real methodological prob-lems as a test of Quillian's theory, and it isweaker evidence against spreading-activation


theory than her first experiment is evidencefor the theory.

Effects of Typicality and Semantic Related-ness in Categorisation Tasks

In recent experiments, Rips et al. (1973),Rosch (1973), and Smith et al. (1974) haveshown that reaction time in a categorizationtask corresponds very closely to ratings ofhow typical the instance is of the category.For example, robins and sparrows are con-sidered typical birds whereas chickens andgeese are not. The effect of typicality onreaction time is quite large even when fre-quency of the particular instances in the lan-guage is controlled. Like Smith et al., wewould argue that the typicality effect is onemore manifestation of the fact that semanticsimilarity speeds up positive decisions andslows down negative decisions. Such aneffect has been found repeatedly (Collins &Quillian, 1969, 1970a, 1972b; Schaeffer &Wallace, 1969, 1970; Wilkins, 1971). WhileLandauer and Meyer (1972) argued thatthe evidence for similarity effects at thattime was either questionable or artifactual,the evidence now seems so overwhelmingthat any viable theory must account forthem. They are very damaging to the cate-gory-search model.

There are two reasons why spreading-activation theory predicts that atypical in-stances will take longer to categorize thantypical instances. The most important rea-son derives from the way evidence is aggre-gated (see Assumptions 8-13). Becausedifferent connections that are found are com-bined as evidence, distinguishing propertiescan slow down a positive decision based ona superordinate connection or on matchingproperties. For example, the decision thata chicken is a bird (i.e., an atypical in-stance) might be made on the basis of asuperordinate connection from "chicken" to"bird," which people learn because chickensare frequently referred to as birds. But thefact that people eat chickens, that they areraised on farms, and that they are ratherlarge are all properties that distinguishchickens from most birds. If these distin-guishing properties are found during the

memory search, as some are likely to be,they act to slow down the positive decision,because they are negative evidence. Simi-larly, matching properties can slow down anegative decision. For example, the deci-sion that a goose is not a duck might bemade on the basis of the difference in theirnecks (a distinguishing property) or simplybecause they are stored as mutually exclu-sive kinds of birds (see Assumption 12),but the matching properties that are found(e.g., their affinity to ponds, their webbedfeet, their large size) will slow down thedecision that they are different. The argu-ment here is similar to that of Smith et al.(1974), which we will discuss in comparingthe two theories.

The second reason for the typicality ef-fects relates to those cases where a super-ordinate connection is found. As we indi-cated earlier, superordinate links differ inaccessibility (or strength), and accessibilitydepends on use. If a person frequently usesthe link that a robin is a bird, and less fre-quently uses the link that a chicken is a bird(assuming approximately equal frequencyfor chickens and robins), then the accessi-bility of "bird" from "robin" will be greaterthan from "chicken." Because of this, ac-cessibility will be highly correlated withtypicality ratings. All the factors acting tomake a chicken or a goose an atypical birdin the real world will also act to make theuse of the superordinate link in a person'smind from "chicken" or "goose" to "bird"infrequent. It is because they are atypicalthat the superordinate link is weak, and thiswill also act to slow down reaction timein making categorization judgments aboutatypical instances.

The way evidence is aggregated in thetheory also explains the common finding(Collins & Quillian, 1970a, 1972b; Holyoak& Glass, 1975; Rips et al., 1974) that peopleare fast to decide that semantically unrelatedconcepts are different (e.g., that a book isnot a dog). In comparing such concepts,there are not likely to be any superordinateconnections, and almost all property con-nections will involve distinguishing ratherthan matching properties. Therefore, almost


all the connections found will constitute neg-ative evidence, and subjects will be quitefast to reach the negative criterion in suchcases. This too is similar to the explana-tion in the Smith et al (1974) model.

Recently, Holyoak and Glass (1975) haveisolated two different cases where semanticrelatedness or typicality does not producethe usual effect on reaction time for nega-tive judgments. One case arises when thedecision depends on what they call a contra-diction and what we have called mutuallyexclusive subordinates. The other casearises when the decision depends on acounterexample.

In the first case, Holyoak and Glass foundthat people are faster to reject sentencessuch as "All fruits are vegetables" or "Somechairs are tables" than sentences such as"All fruits are flowers" or "Some chairs arebeds." In these four sentences the twonouns are mutually exclusive subordinates.The difference between the sentences is that"vegetables" and "tables" are generated withhigh frequency, while "flowers" and "beds"are generated with low frequency, whensubjects are given the frame "All fruitsare . . ." or "Some chairs are . . ." and askedto produce a false sentence. This differenceis in the opposite direction of the usual find-ing that negative judgments are slower whenthe two concepts are more closely relatedsemantically. The explanation for this re-versal according to the theory (and to Holy-oak and Glass) is that people make thesedecisions not on the basis of distinguishingproperties (though some might be consid-ered), but because they are stored as mu-tually exclusive subordinates (Assumption12). Generation frequency in this case is ameasure of the strength of the connectionbetween the two concepts and therefore ofhow long it will take to find the contradic-tion between the two mutually exclusiveconcepts.

The second finding of Holyoak and Glass(1975) involves sentences where people re-ject the sentence by finding a counterexam-ple (Assumption 13). For example, "Allanimals are birds," can be rejected by find-ing another kind of animal, such as a mam-mal. In this case Holyoak and Glass varied

the production frequency of the predicatenoun (e.g., "birds") independently of theproduction frequency of the counterexample(e.g., "mammals"). Their finding was thatreaction time depended not on the produc-tion frequency of the predicate noun (whichis a measure of the semantic relatedness ofthe concepts in the sentence) but on the fre-quency of producing a counterexample.Here again where a decision strategy that isnot based on distinguishing properties isappropriate, the reaction time data do notdepend on the semantic relatedness of thetwo concepts.

The importance of these two findings byHolyoak and Glass (1975), in our view, isthat they demonstrate that different kinds ofevidence can be involved in making cate-gorization judgments. This suggests thatapproaches such as that of Meyer (1970)and Smith et a). (1974), which try to for-mulate a single strategy for making suchjudgments, will inevitably fail.

RELATION OF THE THEORY TO THE MODELOF SMITH, SHOBEN, AND RIPS

Quillian's (1966, 1969) theory was a fore-runner of a number of global theories of se-mantic processing based on network repre-sentations, in particular those of Andersonand Bower (1973), Norman and Rumelhart(1975), and Schank (1972). These theo-rists have made important advances on theQuillian theory (especially in the represen-tation of acts and causes) which in no waycontradict the basic thrust of Quillian's the-ory. There are some differences betweenthese theories and Quillian's, but the basicintent of this paper is to deal with thoseaspects of semantic processing where themodel of Smith et al. (1974) is the majorcompetitor to Quillian's theory.

Unlike the various network models, themodel of Smith et al. represents concepts asbundles of semantic features. Their modelhas the virtues of being quite clear and ex-plicit, and it agrees quite well with the re-action time data for categorization judg-ments, except for the Holyoak and Glass(1975) results. Because it is such an ini-tially compelling model, we want to empha-size how it differs from spreading-activation


theory and point out what we think are itsinherent difficulties.

In the model of Smith et al. (1974), themeaning of a concept is assumed to be rep-resented by semantic features of two kinds:defining features and characteristic features.Defining features are those that an instancemust have to be a member of the concept,and the model assumes that features can bemore or less defining. Characteristic fea-tures are those that are commonly associatedwith the concept, but are not necessary forconcept membership. For example, "wings"might be a defining feature of "birds" and"flying" a characteristic feature, since allbirds have wings but not all fly. In a cat-egorization task, the model assumes that thetwo concepts are first compared in Stage 1with respect to all their features, both char-acteristic and defining. If the match is abovea positive criterion, the subject answers"yes"; if it is below a negative criterion, thesubject answers "no"; and if it is in-be-tween, the subject makes a second compar-ison in Stage 2 based on just the definingfeatures. If the instance has all the definingfeatures of the category, the subject says"yes" and otherwise says "no." If the sub-ject can decide in Stage 1, his reaction timewill be faster than if he decides in Stage 2.

There are several minor differences be-tween the model of Smith et al. (1974) andthe spreading-activation theory that could beminimized by slightly changing their model.The difference in wording between compar-ing features in their model and finding linksbetween properties in our theory is reallya nondifference. But the distinction betweendefining and characteristic features has theinherent difficulty, pointed out through theages, that there is no feature that is abso-lutely necessary for any category.4 For ex-ample, if one removes the wings from a bird,it does not stop being a bird. Furthermore,we doubt if people can make consistent de-

* There is for living things a biologists' taxonomy,which categorizes objects using properties that arenot always those most apparent to the layman.Thus, there are arbitrary, technical definitions thatare different from the layman's ill-defined concepts,but this is not true in most domains. There is notechnical 'definition of a game, a vehicle, or acountry that is generally accepted.

cisions as to whether a feature is definingor characteristic, either from time to time orfrom one person to another. Smith et al.recognized that features are more or lessdefining (or criterial), but they were forcedinto making the artificial distinction betweendefining and characteristic in order to havea two-stage model. Still, the model couldbe revised to work without the two stagesand make essentially the same reaction timepredictions.

The revision is as follows: If features arecompared over time, as in Quillian's (1966)theory, then as the process goes on longer,more features will be compared (assumingfeatures have different accessibilities). Thecomparison process can have a positive cri-terion and a negative criterion just as before,and features can be weighted by their cri-teriality. If the match at any point in timeis above the positive criterion, the subjectsays "yes"; if the match falls below thenegative criterion, the subject says "no" ;and otherwise he goes on comparing fea-tures. Finally, if he is running out of rel-evant information, he says "I don't know."This is simply the Bayesian decision modeldescribed in Assumption 9 of the extendedtheory, where the evidence consists ofmatching and mismatching features as in theproperty comparison of Assumption 11.

Thus, we agree that a decision processsimilar to the one that Smith et al. (1974)postulate does occur for some categorizationdecisions. But there is a fundamental dis-agreement, because they argue that all cat-egorizations judgments are made by compar-ing features of the instance and category,whereas we argue that people use whateverevidence they find, including superordinatelinks.

Because they exclude the use of superordi-nate links, the model of Smith et al. has sev-eral inherent difficulties. The most obviousis the assumption that even when peoplehave superordinate information stored, theydo not use it. While most people may nothave learned some superordinate relations(e.g., that a beaver is a mammal, or a sledis a vehicle), there are many they havelearned (e.g., that a wren is a bird, and abeaver is an animal). Why would they not


use such information if it is stored? Howin fact can they avoid using it? It is anunlikely model which postulates that peopleuse information that is less relevant to makea decision, instead of information that ismore relevant.

Another obvious difficulty with the Smithet al. (1974) model is that people seldomknow the defining properties of concepts.For example, consider whether a whale isa mammal, a sponge is an animal, a bat isa bird, or a wren is a sparrow. In theSmith et al. model, these difficult (and slow)decisions would be made in Stage 2 on thebasis of defining properties. But people gen-erally have no idea what the defining prop-erties of a mammal,' an animal, a bird, or asparrow are. Even if they know that oneof the most criterial properties for being amammal is that it bears its young alive,it seems highly unlikely that they knowwhether whales (or beavers for that matter)bear their young alive. Neither of theauthors has any idea what properties of asponge make it an animal, but if asked inan experiment whether a sponge was an an-imal, we would answer "yes," and we wouldbe comparatively slow about it. The reasonwe would answer "yes" is simply that wewere told at one time that a sponge is ananimal. We were also told that a bat is nota bird, and if we had not been told, we fearwe might have responded "yes" if askedwhether a bat is a bird in a categorizationexperiment. The decision that a wren is nota sparrow would be made because they aremutually exclusive kinds of birds (See as-sumption 12). They are both small song-birds, and it is hard to believe that manypeople know what the defining features of asparrow are that a wren does not have. Thefact that there are cases where people mustuse superordinate information to make cor-rect categorization judgments makes it un-likely that they do not use such informationin other cases where they could make thedecision simply by matching features orproperties. This is one of the strongeslarguments for a hybrid theory.

We would like to close this section byraising the question of why one should adoptsuch a complicated theory when the Smith

et al. (1974) model is simpler and predictsthe reaction time data quite well. We havetried to stress the inherent difficulties thattheir model has in ignoring superordinateinformation and in relying on defining prop-erties. Experimental tests can probably bedevised that will show up those difficulties.We will suggest one such test, but first wemight point out that the results of the Loftus(1973b) categorization experiment describedearlier do not fit the Smith et al. model verywell. If a person is merely comparing fea-tures between the instance and the category,then it should not matter whether the in-stance or category is presented first. It isthe asymmetry in the superordinate connec-tions that predicts the asymmetry Loftusfound in reaction time, and it is hard toimagine how one could have an asymmetryof that kind in comparing features of twoconcepts.

One experiment that might show difficul-ties with the Smith et al. (1974) model isa categorization task. The categories andinstances used are based on their multidi-mensional scaling of birds and animals onthe one hand, and mammals and animals onthe other. As both Collins and Quillian(Note 4) and Rips et al. (1973) report,subjects are faster at deciding that birdnames are in the category "bird" than in thecategory "animal," whereas they are slowerat deciding that mammal names are in thecategory "mammal" than in the category"animal." Collins and Quillian argue thatthis is the way people learn the superordi-nates: that pigeons are birds and lions areanimals. Smith et al. argue that it is basedon shared features, and they show by theirscaling solution that most birds are closerto "bird" than to "animal," and most mam-mals are closer to "animal'' than to "mam-mal." Rut there are several bird names thatare closer to "animal" than to "bird" (inparticular, "goose," "chicken," and "duck";"pigeon" is equidistant), and there are sev-eral mammal names that are closer to "mam-mal" than "animal" (in particular, "deer,""bear," and "lion"; "horse" is equidistant).We would predict that even for those in-stances the above pattern would hold,whereas a pure feature-matching theory,


such as the Smith et al. model, makes theopposite prediction. So this is a possibletest of the two theories. There are un-doubtedly many other tests.

Finally, we want to explain why we havebeen led to such a complicated theory. Intrying to write computer programs that an-swer different types of questions, it becomesapparent that any decision procedure thatgives correct answers must be flexibleenough to deal with many different config-urations of knowledge in memory. This isbecause people have incomplete knowledgeabout the world (see Collins et al., 1975),and they often do not have stored particularsuperordinate links or criterial properties.Any realistic data base for a computer sys-tem will have this same kind of incompleteknowledge. Therefore, perhaps our strong-est criticism of the Smith et al. (1974)model is that it breaks down when peoplelack knowledge about denning features.

While at one level this is a complicatedtheory, at another level it is a simpler theorythan the Smith et al. model. By viewingsuperordinate links as highly criterial prop-erties, the theory becomes a simple Bayesianmodel. It is only in specifying the particularconfigurations of knowledge that constitutepositive or negative evidence for the Bayes-ian process that the theory becomes compli-cated. The difference between the two the-ories is that the Smith et al. model allowsonly one kind of evidence (matching or mis-matching features), whereas the theory pre-sented here allows other kinds of evidenceas well. Thus the theory encompasses arevised version of the Smith et al. model asa special case of a more general procedure.

CONCLUSION

We have extended Quillian's spreading-activation theory of semantic processing inorder to deal with a number of experimentsthat have been performed on semantic mem-ory in recent years. The result is a fairlycomplicated theory with enough generalityto apply to results from many different ex-perimental paradigms. The theory can alsobe considered as a prescription for buildinghuman semantic processing in a computer,though at that level many details are omitted

about decision strategies for different judg-ments that arise in language processing (seeCarbonell & Collins, 1973; Collins et al.,1975; Quillian, 1969). We would arguethat the adequacy of a psychological theoryshould no longer be measured solely by itsability to predict experimental data. It isalso important that a theory be sufficientlypowerful to produce the behavior that itpurports to explain.

REFERENCE NOTES

1. Wilkins, A. J. Categories and the internal lex-icon. Paper presented at the meeting of theExperimental Psychology Society, Oxford, Eng-land, 1972.

2. Loftus, E. F. Haiti to catch a sebra in semanticmemory. Paper presented at the MinnesotaConference on Cognition, Knowledge, and Adap-tation, Minneapolis, 1973.

3. Meyer, D. E., Schvaneveldt, R. W., & Ruddy,M. G. Activation of lexical memory. Paperpresented at the meeting of the PsychonomicSociety, St. Louis, 1972.

4. Collins, A. M., & Quillian, M. R. Categoriesand subcategorics in semantic memory. Paperpresented at the meeting of the PsychonomicSociety, St. Louis, 1971.

REFERENCES

Anderson, J. R., & Bower, G. H. Human associa-tive memory. Washington, D.C.: V. H. Win-ston, 1973.

Anderson, R. C., & McGaw, B. On the represen-tation of meanings of general terms. Journal ofExperimental Psychology, 1973, 101, 301-306.

Bell, A., & Quillian, M. R. Capturing concepts ina semantic net. In E. L. Jacks (Ed.), Associa-tive information techniques. New York: Amer-ican Elsevier, 1971.

Brown, R. W., & McNeill, D. The "tip of thetongue" phenomenon. Journal of Verbal Learn-ing and Verbal Behavior, 1966, 6, 325-337.

Carbonell, J. R., & Collins, A. M. Natural seman-tics in artificial intelligence. Proceedings of theThird International Joint Conference on Arti-ficial Intelligence, 1973, 344-351.

Collins, A. M., & Quillian, M. R. Retrieval timefrom semantic memory. Journal of VerbalLearning and Verbal Behavior, 1969, 8, 240-248.

Collins, A. M., & Quillian, M. R. Does categorysize affect categorization time? Journal ofVerbal Learning and Verbal Behavior, 1970, 9,432-438. (a)

Collins, A. M., & Quillian, M. R. Facilitatingretrieval from semantic memory: The effect ofrepeating part of an inference. Acta Psycho-logica, 1970, 33, 304-314. (b)

Collins, A. M., & Quillian, M. R. Experiments onsemantic memory and language comprehension.


In L. W. Gregg (Ed.), Cognition in learningand memory. New York: Wiley, 1972. (a)

Collins, A. M., & Quillian, M. R. How to makea language user. In E. Tulving & W. Donald-son (Eds.), Organisation of memory. NewYork: Academic Press, 1972. (b)

Collins, A., Warnock, E. H., Aiello, N., & Miller,M. L. Reasoning from incomplete knowledge.In D. G. Bobrow & A. Collins (Eds.), Repre-sentation and understanding: Studies in cognitivescience. New York: Academic Press, 1975.

Conrad, C. Cognitive economy in semantic mem-ory. Journal of Experimental Psychology, 1972,92, 149-154.

Fitts, P. M. Cognitive aspects of information pro-cessing: III. Set for speed versus accuracy.Journal of Experimental Psychology, 1966, 71,849-857.

Freedman, J. L., & Loftus, E. F. Retrieval ofwords from long-term memory. Journal ofVerbal Learning and Verbal Behavior, 1971, 10,107-115.

Grober, E., & Loftus, E. F. Semantic memory:Searching for attributes versus searching fornames. Memory & Cognition, 1974, 2, 413-416.

Hollan, J. D. Features and semantic memory:Set-theoretic or network model? PsychologicalReview, 1975, 82, 154-155.

Holyoak, K. J., & Glass, A. L. The role of con-tradictions and counterexamples in the rejectionof false sentences. Journal of Verbal Learningand Verbal Behavior, 1975, 14, 215-239.

Juola, J. F., & Atkinson, R. C. Memory scanningfor words versus categories. Journal of VerbalLearning and Verbal Behavior, 1971, 10, 522-527.

Landauer, T. K., & Freedman, J. L. Informationretrieval from long-term memory: Category sizeand recognition time. Journal of Verbal Learn-ing and Verbal Behavior, 1968, 7, 291-295.

Landauer, T. K., & Meyer, D. E. Category sizeand semantic memory retrieval. Journal ofVerbal Learning and Verbal Behavior, 1972, 11,539-549.

Loftus, E. F. Activation of semantic memory.American Journal of Psychology, 1973, 86, 331-337. (a)

Loftus, E. F. Category dominance, instance dom-inance, and categorization time. Journal of Ex-perimental Psychology, 1973, 97, 70-74. (b)

Loftus, E. F., & Cole, W. Retrieving attribute andname information from semantic memory. Jour-nal of Experimental Psychology, 1974, 102, 1116-1122.

Loftus, E. F., Senders, J. W., & Turkletaub, S.Retrieval of phonetically similar and dissimilarcategory members. American Journal of Psy-chology, 1974, 87, 57-64.

Loftus, G. R., & Loftus, E. F. The influence ofone memory retrieval on a subsequent memoryretrieval. Memory & Cognition, 1974, 2, 467-471.

Meyer, D. E. On the representation and retrievalof stored semantic information. Cognitive Psy-chology, 1970, 1, 242-300.

Meyer, D. E. Correlated operations in searchingstored semantic categories. Journal of Experi-mental Psychology, 1973, 99, 124-133.

Meyer, D. E., & Schvaneveldt, R. W. Facilitationin recognizing pairs of words: Evidence of adependence between retrieval operations. Jour-nal of Experimental Psychology, 1971, 90, 227-234.

Norman, D. A., & Rumelhart, D. E. Explorationsin cognition. San Francisco: W. H. Freeman,1975.

Quillian, M. R. A revised design for an under-standing machine. Mechanical Translation, 1962,7, 17-29.

Quillian, M. R. Semantic memory. Unpublisheddoctoral dissertation, Carnegie Institute of Tech-nology, 1966. (Reprinted in part in M. Minsky[Ed.], Semantic information processing. Cam-bridge, Mass.: M.I.T. Press, 1968.)

Quillian, M. R. Word concepts: A theory andsimulation of some basic semantic capabilities.Behavioral Science, 1967, 12, 410-430.

Quillian, M. R. The Teachable Language Com-prehender: A simulation program and theory oflanguage. Communications of the ACM, 1969,12, 459-476.

Rips, L. J., Shoben, E. J., & Smith, E. E. Seman-tic distance and the verification of semantic rela-tions. Journal of Verbal Learning and VerbalBehavior, 1973, 12, 1-20.

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

Schaeffer, B., & Wallace, R. Semantic similarityand the comparison of word meanings. Journalof Experimental Psychology, 1969, 82, 343-346.

Schaeffer, B., & Wallace, R. The comparison ofword meanings. Journal of Experimental Psy-chology, 1970, 86, 144-152.

Schank, R. C. Conceptual dependency: A theoryof natural language understanding. CognitivePsychology, 1972, 3, 552-631.

Schvaneveldt, R. W., & Meyer, D. E. Retrievaland comparison processes in semantic memory.In S. Kornblum (Ed.), Attention and perform-ance IV. New York: Academic Press, 1973.

Smith, E. E., Shoben, E. J., & Rips, L. J. Com-parison processes in semantic memory. Psycho-logical Review, 1974, 81, 214-241.

Sternberg, S. High-speed scanning in humanmemory. Science, 1966, 153, 652-654.

Stone, M. Models of choice-reaction time. Psy-chometrika, 1960, 25, 251-260.

Wilkins, A. J. Conjoint frequency, category size,and categorization time. Journal of VerbalLearning and Verbal Behavior, 1971, 10, 382-385.

Wittgenstein, L. Philosophical investigations (G.E. M. Anscombe, trans.). Oxford: Blackwell,1953.

(Received November 4, 1974)

a spreading-activation theory of semantic...

Documents